DeepPCB (defect benchmark)

DeepPCB is a PCB defect dataset of 1,500 image pairs, each pairing a defect-free template image with an aligned tested image annotated with positions of 6 common PCB defect types: open, short, mousebite, spur, pin hole, and spurious copper. Images were captured with a linear scan CCD at around 48 pixels per millimetre and cropped to 640 by 640 sub-images aligned through template matching.

Each defect is annotated as an axis-aligned bounding box with a class ID. The dataset splits 1,000 images as a training set and the rest as a test set, and the README states the work may only be used for research purposes. An annotation tool ships in the repository's tools directory.

Evaluation uses average precision and F-score, with a detection counted correct only when the bounding-box intersection over union with a ground-truth box of the same class exceeds 0.33. The repository ships a modified evaluation script and the test-set ground truth. The README reports the authors' own model at 98.6 percent mAP and 98.2 percent F-score at 62 frames per second.