Janani-V/pcb-defect-yolov8m-dspcbsd

YOLOv8m — DsPCBSD+ Defect Detection

Model Summary

• Model: YOLOv8m
• Task: PCB defect detection (object detection)
• Dataset: DsPCBSD+ (via Roboflow export)
• Classes: 9 defect categories
• Framework: Ultralytics YOLOv8
• Input size: 640 × 640
• Training hardware: Google Colab, Tesla T4 GPU
• Validation mAP@0.5: 0.839
• Validation mAP@0.5:0.95: 0.508
• Companion module: inspector.py — adds severity, root cause, impact, and recommended action per detection

🔍 Part of a two-stage project: see also Janani-V/pcb-defect-yolov8s-deeppcb — a simpler 6-class baseline on DeepPCB.

Model Comparison — Stage 1 vs Stage 2

Note: These two models were trained on different datasets (DeepPCB vs. DsPCBSD+) with different class counts and difficulty levels, so this is not a strictly apples-to-apples benchmark — it's meant to help you choose the right model for your use case.

	Stage 1: YOLOv8s	Stage 2: YOLOv8m
Model	pcb-defect-yolov8s-deeppcb	pcb-defect-yolov8m-dspcbsd (this repo)
Dataset	DeepPCB	DsPCBSD+
Classes	6	9
Train / Val images	1,050 / 150	8,208 / 2,051
Total annotations	~1,003 (val)	4,092 (val)
Image source	Grayscale linear-scan CCD	RGB copper-surface crops (226×226 native)
Model size	YOLOv8s (~11.1M params)	YOLOv8m (~25.9M params)
mAP@0.5	0.985	0.839
mAP@0.5:0.95	0.734	0.508
Precision	0.961	0.807
Recall	0.963	0.810
Best-performing class	copper (mAP50 0.994)	hole_breakout (mAP50 0.984)
Weakest-performing class	short (mAP50-95 0.639)	conductor_foreign_object (mAP50 0.701)

Why the difference in scores?

DsPCBSD+ is a meaningfully harder benchmark than DeepPCB:

• More classes (9 vs. 6) increases inter-class confusion risk
• Class imbalance is more pronounced (spur: 929 instances vs. short: 169, a ~5.5x gap)
• Higher intra-class variability — especially for conductor_scratch and conductor_foreign_object, which vary widely in size, shape, and appearance
• DeepPCB's defects are more visually distinct and the dataset itself is smaller and cleaner by design

A lower mAP on DsPCBSD+ does not mean this model is "worse" — it reflects a genuinely harder detection problem with more real-world defect diversity.

Which model should you use?

• Use the YOLOv8s / DeepPCB model if your defects match DeepPCB's 6 categories (open, short, mousebite, spur, copper, pin-hole) and you're working with grayscale linear-scan imagery — it's faster and more accurate for that specific defect set.
• Use the YOLOv8m / DsPCBSD+ model (this repo) if you need broader defect coverage, including hole breakout, conductor scratches, and foreign object contamination — categories DeepPCB doesn't cover at all.

Model Description

This repository hosts a YOLOv8m object detection model fine-tuned on DsPCBSD+, a large-scale (10,259 image, 20,276 annotation) PCB surface defect dataset covering 9 defect categories across conductors, holes, and base material. This is a substantially harder detection task than DeepPCB — defects are smaller, more varied in shape/scale, and the class distribution is imbalanced.

Alongside the detection weights, this repository includes inspector.py, a companion knowledge-base module providing explanation, severity, root cause, impact, and recommended action per detected defect. This is a deterministic rules layer, not the model itself generating text — best.pt outputs class, bounding box, and confidence only.

Architecture: YOLOv8m (Ultralytics), single-stage anchor-free object detector Base weights: yolov8m.pt (COCO-pretrained, then fine-tuned)

Defect Classes

Class	Abbreviation (paper)	Description
short	SH	Unintended connection between conductors
spur	SP	Sharp protrusion off a conductor edge
spurious_copper	SC	Unwanted copper residue
open	OP	Break in a conductor path
mouse_bite	MB	Small notch/crack at conductor edge
hole_breakout	HB	Hole center deviates from bounding pad
conductor_scratch	CS	Scratch on copper wire/surface
conductor_foreign_object	CFO	Contamination on a conductor
base_material_foreign_object	BMFO	Contamination on bare substrate

---

Evaluation Results

Validation set: 2,051 images, 4,092 annotated instances (official 8:2 train/val split).

Overall

Metric	Value
Precision	0.807
Recall	0.810
mAP@0.5	0.839
mAP@0.5:0.95	0.508

Per-Class

Class	Images	Instances	Precision	Recall	mAP50	mAP50-95
short	126	169	0.875	0.882	0.906	0.592
spur	430	929	0.852	0.786	0.850	0.389
spurious_copper	245	285	0.757	0.779	0.822	0.507
open	274	338	0.803	0.855	0.889	0.532
mouse_bite	391	546	0.851	0.773	0.823	0.410
hole_breakout	271	608	0.917	0.977	0.984	0.830
conductor_scratch	279	448	0.696	0.695	0.731	0.456
conductor_foreign_object	309	423	0.693	0.667	0.701	0.408
base_material_foreign_object	305	346	0.817	0.878	0.848	0.449

Performance Chart

Training Curves

Confusion Matrix

---

Sample Detections

Sample 1

Input	Prediction

Sample 2

Input	Prediction

---

Training Configuration

Parameter	Value
Base model	yolov8m.pt (COCO-pretrained)
Dataset	DsPCBSD+ (9 classes), via Roboflow export
Epochs	50
Image size	640 × 640
Batch size	16
Scheduler	Cosine LR
Hardware	Google Colab, Tesla T4 GPU

---

Usage

from huggingface_hub import snapshot_download
import sys

local_dir = snapshot_download(repo_id="Janani-V/pcb-defect-yolov8m-dspcbsd")
sys.path.append(local_dir)

from inspector import PCBDefectInspector

inspector = PCBDefectInspector(weights_path=f"{local_dir}/best.pt")
result = inspector.inspect("your_pcb_image.jpg")

print(result["summary"])
for f in result["findings"]:
    print(f["class"], "-", f["severity"], "-", f["action"])

---

Limitations

• conductor_foreign_object (mAP50 0.701) and conductor_scratch (mAP50 0.731) are the weakest classes — these defects have high intra-class variability in size, shape, and color, making them inherently harder to detect consistently.
• Trained on 226×226-native PCB crop images; performance on full, un-cropped board images has not been separately validated.
• Class distribution is imbalanced (spur: 929 instances vs. short: 169); rare-class performance may vary more across different data splits.
• The inspector.py explanations are drawn from a static, hand-curated knowledge base — general guidance, not image-specific diagnosis.
• Research/baseline model — not validated for production deployment without further testing on real manufacturing data.

---

Repository Contents

• best.pt — fine-tuned YOLOv8m weights
• inspector.py — companion module with severity/root-cause/impact/action knowledge base
• metrics_chart.png — per-class performance chart
• results.png — training loss/metric curves across all epochs
• confusion_matrix.png — normalized confusion matrix across all 9 classes
• sample*_input.jpg / sample*_predicted.jpg — example detections
• README.md — this file

---

Author

Fine-tuned and maintained by Janani-V.

Citation

Original dataset: Lv, S. et al. "A dataset for deep learning based detection of printed circuit board surface defect." Scientific Data 11, 811 (2024). https://doi.org/10.1038/s41597-024-03656-8

Dataset access: janani-v-sdspd/dspcbsd-plus on Roboflow Universe

Model architecture: Jocher, G. et al. — Ultralytics YOLOv8: https://github.com/ultralytics/ultralytics