AdiiVD/pcb-defect-detection

Automated PCB Defect Detection

This Hugging Face Space deploys the capstone project's PCB defect detector as an online demo. Users can upload a PCB image and receive:

• —Annotated output with color-coded defect boxes
• —Metric cards (detections, latency, input size, classes hit)
• —Inspection verdict banner (pass vs review recommended)
• —Per-class count breakdown and filterable detection table
• —Before/after image comparison gallery
• —Downloadable annotated image, JSON, and CSV results
• —Clean raw PCB sample buttons for each defect class
• —About tab with capstone deployment context and reference offline metrics

Model File

The trained model artifacts are included in:

• —models/best.pt
• —models/best.onnx

The current Space uses the paper-backed YOLO11s-1280 checkpoint from the accepted ESCS'26 experiments. The app loads models/best.pt by default. It also supports setting a custom model path with the MODEL_PATH environment variable.

Reference offline test metrics for this deployed checkpoint:

• —mAP50: 0.902
• —mAP50-95: 0.502
• —recall: 0.866
• —V100 batch-1 inference latency: 12.8 ms/image

Classes

• —Missing hole
• —Mouse bite
• —Open circuit
• —Short
• —Spur
• —Spurious copper

Deployment Note

This online demo satisfies the web-deployment demonstration requirement. It does not claim Jetson/TensorRT deployment. The project report correctly treats TensorRT benchmarking as future work because target hardware/runtime access was unavailable.

Input Image Note

Use clean PCB images for the demo. The bundled sample buttons use raw images from PCB-DATASET-master/images/<class>/. You can also browse the public PCB-DATASET image folders and upload a clean raw .jpg from an images/<class>/ folder. Do not upload YOLO training/validation batch mosaics, screenshots, or images that already contain filenames, class labels, or drawn boxes. Text and overlay labels are out-of-distribution visual noise and can create false positives near letters.