PCB Defect Detection using CNN

Manual PCB inspection is slow, tiring and inconsistent — inspectors miss hairline opens and tiny solder bridges after hours at the microscope, and small fabrication units cannot afford commercial AOI machines. Automated optical inspection compares each board against a golden reference and flags deviations. The DeepPCB dataset (Huang et al., Huawei Noah's Ark Lab) provides 1,500 template/tested image pairs with bounding-box annotations for six defect classes, mirroring real AOI stations: align, difference, classify. This project implements that pipeline with a modern single-shot detector — template alignment via feature matching, a YOLO-style detection head, non-maximum suppression, and a verdict UI mapping each defect to severity and rework action.

1. The tested board photo is registered to its golden template via ORB feature matching and homography.
2. The aligned 640x640 pair feeds a YOLO-style detector with a lightweight backbone (CSPDarknet-tiny/MobileNet).
3. The detection head predicts candidate bounding boxes with class scores across the six defect types.
4. Non-maximum suppression removes overlapping boxes; survivors below the confidence threshold are dropped.
5. Each detection maps to a severity (minor/major/critical) and a rework action for the operator.
6. Training uses mosaic + photometric augmentation on the 1,000-pair DeepPCB training split.

Tech stack:

• Python 3.10, PyTorch
• YOLO-style detector (Ultralytics tooling or equivalent)
• DeepPCB dataset (Huawei Noah's Ark Lab)
• OpenCV (alignment, NMS, visualization)
• Flask (demo backend)
• HTML/CSS/JavaScript (demo UI)
• Git

Dataset & Model Details

• Dataset: DeepPCB — 1,500 template/tested image pairs with bounding-box annotations, 6 defect classes (open, short, mousebite, spur, pinhole, spurious copper); source: T. Huang et al., Huawei Noah's Ark Lab. Standard split: 1,000 train / 500 test.
• Model input: 640x640 aligned board image. Output: bounding boxes + class + confidence per defect.
• Architecture: YOLO-style single-shot detector; lightweight backbone; NMS post-processing.
• Metric: mAP@0.5 per class and overall, measured on the 500-pair test split after training. No figure is claimed before the training run.

• [Six defect classes] Open, short, mousebite, spur, pinhole and spurious copper — the full DeepPCB taxonomy.
• [Bounding-box localization] Every detection ships with a box, class label and confidence, overlaid on the board photo.
• [Template alignment] Tested boards are registered to the golden reference with feature matching before comparison.
• [PASS/FAIL verdict engine] Detections map to severity levels and concrete rework actions (e.g. solder bridge → desoldering wick).
• [YOLO-style detector] Single-shot architecture with a lightweight backbone; 640x640 input, NMS-filtered output.
• [Per-class mAP evaluation] mAP@0.5 measured per class on the 500-pair test split — hardest classes identified honestly.
• [Inspection web demo] Upload a board photo; get boxes, verdict and rework guidance in the browser.

• Python training notebook (detector training, augmentation, evaluation)
• Template-alignment + inference pipeline code
• Trained detector weights
• Test-split evaluation: per-class mAP@0.5 from the actual run
• Inspection web demo (upload, boxes, severity, rework actions)
• Project report PDF (methodology, error analysis, measured results)
• PPT presentation for final review
• Viva Q&A preparation document

• Covers bare-PCB defects only — not assembled-board issues like missing or tombstoned components.
• Needs a roughly top-down, well-lit board photo; extreme angles break the template alignment.
• Tiny defects (pinhole, spur) are the hardest classes — the per-class mAP table shows this honestly.
• Detector training strongly benefits from a GPU; CPU training is very slow.
• Rework actions are guidance for a trained operator, not a repair robot.