PCB defect detection: YOLO11n / YOLO11s, FP32 and INT8
Weights for the pcb-defect-optimization project: YOLO11 trained on PKU-Market-PCB (6 defect classes) at 1280×1280, exported to ONNX and statically quantized to INT8 with ONNX Runtime.
Demo: https://huggingface.co/spaces/michaelandrew35/pcb-defect-detection
Code, method, full results and failure analysis: https://gitlab.com/michael.sucahyo/pcb-defect-optimization
Files
| File | Format | Size | Test mAP@0.5 | Test mAP@0.5:0.95 | CPU ms/img* |
|---|---|---|---|---|---|
yolo11n_img1280.pt |
PyTorch (Ultralytics) | 5.4 MB | 0.967 | 0.557 | 282 |
yolo11n_img1280_fp32.onnx |
ONNX FP32 | 10.6 MB | 0.965 | 0.556 | 261 |
yolo11n_img1280_int8_mixed.onnx |
ONNX INT8, detection head FP32 | 4.7 MB | 0.967 | 0.555 | 217 |
yolo11s_img1280.pt |
PyTorch (Ultralytics) | 18.4 MB | 0.984 | 0.587 | 649 |
yolo11s_img1280_fp32.onnx |
ONNX FP32 | 36.6 MB | 0.984 | 0.586 | 620 |
yolo11s_img1280_int8_mixed.onnx |
ONNX INT8, detection head FP32 | 12.4 MB | 0.982 | 0.583 | 484 |
* Median forward-pass latency, 4 threads, AMD Ryzen 7 4700U, ONNX Runtime 1.20.1. Test split: 102 images.
Classes: missing_hole, mouse_bite, open_circuit, short, spur, spurious_copper. Input: 1×3×1280×1280 RGB, letterboxed, scaled to 0–1. Output: 1×10×33600 (box xywh in input pixels + 6 class scores); apply NMS yourself or load with Ultralytics.
Usage
from ultralytics import YOLO
model = YOLO("yolo11n_img1280_int8_mixed.onnx", task="detect")
results = model.predict("board.jpg", imgsz=1280, conf=0.25)
A dependency-free ONNX Runtime + OpenCV version (letterbox, decode, NMS) is in the demo's app.py.
Why "head FP32"
Quantizing the whole graph scores 0 mAP: the final output concatenates box coordinates (0–1280 px) with class scores (0–1), and a single INT8 scale rounds every score to zero. Keeping the Detect head in FP32 fixes this at a cost of 0.07 (YOLO11n) / 0.37 (YOLO11s) mAP@0.5:0.95.
Limitations
Trained on PKU-Market-PCB, a public research dataset (no explicit license found; used for non-commercial research only). Not validated on production factory images. Weakest on very small spurs and on shorts / spurious copper that resemble normal traces.