PCB Defect Detection using YOLO

Printed circuit boards are inspected for soldering and component defects — opens, shorts, mousebites, spurs, pinholes and spurious copper — that manual visual inspection misses on dense, fast-moving lines. A classifier can only say whether a board image contains a defect; an inspector needs to know where each defect is and what kind it is. Object detection answers both: a YOLO detector predicts where each defect is, drawing a bounding box with a class label and confidence score in a single pass. The difficulty is localizing small, varied defects against complex board backgrounds — copper pours, silkscreen and solder texture all look deceptively defect-like. This project studies the idea end to end, training YOLOv8 on the public DeepPCB dataset (Tang et al. 2019) with its six defect types, and evaluating detection quality with per-class metrics so weak defect types are not hidden behind an average — including a visualization script so every prediction can be verified by eye.

Dataset & model:
Dataset name: DeepPCB (Tang et al., "Online PCB defect detector on a new model of PCB defect dataset", arXiv:1902.06197, 2019).
Source: Official academic release of the DeepPCB dataset.
Task: Object detection (defect localization with bounding boxes).
Classes: Six defect types — open, short, mousebite, spur, pinhole, spurious copper — per the official dataset documentation.
Model: YOLOv8 (Ultralytics) single-pass detector, pre-trained on COCO and fine-tuned on DeepPCB; PyTorch as the training backend.
Input: PCB images resized to the YOLO input size.
Prediction: Bounding boxes with defect class and confidence score, in a single pass.
Output: Boxes above an adjustable confidence threshold, with class labels and scores; box-overlaid annotated images.
Evaluation metrics: Precision, recall, mAP@0.5 and per-class AP on a held-out validation split — computed by the Jupyter notebook during the buyer's build. No measured performance claimed; design target approximately 85% or better mAP@0.5 on the split.

Working:

1. PCB images and defect annotations are loaded from the DeepPCB dataset structure and resized to the YOLO input size.
2. Preprocessing: mosaic and HSV augmentation tolerates board lighting and solder-finish variation.
3. Training phase: a YOLOv8 detector — pre-trained on COCO and fine-tuned on DeepPCB — trains to predict bounding boxes, defect classes and confidence scores in a single pass.
4. Evaluation phase: the Jupyter notebook computes precision, recall, mAP@0.5 and per-class AP on the held-out validation split, with precision-recall curves.
5. Inference phase: the trained detector scans new PCB images and outputs boxes above an adjustable confidence threshold; a batch pipeline exports annotated results for a folder of board images.
6. Output: bounding boxes with class labels and confidence scores, overlaid on the original images by the visualization script for visual verification.

Specifications:
Model | YOLOv8 single-pass detector (Ultralytics), pre-trained on COCO, fine-tuned on DeepPCB
Dataset | DeepPCB (Tang et al. 2019), public research collection with six defect types
Task | Object detection: bounding-box localization of PCB defects
Classes | Open, short, mousebite, spur, pinhole, spurious copper
Input | PCB images resized to the YOLO input size
Output | Bounding boxes with class labels and confidence scores above an adjustable threshold
Augmentation | Mosaic and HSV, tuned for board lighting and solder-finish variation
Evaluation | Precision, recall, mAP@0.5, per-class AP on held-out validation split — computed during the buyer's build
Design target | Approximately 85% or better mAP@0.5 on the split (target, not a measured claim)
Verification | Box-overlay visualization script for image-by-image checking

[Six-Defect Box Localization] (implemented) — Bounding-box localization of the six DeepPCB defect types: open, short, mousebite, spur, pinhole, spurious copper.
[Per-Box Labels and Scores] (implemented) — Each box carries a class label and confidence score with an adjustable detection threshold.
[Mosaic and HSV Augmentation] (implemented) — Augmentation pipeline tuned for board lighting and solder-finish variation.
[Per-Class AP Breakdown] (implemented) — Precision-recall curves and per-class AP show exactly which defect types the detector finds hardest.
[Detection Visualization Script] (implemented) — Overlays predicted boxes, labels and scores on PCB images for visual verification of every prediction.
[Batch Inference Pipeline] (implemented) — Exports annotated results for a whole folder of board images.
[Evaluation Notebook] (implemented) — Computes precision, recall, mAP@0.5 and per-class AP on the held-out validation split during the build.
[Instance Segmentation Masks] (future-scope) — Pixel-level defect masks are future work, not included in the base detection build.

Complete source code (data pipeline, YOLOv8 training, inference and visualization scripts)
Jupyter evaluation notebook (mAP, precision-recall curves, per-class AP, box-overlaid results)
Project report PDF (background, DeepPCB dataset, methodology, evaluation, error analysis)
PPT presentation for final review
Viva Q&A preparation document (YOLO, anchor-free detection, mAP, precision-recall, PCB inspection context)
Setup guide (environment, dataset download, training and inference commands)