// the find
Ixiaohuihuihui/Tiny-Defect-Detection-for-PCB
This is a repository about PCB defect detection.
FPN-based object detection pipeline adapted for PCB defect detection, covering six defect types (missing hole, mouse bite, open circuit, short, spur, spurious copper) using the Peking University PCB dataset. Built on TensorFlow 1.x with ResNet50/101 backbones. Aimed at researchers and engineers working on automated optical inspection who want a working baseline rather than a production system.
The augmented dataset split (9920 train / 2508 test images via 600×600 crops from high-res originals) is well thought out and saves hours of data prep work. FPN with OHEM is a reasonable architecture choice for small defects — the multi-scale feature pyramid actually matters here since PCB defects can be a few pixels wide. Pretrained weights and a demo inference script are both provided, so you can verify the pipeline works before touching your own data. PR curves per defect class are included in the results, which is more honest than reporting a single mAP number.
Hard dependency on TensorFlow 1.12 and CUDA 8.0 — this stack is seven years old and won't run on any modern GPU without painful environment archaeology or Docker gymnastics. The repo is explicitly a fork with 'minor changes' from FPN_Tensorflow; there's no clear description of what was actually changed, so you're partially maintaining someone else's code with no changelog. Dataset is hosted on Baidu Pan and requires contacting the original authors, which is a real barrier if you're outside China or need to reproduce results quickly. No export to ONNX or TFLite, no INT8 quantization path — if you want to deploy this on edge hardware in an actual factory, you're starting from scratch on that part.