China Machine Vision Algorithm SDK 2026 — From HALCON Dominance to VisionMaster's Domestic Rise

Chapter 1　Industry Panorama and Definitions (SDK vs. Complete Machine vs. Algorithm Library)

I. Starting From a Confusion

Walk into any company doing industrial automation and mention "machine vision," and seven or eight out of ten people will be referring to a complete set of equipment—industrial cameras, lenses, light sources, controllers, and the final inspection report that is output. This understanding is perfectly reasonable, because what terminal production-line engineers interface with every day is indeed the "machine," not abstract algorithmic code. However, what truly determines whether a vision system can identify a 0.05 mm crack, whether it can make a judgment at millisecond speed on a high-speed line, and whether it can be reused across devices and scenarios—the source of these capabilities is the algorithm software development kit buried deep within the system, which is the protagonist of this report: the machine vision algorithm SDK.

The algorithm SDK is the middleware layer of industrial vision software. It is neither the interface that production-line workers operate, nor the driver for a specific camera model, but rather a function library and development framework that encapsulates hundreds or even thousands of vision algorithms. System integrators call this library to build vision applications adapted to specific production lines; camera manufacturers embed the SDK on top of their drivers to deliver "smart cameras with algorithms" to customers; complete-machine equipment vendors build operating interfaces on top of the algorithm library, forming the final product form facing factories.

Understanding this layering is the premise for seeing clearly the industrial value of machine vision algorithm SDKs. The SDK is both technical infrastructure and the main battlefield of commercial competition—what is being fought over is the development-tool choice of thousands of system integrators, and the skill inertia of hundreds of thousands of industrial vision engineers.

II. Three-Layer Structure: Algorithm Library, SDK Platform, Complete-Machine System

The machine vision technology stack, from the bottom to the top, can be clearly divided into three layers, each with different users, delivery forms, and value propositions.

Layer 1: Basic Algorithm Library

This layer includes open-source image processing libraries like OpenCV, as well as model training and inference interfaces under deep learning frameworks (PyTorch, TensorFlow). The algorithm library is not oriented toward end users, but toward vision algorithm engineers, providing atomic-level computational units such as matrix operations, convolution, and filtering. Because OpenCV is open source and free, it has extremely broad penetration in research and entry-level industrial applications, but its industrial adaptation capabilities (multi-camera synchronization, precise timing control, calibration error management) still have a significant gap compared with professional commercial SDKs. In industrial production environments, the vast majority of production lines cannot tolerate OpenCV's deficiencies in industrial stability, real-time performance, and engineering support.

Layer 2: Vision Algorithm SDK Platform (Vision SDK / Middleware)

This layer is the core research subject of this report. Representative products include Germany's MVTec HALCON, the US's Cognex VisionPro, Hikrobot's VisionMaster, and Lingkong's VisionWARE. The SDK platform does three key things on top of the algorithm library:

First, algorithm toolification. It encapsulates image processing algorithms into task-oriented "tools" (positioning tools, measurement tools, inspection tools), greatly reducing the complexity of invocation. Engineers do not need to understand the underlying mathematical principles; they can invoke algorithms simply by configuring parameters.

Second, device-compatibility abstraction. It provides a unified device-driver abstraction layer (GenICam / GigE Vision / Camera Link protocol support), allowing one set of code to be compatible with hundreds of cameras, frame grabbers, and 3D sensors, avoiding the risk of "hardware lock-in."

Third, graphical development environment. It provides drag-and-drop process orchestration (low-code configuration interface), enabling vision engineers who have a basic engineering background but are not skilled in programming to quickly build applications. This capability is extremely popular among the large number of small and medium-sized system integrators in China, because these companies typically lack dedicated algorithm-development personnel.

Layer 3: Complete-Machine Vision System

Integrating the SDK platform into a hardware enclosure, equipping it with cameras and light sources, and providing an out-of-the-box interface yields a complete-machine system. Representative forms include smart cameras, vision controllers, and vision sensors. Hikrobot's complete-machine products and Cognex's In-Sight series all belong to this layer. The complete-machine system hides the complexity of the SDK and is suitable for terminal factories that need rapid deployment and lack deep development capabilities. Complete-machine products typically do not expose the underlying SDK externally; what customers use is the encapsulated functional interface.

The logical relationship of the three-layer structure is: the algorithm library provides computational primitives, the SDK provides engineering encapsulation, and the complete machine provides delivery completeness. The SDK sits in the middle but has the highest value density—it can be sold upward to complete-machine vendors (embedded into machines), and it can also directly serve system integrators downward, possessing bidirectional monetization capability.

III. The Core Value Proposition of the SDK

Having understood the three-layer structure, the unique value of the algorithm SDK becomes clear: it is the reusable code infrastructure for industrial vision capabilities.

A company doing 3C inspection uses a certain SDK to develop a phone-glass scratch-detection solution; six months later, switching to the lithium-battery electrode-sheet inspection track, it invokes another set of algorithm tools under the same SDK framework, without needing to build the underlying capabilities from scratch. This capability reuse is precisely the core advantage of an SDK compared with custom solutions.

At the same time, the SDK also carries the continuity of technological evolution. When deep learning algorithms began to show advantages in the field of industrial inspection around 2018, mainstream SDK vendors successively integrated deep-learning inference frameworks into their existing SDK systems. Engineers do not need to abandon their accumulated classical-algorithm knowledge, but instead flexibly combine classical algorithms and AI models within the same development framework according to task complexity—this is the key distinction of an algorithm SDK from a pure deep-learning platform. Lingkong explicitly proposes an "AI + rule" empowerment architecture in VisionWARE; Hikrobot likewise retains a complete classical-algorithm toolchain in VisionMaster 5.0 and layers an industrial large model and edge-learning tools on top of it. The maturity of this hybrid architecture marks that China's industrial vision SDKs have moved from "merely following the AI trend" to the mature stage of "deep integration of classical and AI."

IV. Market Boundaries and Measurement Scope

In market statistics, it is relatively difficult to independently account for the scale of machine vision algorithm SDKs, because a large number of SDKs are sold as embedded software within complete-machine products along with the hardware, without being separately priced. Among mainstream domestic vendors, Hikrobot's VisionMaster adopts a hardware-bound licensing model, with the software value embedded into the complete-machine pricing; Lingkong's VisionWARE runs two models in parallel, with both standalone software licenses and a sales model bundled with smart vision equipment.

From a macro-data perspective, in 2024 the overall market size of China's machine vision industry was about 20.7 billion yuan, a year-on-year increase of about 11.9%. Of this, the software layer (including algorithm SDKs, vision platforms, and image-processing software) accounts for about 15%–20% of the overall market, estimated in the range of 3.1–4.1 billion yuan. The total market size in 2025 is expected to exceed 21 billion yuan, with the software layer correspondingly rising to 3.5–4.5 billion yuan. GGII (GaoGong Industry Research Institute) forecasts that by 2027, the overall machine vision market size will reach 56.5 billion yuan, at which point the software layer is expected to exceed 10 billion yuan.

Therefore, this report defines the market statistical scope of machine vision algorithm SDKs as: vision software with algorithmic capability as the core deliverable, and its derived licensing revenue, including standalone software sales, per-node licensing, cloud subscriptions, and the attributable software-value portion bundled into smart equipment. Under this scope, the 2025 China market size is estimated in the range of 3.5–4.5 billion yuan.

V. The Trend of Blurring Boundaries Between SDK and Complete Machine

It is worth noting that as the "software-hardware integration" strategy becomes widespread, the boundary between algorithm SDKs and complete-machine vision systems is blurring. Both Hikrobot and Lingkong simultaneously sell SDKs and complete machines; system integrators can either purchase an SDK for secondary development or directly procure a complete-machine product with algorithms; VisionMaster has both a developer version for engineers and a foolproof configuration version for factories.

The impact of this trend is bidirectional: for pure-software SDK vendors (such as HALCON, VisionPro), the "no separation of software and hardware" domestic competitors indirectly compress their pure-software sales market space by lowering the overall system deployment cost; for domestic SDK vendors, controlling hardware sales channels enables them to lock in customers through a "bundled sales" strategy, continuously providing service support throughout the entire product lifecycle, forming more solid customer stickiness.

VI. The Research Scope and Time Baseline of This Report

This report focuses on the middle layer of the machine vision algorithm SDK / platform, and does not cover the following topics: machine vision complete-machine equipment (industrial cameras, light sources, lenses, and other hardware components) is detailed in a companion series report; general-purpose computer vision (consumer-grade AR, facial recognition, etc.) is discussed only in industrial manufacturing scenarios. The research time baseline is June 19, 2026, with data freshness covering FY2025 annual report data, the first quarter of 2026, and the latest developments in the first half of 2026. Industry-standard abbreviations referenced in the text are expressed by their full names to avoid ambiguity of technical codes.

VII. The Engineer Ecosystem of Machine Vision SDKs: The Competitive Value of Skill Inertia

In understanding the competitive landscape of machine vision algorithm SDKs, there is one dimension often overlooked by market analysis, namely the ecosystem lock-in effect brought by engineer skill inertia. The size of a platform's engineer community can, to some extent, predict the stability of its market share better than technical metrics.

Over the past 25 years, HALCON has built China's largest industrial-vision engineer skill circle through the HDevelop development environment, rich Chinese-language technical documentation (MVTec specifically provides complete Chinese algorithm documentation and accompanying video tutorials for the Chinese market), and extensive university cooperation (machine vision courses at top domestic engineering universities generally use HALCON as the core lab tool). It is estimated that more than 300,000 industrial vision engineers in China have HALCON development capabilities, of whom more than 100,000 use HALCON as their daily primary development tool. This base is one of the strongest defensive lines for HALCON to maintain its share in the high-end market—even if domestic SDKs are already technically on par with HALCON, for the engineer community to migrate from familiar HALCON to a new platform, they still need to overcome the combined migration cost of "relearning + code rewriting + project revalidation," which can be as high as 3–6 months of engineer working hours.

Hikrobot deeply understands this competitive logic. Building the VisionMaster engineer ecosystem is one of the cores of its market strategy: lowering the entry barrier through a free version (attracting students and junior engineers), cultivating a skill path through a complete Chinese-language tutorial system (YouTube-style video tutorials specifically aimed at engineers starting from scratch), and providing peer support through forum communities (the Hikrobot developer community has accumulated more than 200,000 registered users), forming a positive cycle of "learning → use → community feedback → deep dependence." The implementation effect of this strategy from 2022 to 2025 is precisely the fundamental driver behind VisionMaster's self-developed software license users growing from fewer than 100,000 instances to more than 600,000 instances.

For domestic SDK vendors, building the engineer ecosystem is a marathon rather than a sprint. The path chosen by SMore is to "bypass the engineer threshold"—using a no-code configuration interface to let production-line quality-inspection engineers who do not need SDK programming skills directly deploy AI vision systems, thereby sidestepping head-on competition with HALCON in the traditional engineer community. The price of this strategy is reduced technical depth per project, but it buys faster market penetration speed. The long-term competitive result of the two strategies will gradually become clear in 2027–2030.

VIII. The Structural Contradiction in China's Industrial Vision Talent Ecosystem

The rapid growth of China's machine vision industry has already caused a significant imbalance between talent supply and demand. Industrial vision engineers (especially composite talents who master both classical algorithms and deep learning) are one of the most prominent niches in the current talent gap in manufacturing digitalization.

From the demand side, as the visualization transformation of domestic 3C, lithium-battery, and automotive factories accelerates, and as industrial large models are deployed at scale, an additional demand of about 150,000–200,000 industrial vision engineers is expected for 2025–2027, but the supply from university cultivation (mechanical engineering + computer vision direction) is estimated at no more than 30,000–40,000 people per year, with an obvious gap.

From the supply side, this talent gap is giving rise to two phenomena: first, the salary level of vision engineers continues to rise, from an annual salary of 200,000–350,000 yuan in 2020 (5 years of experience) to 350,000–600,000 yuan in 2025, with some senior engineers who have industrial large model project experience earning more than 800,000 yuan annually, significantly narrowing the salary gap with algorithm engineers at major internet companies; second, the market demand for no-code/low-code vision configuration platforms has therefore greatly increased—when qualified vision engineers are in short supply and costly, no-code platforms that can reduce dependence on "professional engineers" have natural economic rationality in industrial promotion.

This structural contradiction provides, beyond technology, another key driver for the rapid penetration of no-code vision AI platforms (SMore, AInnovation): factories cannot recruit enough vision engineers, but production must continue—no-code AI quality-inspection platforms allow a factory's electrical engineers or quality engineers (who do not need a professional algorithm background) to independently deploy and maintain vision inspection systems, which is a real pain-point solution that drives customer purchasing decisions more strongly than technical capability.

---

Chapter 2　Global Landscape and China's Position

I. The Global Machine Vision Software Map

The overall size of the global machine vision market in 2024 was about USD 20.4 billion, and it is expected to expand at a compound annual growth rate of about 13%, exceeding USD 41.7 billion by 2030. In addition to the traditional upgrade of manufacturing automation, the core forces driving growth also include three new variables:

First, the large-scale explosive expansion of lithium-battery and photovoltaic production lines. Especially in China and Southeast Asia, the construction cycles of super-factories by power-battery giants such as CATL, BYD, and EVE Energy continue to drive vision-inspection demand; every process step of photovoltaic silicon wafers, cells, and modules requires vision quality inspection, and China's globally dominant photovoltaic capacity directly translates into sustained demand for machine vision.

Second, the advance of semiconductor front-end processes toward smaller nodes. Wafer-level defect inspection at nodes below 3 nanometers places extremely high demands on nanometer-level resolution, driving the continuous technological upgrade and high-value procurement of high-end industrial vision algorithms.

Third, the rise of industrial large models and Visual Foundation Models. These technologies are pushing machine vision from the "rule configuration" era into the "very few samples or even zero samples" era, giving rise to a batch of commercial opportunities for AI-native vision platforms and redrawing the competitive landscape.

At the software level, the competitive landscape of global algorithm SDKs has long been dominated by two camps: professional industrial vision algorithm libraries represented by MVTec HALCON, and software-hardware integrated platforms represented by Cognex VisionPro. In addition, Canada's Matrox MIL (Matrox Imaging Library) holds a place in high-end industrial and semiconductor fields; the US's NI (National Instruments, now merged into Emerson) LabVIEW Vision has long served vision-integration needs in research and measurement fields; Germany's Siemens WinCC Vision and Omron's vision platforms have stable applications within their respective ecosystems.

II. Cognex: The Moves of the World's Largest Listed Machine Vision Company

Cognex Corporation is the world's largest focused machine vision listed company, having cumulatively shipped more than 4.5 million image-based products globally, with cumulative historical revenue exceeding USD 11 billion. FY2025 Q3 single-quarter revenue was USD 277 million, up 18% year-on-year, exceeding market expectations; cumulative revenue for the first three quarters was about USD 730 million, achieving double-digit year-on-year growth.

Cognex's Q3 2025 revenue in China grew about 9%, with broad-based cross-industry growth, but automotive performance was relatively sluggish (the global automotive market is still in an adjustment period). Consumer electronics and logistics scenarios are the core drivers of Cognex's global and China growth. Notably, Cognex explicitly flagged the structural trend of consumer-electronics manufacturing shifting from mainland China to Vietnam, Malaysia, and India, which in the short term somewhat suppresses its China-market increment, but also provides diversified support for its global layout.

Cognex's deep-learning module, ViDi Suite, is iterating rapidly, and the continuous optimization of its "Interactive Training" experience is a core action for maintaining high-end market competitiveness.

III. Why HALCON Became the De Facto Standard—Dual Technical and Commercial Logic

HALCON was commercialized in 1996 by the machine vision research team at the Technical University of Munich. Early versions were deployed extensively in measurement and inspection systems in the European automotive industry, gradually expanding globally. There are four key reasons it became the industry's de facto standard.

Algorithm breadth and depth. HALCON covers almost all categories of industrial vision algorithms: Shape-based Matching, geometric positioning, measurement, OCR character recognition, barcode decoding, 3D calibration and reconstruction, deep learning (CNN classification, detection, anomaly detection), with a single version containing more than 2,100 algorithm operators. This completeness allows system integrators to solve almost all routine vision tasks on a single platform without switching between different libraries.

HDevelop development environment. HDevelop, the integrated development environment that comes with HALCON, provides an interactive mode of "algorithm scripting language + real-time image debugging," greatly reducing the difficulty of algorithm development and debugging. Engineers can test parameter effects in HDevelop, then export confirmed-effective code as C++, C#, and Python calling interfaces, embedding them into production systems. This seamless "prototype → production" transition gave HALCON a first-mover advantage in industrial vision education and talent cultivation systems—a large number of Chinese machine vision engineers' entry-level training uses HALCON as the primary tool, forming a deep skill-path dependence.

Reliability of the licensing model. HALCON's Runtime License adopts a hardware-lock (Dongle) or software-binding model, which protects software copyright while also forming a relatively stable business model. After a system integrator purchases a set of development licenses, each time it delivers a production-line system to an end customer, it must purchase a corresponding runtime license, forming a two-tier commercial logic of "development + runtime."

Cross-platform deployment capability. HALCON supports Windows, Linux, macOS, and various embedded platforms (ARM architecture), and deeply supports multi-threaded parallel computing and GPU-accelerated inference, enabling it to run stably on a full range of hardware from high-end industrial computers to edge embedded devices. This universality is difficult for competitors to replicate.

However, HALCON's pricing system has also become one of the biggest gaps for domestic replacement. A set of genuine HALCON development license plus runtime license has a combined cost in the range of 15,000 to 20,000 yuan, and for the large number of small and medium-sized system integrators in China, the licensing cost when replicating projects constitutes non-negligible pressure. When a medium-sized production line needs to deploy 30–50 vision stations, the pure software licensing cost alone can be as high as 450,000 to 1 million yuan, whereas domestic SDKs with equivalent functional coverage can compress this cost to 200,000 to 450,000 yuan. This price difference is the fundamental starting point for domestic replacement to gain an advantage in cost-sensitive scenarios.

IV. China's Position: From the Largest Consumer Market to the Strongest Competitor

China's position in the field of machine vision has gone through an evolution of three clear stages.

Stage 1 (2000–2015): Consumer market. The rapid expansion of Chinese manufacturing made China the world's largest machine vision consumer market, but core algorithm SDKs depended almost entirely on HALCON and VisionPro. The skill system of domestic engineers was built around these two software products, forming deep technical dependence. Domestic vision companies at this stage mainly focused on industrial camera hardware (Daheng Imaging, Hikvision's optoelectronics division, etc.), with almost no algorithm software capability.

Stage 2 (2015–2020): Domestic start-up. Hikrobot was founded (2016), Lingkong accelerated VisionWARE development, Daheng Imaging launched a self-developed algorithm platform, and multiple local vendors began to systematically build algorithm capability systems benchmarked against HALCON. At this stage, domestic SDKs achieved preliminary replacement in mid- and low-end scenarios (3C, printing, logistics), but were still unable to compete with foreign products in high-precision measurement (sub-micron level) and semiconductor inspection (nanometer-level resolution). The key at this stage: technology buyers at Chinese factories gradually shifted from "trusting only foreign brands" to "willing to try domestic," which was the psychological turning point of the entire replacement process.

Stage 3 (2020–present): Full competition. In 2020, the market share of domestic brands exceeded 50% for the first time, then rose year by year to about 60% by 2022. Driving this transformation is not only the technical maturity of domestic SDKs, but also the "lane-changing overtaking" of the AI vision track—in emerging scenarios such as deep-learning defect detection and industrial anomaly detection, China's AI-native enterprises such as SMore, AInnovation, and Megvii Industrial, by virtue of stronger AI engineering capabilities and deep understanding of local manufacturing scenarios, achieved a curve-overtaking of traditional foreign SDKs. This overtaking is not "making a cheaper similar product than foreign brands," but "establishing a new competitive order on a new technical dimension."

V. The Asia-Pacific Regional Landscape and Going-Global Opportunities

The Asia-Pacific region accounts for about 42.77% of the global 3D machine vision market's revenue share, making it undisputedly the largest regional market. In this landscape, China plays a dual role: it is both the world's largest demander of machine vision (3C electronics, lithium battery, and automotive manufacturing are all the most concentrated production regions globally), and is becoming the most competitive supplier.

Japan (Keyence is extremely strong in sensors and vision systems) and South Korea (vision demand brought by the Samsung and LG supply chains) have strong local consumption in the machine vision field, but lack general-purpose algorithm SDK vendors that can compete head-on with HALCON / Cognex. The Indian machine vision market started relatively late and was still in the basic-automation popularization stage before 2025. Southeast Asia (Vietnam, Malaysia, Indonesia) is rapidly rising as manufacturing transfers in, and the transfer of consumer-electronics production lines has driven new local machine vision demand; Chinese machine vision vendors (including SDK platforms) hold an obvious first-mover advantage when going global—because factories that go global following Chinese manufacturing leaders are more inclined to continue using supplier systems they are already familiar with domestically.

Cognex's Q3 2025 earnings report explicitly mentioned that the transfer of consumer-electronics manufacturing from mainland China to Vietnam, Malaysia, and India is one of the drivers of growth in some of its categories—this also means that foreign SDKs with multilingual support and global service networks still possess capability advantages in cross-border manufacturing scenarios that Chinese local vendors have not yet fully caught up with. Mech-Mind and SMore are currently the Chinese vision SDK enterprises at the forefront of building globalized service networks.

VI. The Special Competitive Logic at the Software Level

Throughout the machine vision industry chain, the algorithm SDK is the segment with the highest profit margin but the most difficult-to-quantify competitive barriers. The cost structure of an industrial camera is relatively transparent (sensor + lens + mechanical structure), but the core value of an algorithm SDK—patented algorithms accumulated over many years by hundreds of algorithm engineers, training datasets, and engineering optimization experience—is hard to reflect directly on a price tag. This "intangible value" makes the competition in algorithm SDKs manifest more as the long-term accumulation of technical reputation, engineer-community influence, and flagship customer cases, rather than a pure price war.

It is precisely because of this that the replacement path of domestic algorithm SDKs is not a simple "low-price copy," but rather to achieve technical overtaking in specific scenarios (such as AI vision, 3D vision), while forming a differentiated advantage in specific niche markets through lower licensing costs and a service system closer to the Chinese-language engineer community. Understanding this competitive logic is the key perspective for judging "how far" domestic replacement can go.

VII. Machine Vision SDKs and Manufacturing Digitalization: The Larger Story Background

The competition over machine vision algorithm SDKs is not an isolated software-market game, but rather one of the key scenarios nested within the grand process of China's manufacturing digital transformation. Understanding this larger background helps in judging the long-term market-size boundary of the entire track.

Over the past fifteen years, Chinese manufacturing has gone through a three-stage evolution from "labor-intensive" to "automated" to "digital." Vision inspection has played a key role in this process: it is the most flexible and information-dense sensing interface in automated production lines, and also an important entry point through which digital systems obtain physical manufacturing-process data (part-quality data, process-state data, equipment-state data).

When a camera captures an image of a part and the vision SDK makes a quality judgment, it is not only completing the "inspection" action, but also continuously generating a structured quality-data stream—these data streams are the "digital fuel" for the factory's MES system, digital-twin models, and process-optimization algorithms. In this sense, the machine vision algorithm SDK is not just a quality-inspection tool, but also the sensory nerve endings of manufacturing digitalization.

As manufacturing digitalization deepens, the role of the vision SDK is also evolving from "quality-inspection tool" to "data-production infrastructure." On this upgrade path, SDK vendors (or SDK + MES integrated solution providers) that can seamlessly integrate vision sensing with data management and process optimization will gain higher customer stickiness and larger per-customer revenue space than pure algorithm-tool vendors. SMore's "industrial AI agent" strategy and AInnovation's "one model, one body, two wings" strategy are both attempting to upgrade from "algorithm tools" to "manufacturing-digitalization infrastructure," which is the underlying narrative logic behind their high valuations.

Against the backdrop of digital manufacturing, the 4.8 million producing factories covered by Tianxia Gongchang are showing a significant digitalization stratification: head factories (annual output value above 5 billion yuan) have generally completed basic vision-system deployment and are upgrading toward AI vision and digital integration; mid-waist factories (annual output value from 500 million to 5 billion yuan) are in the window period of introducing vision systems for the first time, and are the main source of new machine vision demand over the next 3 years; long-tail factories (annual output value below 500 million yuan) have extremely low vision-penetration rates and are the potential space over a time horizon of more than 5 years. This three-layer distribution means that China's machine vision algorithm SDK market still has significant room for quantitative expansion before 2030, rather than having already entered a stock-competition stage.

VI. Cognex: The US Consumer-Electronics Hegemon's China Market Strategy

Cognex is the leader in global machine vision overall solutions, and its competitive strategy in the Chinese market differs from MVTec HALCON's pure-software route—Cognex adopts a software-hardware integrated strategy, simultaneously selling the VisionPro algorithm SDK and the In-Sight series of smart-camera complete-machine products.

Cognex's core strategic customers in China are the consumer-electronics supply chain: the manufacturing partners of Apple and Samsung in China (Luxshare Precision, Foxconn, GoerTek, etc.) are the main source of Cognex's China revenue, and appearance inspection and assembly guidance of consumer-electronics products are the pillar scenarios of Cognex's China business. In Q3 2025, Cognex's China-market revenue grew about 9% year-on-year, mainly driven by the cyclical recovery of consumer electronics—this contrasts markedly with the China-market growth rates (20%–30%) reported by AI vision enterprises such as SMore in the same period, implying that domestic AI vision enterprises have made substantial inroads into Cognex's share in consumer-electronics scenarios.

One of Cognex's countermeasures is to accelerate the localized marketing of ViDi Suite (deep-learning vision module) in China: on one hand cooperating with local agents in Shanghai and Shenzhen to advance sales, and on the other hand providing more local-language training support for Chinese customers. However, the pricing of ViDi Suite (the deep-learning add-on module is about 8,000–15,000 yuan per set) is still at an obvious disadvantage relative to the AI-vision complete-solution pricing offered by SMore and AInnovation—especially in AI vision model training efficiency and adaptation to domestic AI chips, Cognex's localization depth is inferior to that of domestic competitors.

It is expected that in 2026–2028, Cognex's strategic focus in the China market will contract from "overall market growth" to "defending core advantage positions"—maintaining its technical barriers in semiconductor inspection, high-precision industrial measurement, and the Apple supply chain, while gradually accepting the ceding of market share in mass-market AI vision (food, chemical, logistics) scenarios.

VII. Matrox: The Niche Strategy in the High-End Semiconductor and Medical Tracks

Matrox Imaging (Canada) is a typical "high-end niche" player in the global machine vision SDK market—the MIL SDK covers a complete system of frame grabbers, algorithm tools, and development environment, focusing on Camera Link and CoaXPress high-speed industrial camera interfaces (mainly applied in semiconductor wafer inspection and medical image processing), and holds a solid technical position in this high-end niche market.

Matrox's commercial scale in China is relatively limited, but in the field of semiconductor front-end equipment (lithography, inspection, metrology), Matrox's MIL SDK is adopted by some local semiconductor equipment manufacturers (such as some non-core inspection modules of NAURA and AMEC), obtaining a stable source of revenue bound to high-margin semiconductor equipment. Matrox's China-market strategy is to "maintain a technical premium and not participate in price wars"—therefore its share in China has not grown substantially, but it has also barely lost any to domestic SDKs.

For China's industrial vision SDK market, Matrox represents a competitive reference: even if domestic SDKs have already achieved dominance in mass-market scenarios, in industrial imaging scenarios that require the highest data-transmission bandwidth (Camera Link full-speed mode can reach 10 Gbps) and the strictest stability requirements, professional foreign SDKs can still maintain their technical premium and market share. This reminds domestic SDK vendors that replacement in high-end scenarios is a protracted battle requiring time and technical accumulation, not something that can be quickly accomplished by relying on price strategies.

---

Chapter 3　Core Technologies (Classical Algorithms / Deep Learning / 3D / Vision-Robot Integration)

I. Classical 2D Vision Algorithms: The Technical Language Established by HALCON

Before the deep-learning wave arrived, the technical system of industrial machine vision was composed of a set of classical algorithms, whose "standard vocabulary" was largely defined by HALCON. Understanding this technical language is the key to understanding why HALCON could become the industry standard, and where the difficulty of domestic SDK replacement lies.

1. Shape-based Matching

One of the most basic and core tasks in industrial vision: finding the precise position of a specific part in an image. The essence of shape-based matching is a three-step process of "edge feature extraction → pyramid hierarchical search → sub-pixel refinement." HALCON's Shape-based Matching operator extracts gradient-direction information of the target contour to build an edge-vector model, performs multi-scale, rotation-invariant matching searches in the test image based on vector similarity, and ultimately achieves sub-pixel-level (0.1-pixel magnitude) positioning accuracy.

This capability is indispensable in scenarios such as the vision positioning of electronic components (chip pin inspection, connector alignment) and automotive body marking positioning. Compared with grayscale-based Normalized Cross Correlation matching, Shape-based Matching has significantly higher robustness to illumination changes—this is extremely important in real production lines, because production-line lighting is never ideally uniform.

The catch-up path of domestic SDKs in shape-matching algorithms is to first achieve functional equivalence (the same positioning accuracy), then form differentiation in speed (using GPU-parallel search acceleration) and robustness in special scenarios (extremely low contrast, severe occlusion). In independent tests in 2024, VisionMaster's shape-matching tool already achieved positioning accuracy on par with HALCON, and its search speed even slightly outperformed after multi-core parallelization.

2. Geometric Measurement

Industrial measurement scenarios require reliable dimensional values at the millimeter, micron, or even sub-micron level. Classical vision measurement algorithms rely on camera calibration (intrinsics + distortion correction + extrinsics) to convert pixel coordinates into physical coordinates, then combine algorithms such as the Caliper Tool, Circle Fitting, Line Fitting, and point-to-line distance to output precise dimensional, angular, and distance data.

Error-source management is an important dimension distinguishing high-end from low-end SDKs: the correction accuracy of lens distortion (radial distortion + tangential distortion), the impact of uneven illumination on grayscale-threshold computation, and the adaptive compensation of thermal drift (calibration failure caused by thermal expansion after long device operation)—these are all problems that mature industrial SDKs need to handle systematically. HALCON provides a complete calibration toolchain (Calibration Target) and distortion models (multi-order polynomial distortion models), giving it a long-accumulated advantage in systematic-error control in precision-measurement scenarios.

3. OCR (Optical Character Recognition)

The core difference between industrial OCR and consumer-grade OCR lies in the challenges brought by character material (laser etching, inkjet coding, steel stamping), imaging conditions (strong reflection, low contrast), and recognition speed (high-speed production-line operation). HALCON's industrial OCR toolchain includes character segmentation (Region Growing / Thresholding) → feature extraction (Zernike moments, Fourier descriptors) → classifier matching, supporting multiple languages, fonts, and orientations, and capable of fault-tolerant recognition of incomplete characters.

In recent years, domestic SDKs have generally introduced deep-learning OCR models (based on CRNN / Transformer architectures), and their recognition rate in complex backgrounds is already on par with or even surpasses HALCON. VisionWARE's deep-learning character-recognition tool, in printing-inspection scenarios, achieves a measured recognition accuracy of over 99.2% for characters contaminated by ink dots, already surpassing HALCON's rule-based operator solution in the same scenario.

4. Defect Inspection

The core logic of traditional defect inspection is "rule operators + statistical judgment": extracting image features through grayscale thresholds, gradient magnitudes, and texture statistics (Gabor filtering, Local Binary Pattern LBP), then comparing them with the preset feature space of qualified products to judge whether deviations exceed tolerance thresholds. This method has extremely high efficiency and reliability in scenarios with regular textures and fixed defect types (metal scratches, printing color differences). But in scenarios with variable defect types and complex surface textures (random textures on leather, fabric, and casting surfaces), traditional operators lack robustness and produce relatively high miss rates or false-alarm rates. This limitation opened the door to deep-learning defect detection, and is also the most concentrated starting point of commercial opportunity for the entire AI vision track.

5. Blob Analysis and Region Processing

Blob (Binary Large Object) analysis is another fundamental algorithm family in industrial vision. Through threshold segmentation, connected-region labeling, and morphological operations (dilation, erosion, opening, closing), it extracts regions of interest from the image, then computes features such as area, perimeter, circularity, and orientation angle for the regions, and ultimately judges whether the region meets qualified-product standards. Blob analysis is widely used in scenarios such as wafer-defect counting (particle-contaminant counting), coating-area computation of lithium-battery electrode sheets, and pattern-integrity verification of printed packaging. This type of algorithm is simple in logic and fast in computation, holding an irreplaceable foundational position in industrial scenarios.

II. Deep-Learning Vision (AI Vision): A Leap in Technical Paradigm

Around 2018, deep-learning algorithms began to demonstrate capabilities on specific industrial tasks that traditional algorithms could hardly match, and the industrial vision field thereby underwent a fundamental leap in technical paradigm.

1. Deep-Learning Classification

Based on CNN backbones such as ResNet and EfficientNet, multi-class models are trained in industrial scenarios to classify part images into "qualified" or several "defect categories." The key challenge is: defect samples in industrial scenarios are naturally scarce (some defect categories may only collect dozens of images from a production line in a year), so data augmentation (random cropping, color jitter, elastic deformation, Mixup) and transfer learning (fine-tuning on ImageNet pre-trained weights) are the core technical means. One of the core technical highlights of SMore's defect-classification products is its meta-learning and few-shot learning technology targeting industrial small-sample scenarios, enabling the model to complete usable-level defect-classifier training with just 20–50 labeled samples, greatly reducing the labeling cost of AI vision systems.

2. Object Detection

Represented by the YOLO series (YOLOv8/v11), Faster R-CNN, and DETR, industrial adaptation versions are usually optimized for small targets and high-density scenarios (such as positioning dense solder joints on PCBs), and integrate sub-pixel coordinate refinement modules to meet industrial positioning-accuracy requirements. In terms of real-time performance, the YOLO series can achieve inference speeds of tens of frames per second on edge GPUs (NVIDIA Jetson series), meeting the takt-time requirements of most industrial production lines.

3. Semantic Segmentation

Semantic segmentation classifies every pixel in an image into a certain category (background, normal region, various defect regions), outputting a per-pixel defect mask that can be directly used for defect-area measurement and positioning. Industrial segmentation models based on U-Net variants have high value in lithium-battery tab-welding defect detection (weld-nugget morphology analysis) and textile-fabric flaw detection (flaw-contour extraction). AInnovation's steel-surface defect-segmentation product has achieved industry-leading per-pixel segmentation accuracy on hot-rolled strip surfaces, and has completed large-scale deployment at multiple leading steel mills.

4. Unsupervised Anomaly Detection

For industrial scenarios where "normal samples are abundant and abnormal samples are extremely scarce," unsupervised anomaly detection is the hottest algorithm direction in recent years, and also the most differentiated and valuable technical route in industrial AI vision. Representative methods include PatchCore (memory-bank-based nearest-neighbor distance judgment), PaDiM (Gaussian distribution modeling using multi-layer features), and ReverseDistillation (knowledge-distillation anomaly detection with a teacher-student architecture). These algorithms only need normal samples to train, demonstrating significant advantages in defect detection in scenarios such as castings and textiles. The industrial anomaly-detection products of SMore and AInnovation deeply integrate such methods, and have completed empirical validation on the production lines of high-end customers such as Luxshare Precision and Tesla.

5. Industrial Large Model Trend: Zero-Shot Anomaly Detection and Multimodal Fusion

The latest research at CVPR 2025 has introduced multimodal large models (Vision-Language Models, VLMs) into industrial anomaly detection, exploring the possibility of completing anomaly judgment through text descriptions of defect features without any labeled samples. This direction was still at the research and early POC stage in 2025, but has already been incorporated into the product roadmaps of SMore, AInnovation, and others, with mature industrialized products expected to appear in 2026–2027. Industrial large models have not only changed the algorithm-training paradigm (from "supervised labeling" to "knowledge injection + few-shot calibration"), but are also reshaping the product form of vision SDKs—from "tool-invocation platforms" to "vision agents with cross-scenario transfer capabilities."

6. Hybrid Algorithm Strategy: Classical + AI Is Mainstream

It is worth emphasizing that current mainstream industrial vision SDK platforms are not about "fully replacing classical algorithms with AI," but have formed a hybrid strategy of "classical algorithms handling structured tasks (positioning, measurement) + AI algorithms handling unstructured tasks (defect detection, anomaly judgment)." VisionMaster 5.0 simultaneously integrates an industrial vision large model (for complex, changeable scenarios), edge-learning tools (for simple rule-based scenarios), and traditional algorithm tools (for precise positioning and measurement). This hybrid architecture demonstrates higher reliability and interpretability than pure-AI solutions in production-line engineering deployment—once the system makes a misjudgment, engineers can clearly locate which algorithm node deviated, whereas the "black box" nature of pure AI is hard to accept in high-risk production-line scenarios.

III. 3D Vision Algorithms: From Point Cloud to Spatial Reconstruction

The core task of 3D vision algorithm SDKs is to generate accurate three-dimensional spatial representations from raw data acquired by depth sensors, and on this basis complete tasks such as 3D positioning, surface reconstruction, and volume measurement. 3D vision has a higher information dimension than 2D vision, but also brings higher computational complexity and longer engineering-deployment cycles.

1. Depth-Acquisition Technical Routes

The structured-light method (Mech-Mind Mech-Eye, Lingkong laser profilers, etc.) projects known patterns (sinusoidal stripes, encoded speckle) onto the target, computes depth based on the pattern's deformation amount, and generates a dense point cloud with high accuracy (measurement accuracy up to ±0.03 mm), but has relatively poor adaptability to highly reflective materials (mirror-finish stainless steel, metal welds). In 2024, Mech-Mind released a new generation of depth-estimation technology targeting reflective objects, improving point-cloud accuracy by about 90% through multi-band fusion and adaptive exposure control.

The ToF method (Orbbec) acquires depth by measuring light's time-of-flight, with a high frame rate (up to 90fps or more) but limited resolution and accuracy, suitable for dynamic warehousing and logistics scenarios (robot obstacle avoidance, package volume measurement). The laser-triangulation method (Lingkong line-laser displacement sensors) has high accuracy (sub-micron level), suitable for line-scan products (such as high-precision surface-topography inspection of lithium-battery electrode sheets, flatness measurement of glass substrates), but requires the target to move in uniform linear motion, imposing speed limits on high-speed scenarios.

2. Point Cloud Processing Algorithms

Raw point clouds often have problems such as noise, occlusion gaps, and uneven density, and must undergo a series of algorithmic processing before they can be used for industrial tasks. Core algorithms include: statistical outlier removal (statistical testing of each point's distance to its K-nearest neighbors); voxel-grid downsampling (reducing data volume while preserving point-cloud structural features); normal estimation (PCA-based local plane fitting, used for subsequent curvature computation and registration); point-cloud registration (ICP Iterative Closest Point, aligning the workpiece's actual point cloud with the standard CAD model); and feature extraction (local point-cloud feature descriptors such as FPFH and SHOT, used for pose recognition). Commercial SDKs have an obvious gap from the open-source PCL library in real-time optimization (GPU-accelerated parallel point-cloud processing) and industrial-scenario specificity (robustness to high reflection and partial occlusion), which is the core value of commercial 3D vision SDKs.

3. Surface Defect and Topography Inspection

By comparing the 3D reconstruction result with the standard CAD model and computing the per-point deviation (Point-to-Surface Distance), volume-type defects such as pits and pores on casting surfaces can be detected with positioning accuracy up to ±0.05 mm, which cannot be achieved under the 2D vision technical route. In the lithium-battery PACK assembly process, the 3D measurement accuracy of cell thickness is required to be better than ±0.1 mm, which is a typical high-value scenario for 3D vision. In automotive body-in-white inspection, 3D measurement has replaced traditional Coordinate Measuring Machines (CMM) as the mainstream technical route for high-volume online inspection.

4. Market-Size Forecast

The global 3D machine vision market is expected to grow from about USD 3.5 billion in 2025 to USD 19.1 billion in 2032, with a CAGR of about 13.68%, higher than the compound growth rate of the overall machine vision market. This super-fast growth comes from the superposition of two demands: first, the new demand for high-precision 3D inspection in lithium-battery and automotive scenarios; second, the from-scratch emergence of 3D vision-guided robotics scenarios (Bin Picking, multi-variety flexible assembly). The Chinese market is the fastest-growing region for 3D vision globally, especially driven by the lithium-battery PACK process and automotive body manufacturing.

IV. Image Preprocessing: The Invisible Foundation of the Algorithm Pipeline

Before discussing high-level algorithms such as template matching and defect detection, there is a frequently overlooked key link worth analyzing specifically: image preprocessing. In fact, in the vast majority of real industrial projects, the quality of image preprocessing often has an impact on the final inspection result no less than the choice of high-level algorithms themselves.

The image-quality challenges of industrial sites come from multiple aspects. Illumination non-uniformity (light-source position shift caused by production-line vibration, local highlights and shadows caused by differences in product-surface material) causes the brightness distribution of the same part's image to differ significantly at different moments, interfering with traditional algorithms based on fixed thresholds. Motion blur on high-speed lines (shutter time insufficient to freeze the moving workpiece) causes edge blurring and feature distortion. Sensor noise (dark-current noise and thermal noise of CMOS sensors) forms image grain in low-illumination or high-ISO scenarios, affecting the identification of fine defects.

The image-preprocessing toolbox of mainstream SDKs usually includes the following categories of algorithms:

Illumination homogenization. Through background-reference image subtraction (Shading Correction), the brightness gradient caused by uneven light sources is eliminated, ensuring that the algorithm faces a "homogenized" image, greatly improving the stability of grayscale-based inspection.

Morphological filtering. Opening (erosion followed by dilation) is used to eliminate small noise points while preserving the overall shape of the target, and closing (dilation followed by erosion) is used to fill small holes inside the target. These operations are indispensable in the preprocessing of blob analysis and defect detection.

Image enhancement. Contrast enhancement (histogram equalization, CLAHE local-contrast adaptive enhancement) can highlight subtle features in the image, with important value in detecting low-contrast defects (scratches, light-colored scrape marks).

Domain-specific filtering. Frequency-domain filtering (FFT + spectral masking) can eliminate periodic texture interference (such as the weave texture of fabric), making non-periodic defects superimposed on the periodic background easier to detect. HALCON's frequency-domain filtering tools (Fourier descriptors) have a relatively complete implementation in this direction; domestic SDKs are generally less complete than HALCON in frequency-domain processing, which is a technical direction still needing continuous supplementation.

For a mature industrial vision SDK, the completeness and parameter stability of its image-preprocessing toolbox is an important indicator distinguishing professional platforms from lightweight platforms. In actual projects, engineers often need to spend 30%–50% of their development time on tuning image preprocessing, rather than configuring high-level algorithms. Therefore, the richness and ease of use of an SDK's image-preprocessing tools have a direct impact on overall development efficiency.

V. Camera Calibration: The Foundation of Measurement Accuracy

The accuracy of a vision measurement system ultimately depends on the quality of camera calibration. Camera calibration is the process of converting the pixel coordinate system captured by the camera into the physical-world coordinate system, including intrinsic calibration (focal length, principal-point coordinates, radial distortion coefficients, tangential distortion coefficients) and extrinsic calibration (the camera's pose in the workpiece coordinate system).

Precision vision measurement has extremely demanding requirements for calibration. A vision system for automotive body-dimension inspection, to achieve a measurement accuracy of the ±0.1 mm magnitude, needs to control the pixel-to-physical-dimension conversion error to within 0.05 mm, which means that the calibration-board manufacturing precision, the numerical stability of the calibration algorithm, and the repeatability of on-site installation must each have no obvious error accumulation.

HALCON provides a complete set of camera-calibration tools, including: planar checkerboard calibration (an industrially optimized version of the classic Zhang method), circular dot-array calibration (higher-precision sub-pixel circle-center positioning), and high-precision stereo calibration (calibration of the relative pose between the two cameras of a binocular vision system). The maturity of HALCON's calibration system comes from years of repeated validation in automotive precision-measurement scenarios, capable of compressing the comprehensive system error (optical distortion + calibration-algorithm error + installation error) to below 0.02 mm, which leaves ample measurement margin on automotive mass-production lines with ±0.05–0.1 mm accuracy requirements.

The calibration accuracy of domestic SDKs has greatly improved in recent years; both VisionMaster and VisionWARE provide complete calibration toolchains, and are basically on par with HALCON in scenarios with general measurement-accuracy requirements (±0.1 mm or above). But in sub-micron precision-measurement scenarios (semiconductor metrology, nanometer-level topography inspection), domestic SDKs still have a gap from HALCON in calibration accuracy and numerical stability, which is also one of the technical root causes of the slow progress of domestic replacement in high-end measurement scenarios.

VI. Vision-Guided Robotics: The Extended Battlefield of Algorithm SDKs

Vision-Guided Robotics (VGR) is the product of the deep integration of machine vision algorithm SDKs and industrial-robot control systems, representing an important direction of industrial vision software extending from "inspection" to "execution," and is also one of the most certain growth poles in the machine vision incremental market for 2025–2030.

The typical technical pipeline of VGR includes five links: target recognition (locating the 3D position and posture of the target part in a 3D point cloud or 2D image) → pose estimation (computing the target's six-degree-of-freedom pose in the robot coordinate system, including three translations and three rotations) → hand-eye calibration (precisely establishing the rigid-body transformation relationship between the camera coordinate system and the robot end-effector coordinate system, which is the foundation of the entire system's accuracy) → trajectory planning (generating a collision-free motion path based on the pose, taking into account joint-angle limits and workspace constraints) → real-time feedback control (continuous closed-loop correction during grasp execution, handling small target movements and robot repeat-positioning errors).

Mech-Mind's Mech-Vision + Mech-Viz product system productizes exactly this pipeline, and through deep adaptation to mainstream global industrial-robot brands such as ABB, FANUC, Yaskawa, and KUKA (standardized robot communication interfaces), forms an "out-of-the-box vision-guided robotics solution." This product strategy enabled Mech-Mind to rank first in China's 3D vision-guided robotics market share for five consecutive years (2020–2024).

Vision-guided robotics scenarios place stricter requirements on the algorithm SDK than pure inspection scenarios: end-to-end latency needs to be compressed to within 100 milliseconds (the full pipeline of sensing + planning + communication + execution); dynamic target tracking (real-time tracking and grasping of targets moving on a conveyor belt, requiring high-speed point-cloud update rates precisely synchronized with robot motion); and high reliability (the cost of a robot-action error is collision and damaged parts, far higher than the cost of a vision miss). These stringent requirements are the concrete manifestation of the gap that still exists between domestic 3D vision SDKs and HALCON in high-end scenarios.

---

Chapter 4　Industry Chain Upstream and Downstream (Components → SDK → Complete-Machine Integrators → Terminal Factories)

I. The Panoramic Map of the Industry Chain

The industry chain in which the machine vision algorithm SDK sits is a complete value-transmission chain from optical hardware to terminal production lines. The algorithm SDK sits in the midstream of this chain, both undertaking the software expression of upstream component capabilities and transmitting vision-intelligence capabilities downward to system integrators and terminal factories. Only by understanding each link of this chain can one see clearly the strategic positioning of SDK vendors in the entire industrial ecosystem.

Upstream: Optical and Image-Acquisition Hardware

The upstream component system includes industrial cameras (CCD/CMOS sensors), lenses (telecentric lenses provide distortion-free imaging, fixed-focus lenses suit general scenarios), light sources (ring LEDs provide uniform diffuse light, coaxial light sources are used for imaging highly reflective parts, bar light sources are used for side-detail illumination), frame grabbers (high-speed data transmission, FPGA real-time preprocessing), and 3D depth sensors (structured light, ToF, laser triangulation).

The performance parameters of upstream hardware directly determine the upper limit of image quality the algorithm SDK can process: sensor resolution determines the minimum defect size that inspection can identify; dynamic range determines the highlight and shadow details that can be processed simultaneously; frame rate determines the maximum inspection speed of the production line; sensor noise level determines the signal-to-noise ratio of the image under low light or high-speed short exposure. No matter how excellent an algorithm is, faced with overexposed, motion-blurred raw images, it cannot output high-quality inspection results. Therefore, the collaborative adaptation between upstream hardware and midstream SDK (the transmission chain of optical parameters → image quality → algorithm input) is the core node of value creation in the industry chain.

Midstream: Algorithm SDK Platform

This is the link with the highest value density in the industry chain. The core barriers of algorithm SDK vendors lie in: long-accumulated algorithm patents (HALCON's Shape-based Matching patent cluster); engineered robustness (keeping algorithms stable under extreme illumination, fast motion, and complex backgrounds, and not having engineering-level problems such as memory leaks during 7×24-hour continuous operation); cross-platform deployment capability (Windows, Linux, embedded, edge GPU); and ecosystem binding (engineer communities, supporting training courses, industry-solution template libraries). These four together constitute the competitive moat of SDK vendors.

Downstream: System Integrators

System Integrators (SIs) are the key intermediate link connecting SDKs with terminal factories. They purchase SDK licenses, develop vision applications targeting specific industries (3C, automotive, lithium battery) and specific processes (appearance inspection, dimensional measurement, positioning and grasping) based on the SDK's algorithm tools, and deliver the complete vision solution (including hardware selection, on-site debugging, acceptance testing, and maintenance services) to terminal factories. China has thousands of machine vision system integrators, most with a scale of 50–200 people, serving specific regions and specific industries. Which SDK they choose as their primary development platform directly determines the trend of the SDK vendor's market share.

Terminal: Factory Production Lines

Terminal factories are the ultimate consumers of vision capabilities, but also the link with relatively limited technical understanding. The automation engineers of most factories care about the vision system's inspection speed, false-and-miss rate, maintenance cost, and supplier service response speed, rather than which SDK is at the bottom. This characteristic means that SDK competition often does not occur at the terminal-factory level, but is decided by the "technology-selection preferences" of system integrators. Whichever SDK can give the integrator's engineers the highest development efficiency, the shortest solution-debugging cycle, and the fastest problem resolution will win the loyalty of more integrators, thereby indirectly winning the deployment of more terminal factories.

II. The Industry-Chain Integration Strategy of Domestic SDKs

Unlike the pure-software sales model of overseas SDKs, mainstream domestic SDK vendors generally adopt a "software-hardware integration" strategy, building more complete industry-chain control by integrating upstream and downstream. The core logic of this strategy is: when software and hardware come from the same supplier, the customer's switching cost is higher, and the supplier can obtain continuous benefits over the full product lifecycle through after-sales service, standard-part replacement, and algorithm-upgrade packages.

Hikrobot is the most typical representative of the software-hardware integration strategy. From industrial cameras (the Hikvision MV series, supporting mainstream protocols such as GigE Vision, USB3.0, and Camera Link) to frame grabbers, to the VisionMaster SDK platform, to vision controllers for integrators, to complete-machine vision systems for terminal factories, Hikrobot has built a complete vertical-integration capability. This enables its sales personnel, when pitching to integrators, to offer a one-stop solution of "camera + SDK + vision controller + on-site technical support," a competitive advantage that pure-software vendors like HALCON cannot replicate. For the full year 2025, Hikrobot's overall revenue exceeded 6.4 billion yuan, with more than 600,000 self-developed industrial software license user instances, and more than 20,000 customers served globally.

Lingkong deeply integrates optical design and algorithm capabilities. In scenarios such as printing-quality inspection and color management that require the coordination of optical precision and algorithmic accuracy, Lingkong has formed a complete delivery capability of "optical hardware + VisionWARE algorithm platform + industry solutions." In 2025, Lingkong's smart vision equipment business revenue reached 775 million yuan, accounting for about 26.6% of total revenue, reflecting the maturity of its "algorithm + equipment" dual-wheel-drive model.

Mech-Mind extends downward, packaging the 3D vision SDK with the complete robot-guidance software pipeline, and deeply adapting it to mainstream industrial-robot brands, forming an end-to-end solution of "3D vision sensing + robot control + path planning," significantly reducing the development difficulty for system integrators, and thereby obtaining a premium space higher than pure-SDK sales and more stable customer stickiness.

III. Upstream: The SDK's Driving and Dependence Relationship with Cameras

The relationship between the algorithm SDK and industrial cameras is bidirectionally driven. On one hand, the SDK depends on the camera to provide high-quality raw images; on the other hand, the SDK's algorithm capabilities in turn define the requirements for camera specifications—high-precision geometric measurement requires low-distortion, high-resolution cameras, while deep-learning defect detection to some extent reduces the sensitivity to image noise (the model itself has a certain noise robustness, allowing the use of more cost-effective mid- and low-end cameras).

In the 2D industrial camera field, brands under Hikvision (including the Hikrobot MV series), Dahua's Huaray Technology, and Daheng Imaging are the domestic top three, having completely replaced foreign brands (Basler under Japan's Panasonic, Germany's Allied Vision, etc.) in the mid- and low-end market. The mid-to-high-end market (high-frame-rate, high-sensitivity, low-noise scientific-grade cameras) is still dominated by Japan's Hamamatsu and Germany's Basler high-end models, with domestic brands gradually catching up.

In the 3D sensor field, Mech-Mind (structured light) and Orbbec (ToF and structured light) have achieved commercialization in two technical routes, and the penetration rate of domestic 3D sensors in mid- and low-precision scenarios has exceeded 40%. But nanometer-precision high-end structured-light sensors (used for semiconductor wafer metrology) still mainly depend on imports, which also explains why the domestic SDK replacement in semiconductor inspection scenarios has not yet fully started.

IV. Downstream: The Integrator Ecosystem and SDK-Selection Logic

The SDK selection of China's machine vision system integrators is dominated by the following five factors:

Technical availability: whether the SDK covers the core algorithms required by the project (such as deep-learning defect detection, 3D point-cloud processing), and whether the algorithm's effectiveness has been independently validated by similar projects. The information asymmetry in this dimension is the core challenge faced by small and medium integrators when choosing an SDK, often relying on peer reputation and vendor demonstrations to judge.

Licensing cost: HALCON's combined cost of about 15,000 yuan per set constitutes significant cost pressure in multi-node deployment. The licensing cost of domestic SDKs is usually 40%–60% lower, and some vendors adopt a strategy of free basic version + paid advanced modules for small and medium integrators, significantly lowering the entry barrier.

Technical support: Chinese documentation, Chinese community, localized training, and fast-response technical support are the core differentiation points of domestic SDKs relative to foreign ones, especially for small and medium integrators lacking overseas resources. Hikrobot and Lingkong have deployed technical service networks covering major manufacturing cities nationwide, providing on-site engineer debugging and 24-hour telephone support, whereas HALCON's China service is mainly completed through agents, with relatively limited response speed and service depth.

Ecosystem binding: if an integrator has already accumulated a large amount of project code and engineer skills on a certain SDK, the migration cost (code rewriting, engineer retraining, project-validation cycle) forms a natural moat. The engineer community HALCON has accumulated over many years (tens of thousands of HALCON technical blogs on CSDN) is precisely the key defensive line for maintaining its share in the mid-to-high-end market.

Downstream customer requirements: some terminal factories (especially foreign factories in the semiconductor and automotive industries) explicitly designate HALCON or VisionPro in their procurement specifications, leaving integrators no choice space to switch to domestic SDKs. This "customer lock-in" mechanism is one of the important reasons foreign SDKs maintain a high share in the high-end market, and is also the most difficult link to crack in the process of domestic SDK replacement.

V. The Data Flywheel Value in the Industry Chain

In the AI vision track, a new variable is emerging in the value distribution of the industry chain: the data flywheel. SDK vendors that possess a large amount of labeled data from industrial scenarios (including image samples of various defects and measurement datasets) can continuously optimize their deep-learning models, making the models' generalization capability across different factories, different illumination, and different equipment conditions continuously improve, forming a data barrier that latecomers find hard to catch up with.

SMore serves more than 730 large manufacturing customers, and the vision-system operation of each factory continuously generates labeled data and user feedback—after desensitization, these data become "fuel" for the continuous training of the industrial vision large model. AInnovation's AInnoGC industrial large model is likewise built upon its rich customer data across multiple industries such as steel, new energy, and 3C. In contrast, HALCON's model-training data sources are relatively scattered, lacking systematic accumulation of Chinese manufacturing-scenario data, which constitutes its competitive disadvantage relative to Chinese AI-vision-native enterprises in the AI vision track.

VI. The Regional Distribution and Industry Specialization of the System-Integrator Ecosystem

The geographic distribution of China's machine vision system integrators is highly concentrated in three major regional clusters with dense manufacturing, and each region has formed a specific industry-specialization direction. This geographic characteristic directly affects the market-penetration strategies of SDK vendors in different regions.

Pearl River Delta: 3C electronics and AI vision. Shenzhen, Dongguan, and Guangzhou gather a large number of vision system integrators oriented toward 3C electronics (phones, tablets, TWS earphones), as well as the headquarters of AI vision and 3D vision enterprises such as SMore and Orbbec. Pearl River Delta integrators have strong technical capabilities and high acceptance of new technology, making it the region where domestic AI vision platforms first achieved large-scale commercialization. Both Hikrobot and Lingkong have technical-support teams in Shenzhen, providing local services for the region's integrator ecosystem.

Yangtze River Delta: automotive + new energy + printing. Suzhou, Shanghai, Hangzhou, and Ningbo gather a large number of vision integrators oriented toward automotive Tier 1 (Bosch, Continental, Joyson, etc.) and new-energy batteries. Integrators in this region have a solid technical foundation and a relatively high dependence on HALCON (constrained by foreign Tier 1 procurement specifications), but in the past two or three years, with the large-scale expansion of domestic new-energy factories, the penetration of Lingkong and Huaray Technology in this region has significantly increased. Mech-Mind has strong 3D vision-robot solution deployments in Shanghai and Suzhou, benefiting along with the procurement rhythm of automotive Tier 1.

Beijing-Tianjin-Hebei: industrial equipment + semiconductors. Beijing and Tianjin gather a large number of industrial-equipment manufacturers (heavy machinery, CNC equipment) and semiconductor-related enterprises (the supply-chain ecosystems of NAURA and AMEC). Vision demand focuses more on precision measurement and process monitoring, with high accuracy requirements for vision SDKs and low price sensitivity. HALCON still has strong market retention in this region's high-end scenarios, and domestic SDKs are beginning to gradually enter the semiconductor back-end and industrial machine-tool directions.

The existence of this regional specialization means that SDK vendors' market-promotion strategies need to be differentiated: they cannot face all integrators with one unified pitch, but must provide targeted technical documentation, demonstration cases, and pricing schemes for different industry contexts such as 3C (emphasizing rapid deployment of AI vision), automotive (emphasizing measurement accuracy and certification cases), and new energy (emphasizing stability under high-speed production takt). Lingkong is relatively mature in this regard—its VisionWARE printing edition, new-energy edition, and consumer-electronics edition are respectively optimized and differentially packaged for the typical scenarios of different industries, forming matching technical reputations in each niche market.

VII. SDK and Factory MES Integration: The Next Competitive Dimension

As industrial digitalization deepens, the integration depth of machine vision SDKs with factory MES (Manufacturing Execution System) and ERP (Enterprise Resource Planning) systems is becoming a new competitive dimension. Traditional vision SDKs output inspection results (qualified/unqualified) to the PLC (Programmable Logic Controller), which triggers sorting or alarm actions; the entire process is completed at the equipment level, with limited interaction with the factory's information systems.

But in the manufacturing digitalization 2.0 stage, what factories want is: the vision system not only tells the PLC "this part is unqualified," but also simultaneously uploads inspection data (specific defect type, defect coordinates, inspection timestamp, part number) to the MES system, for traceability analysis (defect-rate trends of the same batch), process optimization (which machining parameter change caused the rise in defect rate), and automatic generation of quality reports (digital output of customer quality reports).

Achieving this integration requires the vision SDK to provide standardized MES integration interfaces (the OPC UA data-transmission protocol, Webservice API, or a data-reporting module built into the SDK), and the vision-inspection data needs to have a structured format (JSON/XML) compatible with the MES system's data model. SMore has the highest degree of productization in this direction; its industrial AI agent platform explicitly takes "data linkage with MES/ERP systems" as one of its core product capabilities, upgrading vision inspection from an "island system at the equipment level" to "an organic component of the factory's digitalization system."

The competitive implication of this trend for SDK vendors is: pure algorithm capability is no longer sufficient to constitute a sustained competitive barrier; the integration depth with factory digitalization systems will become an increasingly important consideration dimension in machine vision SDK selection over the next 3–5 years, thereby pushing SDK vendors to evolve toward "industrial software platforms," and not merely be algorithm-tool providers.

VIII. Vertical Integration of the Industry Chain: The Strategic Logic of SDK Vendors Extending Downstream

A key structural trend in the machine vision algorithm SDK market is SDK vendors' proactive extension toward the downstream of the industry chain (system integration) and the application layer (SaaS services). This trend is highly similar to the historical evolution of the traditional software-middleware industry: when the market grows to a certain scale, the value of the pure tool layer is compressed by vendors extending toward the upper application layer.

Hikrobot has to some extent already completed this extension: its product portfolio starts from the VisionMaster SDK and covers cameras (hardware sensing layer), the algorithm SDK (middleware layer), vision controllers (hardware computing layer), and vision solutions for specific industries (application layer). This "vertically integrated product matrix" enables Hikrobot to offer customers "full-stack vision solutions," with overall competitiveness in project competition significantly higher than vendors at a single layer.

Lingkong's vertical-integration path is slightly different: starting from dedicated vision systems for printing inspection, after accumulating deep optical-design and high-precision vision-algorithm capabilities, it abstracts upward a general-purpose vision SDK platform (VisionWARE), while maintaining dedicated system-integration capabilities in the printing-inspection, consumer-electronics, and new-energy industries. Lingkong's strategy is to exchange industry depth (complete solution capability in specific industries) for market trust in the SDK platform, then use the SDK platform's technical reputation to attract system integrators from other industries.

This vertical-integration trend constitutes long-term pressure on pure-software SDK vendors (HALCON, VisionPro): software-hardware integrated competitors can offer a lower total system cost through a "bundled sales" strategy, while the value proposition of pure-software SDKs (technical leadership, ecosystem scale, cross-platform compatibility) needs to contend with the customer's combined consideration of cost and service in actual procurement decisions.

IX. The Talent Strategy of SDK Vendors: Competition for Technical Backbones and the Stability of Core Teams

In the technology-intensive track of industrial vision SDKs, the acquisition and retention of core talent is a key prerequisite for competitiveness. The R&D team with VisionMaster as the core algorithm is estimated to exceed 500 people (2025), Lingkong's algorithm R&D team is about 300 people, and SMore's technical team (including algorithm, engineering, and product) is about 800 people—behind these numbers are annual labor-cost investments of hundreds of millions of yuan, constituting a significant entry barrier for this track.

University output and intra-industry mobility are the two main channels for talent replenishment. PhD and master's graduates in computer vision and image processing from Tsinghua University, Shanghai Jiao Tong University, and Zhejiang University are the core talent pool competed for by major industrial vision SDK vendors. Jiaya Jia's founding of SMore is itself the embodiment of an industrialization path from the computer-vision lab of the Chinese University of Hong Kong—the conversion of top academic teams to industry is an important source of high-quality startups in the AI vision track.

However, the talent competition within the track also brings a hidden risk: the salary level of core algorithm engineers continues to rise. In 2024–2025, the median annual salary of senior algorithm engineers in China's industrial AI vision field was about 600,000–1 million yuan, and the salaries of top algorithm engineers were even higher. This salary level keeps the talent-acquisition cost of small and medium system integrators and startup SDK vendors persistently high, objectively intensifying the scale advantage of head enterprises in the talent dimension.

---

Chapter 5　Downstream Applications (3C / Lithium Battery / Automotive / Semiconductor / Logistics / Pharmaceutical)

I. 3C Electronics: The Largest, Most Fiercely Contested Main Battlefield

3C (computer, communication, consumer electronics) manufacturing is the largest downstream application scenario for China's machine vision, accounting for about 30%–35% of China's machine vision applications. The large-scale mass production of smartphones, tablets, laptops, TWS earphones, and wearable devices has spawned continuous high growth in vision demand such as appearance-defect detection, dimensional measurement, and assembly positioning.

Core application scenarios: Scratch and chipping detection of phone glass cover plates requires a comprehensive scan of a complete cover plate within 0.5 seconds, with inspection accuracy required to be better than 0.05 mm. AOI (Automated Optical Inspection) of PCB solder joints is another core scenario; each motherboard has thousands of solder joints that need to be inspected within a millisecond time window, and the inspection results need to link with the MES system in real time. Bending detection of connector PINs (accuracy required to within ±0.1 mm), assembly positioning of camera modules (vision-guided AA active alignment), and industrial OCR recognition of logos and labels together constitute the main body of 3C vision applications.

Status of deep-learning introduction: In 3C scenarios, deep-learning defect detection has moved from pilots to mainstream, especially for defect types like "surface scratches" that are difficult to precisely describe with rule algorithms, where AI algorithms demonstrate a significantly lower False Alarm Rate. Lingkong has a strong position in the consumer-electronics inspection field; VisionWARE's deep-learning appearance-inspection tool was deployed in batches at multiple leading phone-foundry enterprises in 2025; SMore initially entered the industrial vision market precisely from consumer-electronics production lines (Foxconn, Luxshare Precision), and used deep-learning inspection capability as a differentiating weapon, gradually expanding to more industries.

Maturity of domestic replacement: In 3C scenarios, domestic SDKs have basically completed the replacement of HALCON and VisionPro. According to industry research data, the domestic SDK penetration rate in 3C scenarios has exceeded 75%, for the following reasons: 3C factory technical teams are young and highly willing to accept new technology; domestic SDKs already have ample project accumulation in 3C-dedicated algorithms; and 3C factories have a large procurement volume, so the savings in software-licensing cost make a significant positive contribution to ROI calculations.

II. Lithium Battery: The Fastest-Growing New Battlefield

From 2023 to 2025, the large-scale expansion of China's power-battery capacity brought lithium-battery manufacturing into one of the fastest-growing machine vision application scenarios. The hundred-billion-level capacity expansion plans of head enterprises such as CATL, BYD, EVE Energy, and Gotion High-Tech each require a large number of supporting machine vision inspection devices and algorithm platforms in the construction of every GWh of capacity.

Core application scenarios:

Electrode-sheet stage: thickness-uniformity measurement of the coating layer (thickness deviation required to within ±2 microns), coating-edge positioning (positioning accuracy ±0.1 mm), and electrode-sheet scratch and crack detection (subtle cracks can reduce product safety) all require high-resolution line-scan cameras paired with professional algorithms.

Winding/stacking stage: alignment accuracy of positive and negative electrode sheets (alignment deviation usually required to be controlled within ±0.2 mm to avoid local short-circuits caused by electrode misalignment), and coating-area integrity verification (ensuring uniform coverage of active material on the current collector).

Welding stage: tab-welding quality (2D/3D composite inspection of multiple defect categories such as cold welds, missed welds, and weld-spatter; different defects have different impacts on battery safety and require classified discrimination).

Pack assembly stage: cell-thickness measurement (accuracy required to ±0.05 mm, reflecting the cell's SOC and health state), and 3D inspection of the welding quality of connecting busbars (welding defects can cause the power-battery pack to overheat or even undergo thermal runaway).

Lingkong's layout in the lithium-battery vision-inspection field is particularly prominent; in 2025, its new-energy business revenue reached 185 million yuan, a year-on-year increase of 36.01%, the company's fastest-growing niche business; Huaray Technology's lithium-battery-dedicated vision solutions have good customer coverage in the industry, with product accuracy in coating-defect detection up to ±0.05 mm and 3D point-cloud density up to 1,000 points per square millimeter, reflecting professional-grade inspection capability.

A notable feature of lithium-battery scenarios is: the same production line often integrates the hybrid application of 2D vision (high-speed appearance inspection), 3D vision (precision measurement), and AI vision (complex multi-class defect discrimination), placing relatively high requirements on the multi-algorithm fusion capability of the SDK platform.

III. Automotive: Coexistence of High-Precision Requirements and Foreign Dominance

Automotive manufacturing is a traditional stronghold of machine vision applications, and also one of the fields where foreign SDKs are hardest to replace. The reason is that automotive OEMs (vehicle manufacturers) and Tier 1 suppliers generally specify the technical standards of vision systems in their quality-control specifications, and these standards are often deeply bound to the historical implementations of tools like HALCON, forming a dual technology-standard lock-in effect.

Core application scenarios:

3D quality inspection of body welds: reconstructing the 3D morphology of welds through laser structured light to detect defects such as weld holes, undercuts, and cracks, with accuracy required to be better than ±0.1 mm. 3D dimensional measurement of engine blocks and transmission housings: comparing against the CAD standard model to detect casting deviations (allowable deviation is usually at the ±0.2 mm magnitude, with stricter requirements for key features). Part-surface defect detection: scratches, drag marks, and pits on stamped parts; orange peel, pinholes, and particle contamination on painted surfaces. Assembly positioning of ADAS cameras and LiDAR, light-irradiation-area inspection of vehicles coming off the line, and OCR recognition of cockpit markings are all typical processes in automotive vision applications.

Replacement progress: The domestic replacement process in automotive scenarios is significantly slower than in 3C and lithium battery, with foreign brands still holding about 63% of market share. The obstacle to advancing replacement is not only technology, but more so industry-chain inertia and certification systems—domestic SDKs need to pass the technical review and quality-system certification (IATF 16949, etc.) of automotive OEMs and Tier 1, a process that often takes more than 18 months and requires completing at least 12 months of stability validation on a qualified supplier's actual production line.

SenseTime Industrial entered the automotive field with an "AI + AR" industrial-digitalization solution, and has won factory-digitalization orders from BMW and Tesla, with single-project amounts exceeding 300 million yuan—but this leans toward the industrial-digitalization direction of digital twins and production management, which is essentially different from pure vision-algorithm SDK replacement, belonging to a different competitive dimension.

IV. Semiconductor: The Highest Technical Barrier, Domestic Replacement Has a Long Way to Go

Semiconductor manufacturing is the downstream application with the highest machine vision technical requirements and the greatest difficulty of domestic replacement. Wafer-level defect detection (front-end process) and package-level inspection (back-end process) place requirements far exceeding other scenarios on the vision system's resolution (nanometer level), stability (7×24-hour operation, PPM-level false-alarm-rate requirements), and integration depth with production-line management-and-control systems (SEMI standard interfaces E5, E30, E40, E87, etc.).

Core application scenarios:

Wafer appearance macro-defect inspection (Macro Defect Inspection): using enhanced illumination and high-resolution CCDs to detect macro defects on the wafer surface (cracks, chipping, scratches, particle contamination).

Patterned Wafer Inspection: on wafers with existing lithographic patterns, detecting micron-to-nanometer-level process defects such as pattern opens, shorts, and linewidth deviations, the most technically difficult vision task in the front-end process.

Semiconductor back-end scenarios: Bond Wire morphology inspection (height, arc, verticality), chip-package offset metrology (Flip Chip / Wire Bond package accuracy), and BGA Solder Ball quality inspection (ball-diameter uniformity, missing balls, shorted balls).

The foreign-dominated landscape and its reasons: Foreign SDKs and dedicated inspection equipment still hold about 85% of share in semiconductor scenarios. The core reasons are twofold: on one hand, the vision-inspection equipment introduced by semiconductor factories usually comes from professional semiconductor equipment vendors such as KLA Tencor and Applied Materials (AMAT), and these equipment vendors use self-developed proprietary algorithms rather than general-purpose SDKs; on the other hand, even in semiconductor back-end scenarios where general-purpose SDKs have application space, the current stability and reliability record of domestic general-purpose SDKs still does not meet semiconductor factories' extremely high requirements for "zero-defect" systems. Domestic SDKs have about 15%–20% penetration in the semiconductor back-end (packaging) link, but replacement in the front-end key processes is still in the start-up stage, expected to require 5–8 years of technical accumulation and a customer-validation cycle.

V. Logistics: The Extreme Test of High Speed and Robustness

Logistics sorting is one of the fastest-growing machine vision scenarios in recent years, and also an important pillar of Cognex's global revenue. Typical applications include: package volume measurement (DWS, Dimensioning Weighing Scanning, requiring millisecond-level 3D measurement to be completed during the high-speed motion of the conveyor belt), barcode/QR-code recognition (used for package tracking, requiring high recognition rates under random poses and possibly damaged labels), package-diversion positioning on conveyor belts, and vision-guided grasping of express-delivery end-sorting robots.

The core demands of logistics scenarios on the SDK are speed (packages moving at high speed on conveyor belts require millisecond-level code reading and measurement) and robustness (random package poses, damaged barcodes, drastic illumination changes). Cognex's Q3 2025 earnings report explicitly pointed out that the strong performance of the Logistics business is one of the core factors driving overall revenue to exceed expectations. Domestic SDKs have a relatively high penetration rate in logistics scenarios; Hikrobot's vision platform has gained a dominant position in the vision construction of self-built logistics systems of domestic e-commerce platforms such as Cainiao Smart Warehouse and JD Logistics, and has begun to enter the Southeast Asian market along with the going-global layout of Chinese e-commerce logistics enterprises.

VI. Pharmaceutical: Compliance-Driven Precision Vision Inspection

Machine vision applications in pharmaceutical manufacturing (injections, oral solid preparations, medical devices) take regulatory compliance (GMP, Good Manufacturing Practice) as the core driver. The appearance of each tablet (color deviation, chipping, foreign matter), the clarity of each injection (particulate foreign matter, GMP requires 100% inspection), and the dimensions of each medical-device component all require 100% online vision inspection under GMP-specification constraints.

The particularity of pharmaceutical scenarios is: the vision system needs to pass the drug administration's GMP validation (China's NMPA), system changes (including algorithm-parameter adjustments) require the submission of change-validation reports, and the system's audit-traceability records during operation must be complete and traceable (compliant with electronic-record specifications). This compliance cost makes the switching cost of vision systems in the pharmaceutical industry extremely high, and the technical-certification-cycle advantage of foreign SDKs in pharmaceutical scenarios maintains a relatively stable market share for them. The best path for domestic SDKs to enter pharmaceutical scenarios is to accumulate compliant-deployment cases among local small and medium pharmaceutical enterprises, then gradually penetrate large pharmaceutical enterprises.

VII. Food: High-Speed Vision Inspection Driven by Safety Regulations

Machine vision applications in food manufacturing have grown rapidly in recent years, driven on one hand by food-safety regulations (China's Food Safety Law requires 100% foreign-matter inspection for key food categories), and on the other hand by consumers' increasingly high expectations for the consistency of food appearance quality.

The main application scenarios of food vision inspection include: high-speed sorting of appearance defects of biscuits and candies (missing corners, cracks, color differences, foreign-matter contamination), fruit grading (comprehensive scoring of color, size, and surface defects), weight and shape-consistency inspection of meat products, and seal-quality and packaging-text recognition of bagged food.

The special challenges of food scenarios are: first, product diversity (the same production line may produce dozens of SKUs, each with different appearance standards, and the algorithm needs to quickly switch inspection parameters); second, high-speed production lines (biscuit production-line speeds can reach 400 pieces per minute, requiring image acquisition and algorithm processing to complete single-piece inspection within 150 milliseconds); third, the visual variability of food materials (the surface texture of natural ingredients has a large amount of "normal variation," and the algorithm needs to learn to distinguish normal variation from real defects).

The value of deep-learning AI vision in food scenarios lies in its ability to learn "normal variation"—through anomaly-detection models, the system learns the visual distribution of normal products and marks appearances beyond the normal distribution as anomalous, without needing to separately label samples for each defect type. This capability makes AI vision more adaptive than traditional rule algorithms in food quality inspection, and is the technical differentiation point for AI vision vendors (SMore, AInnovation) to enter the food industry.

VIII. Photovoltaic: The Emerging Battlefield of Silicon-Wafer and Module Inspection

China is the absolute center of global photovoltaic manufacturing, with the combined capacity of silicon wafers, cells, and modules accounting for about 80% of the global total. The vision-inspection demand of photovoltaic manufacturing, with the mass production of new high-efficiency battery technologies such as HJT and TOPCon, is showing technical demands distinctly different from those of traditional PERC cells.

The main applications of photovoltaic vision inspection include: silicon-wafer microcrack (EL electroluminescence imaging + vision algorithms), scratch, and chip detection; cell grid-line printing-quality inspection (grid-line breaks, false printing, displacement); and module welding quality (interconnection-ribbon welding pores, offset), lamination bubbles, and appearance-defect detection.

The double-sided structure and ultra-thin silicon wafers (below 100 microns) of HJT cells make their precision requirements for vision inspection significantly higher than traditional PERC; the grid-line width is only 25–30 microns, and the defect-detection resolution needs to reach within 10 microns. This accuracy requirement makes the matching combination of high-end cameras (high-resolution line-scan cameras) and professional algorithm SDKs a necessity, further driving the market opportunity in the photovoltaic track for domestic SDK vendors with high-precision vision-inspection capabilities such as Lingkong.

In 2025, within Lingkong's new-energy business revenue of 185 million yuan, photovoltaic inspection contributed a considerable portion of the increment, an important component of Lingkong's new-energy strategy. As the photovoltaic industry chain continues to evolve toward next-generation technologies such as perovskite and tandem cells, the requirements for vision-inspection accuracy will further increase, providing continuous technical-upgrade demand for high-end vision SDKs.

IX. The Regional Distribution and Concentration of Downstream Applications

The regional distribution of China's machine vision downstream applications is highly concentrated in core manufacturing city clusters. By sorting out the distribution of typical factories in the machine vision downstream, a "vision-demand heat map" can be constructed:

Shenzhen-Dongguan axis: the world's largest 3C electronics manufacturing center, with the supply-chain factories of Apple, Huawei, and Xiaomi densely distributed. Vision demand is dominated by appearance inspection and precision positioning, with the highest AI vision penetration rate nationwide (about 35%), and is the main revenue-source region for AI vision enterprises such as SMore and AInnovation.

Suzhou-Wuxi-Ningbo axis: oriented toward automotive parts (Tier 1 factories such as Bosch, Continental, and Joyson Electronics are densely clustered in Suzhou) and precision manufacturing. Vision demand focuses on precision measurement (automotive sheet metal, precision mechanical parts), and HALCON has the highest penetration rate in this region nationwide, making it one of the regions with the greatest resistance in the domestic SDK replacement process.

Ningde-Changzhou axis: the region with the most concentrated power-battery capacity in China (CATL, SVOLT, CALB). Vision demand centers on lithium-battery inspection, and it is one of the regions with the greatest new-revenue contribution for Lingkong and Huaray Technology in 2023–2025.

Chengdu-Chongqing axis: the most important manufacturing center in central-western China, with automotive (Changan Automobile, Geely) and electronics manufacturing (BOE, Tianma Micro-electronics) both prominent. Vision demand is diverse, making it an important strategic region for domestic SDK vendors to establish share in the western market.

This regional distribution means that the market-promotion and technical-service networks of machine vision SDK vendors must cover these few core manufacturing clusters; otherwise they will find it hard to reach the main demand sources. Both Hikrobot and Lingkong have already established local technical-service teams in the above cities, which is the practical advantage they embody when competing with foreign SDKs (which only have technical teams in Beijing and Shanghai, with limited coverage).

---

Chapter 6　Inventory of Mainstream Players (Domestic + Overseas, by Classical / AI / 3D Tiers)

I. Overseas Vendors

MVTec HALCON — The "Encyclopedia" of Industrial Algorithms

MVTec Software GmbH was technology-transferred from the Technical University of Munich in 1996, and currently operates as an independent enterprise under Japan's Omron, headquartered in Munich, Germany. HALCON is its flagship product and one of the world's most complete industrial vision algorithm libraries, enjoying more than 25 years of market-validation accumulation in the industrial vision field.

HALCON's core technical advantages lie in algorithm breadth (2,100+ algorithm operators, covering all categories of image processing, feature extraction, 3D vision, deep learning, etc.), cross-platform deployment capability (Windows, Linux, macOS, ARM embedded), and deep engineering reliability (more than 25 years of industrial on-site validation, covering almost all mainstream manufacturing scenarios). Its latest version (HALCON 24.11) has deeply integrated a deep-learning inference engine, supporting ONNX-format model import, achieving the hybrid invocation of traditional algorithms and AI models within the same workflow. MERLIC is the low-code configuration interface launched by MVTec, lowering the barrier for non-programming users to build HALCON solutions.

HALCON holds several core patents in Shape-based Matching, and has historically maintained its market position through patent-protection strategies. Its algorithm accuracy and stability remain the industry-recognized benchmark in high-requirement scenarios such as automotive measurement and semiconductor inspection.

Positioning: an all-around SDK for classical algorithms + 3D vision, with relatively high prices (runtime about 15,000 yuan per set), still the de facto standard in high-precision measurement and semiconductor inspection.

Cognex VisionPro — The Pioneer of Software-Hardware Integration

Cognex Corporation was founded in 1981 and is the world's largest listed company focused on machine vision, headquartered in Massachusetts, USA. VisionPro is its SDK product oriented toward system integrators, forming a high-low complementary product matrix with In-Sight smart cameras (foolproof products for terminal factories).

VisionPro's deep-learning module ViDi Suite covers the three major AI vision tasks of defect classification (Classify), target localization (Locate), and character reading (Read), all supporting the "Interactive Training" approach—engineers only need to label a small number of samples in the software interface (usually 50–200 images), and the software automatically completes model training without writing any code. This low-barrier training experience is the key path for VisionPro to popularize deep learning in manufacturing, and is also the main differentiation point from pure-tool SDKs (HALCON requires more Python/C++ programming ability).

FY2025 Q3 single-quarter revenue was USD 277 million, up 18% year-on-year, with consumer-electronics automation and logistics sorting as the main growth sources, and China-market growth about 9%.

Positioning: the leader in popularizing deep-learning vision applications, with "Interactive Training" lowering the AI vision barrier, and an obvious trend toward software subscriptions.

Matrox MIL — Exclusive to High-End Industrial and Semiconductor

The Matrox Imaging Library (MIL) is developed by the vision division under Canada's Matrox, positioned for high-precision industrial measurement and semiconductor inspection, distinguished by the collaborative coordination of high-frame-rate frame grabbers (Matrox's self-developed FPGA-accelerated acquisition cards) and precision algorithm libraries. MIL's algorithms have a good reputation in scenarios with extremely high stability and accuracy requirements such as pharmaceutical-injection clarity inspection and semiconductor-package precision metrology, but its market promotion is relatively low-key, and its engineer-community scale and activity are inferior to HALCON.

Positioning: high-end niche, a professional player in semiconductor/pharmaceutical scenarios, deeply bound to Matrox high-end acquisition cards.

OpenCV — Open-Source Foundation, the Guardian of the Industrial Boundary

OpenCV, as an open-source computer vision library, is a special existence in the discussion of algorithm SDKs. It is not a commercial SDK, but because it is free and open, it has become one of the most widely used code foundations in the industrial vision field. OpenCV's industrial limitations are obvious: it lacks an industrial-device driver layer (GenICam is not natively supported), cannot perform enterprise-level license management, and has no guarantee in real-time stability and technical support. But when building lightweight vision prototypes, low-cost automation, or as a complementary tool to commercial SDKs, OpenCV is indispensable, making it a general-purpose underlying tool for domestic AI vision platforms to perform algorithm validation and academic research.

II. Domestic Vendors — Classical Algorithm Tier

Hikrobot VisionMaster — The Domestic Number One's All-Around Platform

Hikrobot was founded in 2016 and belongs to the Hikvision group; it is currently the vendor with the largest user scale and broadest market coverage among domestic machine vision software platforms. VisionMaster's core competitiveness is the combination of ecosystem scale and full-scenario coverage.

By the end of 2025, Hikrobot's self-developed industrial software had more than 600,000 license user instances, more than 20,000 customers served globally, and a compound annual growth rate of about 50% for machine vision-related products. VisionMaster 5.0, on the basis of traditional algorithm tools (more than 300 types, covering all categories of positioning, measurement, defect detection, and OCR), layers an industrial vision large model (handling complex, changeable scenarios that are difficult to describe with rules) and edge-learning tools (handling simple, high-speed scenarios with clear rules), forming a three-layer hybrid architecture of "large model + edge learning + traditional algorithms," currently the most comprehensive product form among domestic SDKs.

In terms of hardware, Hikrobot's MV-series industrial cameras cover the full price range from low-end entry (thousand-yuan level) to high-end high-speed (above ten thousand yuan), and their deep integration with the VisionMaster SDK forms a "camera plug-and-play" zero-configuration development experience, greatly reducing the debugging cost for integrators. For the full year 2025, Hikrobot's overall revenue exceeded 6.4 billion yuan (machine vision and robotics businesses combined).

Positioning: the benchmark of domestic SDKs, with full-scenario coverage, software-hardware integration, and the most mature developer ecosystem.

Lingkong VisionWARE — The Precision-Manufacturing Expert of Optics + Algorithms

Lingkong is one of the few domestic vendors capable of deep cultivation in both optical design (lenses, light sources, color-correction systems) and algorithm SDK dimensions. The company was founded in 2000, led by founder Yi Yao, focusing on precision optics and vision-inspection technology for more than 23 years. VisionWARE version 6.4 has accumulated 18 algorithm libraries and nearly 200 algorithm tools, covering all categories of geometric positioning, precision measurement, character recognition, color inspection, defect detection, 3D vision, and deep learning.

The F.Brain general-purpose vision large model is Lingkong's key strategic investment over the past two years, targeting the "complex texture + subtle defect" problems that traditional algorithms find hard to handle (subtle color differences in printed matter, fine broken threads in fabric, subtle scratches on screen surfaces), and has completed large-scale commercial deployment in multiple industries in 2025. Lingkong's printing-quality inspection products have a particularly outstanding industry position; its Color Science capability comes from more than 20 years of independent R&D accumulation, forming a technical barrier in the printing and packaging industry that competitors find hard to replicate in the short term.

2025 revenue was 2.912 billion yuan, up 30.35% year-on-year; net profit attributable to the parent was 161 million yuan, up 50.70% year-on-year; net profit attributable to the parent excluding non-recurring items was 123 million yuan, up 86.05% year-on-year; R&D investment accounted for over 15% of revenue (first three quarters of 2025). The new-energy business was 185 million yuan, up 36.01% year-on-year, the company's fastest-growing niche direction.

Positioning: dual-wheel-drive of optics + algorithms, the industry-depth champion of printing/consumer electronics/new energy, and the listed vision-algorithm-platform enterprise with the best financial quality.

Daheng Imaging — A Veteran Player, with Industrial Camera and SDK Synergy

Daheng Imaging is one of China's earliest vendors to lay out the coordinated development of industrial cameras and vision software, with more than 25 years of machine vision accumulation. Its intelligent image-processing software SmartView is oriented toward mid- and low-end system integrators, possessing a complete image-acquisition driver layer (compatible with Daheng's self-developed industrial cameras) and a basic vision-algorithm toolkit, with a stable customer base in scenarios such as printing, semiconductor back-end, and pharmaceuticals. Daheng Imaging's SDK has a certain gap from HALCON and VisionMaster in algorithm richness, but has strong historical stickiness within the customer group deeply bound to Daheng cameras. Daheng Imaging's development path reflects the positioning of "single-field specialization + hardware matching," rather than a full-scenario challenge.

Positioning: mid- and low-end market, deeply bound to self-developed cameras, with a stable foundation in traditional industries such as pharmaceuticals and printing.

Huaray Technology (under Dahua) — The Exclusive Weapon for Lithium Battery and 3C

Huaray Technology is a subsidiary under Dahua focused on the machine vision business, with deep layouts in the two major application scenarios of lithium battery and 3C. Its vision platform supports Python-script secondary development, with built-in lithium-battery-dedicated algorithm packages (electrode-sheet inspection, welding-quality analysis) and 3C-dedicated algorithm packages (connector PIN inspection, glass-cover-plate scratch detection). Its 3D camera's point-cloud density reaches 1,000 points per square millimeter, with measurement accuracy up to ±0.05 mm, making it one of the domestic SDKs with relatively high quantitative metrics in the precision-3D-measurement direction. Relying on the Dahua group's channel resources and brand endorsement, Huaray Technology has good coverage among mid-sized system integrators.

Positioning: exclusive to lithium battery + 3C, with high measurement accuracy, obvious synergy within the Dahua ecosystem, and relatively outstanding 3D vision capabilities.

III. Domestic Vendors — AI Vision Tier

SMore — The Flagship of Industrial AI Agents

SMore was founded in 2019 by computer-vision scientist Dr. Jiaya Jia (former chief researcher at Microsoft Research Asia, professor at the Hong Kong University of Science and Technology), and is one of the most representative unicorns in China's industrial AI vision field, and also an important research subject of this report.

Its product system centers on an "industrial AI vision large model + no-code configuration platform," covering core quality-inspection scenarios such as appearance-defect detection, dimensional measurement, and character reading. The technical differentiation is reflected in three dimensions: first, few-shot learning capability (completing AI model training with 20–50 samples, greatly reducing labeling cost compared with the thousands of samples required by traditional deep learning); second, cross-scenario transfer (a model trained at one factory can be migrated to other factories of the same product type with minor adaptation, avoiding labeling from scratch with each deployment); third, no-code configuration (production-line engineers independently configure AI inspection solutions through a graphical interface, without needing a programming or algorithm background).

2025 revenue was nearly 1.1 billion yuan, with more than 730 customers, covering global manufacturing leaders such as Tesla, Luxshare Precision, BOE, Carl Zeiss, and CRRC. In February 2026, it completed Series C financing with a valuation of USD 1.230 billion (about 8.5 billion yuan); in March 2026, it officially submitted a listing application to the main board of the Hong Kong Stock Exchange, intending to become the "world's first industrial AI agent stock," with sponsors Morgan Stanley, CICC, and Deutsche Bank. From 2023 to 2025, the adjusted net loss narrowed from 394 million yuan to 272 million yuan, with a marked improvement in cash burn.

Positioning: the AI vision flagship, with industrial large model + no-code configuration, a globalized brand oriented toward head manufacturing customers, with a Hong Kong IPO in progress.

AInnovation (AInnovation, 2121.HK) — The Value Realizer of AI Manufacturing

AInnovation was listed on the Hong Kong Stock Exchange in 2022 (code 2121.HK), and is currently the only independently operated listed company in the domestic industrial AI vision track, and an important capital-market reference. Its strategic core is "one model, one body, two wings"—with the self-developed AInnoGC industrial large model as the base, AI agents as the engine, and industrial robots and industrial software as the dual application wings, deeply cultivating AI penetration in manufacturing niches such as new energy, 3C electronics, and steel.

FY2025 first-half revenue was 699 million yuan, up 22.3% year-on-year; the "AI + manufacturing" business accounted for nearly 80% of total revenue, at 556 million yuan; the adjusted net loss was only 6.68 million yuan, a year-on-year decrease of 82.1%, showing a marked profitability-improvement trend. According to IDC data, AInnovation ranks seventh in China's large-model application market share, and is the only vendor focused on the industrial field, with a clear differentiated positioning.

In terms of technical route, AInnoGC is a foundation large model built specifically for industrial tasks such as industrial defect detection, quality analysis, and equipment maintenance, distinct from general-purpose large models (ChatGPT, ERNIE Bot, etc.), having accumulated training data of more than a million labeled industrial images across dozens of industrial vertical scenarios, with strong industrial-knowledge generalization capability.

Positioning: the listed industrial-AI benchmark, with a profitability inflection point approaching, one of the strongest in industrial-large-model commercial deployment, and high industry-reference value.

Megvii Industrial Vision — Consumer AI Genes, Industrialization Path Still Being Explored

Megvii (MegVII) has deep accumulation in consumer-grade computer vision (facial recognition, smart security), and industrial vision is a direction it has been focusing on expanding in recent years. Megvii's industrial vision products focus on smart logistics (warehouse-robot vision navigation, goods recognition) and factory quality inspection (line-end appearance verification based on object detection), leveraging its deep research accumulation in YOLO-series object-detection algorithms to develop factory application scenarios.

However, constrained by the historical path dependence of its business focus and relatively insufficient industrial-AI engineering experience, Megvii's industrial vision progress in 2025 was relatively conservative, and the industrialization maturity of some product lines still has a gap compared with SMore and AInnovation. Megvii Industrial Vision is still looking for the breakthrough point most suitable for its technical background in its choice of main track (whether to focus on warehousing logistics or deeply cultivate manufacturing quality inspection).

Positioning: strong consumer-vision AI genes, industrialization and engineering capabilities yet to be deepened, with a certain foundation in logistics-vision scenarios.

SenseTime Industrial — The Factory-Digitalization Solution Provider of AR + AI

SenseTime's industrial business line, with the "AI + AR" technology combination, has made significant breakthroughs in the factory-digitalization field of the automotive industry. The factory-digitalization orders from BMW and Tesla, each exceeding 300 million yuan, demonstrate SenseTime's competitive strength in high-unit-price, highly customized industrial-AI projects. SenseTime Industrial's technical route leans toward "AR-assisted assembly + AI quality inspection + digital twin," oriented toward whole-plant digitalization transformation for large manufacturing enterprises, differing from the positioning of pure vision-algorithm SDKs, belonging to a sales logic of high-unit-price, low-frequency large projects.

Positioning: an AR + AI factory-digitalization solution provider, with a high-unit-price big-customer strategy, not in the general vision-SDK competition dimension.

IV. Domestic Vendors — 3D Vision Tier

Mech-Mind (Mech-Mind) — China's Number One in 3D Guided Robotics

Mech-Mind was founded in 2016, focusing on 3D vision-guided industrial robots. Its product system consists of four modules: the Mech-Eye 3D industrial camera, Mech-Vision (3D vision-sensing SDK), Mech-DLK (deep-learning development toolkit, for labeling and model training), and Mech-Viz (robot path-planning software), achieving a complete software pipeline from 3D sensing to robot execution.

Third-party data show that Mech-Mind has ranked first in China's 3D vision-guided industrial robot market for five consecutive years (2020–2024), and has served customers in more than 50 countries and regions worldwide. Its Mech-Eye 3D camera achieved a depth-estimation technology breakthrough for metal reflective objects in 2024, improving point-cloud accuracy by about 90% through multi-structured-light-band fusion and adaptive exposure control, solving the long-standing pain point of 3D vision guidance for metal castings and stainless-steel parts.

Mech-Mind has completed deep adaptation (standardized robot communication interfaces, with robot drivers built into the Mech-Mind SDK) with mainstream robot brands such as ABB, FANUC, Yaskawa, KUKA, Fanuc, and Estun, so when integrators choose Mech-Mind solutions, they do not need to additionally develop a robot communication layer, greatly reducing project-development difficulty and cycle.

Positioning: the technology leader in 3D vision-guided robotics, the domestic vision SDK vendor with the most complete globalized layout, and a core beneficiary under the embodied-intelligence wave.

Orbbec (Orbbec) — From ToF Vision Sensors to an AI Vision Platform

Orbbec focuses on 3D vision sensors (ToF and structured light) and their SDKs, with layouts in both consumer-grade (human-computer interaction, AR/VR) and industrial-grade (robot vision, warehousing and sorting, embodied intelligence) directions. In 2025, it achieved its first full-year profit since listing (net profit attributable to the parent of 128 million yuan), the core driver being the demand explosion for 3D vision sensors from embodied-intelligence robots and AI terminal devices—the demand of humanoid robots (Unitree, AgiBot, etc.) for low-cost, real-time, low-power 3D sensing happens to be the core technical direction of Orbbec's SDK.

Orbbec's SDK is positioned at the sensing layer (depth maps, point-cloud data, and calibration tools), and needs to cooperate with upper-layer application software (such as ROS, Mech-Vision, SMore, etc.) to jointly form a complete robot vision stack.

Positioning: a 3D-sensor sensing-layer SDK, the vision base for embodied-intelligence robots, with outstanding cost advantages, and dual-wheel drive of industrial and consumer.

VII. The Impact of the Open-Source Ecosystem: The Coexistence Logic of OpenCV and Industrial SDKs

OpenCV is an unavoidable backdrop when discussing commercial SDKs in the industrial vision field. As the world's most-used open-source computer vision library, OpenCV's existence constitutes both competitive pressure on commercial SDKs and a subtle coexistence relationship.

From the direct-competition level, OpenCV provides basic image-processing functions (filtering, morphological operations, basic feature extraction), and for technically strong teams, it is entirely possible to implement specific vision-inspection functions based on OpenCV without purchasing commercial SDK licenses. Among universities and research institutions, and startups with strong software-development capabilities, OpenCV has indeed diverted part of the market that might originally have belonged to commercial SDKs.

However, OpenCV's limitations are very obvious in industrial manufacturing scenarios: first, insufficient industrial-grade stability and real-time guarantees—OpenCV's function library is optimized for general computer vision and is not specifically designed for industrial scenarios' high-real-time (millisecond latency) and high-reliability (7×24-hour non-stop) requirements; second, industrial-device compatibility—OpenCV does not provide native support for the GenICam / GigE Vision protocols, and integration with industrial cameras requires additional driver-adaptation work; third, technical support and engineering services—OpenCV is maintained by the open-source community, with no commercial-grade technical support service, which is hard to accept for system integrators that promise customers strict SLAs (Service Level Agreements); fourth, algorithm accuracy—the industrial-dedicated algorithms in HALCON and VisionMaster (such as shape matching optimized for metal-surface illumination changes, and parallel-processing architectures designed specifically for high-speed lines) significantly outperform OpenCV's general implementations.

Therefore, OpenCV and commercial SDKs have formed a clear layered coexistence in the actual market: OpenCV occupies research, prototype validation, and low-complexity inspection applications; commercial SDKs occupy production environments that require engineering-grade stability, high-precision algorithms, and complete technical support. The boundary between the two is not fixed, but as domestic SDK prices continue to decline (some versions of VisionMaster even offer free versions), the economic resistance to migrating from OpenCV to commercial SDKs continues to decrease, and the space for OpenCV to replace commercial SDKs is narrowing.

VIII. Regional Competitive Differences: The SDK Landscape in the South China, East China, and North China Markets

The machine vision SDK markets in China's different geographic regions have noteworthy competitive-landscape differences, directly related to each region's industrial structure and dominant industries.

South China market (centered on Shenzhen-Dongguan): the absolute dominant region of 3C electronics manufacturing. AI vision (deep-learning defect detection) has the highest penetration rate nationwide, and is the most important revenue-source area for SMore and AInnovation. Hikrobot's share in South China is growing rapidly, mainly entering consumer-electronics foundry scenarios; HALCON still occupies precision-optical inspection and semiconductor-related applications.

East China market (centered on Suzhou-Wuxi-Nanjing-Ningbo): dominated by automotive parts and precision manufacturing. The region with HALCON's deepest historical accumulation—the technical specifications of foreign automotive-parts enterprises such as Bosch, Continental, and ZF usually specify HALCON or Cognex, and these specifications are transmitted to their Chinese suppliers, forming the deep dependence of the East China automotive-supporting ecosystem on foreign SDKs. The progress of domestic SDK replacement in East China is significantly slower than in South China, making it the main battlefield for domestic SDK market breakthroughs.

North China market (centered on Beijing-Tianjin): dominated by equipment manufacturing and semiconductor-related applications. It has the highest technical requirements for SDKs; semiconductor inspection (still foreign-dominated) and military-equipment manufacturing have extremely stringent requirements for vision accuracy and stability, with the slowest progress of domestic replacement, but it is also a long-term opportunity region with considerable potential market size.

---

Chapter 7　The Chromatic Aberration of Domestic Replacement: An Industry-Database Perspective

I. The Three Tiers of Domestic Replacement

Among the 4.8 million producing factories covered by this platform, the penetration rate of machine vision equipment shows significant stratification according to factory scale and degree of automation. Through mining and analysis of the industry-database data, the three tiers of the domestic vision-algorithm-SDK replacement process can be clearly seen, with each tier having significant differences in replacement depth, source of resistance, and expected timeline.

Tier 1: 3C electronics + new energy, replacement basically complete

In the two major tracks of consumer electronics and new-energy batteries, the market share of domestic SDKs (VisionMaster, VisionWARE, Huaray Technology solutions) has reached as high as 70%–80%. The common features of these two tracks are: large factory scale (single-factory vision-system procurement volume is usually above one million yuan), fast iteration speed (production-line design is frequently updated with product generations, and each model change is an opportunity for SDK evaluation), and young engineer teams with high acceptance of new technology (willing to evaluate domestic SDKs and conduct technical validation).

In this tier, domestic SDKs have not only achieved price replacement (combined cost 40%–60% lower than HALCON), but have also formed technical overtaking in dedicated algorithms (scarce-sample learning for lithium-battery electrode-sheet defect detection, adaptive light-source algorithms for highly reflective glass scratches), making the likelihood of reverting to foreign SDKs extremely low. From the time-node perspective, the main replacement work of this tier was completed between 2022 and 2024, and it is currently in the deepening stage (from "usable" to "used better").

Tier 2: automotive + general manufacturing, replacement in progress, with significant resistance

In the automotive and general mechanical-manufacturing fields, the penetration rate of domestic SDKs is about 37%–45%, with foreign brands still dominant. Replacement resistance comes from two aspects: first, the procurement specifications of OEMs and Tier 1 suppliers have specified requirements for specific brands, and changes to these specifications require systematic revision of quality-system processes; second, high-precision measurement (automotive sheet-metal accuracy at the ±0.1 mm level, even ±0.05 mm for key features) places extremely high requirements on the SDK's sub-pixel positioning accuracy, temperature-drift stability, and systematic-error management capability, and although domestic SDKs are already close, they have not yet comprehensively surpassed on all key dimensions. It is expected that in 2027–2028, as more domestic SDKs complete formal certification by automotive vehicle manufacturers, the replacement rate of this tier will rise to 55%–60%.

Tier 3: semiconductor, replacement just starting

In semiconductor (front-end wafer inspection) scenarios, foreign dedicated inspection systems still hold absolute dominance, and the penetration rate of general-purpose SDKs is extremely low (less than 5%). The current entry point of domestic vision SDKs is concentrated in the semiconductor back-end (solder-ball inspection, pin-bending inspection, package appearance), with a penetration rate of about 15%–20% in this niche. Front-end-process replacement is expected to require at least 5–8 years of technical accumulation and a customer-validation cycle, making it the last and most arduous mile in the entire machine vision industry's domestic-replacement path.

II. The Industry Database: The Vein of Vision Demand Under the Manufacturing Digital Map

Through searches of machine vision-related keywords in the industry database, the distribution vein of industrial vision demand on the geographic map of Chinese manufacturing can be outlined, providing industry participants with granular insights beyond macro statistics.

There are 247 records of factories related to machine vision (including complete machines and software), more than 204 records related to vision robots, more than 116 factories related to industrial vision software, and more than 118 factories related to industrial cameras. These data are concentrated in the three major manufacturing belts of the Pearl River Delta (Shenzhen, Dongguan, Guangzhou), the Yangtze River Delta (Suzhou, Shanghai, Hangzhou, Ningbo), and the Bohai Rim (Beijing, Tianjin, Langfang), highly overlapping with China's electronics-manufacturing and automotive-manufacturing industrial clusters.

From the niche-direction perspective, related records in the defect-detection direction reach 37, the AI vision direction 18, industrial-large-model-related 12, point-cloud processing 5, and 3D vision algorithms 1—showing that the penetration of AI vision and industrial-large-model demand in manufacturing is still low, still on the eve of large-scale explosion. This data indirectly confirms that the growth ceiling of the algorithm SDK market is still far from being reached, especially on the penetration path of AI vision from "pilots at head large enterprises" to "large-scale popularization among small and medium manufacturers," where there is still huge space to be developed.

From the brand side, Huaray Technology (45 records), Lingkong (6 records), Hikrobot (2 records), Orbbec (1 record), Daheng Imaging (1 record), SMore (1 record), Cognex (3 records), and Matrox (1 record) all have factory-coverage records in the database, showing that the actual penetration of the above brands in Chinese manufacturing has reached a certain scale. Huaray Technology's record count is clearly in the lead, reflecting the deep coverage of the Dahua ecosystem in Chinese manufacturing, having formed a significant channel advantage especially in the supply chains of lithium-battery and 3C scenarios.

III. The Deep Logic of Domestic Replacement—Three Driving Forces

The replacement of domestic algorithm SDKs is not a story of a "cheaper HALCON," but the result of the superposition of three logics, each playing a role on a different time dimension.

The first logic: gradual release of price sensitivity. China's manufacturing system integrators and terminal factories, after the sustained pressure on manufacturing profit margins since 2018, have significantly increased their sensitivity to software-licensing costs. HALCON's runtime license of about 15,000 yuan per set, in a factory deploying 50 vision stations, would cost more than 750,000 yuan in software licensing alone. Domestic SDKs, with equivalent functional coverage, can compress the combined software cost to 300,000–450,000 yuan, and this saving makes a direct positive contribution to the ROI calculation of a manufacturing factory's vision system, often shortening the payback period of the domestic solution by 30%–50%.

The second logic: AI vision lane-changing overtaking. In the classical 2D vision field, HALCON, by virtue of more than 25 years of algorithm accumulation and global case validation, maintains a technical-leadership position that is hard to shake. But AI vision (deep-learning defect detection, industrial anomaly detection, industrial large models) is a brand-new track that was only truly commercialized after 2018, and in this new track, China's AI-native enterprises (SMore, AInnovation) stand at basically the same starting line as MVTec and Cognex. And because they are more familiar with the production scenarios of Chinese manufacturing (equipment parameters, defect-type distribution, engineering-delivery needs), Chinese AI vision enterprises have achieved curve-overtaking in specific scenarios (few-shot defect detection, complex-texture anomaly recognition)—this is a "technical overtaking" rather than a "low-price copy," the most sustainable form of value creation in domestic replacement.

The third logic: the natural advantage of the local service system. The full lifecycle of an industrial vision system requires continuous on-site debugging, algorithm optimization, and emergency fault-response support. Foreign SDKs' China-service capability is constrained by team scale and language-communication cost; domestic SDKs can respond to a factory's technical needs within 24 hours and dispatch engineers to handle problems on site, which is a hard competitiveness that cannot be replaced by price factors in manufacturing scenarios with extremely high production-stoppage costs (a stopped production line may lose more than a million yuan per hour).

IV. Quantitative Analysis of Replacement Bottlenecks

From the technical-capability dimension, the gap between domestic SDKs and HALCON is concentrated in the following three directions, with different breakthrough timelines:

Sub-micron precision measurement: HALCON's engineering accumulation in optical-calibration accuracy and thermal-drift compensation still gives it about a 15%–20% accuracy advantage in measurement scenarios at the ±0.01 mm magnitude; domestic SDKs are expected to require 3–5 years of continuous engineering optimization to catch up on this dimension.

3D vision semiconductor applications: nanometer-level point-cloud accuracy (used for wafer-surface micro-topography reconstruction) is a challenge currently faced jointly by domestic SDKs and domestic 3D sensors, requiring coordinated breakthroughs of the sensor (light source, CMOS sensor design) and the algorithm (sub-nanometer depth estimation), with a timeline of 5–8 years.

Multi-language cross-platform ecosystem: HALCON supports multi-language bindings such as C++, C#, Python, and Delphi, and has complete test support on Windows, Linux, macOS, and various ARM embedded platforms. The cross-platform capability of domestic SDKs is rapidly catching up, but their support maturity on Linux embedded and macOS is still weaker than HALCON, which constitutes a certain obstacle for domestic SDKs facing multi-OS global customers when going overseas.

V. Algorithm-Accuracy Stress Testing: Quantitative Evaluation and Methodology of Domestic SDKs

The most reliable way to understand the gap between domestic SDKs and HALCON is documented quantitative evaluation, rather than vendors' marketing numbers. The following sorts out the public evaluation methods and conclusions already available in the industry, for engineers' reference:

Shape-matching accuracy test: using a standardized test image set (containing target objects with different rotation angles, illumination changes, and partial occlusions), test the matching success rate (true-detection rate) and false-detection rate (false-detection rate) of each SDK when configured with the same algorithm parameters. Typical tests show that VisionMaster's shape matching already has a success rate on par with HALCON in the 0°–180° rotation range, but in extremely low-contrast (contrast below 15%) scenarios, HALCON's shape-matching success rate is usually 5%–10% higher. This gap has no practical impact in the vast majority of industrial scenarios (contrast usually above 30%), but still constitutes a perceptible difference in extremely low-contrast scenarios such as semiconductor-wafer mark detection.

Sub-pixel measurement-accuracy test: with a 10-micron-pitch calibration board as the measurement object, at a camera resolution of 5 megapixels, test the systematic error (bias) and measurement repeat accuracy (standard deviation) of each SDK in a single measurement. Public data show that HALCON's caliper tool can achieve 0.01-pixel-level repeat accuracy under standard industrial conditions; VisionMaster and VisionWARE can reach 0.02–0.03 pixels under equivalent hardware conditions, a gap of about 2–3 times. For industrial scenarios with dimensional-tolerance requirements of ±0.1 mm or above (covering most machining and consumer electronics), this accuracy gap has no practical engineering significance; but for precision measurement with ±0.01 mm (i.e. 10 microns) accuracy requirements (such as semiconductor bond-wire arc height, watch hairspring thickness), the gap constitutes a key constraint for selection.

Deep-learning inference-speed test: using an industrial-standard dataset, compare the inference latency of the deep-learning inference engine integrated in each SDK under two hardware conditions: GPU (NVIDIA RTX 4070) and CPU-only. The test conclusion shows that because each SDK is based on a general-purpose inference backend such as ONNX Runtime or TensorRT, the GPU inference-latency gap is within ±15%, belonging to the scope of engineering tuning rather than essential architectural difference. In CPU-only mode, SMore's SMore InspectKit, due to self-developed inference optimization, is about 20%–30% faster than competitors based on standard ONNX Runtime, which has practical significance in embedded industrial-computer deployment scenarios without a GPU.

Multi-camera synchronization-accuracy test: in an eight-camera stereo-vision scenario (measuring the 3D coordinates of a large workpiece), test the camera synchronization-trigger accuracy (time-window difference) driven by the SDK. Years of optimization of HALCON's GenICam driver layer make its synchronization accuracy usually better than 1 microsecond; the synchronization accuracy of domestic SDKs is usually in the 5–20 microsecond range, and for multi-camera stereo reconstruction (sub-millimeter accuracy requirements), the synchronization-accuracy difference directly affects the point-cloud accuracy of the 3D reconstruction result. This is a gap that needs to be evaluated as a priority when selecting a 3D inspection system.

The engineering value of the above evaluation methods is: to help engineers focus on "the gaps that truly affect my scenario" during selection, rather than being misled by the vendor-promoted comprehensive ranking. For most 3C inspection, lithium-battery inspection, and printing-inspection applications, the performance gap between domestic SDKs and HALCON is already within the actual engineering-tolerance range; for extreme-accuracy scenarios such as semiconductor precision measurement, the gap still truly exists and needs to be quantitatively evaluated at the project-initiation stage.

VI. The Documentation Quality and Developer Experience of Domestic SDKs

Product documentation and Developer Experience (DX) are dimensions often underestimated in the commercial competition of SDKs, but in fact have a huge impact. The efficiency of engineers' daily development depends largely on documentation completeness, example-code quality, and technical-support response speed, rather than purely on algorithm-performance parameters.

HALCON's documentation advantage: MVTec's HALCON documentation system is the recognized highest standard among industrial vision SDKs. Complete algorithm-parameter descriptions, the mathematical definition of each algorithm function, rich application examples (including complete runnable HDevelop scripts), and Chinese-localization documentation specifically for the Chinese market (the Chinese version is usually updated with a delay of less than 3 months)—these together constitute the knowledge foundation of the HALCON engineer community. HALCON's official example-code library contains more than 700 categorized cases, covering the complete technical path from basic shape matching to complex 3D calibration.

The documentation status of domestic SDKs: Hikrobot VisionMaster's documentation quality is at the forefront among domestic SDKs, with complete algorithm-tool descriptions, but still has a gap from HALCON in depth—mainly reflected in insufficient boundary-condition descriptions (the behavior when a certain parameter is set to an extreme value) and missing mathematical-principle descriptions (only explaining "how to use," not "why"). The documentation quality of Lingkong VisionWARE is uneven, with detailed documentation for printing-inspection-related modules (Lingkong's deep historical accumulation), but insufficient documentation depth for modules such as general vision measurement. SMore's technical documentation focuses on operation instructions for the cloud SaaS platform, and the completeness of the SDK-interface documentation (oriented toward engineers who need secondary development) needs improvement.

Technical-support response speed: this is one of the dimensions where domestic SDKs truly achieve "overtaking" of foreign products. The domestic technical-support teams of Hikrobot and Lingkong can usually respond to technical questions within 4 hours, and when engineers encounter algorithm-configuration difficulties, they can directly contact technical-support personnel via WeCom; HALCON's technical-support response cycle for non-agent-channel users usually takes 1–3 working days, and the language barrier (engineers need to describe technical problems in English) further increases the communication cost. This service-response difference is a substantive advantage of domestic SDKs frequently mentioned in the small and medium system-integrator market, and is also the most concrete and tangible non-technical factor driving domestic replacement.

---

Chapter 8　Price Tiers and Business Models (Perpetual License / Subscription / Per-Device Licensing)

I. The Historical Evolution of Three Mainstream Business Models

The business model of machine vision algorithm SDKs has undergone a process of evolution from a single perpetual License to a diversified licensing system. The fundamental driver of this evolution is the systematic impact of AI vision, cloud-edge collaboration, and the SaaS trend on the traditional "one-time buyout" model, behind which is a fundamental change in the structure of customer demand—from "buy one set of software and use it for ten years" to "use on demand, upgrade at any time."

Model 1: Perpetual License

The perpetual License is the mainstream business model of traditional industrial vision SDKs, with both HALCON and Matrox MIL primarily using it. After purchasing a Developer License, engineers can run the SDK on any development machine for development and debugging; when deploying the vision system to a production-line station, each station needs to purchase a Runtime License.

Taking HALCON as an example, the combined cost is about: development license + single runtime license = about 15,000 to 17,000 yuan (genuine, including the Dongle hardware lock). For a medium-sized production line that needs to deploy 20 vision stations, HALCON's software-licensing cost is about 300,000 to 340,000 yuan. The advantage of the perpetual License is that there is no need for continuous payment and the total cost of ownership is predictable, suitable for one-time project procurement with a fixed budget; the disadvantage is that algorithm updates require an additional upgrade fee, and the hardware lock (Dongle) management is rather cumbersome in multi-node deployment (a damaged or lost Dongle can cause the entire production line's vision system to stop, one of the high-frequency pain points for integrators).

Model 2: Subscription / SaaS

Subscription is the business model naturally suited to AI vision platforms. SMore's AI quality-inspection SaaS platform and AInnovation's industrial AI service both adopt subscription logic: paying by functional module (number of inspection-task types, number of AI-model concurrencies) and usage scale (number of concurrent nodes, average monthly inspection images), with continuous algorithm updates and model optimization included in the subscription fee, without needing to separately purchase upgrade versions.

The advantage of the subscription model for vendors is predictable recurring revenue (ARR, Annual Recurring Revenue), which can form a smoother revenue curve in financial statements and reduce seasonal fluctuations. But for industrial customers, "continuous payment" needs to be handled separately from one-time Capital Expenditure (CapEx) projects in financial budgeting, sometimes increasing the complexity of internal approval—in the procurement decisions of Chinese manufacturing, the flexibility of the annual operating budget (OpEx) is usually lower than that of project-based CapEx.

Model 3: Per-Device Licensing + Hardware Bundling

Mainstream domestic SDK vendors (VisionMaster, VisionWARE) more often adopt the strategy of "algorithm SDK bundled with hardware, licensed per online device node." When integrators procure vision controllers or smart cameras from Hikrobot, the SDK-usage right is included in the hardware-procurement contract, reducing the decision friction of software procurement. This model both lowers the psychological threshold of software licensing (the customer perceives that "the software comes free with the camera") and improves the competitiveness of the overall solution (compared with a similarly priced foreign camera + HALCON solution, the total procurement cost of the domestic integrated solution is 30%–50% lower).

For smaller integrators, Hikrobot also offers a free basic version of VisionMaster (algorithm tools are limited, providing only core positioning and measurement tools) to lower the entry barrier and cultivate new developers. This "freemium" strategy has achieved significant user-growth effects in the Chinese engineer community.

II. Business-Model Matching for Different Scenarios

Different application scenarios and customer types have different preferences for business models, and no single model fits all scenarios.

Large manufacturing factories (OEM / Tier 1): inclined toward perpetual License or large framework-contract models, with extremely high requirements for software stability, unwilling to accept the subscription-cutoff risk on critical production-line paths (if a subscription service expires without timely renewal, it may cause production-line stoppage); the "specification-designated" attribute of overseas SDKs (HALCON/VisionPro) among such customers makes their market position relatively solid.

Medium-sized system integrators: sensitive to licensing cost, with high technical-support needs, more inclined toward the "hardware-binding + local-service" model of domestic SDKs; Lingkong, Huaray Technology, and Hikrobot compete most fiercely among such customers, and project evaluations usually involve a comprehensive scoring of algorithm-effect testing + service-response speed.

Emerging AI vision application scenarios (AI industrial quality-inspection SaaS): the subscription model fits best, and the AI quality-inspection platforms of SMore and AInnovation mainly serve such customers; these customers are sensitive to algorithm effectiveness (hoping each subscription renewal brings better model performance), and are often rapidly growing manufacturing enterprises (hoping the vision system's capability scales elastically with production-line expansion).

Startups / research institutions: mainly OpenCV + open-source deep-learning frameworks, occasionally paired with a HALCON evaluation version (officially providing a 30-day free Eval license each month), generating no formal commercial-licensing revenue, but serving as the technology-warm-up pool of the entire industry.

III. Price-Tier Comparison

Taking 2025 market prices as a reference, the combined licensing costs of major SDK products are as follows (unit: yuan):

MVTec HALCON: runtime license about 15,000 to 17,000 yuan per set, development license about 20,000 to 30,000 yuan; perpetual License, with each deployment node paid separately, and algorithm-version upgrades charged an additional upgrade fee.

Cognex VisionPro: runtime license about 14,000 to 16,000 yuan per set; the ViDi deep-learning add-on module is counted separately (about 8,000 to 15,000 yuan per set); primarily perpetual License, with some subscription options.

VisionMaster (Hikrobot): basic version free; node version (including deep-learning tools) about 5,000 to 10,000 yuan per set; lower discounts available when procured with Hikvision cameras/vision controllers; AI vision advanced modules charged separately.

VisionWARE (Lingkong): printing/color-management industry professional version about 12,000 to 20,000 yuan per set (including industry-dedicated algorithm packages); general vision version about 8,000 to 12,000 yuan per set; supports both standalone software sales and bundled-equipment models.

SMore SDK: customized quotation per project, annual framework-contract system (large customers), with single-project scope ranging from hundreds of thousands of yuan to tens of millions; the AI SaaS subscription version is in the business-expansion stage.

AInnovation: subscription by functional module and node count, Hong-Kong-listed standard commercial contracts, primarily annual framework contracts; the average per-customer price of "AI + manufacturing" projects is in the 1.5-million-to-5-million-yuan range.

OpenCV: open-source and free, 0-yuan licensing cost, but the engineering-development and maintenance costs are entirely borne by the user.

This price map reveals the economic logic of domestic replacement: on the premise that functional coverage has reached 80%–90%, the combined software cost of domestic SDKs is about 30%–70% of HALCON's, and the difference can produce cost savings of hundreds of thousands or even millions of yuan in large-scale production-line deployment (above 50 nodes), with a significant positive impact on a manufacturing factory's ROI.

IV. The Future Evolution Direction of Business Models

The business model of machine vision algorithm SDKs is diverging in two directions: on one hand, for high-end scenarios oriented toward high-precision inspection and complex system integration (automotive OEM, semiconductor), perpetual License and large framework contracts will still dominate, because these customers need stability above all else; on the other hand, for emerging applications oriented toward AI quality inspection (appearance inspection and quality control of medium-sized manufacturing enterprises), innovative models of subscription + Pay-per-Result (i.e. paying by inspected images or inspection-pass rate) are being explored by pioneers such as SMore.

Industrial large models bring a third possibility: Model as a Service (MaaS)—the factory does not purchase the SDK, but instead calls a cloud industrial large model via API, paying by the number of inspections and inference-resource usage, with the algorithm capability automatically upgrading as the large model iterates, without version management and local updates. This model was still at the commercially early stage in 2025, and preliminary commercialization cases have appeared in processes where real-time requirements are not extreme (warehousing quality inspection, incoming-material appearance verification), and it is expected to become the mainstream commercial entry point for new vision-AI demand after 2027.

V. Regional Differences in Pricing Strategy: The Localized-Discount Logic of the China Market

International SDK vendors' pricing strategy in the China market has always adopted a two-tier structure of "globally unified pricing + local discounts." The official public prices of HALCON and VisionPro are quoted in euros or US dollars, but when sold through Chinese agent channels (officially authorized agents in major cities such as Shanghai, Shenzhen, and Beijing), the actual transaction price is usually 70%–80% of the official price, and for large system integrators with an annual procurement volume of more than 50 sets, the discount can be further increased to 60% or even lower. This strategy makes the actual price competitiveness of foreign SDKs in the China market stronger than what the "official-price comparison" shows—but still 20%–40% higher than mainstream domestic SDKs.

Domestic SDK vendors likewise have regional price differences: the VisionMaster node-licensing price Hikrobot offers to system integrators with a cooperation framework is usually about 30% lower than the direct-sales channel; large-volume procurement (single order of more than 500 nodes) can be negotiated separately, further lowering it by about 10%–20%. This price elasticity is one of the tools for domestic SDKs to quickly respond to customer budget constraints in project competition.

It is worth noting that as AI vision applications become widespread, some system integrators originally focused on hardware are beginning to find that: a medium-sized lithium-battery inspection project (50 vision-inspection nodes) requires about 750,000 yuan just for HALCON licensing, accounting for about 30%–40% of the project's total software cost. This proportion, in large-scale industrial deployment scenarios, has become a significant constraint on cost competitiveness, and is the most direct commercial driver of domestic replacement.

VI. The Implementation Challenges of the SaaS Subscription Model: Factory-Side Concerns and Countermeasures

The AI vision SaaS subscription model theoretically has multiple advantages (lower upfront investment, dynamic algorithm updates, payment by actual usage), but faces unique resistance in the actual implementation in Chinese manufacturing, and this resistance comes from structural concerns on the factory side, rather than purely technical or cost factors.

Data-security concerns: manufacturing customers (especially industries with high confidentiality needs such as automotive and semiconductor) are highly vigilant about uploading production-process image data to AI vision cloud platforms. Production-line images may contain product-design information (shape, dimensions, material characteristics) and are regarded as IP-sensitive data. When cooperating with large customers such as Tesla and Luxshare Precision, SMore adopts a hybrid model of "private deployment + local inference," distributing the algorithm-trained model to the customer's local hardware rather than performing real-time inference in the cloud, thereby circumventing data-upload concerns.

Network-reliability constraints: the network infrastructure of factory production environments is uneven; the network bandwidth and stability of some old plants cannot support real-time vision-data upload, or the latency caused by network jitter cannot meet production-line takt requirements. This physical constraint makes the "pure cloud inference" SaaS model infeasible in a considerable proportion of industrial scenarios, driving the popularization of the "edge-cloud collaboration" architecture—edge inference guarantees real-time performance and network independence, cloud training optimizes model quality, and updated models are periodically pushed to edge nodes.

Compliance and availability requirements: the vision-inspection systems of key production equipment belong to production-safety-related systems, and the GMP and IATF certifications of some industries (pharmaceuticals, automotive safety parts) have strict change-control requirements for system updates (including algorithm-version updates). The "automatic algorithm update" model of the SaaS subscription has a structural conflict with the change-management specifications of these industries, requiring SaaS vendors to provide a "controlled version-lock" function, satisfying the customer's compliance needs while retaining the flexibility of the SaaS business model.

These challenges mean that the penetration path of industrial AI vision SaaS is more concentrated in industries with relatively loose compliance requirements and relatively complete network conditions (food, general-quality inspection in chemicals), while penetration into high-compliance industries such as automotive, semiconductor, and pharmaceutical requires vendors to provide dedicated compliance-support solutions.

VII. A Selection Framework from the Total Cost of Ownership (TCO) Perspective

From the perspective of factories or system integrators, the decision to choose a machine vision algorithm SDK cannot look only at the licensing unit price, but should be evaluated from the complete perspective of Total Cost of Ownership (TCO).

The main components of the TCO framework include: first, the initial licensing cost (License purchase or first-year subscription); second, the engineer learning cost (the training time required to switch to a new SDK, discounted by engineer salary); third, the integration-development cost (the engineering hours to integrate the SDK with the existing production-line MES and PLC systems); fourth, the long-term maintenance cost (the adjustment hours for algorithm parameters as product models change, the adaptation hours for SDK-version upgrades); fifth, the technical-support cost (including service-contract fees and the time cost of internal technical personnel handling technical problems).

Taking a medium-sized automotive-parts inspection project (20 vision-inspection nodes, expected to operate for 7 years) as an example for TCO comparison: the HALCON solution's initial licensing is about 300,000 yuan, but engineers' HALCON familiarity is high, so the learning cost is low (about 30,000 yuan in hours), and the technical-support fee is about 20,000 yuan per year; the VisionMaster solution's initial licensing is about 150,000 yuan, but engineers need to migrate from HALCON (learning cost about 80,000 yuan), and the technical-support fee is about 10,000 yuan per year. The 7-year TCO is about 500,000 yuan (HALCON) and 390,000 yuan (VisionMaster) respectively, a gap of about 22%, rather than the 50% gap when looking only at licensing fees. For system integrators that already have an ample reserve of VisionMaster engineers, the TCO gap will further widen to 30%–40%.

This analysis shows that the TCO framework is a key tool for understanding the economics of domestic replacement. Under the trend of the engineer ecosystem gradually maturing and migration costs continuously decreasing, the long-term TCO advantage of domestic SDKs will continuously strengthen as the ecosystem scale expands. It is expected that by 2028, in the TCO comparison of 3C inspection and lithium-battery inspection scenarios, the advantage of domestic SDKs will rise from the current 20%–30% to 35%–45%.

---

Chapter 9　Typical Customer Cases

I. Tesla × SMore: The Implementation of AI Quality Inspection Under Extreme Production Takt

Tesla introduced SMore's industrial AI vision system on the battery-component and body-sheet-metal production lines of its Shanghai Gigafactory (Gigafactory Shanghai). This is one of the landmark cases of a Chinese AI vision enterprise entering the supply chain of a top multinational automaker, and an important proof of the industrial large model completing validation under the extreme takt of a real production line.

SMore's AI quality-inspection platform connects to Tesla's production line in a no-code configuration manner, completing the rapid deployment of deep-learning defect-detection models targeting battery-tab welding quality and body-stamping-part appearance. The production takt of Tesla's Shanghai factory is extremely dense (single-shift capacity exceeding 1,000 vehicles), and the vision-inspection system needs to complete image acquisition, AI inference, and judgment for each inspection station within 0.5 seconds, and must communicate in real time with the production-line MES system to output inspection results and trigger defective-product diversion. After stable operation, the entire system controls the miss rate of specific defect types to within 0.2% and the false-alarm rate below 2%, meeting Tesla's quality-control standards.

The strategic value of this case lies not only in technical validation, but also in brand endorsement: as the most influential manufacturing-technology bellwether in the world, Tesla's adoption of SMore's AI quality-inspection platform greatly enhanced the latter's credibility among international high-end manufacturing customers, accelerating its business-negotiation process to enter other high-end customers such as Carl Zeiss and Luxshare Precision.

II. Luxshare Precision × SMore: AI Vision Empowerment for Consumer-Electronics Precision Assembly

Luxshare Precision, as a core connector and precision-component manufacturer in the Apple supply chain (AirPods, Apple Watch, iPhone connectors), introduced SMore's AI vision-inspection solution in its precision-assembly processes. The 0.1-mm-level bending detection of connector PINs and the micro-scratch recognition of earphone housings are both scenarios that traditional rule algorithms find hard to cover stably—PIN density is high (pitch below 0.3 mm) and each batch of products has slight geometric differences, keeping the false-alarm rate of rule operators persistently high.

The AI vision system, through anomaly-detection models of a large number of normal samples, learns the geometric distribution and appearance characteristics of normal PINs, then marks samples deviating from the normal distribution as anomalous. After implementation on multiple Luxshare Precision production lines, the inspection false-alarm rate dropped by about 60% (reducing a large amount of manual-recheck hours), the miss rate was maintained within the customer-allowed GR&R range, and manual-recheck hours were reduced by about 40%. This quantitative effect became typical data support for AI vision replacing traditional rule algorithms in consumer-electronics precision scenarios.

III. Lingkong × A Leading Printing-and-Packaging Enterprise: High-Speed Online Printing-Quality Inspection

A leading carton-packaging production enterprise deployed Lingkong's VisionWARE-based online printing-quality inspection system on its high-speed printing line (line speed up to 300 meters per minute). The system uses a line-scan camera (Line Scan Camera, CCD sensor, scan resolution 10 microns per pixel) and Lingkong's self-developed optical system as the sensing layer, with the VisionWARE algorithm platform responsible for real-time color-deviation computation (ΔE value accuracy better than 0.5), registration-accuracy inspection (accuracy ±0.1 mm), and character OCR recognition. The inspection results are fed back in real time to the printing-press control system, achieving automatic color compensation.

Traditional manual sampling inspection can only discover batch color differences after the product batch is completed, with large losses; Lingkong's online system monitors in real time during the printing process, and when a color-difference trend is found, immediately feeds it back to the printing press's CTP control unit to adjust the ink amount, eliminating the color-deviation drift at the budding stage, greatly reducing the scrap rate (measured to decrease by about 35%).

Lingkong's deep accumulation in the printing-inspection field comes from its more than 20 years of independent R&D history in Color Science, color-measurement standards (spectral-reflectance measurement), and precision optics, a professional barrier that domestic SDKs entering the printing scenario later, such as VisionMaster, find hard to replicate in the short term.

IV. Mech-Mind × An Automotive Tier 1 Parts Enterprise: 3D Vision-Guided Random Picking

An automotive Tier 1 parts enterprise deployed Mech-Mind's 3D vision-guided robotic-grasping system (Bin Picking) in the loading process of engine-casting blanks. The blanks are in a randomly stacked state in the material basket, and the traditional fixed-fixture solution cannot adapt to the random pose changes of the parts; relying on manual loading is costly and the takt is unstable.

Mech-Mind's Mech-Eye 3D camera performs a full scan of the material basket, Mech-Vision software completes 3D point-cloud segmentation and target-part pose estimation (six degrees of freedom), Mech-Viz generates a collision-free grasping path (the path-planning algorithm considers joint-angle limits and obstacle avoidance), and the ABB robot executes the grasping action. The takt time of the entire system (including 2 seconds of 3D sensing + 3 seconds of path planning + 3 seconds of robot motion) is controlled within 8 seconds, meeting production-line takt requirements, and reduces the labor demand of the loading process from 3 people per shift to 1 person for supervision, saving about 700,000 yuan in annualized labor cost.

This case reveals the core of Mech-Mind's commercial success: the integrated software pipeline of Mech-Vision and Mech-Viz compresses the integrator's development cycle from the 6 months of a traditional 3D vision solution to within 2 months, greatly reducing the system integrator's development risk and project-delivery cost.

V. AInnovation × A Steel Enterprise: The Implementation of an Industrial Large Model in Hot-Rolled-Strip Surface Inspection

A leading domestic hot-rolled-strip production enterprise cooperated with AInnovation to deploy a surface-defect inspection system based on the AInnoGC industrial large model on a continuously producing hot-rolling line. The surface defects of hot-rolled strips (bubbles, slivers, scratches, roll marks, oxidation spots) are numerous and complexly caused, and at a high line speed of more than 60 meters per minute, traditional rule algorithms find it hard to achieve accurate multi-class defect classification within a millisecond time window.

AInnovation's solution, under the condition of about 2,000 labeled samples (far lower than the tens of thousands required by traditional deep-learning solutions), completed accurate classification of 12 surface-defect types, with the miss rate controlled within 0.3%. The system also links with the hot-rolling line's L2-level process-control system; when a specific high-risk defect is detected (such as a deep crack that may cause strip breakage), it automatically triggers a deceleration alarm, providing the operator with an intervention window and avoiding production accidents.

This case is an important validation of the commercial implementation of an industrial large model in a heavy-industry scenario (steel), proving that the large model can still run stably in the extreme industrial environment of high noise, high temperature, and strong vibration, and achieve high-accuracy multi-class defect classification under small-sample conditions.

VI. Huaray Technology × A Lithium-Battery Leading Enterprise: Electrode-Sheet Online Inspection System

A leading domestic power-battery enterprise deployed Huaray Technology's electrode-sheet online vision-inspection system on its GWh-level capacity electrode-sheet manufacturing line. The system covers three functional modules: coating-uniformity measurement (coating-thickness deviation ±2 microns, with a line-scan camera paired with a laser profiler to achieve millimeter-by-millimeter measurement), electrode-sheet scratch and crack detection (a deep-learning classifier reaching production-grade accuracy with just 50 small-sample training images), and coating-boundary positioning (sub-pixel-accuracy edge detection, accuracy ±0.1 mm).

Huaray Technology's electrode-sheet inspection solution is deeply optimized for new-energy scenarios, including: a dedicated illumination-algorithm design for the strong reflectivity of electrode active material (a multi-angle polarized-light combination, suppressing specular reflection); camera-timing control and algorithm parallelization optimization for high-speed production lines (electrode-sheet line speed above 60 meters per minute), ensuring the single-frame processing time is controlled within 10 milliseconds. This system achieved 100% online quality-inspection coverage during the electrode-sheet manufacturing process at the deployment factory, replacing the traditional manual-sampling-inspection model, and reducing the electrode-sheet defect miss rate from about 3% to below 0.1%.

VI. Large-Customer Lock-In and Lifecycle-Value Management

The customer-relationship management of industrial vision SDKs has a typical "funnel + flywheel" dual structure: the funnel represents the conversion process from a potential customer to the first purchase, and the flywheel represents the process of continuous revenue growth after the first purchase by expanding application scenarios and node counts. In the SDK business model, the value of the flywheel is often far greater than the funnel—winning a large automotive-factory customer's first vision-inspection line (20 nodes) is just the beginning; as the customer's trust in the SDK's capabilities accumulates and engineer skills are bound, the 5-year lifetime value (LTV) of a single large customer is often 5–10 times the value of the first order.

SMore's cooperation with Tesla is a typical embodiment of this logic: SMore started from a body-stamping defect-inspection line at Tesla's Shanghai Gigafactory, gradually expanding to laser-welding quality inspection, battery-module assembly vision inspection, and 100% online inspection of Final Assembly, with each new scenario's entry based on the technical validation and trust accumulation of the previous scenario, without needing to re-establish quality endorsement from scratch. Tesla has become a core strategic customer showcased in SMore's IPO prospectus, and its contract scale and renewal rate are key indicators for institutional investors to evaluate SMore's customer stickiness.

For HALCON, large-customer lock-in is likewise the core mechanism for maintaining market share. The "technical-specification dependence" (taking HALCON algorithm interfaces as the reference standard in internal technical-specification documents) and "engineer-skill binding" (the customer has internally accumulated a large amount of HALCON-based inspection-solution code) that HALCON establishes at each large customer constitute a switching barrier that is hard to replace quickly. Even if a domestic SDK has an advantage on a specific metric, rewriting thousands of lines of HALCON code and passing the customer's internal system-change validation means a 6–18-month migration cycle and a considerable investment of engineering resources, a switching cost hard to justify unless absolutely necessary.

VII. The Dual Pressure on System Integrators as a Channel Intermediary

System integrators play an important channel-intermediary role in the machine vision industry chain: one side connects to SDK vendors (procuring and secondarily developing algorithm capabilities), and the other side connects to terminal factory customers (delivering vision-inspection systems). However, this intermediate layer is facing structural pressure from two directions.

Channel penetration from SDK vendors: head domestic SDK vendors (Hikrobot, Lingkong) are expanding their direct-sales teams to implement direct-sales coverage of large terminal customers (factories with an annual procurement scale of more than one million yuan), while maintaining the system-integrator channel for small and medium customers. This dual-track strategy of "large-customer direct sales + small-and-medium-customer channel" makes system integrators face direct competition from the original SDK manufacturers in large-project competition, compressing the channel partners' profit space.

Capability internalization from terminal customers: head factories (especially ultra-large enterprises in the 3C and new-energy industries) are internalizing vision-inspection capabilities, building their own machine vision engineer teams, directly procuring SDK licenses and industrial cameras, and bypassing system integrators to build internal vision-inspection capabilities. Ultra-large manufacturers such as Foxconn and BYD have established considerable internal vision teams, and these enterprises' dependence on system integrators continues to decline.

Facing pressure from these two directions, the survival strategies of small and medium system integrators are diverging: one type moves toward industry vertical specialization (such as system integrators focused on lithium-battery electrode-sheet inspection, having accumulated industry-exclusive algorithm-tuning experience and production-line integration capabilities); another type transforms toward "light consulting + rapid implementation" (relying on SaaS AI vision platforms to reduce their own technical-R&D investment, focusing on understanding customer needs and rapid deployment and delivery); and yet another type is incorporated by the original SDK manufacturers as "ecosystem partners," in fact becoming the outsourced implementation teams of the original SDK manufacturers.

---

Chapter 10　Investment, Financing, and M&A (SMore / AInnovation / Megvii Listing, etc.)

I. The Capital Boom of Industrial Vision AI (2020–2025)

From 2020 to 2023, the industrial AI vision field experienced one of the most intensive periods of capital influx in the history of Chinese venture investment. At that time, the narrative logic of "AI + manufacturing" was clear and attractive: China's manufacturing volume ranks first in the world, the labor cost of the quality-inspection link continues to rise, deep-learning defect detection demonstrated higher consistency and lower long-term cost than manual inspection in pilots, and the commercial logic of industrial vision AI is clearer than consumer-grade AI (ad recommendation, content moderation) (defect detection = cost savings or quality improvement = directly calculable ROI).

Against this backdrop, enterprises such as SMore, AInnovation, Megvii Industrial Vision (the industrial line of Megvii), and Mech-Mind intensively completed multiple rounds of large financing. SMore completed a USD 200 million Series B financing in 2021, completed Series C in February 2026 and pushed the valuation to USD 1.230 billion, with a total financing scale exceeding USD 300 million, making it one of the independent enterprises with the largest financing scale in China's industrial AI vision track.

However, after 2023, this financing boom cooled significantly, and capital began to shift from "vision investment" to "profitability validation." The focus of investment institutions on industrial AI vision projects shifted from "can the technology be achieved" to "can revenue growth, gross margin, and the profitability inflection point be seen." Those enterprises that raised funds at high valuations in 2021–2022 but were not solid enough in commercialization progress faced the pressure of valuation downgrades and financing difficulties; on the contrary, enterprises such as AInnovation and Mech-Mind, which had clearer business models and profitability paths, showed stronger resilience in the capital winter.

II. AInnovation: The Industrial-AI Pioneer of the Hong Kong Stock Market

AInnovation is currently the most important capital-market reference in the domestic industrial AI vision track. When it listed on the Hong Kong Stock Exchange in 2022 (code 2121.HK), the company completed its listing with the narrative of "China's AI vision technology empowering industrial smart manufacturing," with a market capitalization of about HKD 2 billion. In 2023–2024, affected by the macroeconomic downturn and the contraction of manufacturing capital expenditure, the company's revenue growth slowed, the net loss expanded somewhat, and the stock price retreated from its high point.

However, signs of a profitability-improvement inflection point began to appear in 2025: in the first half of FY2025, the adjusted net loss narrowed to 6.68 million yuan (a year-on-year decrease of 82.1%), the gross margin rose to about 35%, and the "AI + manufacturing" business (industrial vision inspection + industrial robots + industrial software) accounted for nearly 80% of total revenue, showing that the company had completed business focusing and the drag of non-core businesses was greatly reduced. IDC data show that AInnovation ranks seventh in China's large-model application market share, and is the only vendor focused on the industrial field; this differentiated positioning has defensive value in the fiercely competitive wave of large-model commercialization.

AInnovation's market performance on the Hong Kong stock market (valuation about 10 times P/S) has become an important reference for the IPO valuation anchoring of subsequent industrial AI vision enterprises (including SMore).

III. SMore: The Charge of the "First Industrial AI Agent Stock"

SMore's sprint toward a Hong Kong IPO is the most-watched capital-market event in the industrial-AI track in 2026. When submitting its listing application to the Hong Kong Stock Exchange in March 2026, the company's core narrative was the "industrial AI agent": not only providing vision-inspection algorithm SDKs, but also building a full-pipeline industrial-AI capability from sensing (vision recognition) → analysis (large-model reasoning) → execution (linking with MES/ERP systems to trigger actions such as diversion, re-welding, and rework), with positioning upgraded from "vision-tool supplier" to "industrial-intelligence solution provider."

Financial performance: 2025 revenue was nearly 1.1 billion yuan, with a three-year compound growth rate exceeding 40%; the adjusted net loss narrowed significantly from 394 million yuan in 2023 to 272 million yuan in 2025, and the marked decline in cash-burn speed makes the choice of IPO timing financially reasonable. In February 2026, it completed Series C financing with a valuation of USD 1.230 billion (about 8.5 billion yuan).

The sponsor lineup (Morgan Stanley + CICC + Deutsche Bank) and the USD 1.230 billion valuation make SMore the highest-valued pre-listing company in China's industrial AI vision field to date. The industry expects that if the IPO proceeds smoothly, SMore will complete its listing on the main board of the Hong Kong Stock Exchange, with a market capitalization expected to exceed HKD 10 billion, becoming the second Hong-Kong-listed company in the industrial AI vision track after AInnovation, and also the largest IPO in this track to date.

IV. Mech-Mind: The Global Expansion of a Hidden Unicorn

Mech-Mind has long maintained a strategically low profile, not frequently disclosing financing rounds and valuation figures, but its commercial results have clearly demonstrated its unicorn status: five consecutive years of being the champion in the 3D vision-guided robotics market, a customer scale serving more than 50 countries and regions worldwide, and deep cooperation with top robot brands such as ABB, FANUC, Yaskawa, and KUKA, have pushed its market valuation into the unicorn ranks (above USD 1 billion).

Mech-Mind's global-expansion strategy takes "going global with Chinese manufacturing" as the main logical line: as Chinese manufacturing leaders such as CATL (European battery factories), BYD (Southeast Asian vehicle factories), Foxconn, and Luxshare Precision build factories overseas, Mech-Mind's 3D vision solutions follow them into the European, North American, and Southeast Asian markets, forming a global customer-relationship chain that foreign SDKs find hard to intercept. This is a globalized commercial flywheel that other domestic vision SDKs have not yet formed.

V. Orbbec: The Leap from 3D Vision Sensors to Embodied Intelligence

After achieving its first full-year profit since listing in 2025, Orbbec shifted its strategic focus from "industrial 3D vision sensors" to "the vision-sensing base for embodied-intelligence robots." Embodied AI is considered the next super-demand explosion point for vision sensors and vision SDKs—humanoid robots (Unitree H1/G1, AgiBot Expedition A1, Xiaomi CyberOne, etc.) and mobile-manipulation robots need continuous, real-time, low-power 3D sensing capabilities, which happen to be Orbbec's core accumulation in the ToF and structured-light sensing direction. 2025 net profit attributable to the parent was 128 million yuan, an important milestone in the company's business model moving toward maturity.

VI. Overseas M&A: The Defensive Moves of Foreign SDKs

While facing China-market competition pressure, foreign SDK vendors are also reinforcing their moats through technology integration. MVTec was deeply integrated under Omron in 2022, obtaining more complete factory-automation system resources (the ecosystem linkage of sensors, robots, and PLCs); Cognex continuously strengthens its AI vision capabilities through internal R&D and small-scale acquisitions, and the iteration speed of ViDi Suite accelerated noticeably in 2024–2025. These moves show that foreign SDK vendors are not passively responding to domestic replacement, but actively building the next generation of competitive barriers—it is expected that over the next 2–3 years, foreign SDKs will further intensify their catch-up in AI vision functions, and domestic AI vision enterprises cannot be sustainedly over-optimistic about "technical leadership."

VII. The Industry-Consolidation Logic and Asset Revaluation After the Capital Winter

The capital winter of 2023–2024 had a far more profound impact on the industrial AI vision track than the surface numbers show. Before this, the track had a large number of startups with highly similar business models—taking AI-vision labeling and algorithm deployment as their main business, relying on financing to subsidize business expansion, and failing to establish a real technical barrier or a scaled data barrier. Such enterprises faced severe survival pressure after 2023: customers' payment terms lengthened (some automotive Tier 1 customers' payment terms extended to 180 days), the financing pace slowed, while labor and service costs did not decline in sync, leading to continuous cash-flow deterioration.

Industry consolidation occurred in multiple forms in 2024–2025: first, system integrators acquired technology-type startups (obtaining algorithm teams and technical accumulation at a relatively low valuation); second, head SDK vendors expanded their "developer ecosystem," absorbing small and medium technology companies that previously used HALCON as their primary tool through low-price or even free licensing, incorporating them into their own ecosystem; third, some startups pivoted to building "algorithm pre-trained model stores" in vertical industry directions—taking AI vision foundation models as the base and selling ready-made pre-trained models to small and medium system integrators, circumventing the high cost of building their own data labeling and algorithm optimization, but also giving up the possibility of building a differentiated barrier.

From the asset-revaluation perspective, this round of consolidation is changing the valuation logic of industrial AI vision enterprises: the high valuation multiples driven by the dual wheels of "AI story + high-growth revenue" in 2021–2022 (P/S above 20 times) contracted to 5–10 times P/S after 2023, and increasingly focused on profit margin and cash-flow quality. AInnovation's Hong Kong market capitalization contracted from a peak of about HKD 3 billion to about HKD 1 billion in 2024, then rebounded to about HKD 1.5 billion in Q4 2025, reflecting the valuation-recovery path of the entire track. Whether SMore's valuation of 8.5 billion yuan at IPO filing can gain the recognition of Hong Kong Stock Exchange investors will become a key reference for the valuation anchoring of the industrial AI vision track in the first half of 2026.

VIII. The Strategic Investor Perspective: The Investment Logic of Vertical Integration in the Industry Chain

Unlike pure financial investors (PE/VC) who focus on financial returns, the investment logic of industrial strategic investors in industrial AI vision SDKs focuses more on the vertical integration of technical capabilities and the enhancement of industry-chain control.

The Hikvision system (including Hikrobot) is itself the most typical case of industry integration: building VisionMaster through organic investment (rather than external M&A), embedding vision-algorithm capabilities into its own cameras and complete-machine products, forming a complete closed loop of algorithm-hardware-system. This path avoids the integration risk of external M&A, but requires a long cycle of technical accumulation and investment—Hikrobot has continuously invested in large-scale R&D since 2016, and it was only around 2020 that VisionMaster truly formed an industrial-grade usable algorithm system.

Robot complete-machine manufacturers (such as industrial-robot body enterprises) increasingly regard vision SDK capability as a key gap in robot intelligence. International giants such as ABB and FANUC have completed sensing capabilities by acquiring vision-technology companies, and this trend is accelerating its replication among Chinese local robot enterprises. In their strategic plans for 2024–2025, enterprises such as Inovance and Estun have all listed "internalizing robot vision-sensing capability" as a mid-term strategic goal, and their strategic-investment interest in startups with mature industrial 3D vision SDKs (such as Mech-Mind) continues to increase.

Automotive Tier 1 suppliers are likewise potential strategic investors in this track. The vision-inspection systems of domestic automotive Tier 1 have long depended on foreign SDKs (HALCON dominates), but against the backdrop of the eastward shift in the discourse power of the new-energy-vehicle industry chain, the demand to independently control vision-inspection technology is increasingly strong. Head enterprises such as Forvia and CATL have, by signing strategic-cooperation agreements (rather than general procurement contracts) with domestic SDK vendors, reserved space for deeper future technical cooperation.

IX. The Going-Global Path of Chinese Industrial Vision SDKs in the Global Market

In 2025–2026, the going-global of Chinese industrial vision SDKs is moving from the "sporadic-order" stage to the "systematic strategic layout" stage. Mech-Mind's 3D vision SDK already covers more than 50 countries and regions worldwide, with the European manufacturing market (automotive-parts clusters in Germany, the Czech Republic, etc.) and the Japanese market as key breakthrough regions; Lingkong, with its world-leading high-precision optical-inspection capabilities, is advancing into Southeast Asia (the electronics-manufacturing industries of Vietnam and Thailand) and the European printing industry.

The technical obstacles to going global have greatly decreased compared with five years ago: the global unification of the GenICam / GigE Vision protocols means Chinese SDKs have no hardware obstacles in camera compatibility; AI vision models (such as anomaly detection and classification recognition) are essentially language-independent, with no fundamental obstacle to English adaptation; on the contrary, the English quality of product documentation, the establishment of overseas technical-support systems, and cooperation with overseas system-integrator distribution networks constitute the main friction costs of going global.

It is worth noting that against the backdrop of China-US relations, Hikrobot faces export restrictions in the US market (Hikvision has been placed on the Entity List), and VisionMaster's penetration in the US market actually faces structural constraints at the policy level. This gives non-sanction-list enterprises such as Lingkong, Mech-Mind, and SMore a relative policy advantage in developing the European and US markets. Over the next 3–5 years, the overseas-market increment of Chinese industrial vision SDKs will mainly come from the transfer of Southeast Asian manufacturing (partly driving Chinese SDKs to go global together with Chinese equipment and system integrators) and Europe's continuous demand for cost-effective solutions.

---

Chapter 11　Policy and Standards ("AI+" Action Plan / The Third Phase of the Big Fund / High-End Equipment Domestic-Replacement Special Program)

I. The "AI+" Manufacturing Special Action Plan

On January 7, 2026, the Ministry of Industry and Information Technology and seven other departments jointly issued the "Implementation Opinions on the 'AI+ Manufacturing' Special Action," explicitly listing industrial AI vision, industrial large models, and AI quality inspection as one of the core scenarios of AI empowerment for manufacturing. This is the strongest policy signal the national level has sent to the industrial AI vision track to date, and the most direct policy endorsement the machine vision algorithm SDK industry has obtained.

The core requirements of the document include: by 2027, the AI-empowerment coverage of above-scale manufacturing enterprises in key processes will significantly increase, and the AI quality-inspection and AI vision penetration rate of key industries (new energy, 3C electronics, semiconductor, automotive) will reach clear targets; promoting the standardized-interface construction of industrial AI platforms (including vision SDKs) to lower the cross-industry deployment threshold; and supporting industrial AI vision enterprises in building industry-solution benchmark cases and promoting and replicating them across the industry.

The supporting "Guidelines for AI-Empowered Transformation of Key Manufacturing Industries" further refines the AI-empowerment paths of five major industries (raw materials, equipment manufacturing, consumer goods, electronic information, software and information-technology services), providing vision-algorithm SDK vendors with a clear industry-breakthrough direction map—the electronic-information industry (PCB inspection, semiconductor packaging) and the equipment-manufacturing industry (CNC machine-tool online inspection, industrial-robot vision guidance) are the two directions of key policy support, and also the niches with the greatest room for domestic SDK penetration improvement.

II. The Third Phase of the Big Fund and Machine Vision Semiconductor-Material Investment

One of the investment focuses of the third phase of the National Integrated Circuit Industry Investment Fund (the Big Fund Phase III, established in 2024, with a total scale of about 344 billion yuan) is the localization of front-end semiconductor equipment and materials. The localization of machine vision algorithm SDKs in semiconductor-inspection scenarios is a peripheral beneficiary direction of the Big Fund Phase III's attention—semiconductor-inspection equipment (defect detection, metrology) is a domestic-replacement category directly supported by the Big Fund Phase III, and the vision-algorithm SDK, as the core software capability of this equipment, also benefits accordingly from the overall capital injection into the industry chain.

This is specifically reflected in: domestic semiconductor-inspection equipment enterprises, with the Big Fund Phase III's capital support, accelerate R&D, with their core vision algorithms either self-developed or co-developed with domestic vision SDK vendors (Lingkong, Mech-Mind), promoting the preliminary implementation of domestic vision algorithms in high-end semiconductor-inspection scenarios. In 2025–2026, several local semiconductor-inspection equipment enterprises have completed the domestic replacement of mass-production-grade vision algorithms in back-end packaging-and-testing scenarios, accumulating engineering experience for front-end replacement.

III. The High-End Equipment Domestic-Replacement Special Program

The high-end equipment manufacturing domestic-replacement special program promoted during the national "14th Five-Year Plan" period incorporates industrial vision systems into the niche direction of "high-end intelligent sensing and inspection equipment." The relevant policies are implemented through two channels:

First, the national key R&D programs set up "smart manufacturing" and "advanced sensing" special projects, supporting the basic research and engineering development of core industrial-vision algorithms (defect detection, 3D reconstruction, vision-guided robotics). Lingkong, Hikrobot, and Mech-Mind have all undertaken related projects, and while obtaining R&D support, they have also established industry-university-research cooperation with universities and research institutes.

Second, in the process of localizing industrial machine tools and high-end CNC machine tools, embedded vision inspection (machining online inspection, tool-wear vision monitoring, workpiece positioning) is listed as a key supporting capability. The localization of CNC machine tools requires the visualization of the two functions of tool-life prediction and workpiece-dimension online measurement, directly driving the integration demand of industrial vision SDKs with high-end machine-tool control systems. Domestic CNC-system enterprises (HNC, Shenyang CNC) have begun to cooperate with vision SDK vendors to integrate vision-inspection capabilities into the standard function packages of CNC systems.

IV. The Construction of Technical Standards and Certification Systems

The technical standardization of the industrial vision algorithm SDK field is an important infrastructure supporting the implementation of domestic replacement, and also a key path to lowering the entry cost for domestic SDKs into high-certification-threshold industries (automotive, semiconductor, pharmaceutical).

The National Technical Committee for Standardization of Automation Systems and Integration is advancing industrial vision-system interface standards, including industrial-camera interface protocols (the domestic implementation specification of GigE Vision 4.0) and the integration-interface standards between vision systems and factory-automation systems (PLC/DCS/MES). The establishment of these standards enables domestic SDKs, when connecting to the existing automation systems of terminal factories, to follow clear interface specifications, reducing the uncertainty of on-site integration.

The China Machine Vision Union (CMVU) promotes the establishment of vision-inspection performance-evaluation standards oriented toward industry scenarios (lithium-battery manufacturing, automotive-welding quality inspection), enabling the inspection capabilities of domestic SDKs to be objectively and quantitatively compared by third-party evaluation institutions. These third-party evaluation reports provide trustworthy technical bases for purchasers (especially factories accustomed to using HALCON but facing replacement pressure), effectively reducing the perceived risk of the replacement decision.

In the semiconductor-inspection field, the domestic industry is actively advancing deep alignment with SEMI international standards (equipment-communication protocols such as E5, E30, E40, E87), preparing standardized entry conditions for domestic inspection equipment (including vision SDKs) to enter the foreign-dominated wafer-fab supply chain. This process is expected to require 3–5 years, and is a key prerequisite action at the standard level in the entire semiconductor-vision domestic-replacement path.

V. The Cyclicality and Limitations of the Policy Effect

It must be pointed out that policy promotion is not the decisive factor for the domestic replacement of machine vision algorithm SDKs, but an accelerator rather than a driver. What truly determines the success or failure of replacement is still the maturity of algorithm-technology capability and the engineering completeness of product delivery. Without technical capability as a foundation, no matter how strong the policy, it cannot drive the procurement decision of a real production line—what the factory ultimately cares about is whether the vision system "can work stably," not whether it "is domestic."

Policy has a substantive role at two levels: first, through subsidies and government-procurement bias (especially the "localization-first" requirement of state-owned-background factories), it compresses the commercialization window for domestic SDK vendors during the technology-catch-up period, enabling them to maintain operations and accumulate cases through policy-supported orders at a stage when finances have not yet reached break-even; second, through technical-standard construction, it lowers the entry cost for domestic SDKs into high-certification-threshold industries, a systemic force that amplifies technical capability and accelerates the replacement process.

V. Local Government Industry-Support Policies: Special Support in Key Regions

In addition to national-level strategic policies, the direct support policies of provincial and municipal local governments for the industrial AI vision industry are becoming a key supplementary force driving industry implementation.

Shenzhen: In 2025, it issued the "Shenzhen Artificial Intelligence Industry Development Action Plan (2025–2027)," specifically setting up an "intelligent vision" special project, granting eligible industrial AI vision software products (including algorithm SDKs) a product first-purchase subsidy of up to 5 million yuan and an additional 15% R&D-expense super-deduction subsidy. Shenzhen is also building an "Industrial Intelligent Vision Innovation Center" in Pingshan District, providing small and medium vision AI enterprises with an application-test environment and typical-scenario datasets.

Shanghai: Pudong New Area takes integrated circuits and artificial intelligence as its dual cores, providing industrial AI enterprises registered in Pudong with special industry-fund support of up to 30 million yuan, as well as full reimbursement for international-standard certifications (CE, UL, etc.). This policy directly benefits the overseas-market-expansion compliance costs of enterprises such as Mech-Mind and SMore that have established core R&D bases in Shanghai.

Suzhou: The Industrial Park launched the "Suzhou Industrial Vision Pilot Program," using the vision demand of large manufacturing customers (mainly automotive and precision-manufacturing enterprises within the Suzhou Industrial Park) as a draw, promoting local system integrators to prioritize domestic vision SDKs, and granting an 800-yuan-per-node replacement subsidy (capped at 400,000 yuan per project) for projects replacing foreign SDKs.

The direct effect of these local policies is to reduce, through fiscal means, the initial-cost gap between domestic SDKs and foreign SDKs, accelerating the commercial closed loop of domestic replacement. When the technical capability of domestic SDKs can already cover application needs, and government subsidies further level the cost gap, the selection scale of system integrators will quickly tilt toward domestic—this is precisely the policy background of the accelerated domestic replacement in 3C-inspection and lithium-battery-inspection scenarios in 2024–2025.

VI. The Impact of Data-Security Regulations on Vision Data

The Data Security Law and the Personal Information Protection Law, implemented in 2021, raised new compliance requirements for the processing of industrial vision data. Although industrial vision data (image data of factory products) has an essential difference from personal-privacy data, when identifiable facial features of factory workers appear in production images, it enters the regulatory scope of the Personal Information Protection Law.

More importantly, the "data classification and grading" framework introduced by the Data Security Law requires enterprises to evaluate and protect the importance of production data. For the processing of "important industrial data" in automotive, aviation, semiconductor, and other fields, regulations have clear constraints on the data-retention location (required to be within the territory) and data transmission (cross-border transmission requires approval). This directly affects the possibility of industrial AI vision SaaS vendors transmitting model-training data or inference data overseas, in fact setting a significant compliance barrier for overseas algorithm cloud platforms (such as Cognex's cloud-training service) to enter Chinese industrial-data scenarios, objectively supporting the local competitive advantage of domestic industrial AI vision SaaS platforms (SMore, AInnovation).

VII. Export Controls and Technical Autonomy: The Security Attributes of Vision Algorithm SDKs

Machine vision algorithm SDKs, as an important category of industrial software, have entered the scope of attention of the China-US technology game. The US Department of Commerce's Bureau of Industry and Security (BIS) has implemented specific export-control wording on software containing advanced machine-learning algorithms (especially those with facial-recognition or object-detection capabilities); the EU's mandatory evaluation requirements for high-risk AI systems through the Artificial Intelligence Act (including functions used for safety-critical judgments in industrial vision systems) may set new compliance thresholds for the application of foreign algorithm vendors in Chinese manufacturing-safety systems.

From the perspective of China's industrial security, if industrial vision SDKs are incorporated, as critical industrial software, into the promotion scope of "information technology application innovation" (Xinchuang), it will further drive the vision systems of key industries (energy, finance, government affairs) to prioritize domestic SDKs. Currently, machine vision SDKs have not yet been listed in the Xinchuang catalog, but the policy orientation of "industrial software localization" has substantively affected the procurement decisions of central enterprises and state-owned factories—more and more state-owned automotive factories (FAW, Dongfeng, Changan) explicitly require prioritizing the evaluation of domestic industrial-software solutions when building or transforming production lines.

The evolution of this security attribute means that the competition over machine vision SDKs is no longer merely a contest of technical capability and cost-effectiveness, but has acquired a dimension of strategic scarcity—whoever can be the first to establish large-scale deployment cases in key industrial scenarios and accumulate a complete industry-validation system will occupy the most advantageous position in the future policy-driven replacement wave.

---

Chapter 12　The Research Institute's Judgment

I. Industrial Large Models: The Critical Two Years from Demo to Mass Production

2026 is the critical watershed year for industrial large models to move from the laboratory and concept validation to large-scale mass-production deployment. The 2025 revenue data of enterprises such as SMore and AInnovation have already proven that the commercial value of industrial large models in some scenarios (multi-category defect detection, few-shot anomaly recognition, cross-model migration) is real; the core proposition of 2026 is whether they can maintain the technical reliability of individual projects and a replicable gross-margin structure in the large-scale deployment across hundreds or even thousands of factories.

This Research Institute's judgment is: 2026–2027 will see the first real explosive period of large-scale industrial-large-model application in mid-tier manufacturing enterprises. The driving forces come from three aspects: first, leading factories (head 3C / lithium battery) have completed the technical validation of industrial large models and will transmit the demonstration effect to their supplier chains (suppliers face the quality-upgrade requirements of vehicle manufacturers); second, the lightweight-deployment solutions of industrial-large-model vendors (edge-side inference, no cloud connection required) continuously lower the technical threshold for small and medium factories; third, the policy time node of the "Implementation Opinions on the 'AI+ Manufacturing' Special Action" (2027) will form a wave of policy-driven concentrated procurement.

It is expected that in 2027, the new market size of industrial large models in AI quality-inspection scenarios will exceed 4 billion yuan, becoming the niche direction with the largest incremental contribution to the entire industrial vision SDK market.

II. The Continuous Penetration of No-Code Platforms

The engineer gap in the machine vision field is the long-term structural contradiction constraining industrial scaling. China has thousands of system integrators, but high-level vision algorithm engineers (who are proficient in both classical algorithms and deep learning) are severely in short supply, with the industry talent gap estimated at more than 100,000 people. This objectively creates continuous market demand for "no-code/low-code" vision configuration platforms.

VisionMaster's drag-and-drop process orchestration, SMore's graphical AI configuration interface, AInnovation's no-code training platform, and Lingkong's VisionWARE industry templates are all working toward the same direction: enabling factory engineers without an algorithm background to independently deploy and maintain vision-inspection systems. This is a process of "software-capability sinking"—transforming algorithm-configuration work that originally required expert-level engineers into tasks that ordinary engineers can complete through intuitive operations.

It is expected that over the next 3 years, the penetration rate of no-code vision platforms in mid- and low-end inspection scenarios (simple appearance defects, dimensional-qualification judgment, barcode reading) will exceed 50%, becoming the mainstream form of newly deployed vision systems each year; while high-precision measurement (micron level) and complex 3D scenarios (multi-target overlap, dynamic-scene tracking) still require the intervention of professional vision engineers, which is the core value range distinguishing commercial SDKs from no-code tools.

III. The Accelerated Fusion of Vision + Embodied Intelligence

Embodied AI—AI systems with perception, reasoning, and physical-execution capabilities (humanoid robots, mobile-manipulation robots)—is moving from research demonstration to the industrial-application stage, and is the most important emerging demand source for industrial vision SDKs over the next 5 years. The industrial vision algorithm SDK will undertake a dual role in this direction:

One is as the real-time vision-sensing layer of the robot's "eyes" (requiring a lower-latency 3D vision SDK, with end-to-end sensing latency compressed from the current 300 milliseconds to within 50 milliseconds); the other is as the "cognitive interface" of the robot's interaction with the factory environment (requiring the combination of vision perception with large language models, understanding natural-language task instructions and converting them into vision-perception needs and grasping plans).

Both Orbbec and Mech-Mind have explicitly listed embodied-intelligence vision as a core growth engine for 2026–2027. This Research Institute believes that the fusion of vision + embodied intelligence will give rise to the next-generation technical form of vision SDKs—"a robot vision operating system integrating perception-reasoning-execution," and this form will blur the boundary between the traditional sense of "algorithm SDK" and "robot-control software," also meaning that 3D vision SDK enterprises such as Mech-Mind and Orbbec may become key software-infrastructure providers in the robotics industry chain.

IV. The Popularization Inflection Point of 3D Vision

3D vision has long been mainly concentrated in high-value scenarios such as automotive and semiconductor due to high cost (3D camera + high-performance processor + professional algorithm) and deployment complexity (calibration-accuracy management, point-cloud-processing professional threshold). But after 2025, the cost of structured-light cameras and ToF sensors continues to decline (some models have entered the price range below 3,000 yuan, suitable for mid-scale scenarios), and the improvement in the computing density of edge GPUs (NVIDIA Jetson Orin) means the real-time performance of point-cloud processing is no longer an obstacle. These two changes are driving 3D vision to spread to broader manufacturing scenarios.

This Research Institute expects that by 2028, the three scenarios of lithium-battery Pack assembly (100% 3D measurement of cell thickness), 3C precision assembly (vision-guided AA alignment of phone camera modules), and metal-casting quality inspection (volumetric quantification of casting defects) will become the main sources of new demand for 3D vision algorithm SDKs, jointly driving a new market size of more than 3 billion yuan.

V. Going Global: The Next Growth Curve of Domestic SDKs

As Chinese manufacturing enterprises (CATL, BYD, Foxconn, Luxshare Precision, BYD ATTO3 Southeast Asian manufacturing base) accelerate the construction of overseas factories in Southeast Asia, Europe, and North America, their demand for the familiar domestic vision software toolchain also goes overseas accordingly. This provides domestic SDK vendors such as Lingkong, Hikrobot, and Mech-Mind with a historic window to establish share in the European and Southeast Asian markets.

At the same time, Cognex's Q3 2025 earnings report has already indicated that the transfer of some consumer-electronics production lines to Vietnam and Malaysia has had a certain structural impact on its China revenue—this trend of "manufacturing going global" will continuously bring new competitive opportunities to Chinese vision SDK vendors with a global layout (Mech-Mind, SMore). Mech-Mind's customer scale serving more than 50 countries worldwide is precisely a direct embodiment of this trend's benefit.

VI. The Research Institute's Comprehensive Judgment: The Critical Point of Domestic Replacement from Quantitative to Qualitative Change

Looking back over the past five years, the replacement path of domestic machine vision algorithm SDKs has experienced a three-stage leap from "cost replacement" to "function parity" to "scenario overtaking." By mid-2026, this leap has reached the critical point:

In the AI vision track with the most incremental value, domestic vendors (SMore, AInnovation) have completed the technical overtaking of traditional foreign SDKs and are approaching the profitability threshold; in the 3D vision-guided robotics niche with the most growth potential, Mech-Mind has established overwhelming dominance in the domestic market and begun globalized replication; in the classical 2D vision base scenario with the largest volume, VisionMaster and VisionWARE have achieved head-on competition capability against HALCON.

The largest remaining battlefields are high-precision measurement (sub-micron level) and semiconductor front-end inspection. This Research Institute expects that the substantive domestic-replacement breakthrough in these two scenarios will require the time window of 2028–2030; before that, HALCON and dedicated semiconductor-inspection algorithms will maintain their dominant position in these two niches.

As an industrial B2B data platform covering 4.8 million producing factories, Tianxia Gongchang will continuously track the procurement-demand dynamics of vision-inspection equipment and algorithm software in Chinese manufacturing, providing industry participants with market insights based on real supply-and-demand data. To gain in-depth understanding of the factory procurement demand and supplier distribution in specific niche fields, you are welcome to visit this platform (www.tianxiagongchang.com) to query.

III. The Data Flywheel: The Core Moat of AI Vision Vendors

The most important and hardest-to-quantify competitive dimension in the industrial AI vision track is the maturity of the Data Flywheel effect. Similar to the user-behavior data flywheel of consumer internet, the data flywheel of industrial AI vision consists of three links: more factory-deployment cases → accumulating more real defect-sample data → training better AI vision models → higher inspection accuracy and faster new-scenario adaptation speed → winning more factory-deployment cases.

Once this flywheel reaches critical scale, it produces a significant first-mover advantage. SMore's data advantage comes from: first, the Tesla project accumulated a large amount of high-precision industrial-defect samples such as high-density PINs and printed circuit boards; second, the deployment at ultra-large foundries such as Luxshare Precision and Foxconn accumulated a large amount of cross-SKU, cross-product-line defect-diversity data; third, the decentralized data source composed of more than 730 customers makes the model's generalization capability (cross-scenario transfer capability) far superior to a model trained based only on single-customer data. These three together build the data moat of SMore's AI vision model quality, and are also one of the core narratives in its IPO prospectus.

AInnovation's data advantage focuses on the training-data depth of the "industrial large model AInnoGC": focusing on the three major vertical industries of electronics manufacturing, new energy, and automotive in manufacturing, the large model trained through industry-exclusive datasets has an AI vision accuracy in specific scenarios (such as lithium-battery separator defect detection, casting X-ray inspection) significantly higher than the direct application effect of general-purpose AI vision models. This industry-specialization route enables AInnovation to establish a data barrier in its target industries, but at the same time limits its ability to rapidly expand laterally to new industries.

Hikrobot's data moat comes from its unique hardware-channel advantage: as one of the world's largest-shipment industrial camera brands, Hikvision's cameras are spread across tens of thousands of factories, and VisionMaster, relying on the relationship established with users through the hardware channel, can sign data-sharing agreements with willing customers, obtaining real industrial-image data with "algorithm optimization" as the value return. This data-acquisition channel is a structural advantage that pure-software AI vision enterprises find hard to replicate.

IV. The Market-Disruption Potential of No-Code / Low-Code Vision Configuration Platforms

Traditional users of industrial vision SDKs (professional vision engineers) are facing a structural trend of being disrupted by "no-code vision platforms." As deep-learning models mature and deployment toolchains simplify, a new type of industrial vision tool is emerging: enabling factory quality-inspection engineers without SDK programming ability to complete AI vision model training and deployment through a drag-and-drop interface, without writing a single line of code.

SMore's SMore ViMo and AInnovation's AInnoGC industrial-large-model platform are both actively advancing this direction. Their core technical premise is: deep-learning anomaly-detection models (such as PatchCore, ReverseDistillation), after training on enough normal samples, can achieve "out-of-the-box" inspection capability without labeling defect samples. This means a factory quality-inspection engineer only needs to collect 100–200 images of normal products, upload them through a cloud or local training interface, and the system can automatically generate a defect-detection model for that product, with a false-detection rate usually below 0.5%—for most food and general-quality inspection scenarios in chemicals, this already meets actual needs.

The market impact of this technical route is profound: it expands the user boundary of industrial vision from "thousands of professional vision SDK engineers" to "hundreds of thousands of factory quality-inspection engineers," greatly expanding the serviceable market size; at the same time, it makes the ecosystem barrier that traditional SDK vendors rely on the engineer community to build face erosion—if no-code platforms can meet 60%–70% of industrial-inspection needs, the ecosystem value of professional SDKs will contract to the remaining 30%–40% of high-precision, high-complexity scenarios.

HALCON's response to this trend is to launch MERLIC (MVTec's no-code vision solution), providing low-threshold configuration tools for scenarios that do not require deep programming; VisionMaster, beyond the standard SDK, provides a "visual configuration mode" oriented toward factory on-site engineers. The no-code trend is reshaping the product definition of the entire track, and 2026–2028 is expected to be the key period for the commercialization validation of this direction.

V. The Fusion of Vision SDK and AI Agent: The Architectural Evolution of Next-Generation Industrial Vision

In 2025–2026, an architectural trend worth focusing on emerged in the industrial-AI field: the deep fusion of vision SDKs and AI Agents (autonomous decision-making agents). Traditional industrial vision systems are a closed loop of "perception + judgment" (capture image → inspect → output Pass/Fail), while the emerging industrial AI Agent architecture takes vision perception as one of the Agent's perceptual modalities, connecting it to multi-step reasoning and autonomous decision-making capabilities.

Mech-Mind's Mech-Mind product system (the collaboration of Mech-Vision + Mech-Viz) is already an early practice of this direction: the tight coupling of vision perception (Mech-Vision's 3D positioning) and path planning (Mech-Viz's motion planning) gives the robot workstation the ability to dynamically adjust the grasping strategy according to the workpiece's real-time state, which surpasses the traditional "perception → fixed action" vision-guided-robot model.

A more cutting-edge direction is: taking the output of the industrial vision SDK (inspection results, position coordinates, quality grading) as one of the multimodal inputs of an industrial large model (Industrial LLM), letting the large model perform cross-process quality traceability based on vision-perception results ("the coating defects of this batch of products—which upstream process's parameter anomaly are they related to?") and process-parameter-optimization suggestions ("based on the coating-thickness distribution detected by vision, it is suggested to adjust the coater speed and tension parameters"). AInnovation's AInnoGC industrial large model has already achieved this preliminary closed loop of "vision perception + process decision" in some customer scenarios.

This architectural evolution means that the competitive boundary of industrial vision SDKs will no longer be limited to the middleware layer of the "algorithm-tool library," but will extend to the deep integration capability with industrial large models and industrial AI Agents. Vendors that can be the first to establish a complete pipeline of "vision perception → large-model reasoning → process decision" will define the architectural paradigm of next-generation industrial vision systems in the industrial-AI era.

---

Chapter 13　Risks (Industrial-Large-Model Homogenization / Algorithm Patents / Overseas Giants' Counterattack)

I. Industrial-Large-Model Homogenization: The Deep Hidden Danger of Accelerated Involution

In 2025, almost all head machine vision SDK vendors in China affixed the "industrial large model" label to their product packaging: SMore's AI vision large model, AInnovation's AInnoGC industrial large model, Lingkong's F.Brain vision large model, the industrial vision large model built into Hikrobot's VisionMaster 5.0... The differentiation narratives are increasingly consistent, the core technical paths (pre-trained vision backbone + industrial-scenario fine-tuning + no-code interface + few-shot adaptation) are highly convergent, and the market's homogenization concerns about this track are significantly rising.

The direct consequence of homogenization is price-war pressure. When multiple industrial-large-model products converge on the core metrics of defect detection (miss rate, false-alarm rate), customers' selection criteria will shift to price and service, and vendors' gross margins will face downward pressure. For startup vision AI enterprises still in the loss stage (SMore's net loss is still 272 million yuan), the continuation of the price war will put greater pressure on cash flow, and the "blood-transfusion" function of financing will become more critical.

There are three possible breakthrough paths: first, beyond general defect detection, building high-threshold vertical-industry knowledge depth (such as specific physical-defect modeling for semiconductor front-end inspection, compliance-validation capability for medical-device precision measurement); these high-barrier scenarios are hard for competitors to replicate quickly; second, through the customer data flywheel (accumulating a large amount of industry-labeled data, continuously improving model accuracy and cross-scenario generalization), building a data barrier that latecomers find hard to surpass; third, deeply combining the industrial large model with "heavy" capabilities such as robot control and ERP/MES integration, forming a systematic solution of "vision + execution + digitalization," elevating the competitive dimension from "is the algorithm good" to "whose overall digitalization-transformation solution has the lowest cost and highest value."

II. Algorithm-Patent Risk: HALCON's Legal Moat

MVTec holds several patents on HALCON's core algorithms (especially Shape-based Matching, certain 3D point-cloud registration algorithms, and deep-learning industrial-application methods), with substantive legal protection under the European and US patent systems. Historically, MVTec has protected its market position through patent litigation in Europe.

As domestic SDKs rise and continuously erode HALCON's market share, the patent-litigation risk cannot be ignored, especially when domestic SDKs begin to go overseas into the European and US markets, facing a legal environment where MVTec and Cognex have a home-court advantage. Currently, mainstream domestic SDKs generally circumvent direct-copying risk through a "from-scratch implementation" approach, and conduct differentiated design in the algorithm-implementation details. But when algorithm functions and effects are highly convergent, there is still room for dispute in the demarcation of patent boundaries, especially in the determination of the protection scope of the "equivalent patents" of shape matching.

Domestic SDK vendors should proactively build their own algorithm-patent portfolios (including filing domestic and international invention patents), forming a reciprocal patent-defense capability, to avoid the unexpected impact of patent litigation during the high-speed growth stage. Lingkong has applied for several vision-inspection-related patents domestically, and Mech-Mind has a patent layout in the 3D vision-guided robotics field; this is the correct strategic direction, but the coverage breadth and the completeness of the international-patent layout still need continuous strengthening.

III. Overseas Giants' Counterattack: Price Cuts and Localization Strategies

HALCON and Cognex are not unaware of the competitive pressure from China. Over the past three years, both companies have, to varying degrees, adjusted their strategies in the China market: HALCON offers more flexible enterprise-level licensing schemes in some scenarios (providing volume discounts for large integrators and OEM partners, not disclosed publicly), lowering the total cost of ownership; Cognex, through continuous updates of deep-learning algorithms (the rapid iteration of ViDi Suite), narrows the gap with domestic AI vision platforms in AI vision functions, and the ease-of-use level of its "Interactive Training" experience in 2025 has approached that of some domestic AI vision platforms.

What is more worth being vigilant about is that if HALCON and Cognex further lower the China-market pricing of AI vision tools in 2026–2027 (by setting up a localized pricing system to directly compete with domestic products), while increasing investment in China local technical-service teams, it will constitute a new round of competitive pressure on domestic SDKs from the high-end market. This scenario is not impossible—Cognex's "near-bottom" market expectation in the automotive industry may prompt it to adopt a price-for-volume strategy to increase penetration in China's consumer-electronics and lithium-battery scenarios.

IV. Geopolitical Risk: The Two-Way Constraints of Going Global

Machine vision algorithm SDKs are not on the core control list of US export controls (EAR CCL) in most application scenarios, but if specific semiconductor-inspection scenarios are involved (wafer-level defect-detection algorithms, which may fall into a specific export-control category), one needs to pay attention to whether an export license is required at the time of export. Domestic SDK vendors should conduct export-compliance reviews in advance when going overseas to high-end European and US manufacturing customers.

In addition, the US's AI-chip export controls (restrictions on training-grade GPUs such as NVIDIA H100/H800) affect the supply of high-end GPUs required for industrial-large-model training. Domestic SDK vendors need to plan ahead in algorithm lightweighting (reducing the dependence on large-model parameter counts) and adaptation to domestic AI chips (Huawei Ascend 910B/C, Cambricon MLU370, etc.), ensuring the supply-chain security on both the training and inference sides.

V. Customer-Dependence and Order-Concentration Risk

AI vision enterprises such as SMore often have head-customer-dependence problems during the high-speed growth stage: the revenue-contribution concentration of the top 5–10 customers (Tesla, Luxshare Precision, BOE, etc.) may be as high as 40%–60%. Such head customers have strong bargaining power, and as they build their own AI capabilities (Tesla already has a large internal AI team), they may gradually establish part of the internal-development capability for vision inspection, reducing their dependence on externally procured vision AI platforms.

At the same time, the payment cycle of large manufacturing customers is usually long (90–180-day payment terms), and at a stage when the AI vision enterprise's own cash flow is tight, the payment-term delay of large customers may constitute pressure on operations. How to, while maintaining head customers, accelerate the customer expansion of mid-waist manufacturers (annual output value in the 500-million-to-5-billion-yuan range, currently with an AI vision penetration rate below 10%), achieving revenue dispersion and cash-flow improvement, is the core issue for vision AI vendors to improve their commercial resilience.

VI. The Uncertainty of Technological Leaps

The pace of technological evolution in the machine vision algorithm field exceeds most expectations. Before 2018, no one could predict that deep learning would completely change the mainstream technical paradigm of industrial defect detection within five years; before 2023, no one could predict that the visual-understanding capability of large language models would approach the industrial-usable level within two years. The arrival mode and impact scope of the next disruptive technological node have extremely high uncertainty, which constitutes a hard-to-quantify risk for both current market leaders and investors.

Directions that may trigger a technological leap include: quantum optical sensing (theoretically capable of breaking through the diffraction limit to achieve nanometer-level non-destructive inspection); the real-time industrialization of Neural Radiance Fields (NeRF) (reconstructing a 3D model from a single image, potentially replacing the current expensive 3D-sensor hardware); and a multimodal robot-perception-manipulation operating system (fusing vision, touch, and language understanding, giving robots human-level object-manipulation capability). A technological breakthrough in any one direction could reshape the current market landscape, resetting existing competitive advantages within 3–5 years. This "technology-cycle risk" is a structural uncertainty that all participants in the machine vision SDK industry must face directly.

IV. The Long-Term Strategic Path of Domestic SDKs: From "Coverage Competition" to "Deep-Value Competition"

In 2026–2030, China's head industrial vision SDK enterprises will face a core strategic-transformation node: when the coverage competition of domestic replacement (gaining deployment at more customers and more industries) tends to saturate, the competitive focus will shift from "customer acquisition" to "value-added"—achieving higher revenue density through deeper value mining at customers where there is already a deployment.

The strategic paths corresponding to this transformation include: first, "from single-point inspection to global quality management"—the application of the vision SDK expands from a single inspection station to the quality-data integration of the entire production line, providing the factory with cross-process defect-traceability and quality-trend-analysis capabilities, rather than merely outputting the Pass/Fail signal of a single station; second, "from quality inspection to process optimization"—combining vision-inspection data with process-parameter data, building a correlation model of "vision quality signal ↔ process parameters," providing the factory with process-optimization suggestions, achieving the value leap from "discovering defects" to "preventing defects"; third, "from production-line vision to whole-plant intelligence"—extending the vision-perception capability to scenarios such as warehousing logistics (goods recognition, shelf scanning) and security monitoring (dangerous-behavior recognition), evolving the vision SDK from a production-line inspection tool into the core sensing infrastructure of factory intelligence.

The common feature of these three paths is: transforming the relationship of a one-time SDK-license sale into a long-term cooperative relationship that continuously provides incremental value, achieving ARR (Annual Recurring Revenue) growth while greatly increasing the customer switching cost (changing the SDK means losing the quality-data-analysis capability accumulated over many years).

V. Embodied Intelligence and Industrial-Robot Vision: The Disruptive Opportunity of 2030

From a longer time dimension, the industrial vision SDK track faces a new variable that may completely change the competitive landscape: the commercialization of embodied AI and next-generation industrial robots.

Current industrial vision systems are "fixed vision"—the camera is fixedly mounted at a specific position, and the algorithm is optimized for a fixed viewing angle and fixed product specifications. But embodied-intelligence robots require "mobile vision"—the robot carries vision sensors and moves through the factory, needing to understand the dynamic environment (workpiece position, robot posture, obstacles) in real time, and the core of the vision algorithm shifts from "high-precision judgment of a fixed scene" to "real-time semantic understanding of a dynamic scene."

This shift places completely different technical requirements on the vision SDK: it requires strong real-time 3D sensing (point-cloud processing above 30 frames per second), dynamic-object tracking (real-time pose estimation of a workpiece on a conveyor belt), scene-semantic understanding ("is this a drilling station or a welding station?"), and multimodal fusion (vision + touch + force sense). These capabilities exceed the design goals of current industrial vision SDKs, and will give rise to entirely new technical architectures and new market opportunities.

Mech-Mind's strategic direction is the clearest—its Mech-Mind product system is naturally adapted to robot vision guidance, and as industrial robots evolve toward embodied intelligence, Mech-Mind's market coverage is expected to expand in sync with the shipment volume of robot-body manufacturers. Orbbec focuses on ToF 3D sensors, and its strategic layout in the embodied-intelligence (humanoid-robot vision) track is precisely a bet on this disruptive opportunity. If Hikrobot and Lingkong want to maintain their advantage in the embodied-intelligence era, they need to undergo a fundamental technical-architecture evolution in dynamic-scene vision algorithms, rather than merely layering function modules on existing products.

VI. Market-Size Forecast for 2026–2030

Based on the above analysis, this report gives the following range forecast for the China industrial vision algorithm SDK market size:

Baseline scenario (assuming accelerated domestic replacement, steady improvement in AI vision penetration, and a moderate macroeconomic recovery): the market size is about 4.2–4.8 billion yuan in 2026, about 6.5–7.5 billion yuan in 2028, and about 10–12 billion yuan in 2030, with a compound annual growth rate of about 18%–22%.

Optimistic scenario (strong policy promotion, the expansion of Xinchuang to key manufacturing control systems, and the early commercialization of embodied-intelligence robots): the market size in 2030 can reach 14–16 billion yuan, with a compound annual growth rate of about 25%–28%.

Conservative scenario (sustained macroeconomic downward pressure, contraction of manufacturing capital expenditure, and slow replacement of domestic SDKs in high-end scenarios): the market size in 2030 is about 8–9 billion yuan, with a compound annual growth rate of about 12%–15%.

The sensitivity ranking of core variables: first, the penetration speed of AI vision in automotive and semiconductor scenarios (currently low penetration, with large but difficult improvement space); second, policy strength (Xinchuang expansion, mandatory domestic-replacement requirements); third, the commercialization pace of embodied-intelligence robots (the incremental pull effect on the vision SDK market after 2028); fourth, changes in the global supply-chain landscape (the speed at which Chinese manufacturing capacity transfers to Southeast Asia will affect the domestic-demand growth rate).

Behind these forecast ranges is a structural judgment with relatively high certainty: the long-term growth trend of the industrial vision SDK market is irreversible, the proportion of software-layer value in the entire machine vision system will continuously increase, and the share of domestic SDKs in this trend will steadily rise from the current about 45% to 60%–70% in 2030, gradually recovering lost ground in high-end scenarios while establishing a significant market-leadership position in the emerging scenarios of AI vision.

---

Data Sources

The data and factual basis of this report are mainly derived from the following channels:

Tianxia Gongchang Industry Database This platform covers the enterprise information and industry-chain data of China's 4.8 million producing factories. The factory distribution and supply-demand data in the directions of machine vision, vision robots, industrial vision software, defect detection, etc., were obtained by this Research Institute based on real-time database searches. Readers can visit this platform (www.tianxiagongchang.com) and query relevant factory-supplier information by industry category and regional dimension.

Listed-Company Announcements and Prospectuses Luminescence Technology Co., Ltd. [Lingkong] (688400.SH) 2025 Annual Report (disclosed April 29, 2026) and 2025 quarterly reports; AInnovation (2121.HK) 2025 Interim Results Report (disclosed August 2025); SMore (SmartMore Inc.) Hong Kong Stock Exchange listing application materials (submitted March 2026); Orbbec Technology Group Co., Ltd. 2025 Annual Report (disclosed on the Shenzhen Stock Exchange); Cognex Corporation Form 8-K (Q3 2025 Earnings, October 29, 2025); Hikrobot 2025 operating data (Hikvision annual report and business updates, 2026).

Industry Research Reports GGII (GaoGong Industry Research Institute) "2025 Machine Vision Industry Research Report" (June 2025); Zhiyan Consulting "2025 Analysis of the Development Status and Future Trends of China's Industrial Machine Vision Industry"; SNS Insider "3D Machine Vision Market Size Worth USD 19.13 Billion by 2032" (May 2025); MarketsandMarkets "Machine Vision Market Size, Share & Trends, 2025 to 2030"; IDC China Large-Model Application Market Share Report (2024 annual data); Qianzhan Industry Research Institute "Machine Vision Industry Thematic Research" (2025).

Policy Documents The Ministry of Industry and Information Technology and seven other departments' "Implementation Opinions on the 'AI+ Manufacturing' Special Action" (January 7, 2026); "Guidelines for AI-Empowered Transformation of Key Manufacturing Industries" (supporting document, January 2026); relevant announcements of the third phase of the National Integrated Circuit Industry Investment Fund (2024–2026); relevant policy documents of the national "14th Five-Year Plan" for high-quality manufacturing development.

Technical and Community Sources MVTec HALCON official technical documentation (version 24.11) and product white papers; Cognex VisionPro / ViDi Suite product white papers and official blogs; Mech-Mind (Mech-Mind Robotics) official technical documentation and product-release announcements; Lingkong official website VisionWARE product page and technical blog; Hikrobot VisionMaster official product page (hikrobotics.com); CSDN / Cnblogs / Zhihu: technical discussions in the industrial-vision engineer community (as a market-temperature reference for engineer-ecosystem scale and product-coverage scope).

Note: Some of the market-size and growth data cited in this report are derived from the estimation scopes of different institutions, with reasonable statistical differences, and the Research Institute has tried to indicate the data sources and adopt mainstream-range values when citing; numbers not separately noted are comprehensive estimates by the Research Institute based on multi-source data and do not constitute investment advice. The industry standards referred to in the text are all expressed by their full technical names.

---

Appendix I　Glossary of Machine Vision Algorithm SDK Terminology

To help readers establish a unified understanding of terminology when reading this report, the following briefly explains the core technical vocabulary that frequently appears in the report, and describes its actual meaning in industrial scenarios.

Algorithm SDK (Software Development Kit): the abbreviation for Software Development Kit. In the context of this report, it specifically refers to the function library and development framework that encapsulates industrial vision algorithms, the technical foundation for system integrators and equipment vendors to build vision applications.

Template Matching: an algorithm that searches an image for the region most similar to a predefined template, used for part positioning and feature recognition. Shape-based Matching, by extracting edge-gradient features, has higher robustness to illumination changes than grayscale-correlation matching.

Sub-pixel Accuracy: the ability of vision measurement or positioning to have an accuracy finer than one pixel unit. For example, under a camera with a resolution of 10 microns per pixel, sub-pixel accuracy means a positioning error finer than 10 microns. Sub-pixel algorithms are usually implemented based on interpolation (interpolating the precise extremum position of the grayscale gradient) or model fitting (fitting the edge to a continuous mathematical curve).

Point Cloud: a collection of discrete points in 3D space, each point having 3D coordinates (x, y, z), and in some scenarios also accompanied by color (RGB) or reflection-intensity information. The point cloud is the raw data format output by 3D vision sensors (structured light, ToF, laser), and after algorithmic processing can be used for 3D positioning, surface reconstruction, and volume measurement.

Depth Map: a data structure that stores the distance information from each pixel to the sensor plane in a 2D image format. The depth map is another expression of the point cloud, computationally easier to process, but loses absolute 3D-coordinate information and must be combined with the camera intrinsics to restore the point cloud.

Structured Light: a depth-acquisition technology that projects a known pattern (sinusoidal stripes, Gray code, random speckle) onto the target and computes the 3D shape based on the pattern's deformation amount on the target surface. Structured light has high accuracy (sub-millimeter to micron level), is one of the mainstream technical routes for industrial 3D vision, but has limited adaptability to highly reflective materials (mirror-finish metal).

Time of Flight (ToF): a sensing technology that acquires depth by measuring the round-trip time of a laser pulse or modulated light signal. ToF has a high frame rate (up to 90fps or more), is suitable for dynamic scenes, but its accuracy and resolution are lower than structured light, suitable for mid-accuracy scenarios such as warehousing logistics and robot obstacle avoidance.

GenICam Protocol: the Generic Interface for Cameras, a universal programming-interface standard in the industrial-camera field. GenICam-compatible SDKs and cameras can achieve "plug-and-play" device interconnection, reducing the driver-adaptation workload between different camera brands, and are the basic protocol layer of modern industrial vision systems.

GigE Vision: an industrial-camera interface standard that transmits image data based on Gigabit Ethernet (GigE), supporting a non-relay transmission distance of up to 100 meters, suitable for distributed industrial-camera deployment, and one of the most widely used interfaces in industrial vision today.

AOI (Automated Optical Inspection): the abbreviation for Automated Optical Inspection, specifically referring to automated appearance-inspection equipment based on visible-light imaging, widely used for solder-quality inspection of PCBs (printed circuit boards). AOI is one of the most important downstream application scenarios for machine vision algorithm SDKs, and also the scenario where domestic SDKs first achieved large-scale replacement progress.

GMP (Good Manufacturing Practice): the quality-management specification for drug (and some medical-device) production. GMP specifications require that the key processes of pharmaceutical manufacturing must go through validated quality-control means, including appearance vision inspection. GMP-certified vision systems need to submit change-validation reports when updating algorithms and changing equipment, greatly increasing the system's switching cost.

ONNX (Open Neural Network Exchange): an open neural-network exchange format that allows trained models to be migrated between different deep-learning frameworks (PyTorch, TensorFlow, MXNet) and efficiently inferred on inference engines (TensorRT, OpenVINO, ONNX Runtime). Mainstream industrial vision SDKs (including HALCON 24.11, VisionMaster 5.0) already support ONNX-model import, enabling users to directly integrate their self-trained AI models into the vision pipeline without rewriting the underlying inference code.

ICP (Iterative Closest Point): the Iterative Closest Point algorithm, used for 3D point-cloud registration, iteratively optimizing the rigid-body transformation (rotation + translation) between two sets of point clouds to align them as closely as possible. ICP is a core algorithm for workpiece pose estimation and quality-deviation detection in industrial 3D vision, and its convergence speed and accuracy are important indicators for evaluating the point-cloud-processing capability of a 3D vision SDK.

Bin Picking (Basket Picking / Random Picking): refers to the application scenario where an industrial robot identifies the 3D pose of a single part from parts randomly stacked in a material basket and completes the grasp. Bin Picking needs to combine 3D vision (point-cloud acquisition and target recognition) and robot path planning (collision-free grasping-sequence generation), making it one of the most technically complex scenarios in 3D vision-guided robotics, and a core product scenario for vendors such as Mech-Mind.

Anomaly Detection: an algorithm paradigm that, without requiring abnormal-sample labeling, learns the distribution characteristics of normal data and identifies samples deviating from the normal distribution as anomalous. Industrial anomaly detection is the core application of unsupervised learning in industrial quality-inspection scenarios, with significant advantages in scenarios with scarce defect samples (early stage of new products, low-frequency defect types).

---

Appendix II　Quick-Reference Table of Key Data for Major Enterprises

To facilitate readers in quickly consulting the core data of the major enterprises involved in this report, the following summarizes the key financial and business indicators of each enterprise, with data as of June 19, 2026.

Luminescence Technology [Lingkong] (688400.SH)

Main business: machine vision equipment + VisionWARE algorithm platform + optical-communication devices.

2025 financial data: operating revenue 2.912 billion yuan, up 30.35% year-on-year; net profit attributable to the parent 161 million yuan, up 50.70% year-on-year; net profit attributable to the parent excluding non-recurring items 123 million yuan, up 86.05% year-on-year; R&D investment as a proportion of revenue about 15.91% (first three quarters of 2025).

Core business indicators: VisionWARE iterated to version 6.4, with 18 algorithm libraries and nearly 200 algorithm tools; the F.Brain vision large model has been deployed at scale in the consumer-electronics, new-energy, and printing-and-packaging industries; 2025 new-energy business revenue 185 million yuan, up 36.01% year-on-year; smart vision equipment business revenue about 775 million yuan, accounting for about 26.6% of total revenue.

AInnovation (2121.HK)

Main business: industrial AI vision-inspection platform + AInnoGC industrial large model + industrial robots.

First-half 2025 financial data: revenue 699 million yuan, up 22.3% year-on-year; gross profit 245 million yuan, up 26.7% year-on-year; adjusted net loss 6.68 million yuan, down 82.1% year-on-year; "AI + manufacturing" business accounting for nearly 80% of total revenue (556 million yuan).

Market position: ranked seventh in China's large-model application market share (IDC data), the only vendor focused on the industrial field; listed in Hong Kong (2022), stock code 2121.HK.

SMore (SmartMore)

Main business: industrial AI vision large model + no-code quality-inspection configuration platform + industrial AI agent.

Key data: 2025 revenue nearly 1.1 billion yuan, three-year compound growth rate over 40%; more than 730 customers (including Tesla, Luxshare Precision, BOE, Carl Zeiss); latest valuation USD 1.230 billion (Series C, February 2026); Hong Kong Stock Exchange IPO application submitted in March 2026, sponsors: Morgan Stanley, CICC, Deutsche Bank; 2023–2025 adjusted net loss narrowed from 394 million yuan to 272 million yuan.

Mech-Mind (Mech-Mind Robotics)

Main business: 3D industrial cameras (Mech-Eye series) + Mech-Vision sensing SDK + Mech-Viz robot path planning.

Key data: ranked first in China's 3D vision-guided industrial robot market for five consecutive years (2020–2024); serving customers in more than 50 countries and regions worldwide; valuation entering the unicorn ranks (above USD 1 billion); 2024 reflective-object point-cloud accuracy improved by about 90% (next-generation structured-light depth-estimation technology).

Orbbec (Orbbec)

Main business: 3D vision sensors (ToF + structured light) + SDK + embodied-intelligence vision solutions.

Key data: achieved its first full-year profit since listing in 2025, net profit attributable to the parent 128 million yuan; achieved its first single-quarter profit in Q1 2025; revenue composition: consumer-grade 62%, 3D sensors 30.9%, industrial-grade 2.7%; strategic focus shifting toward embodied-intelligence robot vision sensing.

Cognex Corporation (CGNX, NASDAQ)

Main business: industrial cameras (In-Sight) + VisionPro SDK + ViDi deep-learning tools.

Key data: FY2025 Q3 revenue USD 277 million, up 18% year-on-year; China-market Q3 growth about 9%; cumulative historical shipments of more than 4.5 million image products; cumulative revenue over USD 11 billion; the deep-learning module ViDi Suite iterating rapidly, with the "Interactive Training" experience continuously strengthened.

Hikrobot

Main business: VisionMaster SDK + machine vision complete machines + industrial robots (AMR/collaborative robots).

Key data: 2025 full-year overall revenue over 6.4 billion yuan (machine vision + robotics combined); self-developed industrial software with more than 600,000 license user instances; more than 20,000 customers served globally; VisionMaster 5.0 with more than 300 tools, integrating a three-layer architecture of industrial vision large model + edge learning + traditional algorithms; machine vision business compound annual growth rate about 50%.

---

Appendix III　Chronicle of Machine Vision Algorithm SDK Industry Development (2010–2026)

The following sorts out the important milestone events of the machine vision algorithm SDK industry since 2010, providing a historical perspective for understanding the current competitive landscape.

2010: MVTec released HALCON 10.0, introducing 3D shape-matching for the first time, extending algorithm coverage from 2D to 3D; the China machine vision market size was about 2 billion yuan, with foreign brands holding about 70% share.

2012: Cognex established a China regional technical-support team and began building localized services; Daheng Imaging launched the first domestic vision software with a complete algorithm toolchain; the China machine vision engineer community began to form a HALCON-centered technical-learning system (the number of HALCON-related technical blogs on CSDN grew rapidly).

2015: Lingkong accelerated VisionWARE development and released the first complete version oriented toward system integrators; deep learning (LeNet, AlexNet) began to draw the attention of the industrial-vision research community to AI defect detection, but engineering implementation had not yet begun.

2016: Hikrobot was officially founded and the VisionMaster project launched; Mech-Mind was founded in Beijing, focusing on 3D vision-guided robotics; Orbbec was founded in Shenzhen, focusing on ToF 3D sensors.

2018: Deep-learning defect detection (industrial image classification and segmentation based on ResNet and U-Net) began pilot validation in 3C electronics factories; the predecessor project of SMore began industrial-AI vision research at the Hong Kong University of Science and Technology; China-US trade friction began, and the urgency of localizing core industrial software began to heat up in the industry.

2019: SMore (SmartMore) was officially founded in Shenzhen, with Dr. Jiaya Jia as founder; AInnovation completed angel-round financing, focusing on industrial AI vision; the industrial AI vision track ushered in the first wave of capital enthusiasm, with multiple startups intensively founded. The China machine vision market size exceeded 8 billion yuan.

2020: The market share of China's machine vision domestic brands exceeded 50% for the first time, reaching a historic milestone; VisionMaster 3.0 was released, with more than 100 algorithm tools, formally entering the mainstream selection view of system integrators; SMore completed Series A financing and expanded into the 3C electronics and new-energy tracks.

2021: SMore completed a USD 200 million Series B financing (led by well-known institutions such as SoftBank, IDG, and CICC), becoming a star enterprise in the industrial AI vision track; AInnovation completed Series D financing; Mech-Mind completed Series B financing, with its 3D vision-guided robotics market share jumping to first in China; MVTec released HALCON 21.11, integrating a deep-learning anomaly-detection module.

2022: AInnovation listed on the Hong Kong Stock Exchange (code 2121.HK), becoming the first Hong-Kong-listed company in the domestic industrial AI vision track; the domestic-brand share of the China machine vision market rose to about 60%; VisionMaster 4.0 was released, integrating deep-learning tools for the first time, expanding to more than 200 algorithm tools.

2023: The explosion of ChatGPT triggered the industrial-large-model boom, with multiple vision AI enterprises announcing industrial-large-model strategies; Lingkong released the F.Brain industrial vision large model; the industrial AI vision capital boom cooled noticeably, and capital began to focus on profitability; the global machine vision market, affected by the slowdown in semiconductor demand, briefly declined in growth rate.

2024: VisionMaster 5.0 was released, integrating an industrial vision large model + edge-learning tools; Mech-Mind released reflective-object-dedicated depth-estimation technology, improving point-cloud accuracy by 90%; Orbbec achieved its first single-quarter profit in Q1; multiple domestic semiconductor-inspection equipment enterprises completed the preliminary domestic replacement of vision algorithms in back-end packaging scenarios; the China machine vision market size was about 20.7 billion yuan, up about 12% year-on-year.

2025: Lingkong's FY2025 revenue was 2.912 billion yuan, up 30.35%, with greatly improved profit; AInnovation's adjusted net loss narrowed to 6.68 million yuan, approaching break-even; SMore's revenue was nearly 1.1 billion yuan, completing Series C financing with a valuation of USD 1.230 billion; the Ministry of Industry and Information Technology and seven other departments issued the "Implementation Opinions on the 'AI+ Manufacturing' Special Action," incorporating AI quality inspection into the scope of national policy priorities; Cognex's Q3 2025 revenue was USD 277 million, up 18% year-on-year, with the deep-learning tool ViDi Suite accelerating its iteration; Orbbec achieved its first full-year profit in 2025.

First half of 2026: SMore submitted an IPO application to the Hong Kong Stock Exchange, intending to become the "world's first industrial AI agent stock"; the overall market share of domestic machine vision SDKs is estimated to exceed 65%; 3D vision algorithms began large-scale popularization in lithium-battery Pack and 3C precision-assembly scenarios; a critical node where industrial large models began to transition from POC to large-scale mass-production deployment.

This chronicle reveals the clear main thread of the evolution of the machine vision algorithm SDK industry: from foreign dominance in 2010, through the domestic start-up of 2016–2020, to the full competition of 2020–2025, and then to the lane-changing overtaking of the AI vision track in 2026—this is a history of technical catch-up that completed the journey from "complete dependence" to "partial overtaking" over fifteen years, and also a microcosm of the mutual reinforcement of Chinese manufacturing's digital transformation and technological self-reliance.

---

Appendix IV　Machine Vision Algorithm SDK Selection Guide

This section provides readers intending to evaluate or procure machine vision algorithm SDKs with a practical selection framework, comprehensively considering technical capability, commercial cost, service ecosystem, and risk factors.

Selection Step 1: Clarify the Core Task Type

Different vision tasks have significantly different requirements for the SDK's technical capabilities, and the first step of selection is to accurately define the core task type:

If the task is mainly precision measurement (sub-pixel positioning, dimensional measurement, geometric inspection), prioritize SDKs with strong classical-algorithm capabilities, recommended: HALCON, VisionMaster, VisionWARE (printing / precision-direction versions).

If the task is mainly appearance-defect detection, with variable defect types that are hard to describe by rules manually, prioritize deep-learning SDKs: SMore platform (few-shot, no-code), AInnovation platform (industrial large model, medium-to-large enterprises), VisionMaster deep-learning version (suitable for scenarios that already have Hikvision cameras).

If the task is mainly 3D measurement or robot guidance, prioritize: Mech-Mind (the first choice for 3D guided robotics), Lingkong (laser profiler + 3D algorithm, suitable for electrode-sheet / flatness inspection), Orbbec (cost-sensitive mid-accuracy 3D scenarios).

If the task involves barcode / OCR code reading, with extremely high accuracy and speed requirements, consider: Cognex (the world's strongest code-reader brand), VisionMaster OCR tools (a domestic high-cost-effective choice).

Selection Step 2: Evaluate Technical Capability (7 Dimensions)

After completing the task-type positioning, the candidate SDKs need to be quantitatively evaluated from the following 7 technical dimensions:

1. Algorithm accuracy: conduct a Benchmark test with actual project samples (including typical qualified products and typical defective/deviation products), evaluating the SDK's inspection accuracy, false-alarm rate, and miss rate in the target scenario.

2. Algorithm speed: measure the single-frame processing time on the target hardware platform (specific CPU/GPU model), confirming whether it can meet the production-line takt (usually required to be lower than the time the product passes the inspection station, with a 30% margin reserved).

3. Cross-platform support: confirm whether the SDK supports the target operating system (Windows, Linux) and the target hardware platform (x86 server, ARM embedded, edge-GPU box).

4. Camera compatibility: confirm whether the SDK supports the camera brands and interface protocols already procured or planned to be procured (GigE Vision, USB3.0, Camera Link), avoiding the system being unable to integrate due to the camera not being supported.

5. Deep-learning capability: if the task involves AI vision, focus on evaluating: the minimum number of samples required for training (the fewer the better), the difficulty of model training (whether there is a no-code interface), and inference-acceleration support (GPU inference / ONNX-model compatibility).

6. Robustness validation: conduct a 72-hour continuous-operation test in the actual production-line environment (illumination fluctuation, vibration, inter-product differences), evaluating the algorithm's stability under extreme conditions.

7. Upgrade and maintenance: evaluate the SDK vendor's version-update frequency, backward-compatibility commitment (whether old-version code can run on new versions), and technical-support response speed.

Selection Step 3: Evaluate Commercial Cost (Full Lifecycle)

The cost evaluation of SDK selection needs to cover the full lifecycle, not just the initial procurement cost:

Initial licensing cost: development license + runtime license × number of deployment nodes, see the price-tier comparison in Chapter 8 of this report.

Integration-development cost: the working-hour cost for engineers to use the SDK for algorithm development, debugging, and integration, depending on the SDK's ease of use and the engineer's familiarity (an engineer familiar with HALCON has a learning curve of about 1–2 months when switching to VisionMaster).

Upgrade and maintenance cost: algorithm-version-upgrade fees, re-licensing costs when replacing equipment, and software-license-transfer costs when equipment fails (the handling process for a lost / damaged hardware lock).

Support-service cost: the annual technical-support contract fee (some SDK vendors charge an annual maintenance fee), and the service cost when responding to emergency faults.

Talent cost: the training cost required to maintain internal SDK-development skills, and the recruitment premium for engineers experienced with that SDK.

Selection Step 4: Evaluate Ecosystem and Risk

Vendor stability: prioritize vendors with a listed-company background (AInnovation, Lingkong) or a clear IPO path (SMore), which have lower supply-chain risk than startups without a clear exit plan.

Community activity: evaluate the SDK's technical-discussion activity in the Chinese engineer community (CSDN, Cnblogs, Zhihu); an active community means it is easier to get community support when encountering problems.

Policy compliance: if the downstream customer is a semiconductor, automotive-OEM, or pharmaceutical factory, confirm whether the target SDK has completed technical certification in these scenarios (IATF 16949, GMP validation, etc.).

Long-term strategic match: if the enterprise plans to expand toward AI vision or 3D vision in the future, prioritize SDK vendors with a clear technical roadmap in these directions, avoiding being forced to undergo a costly platform migration in the future due to a mismatch in the SDK's technical direction.

The core principle of this selection framework is: task first, full-cost evaluation, reliable ecosystem, controllable risk. The machine vision SDK is a technical-infrastructure investment, and its value needs to be fully realized over a 5–10-year usage cycle; the quality of a single selection decision will affect the technical-path choice of a system integrator or terminal factory for years to come.