A consumer electronics factory in Shenzhen inspects 3,000 PCBs per hour — one every 1.2 seconds — for solder joint quality, component presence, and placement accuracy. At this rate, human visual inspection is not feasible: a trained human inspector can reliably examine approximately 200 to 400 units per hour before fatigue degrades accuracy. Machine vision inspection at 3,000 units per hour requires a specific hardware and software architecture that is now mature in Chinese factory deployments. This article covers how it works.
The Hardware Stack
Camera: For PCB solder joint inspection, the standard is a 12 or 25 megapixel monochrome GigE Vision camera (5 or 10 Gbps interface) with a resolution sufficient to image a 0.3 mm BGA solder ball as at least 3 to 4 pixels across. At 1.2-second cycle time, the camera must complete image acquisition, transfer, and buffer clearing in under 200 ms — the remaining 1,000 ms is for processing and mechanical transport. High-speed industrial cameras from Hikrobot (海康机器人), Huaray, and Daheng Imaging dominate Chinese factory deployments at 30 to 70% lower cost than equivalent Basler or FLIR units.
Lighting: Solder joint inspection uses structured light — specifically, multi-angle LED ring illumination with programmable phase sequences. Bright-field illumination from above reveals surface texture. Dark-field illumination from low angles reveals height variation and bridging defects. Running both illumination modes per inspection cycle adds 80 to 100 ms to cycle time but doubles defect detection coverage.
Optics: Telecentric lenses are mandatory for dimensional measurement — standard lenses produce perspective distortion that makes near objects appear larger than far objects, introducing measurement error. Telecentric lenses maintain constant magnification regardless of object distance within their depth of field. For ±0.05 mm dimensional tolerance inspection, telecentric optics are the only option.
The Software Stack
Two architectures dominate Chinese factory machine vision deployments in 2025. The first is rule-based: engineers define the inspection criteria (minimum solder area, maximum component offset, acceptable color range) as explicit parameters. Fast, deterministic, and explainable — but requires re-configuration for each new product variant. The second is deep learning-based: a convolutional neural network trained on labeled images of good and defective parts. Slower to set up (requires 1,000 to 5,000 labeled training images per defect class) but generalizes to new product variants with minimal re-training and handles defect types that are difficult to specify as rules (subtle cosmetic defects, partial solder bridging).
Chinese machine vision software vendors dominating factory deployments: Mech-Mind (梅卡曼德) for 3D bin picking, Hikrobot for 2D inspection, and VisionPro (Cognex, US) for high-precision dimensional measurement. Mech-Mind’s 3D vision systems are specific to robot-guided operations — the point cloud output drives robot arm path planning for picking randomly oriented parts. For 2D inspection tasks at 3,000 parts per hour, Hikrobot’s MVS software environment runs on standard industrial PCs at cycle times under 150 ms per part.
Total System Performance at 3,000 PPH
A complete inspection station for 3,000 PPH PCB inspection uses: one 25 MP camera, multi-angle LED ring at 3 illumination angles, telecentric 0.3x lens, GigE Vision 10 Gbps cable, industrial PC with NVIDIA RTX A2000 GPU for deep learning inference, conveyor with precision stop mechanism (±0.1 mm repeatability). Total hardware cost: $15,000 to $25,000 per station. Defect detection rate: greater than 99.5% on trained defect classes. False positive rate: 0.3 to 0.8% depending on training data quality.
For more on industrial inspection automation in China see our industrial inspection equipment page and innovations in China’s inspection landscape.
Sources
- Hikrobot: MVS machine vision software documentation (2025)
- Mech-Mind: 3D vision system technical specifications (2025)
- China Machine Vision Industry Alliance: Market report 2024
