PCB Defect Detection (YOLOv8s)
YOLOv8s fine-tuned to detect six PCB manufacturing defects: missing_hole, mouse_bite,
open_circuit, short, spur and spurious_copper.
Usage
from huggingface_hub import hf_hub_download
from ultralytics import YOLO
model = YOLO(hf_hub_download("Sulojan/pcb-defect-yolov8s", "best.pt"))
results = model.predict("board.jpg", conf=0.3)
results[0].show()
Training
- Data: Kaggle akhatova/pcb-defects (693 images, 10 board designs), split 80/10/10 with seed 42
- Input 640×640, 180 epochs, AdamW (lr0 0.001), mixup 0.3; best checkpoint at epoch 116
Results (held-out test set, 70 images, 286 boxes)
| Class | mAP50 | mAP50-95 |
|---|---|---|
| all | 0.941 | 0.503 |
| missing_hole | 0.995 | 0.657 |
| mouse_bite | 0.922 | 0.512 |
| open_circuit | 0.992 | 0.538 |
| short | 0.975 | 0.502 |
| spur | 0.904 | 0.448 |
| spurious_copper | 0.861 | 0.362 |
Limitations
The model was trained on images of 10 board designs under controlled lighting. Confidence drops on boards, cameras and lighting that differ from the training data.
Source code: https://github.com/Sulojan2027/PCB_Defect_Detect
- Downloads last month
- 53