CSUF Master's Capstone — Online Deployment Demo

Automated PCB Defect Detection

Upload a PCB inspection image and run the exported YOLO11s detector for six defect classes. The demo returns annotated boxes, per-class counts, latency, and downloadable results for presentation or QA review.

6 defect classes YOLO11s @ 1280px PyTorch + ONNX export Gradio web demo

Model ready: YOLO11s (best.pt, PyTorch) · device: CPU · backend: PyTorch

Input

Upload a clean PCB image or load a clean sample. Do not use screenshots, YOLO validation mosaics, or images that already contain text labels.

Upload PCB image

Confidence threshold

0.05 0.95

NMS IoU threshold

0.2 0.9

Inference image size

Clean sample inputs (auto-run)

Need more test images? Browse the public PCB-DATASET image folders and upload a clean raw JPG from an images/class folder.

Missing hole

Mouse bite

Open circuit

Short

Spur

Spurious copper

Output

Annotated result, metrics, and exports.

Annotated detections

Detections

above threshold

Latency

milliseconds

Input size

1280

pixels

Classes hit

of 6 defect types

Ready. Upload a clean PCB image or choose one of the clean sample inputs.

Class breakdown appears after detection.

No detections yet.

Before / after comparison (original, annotated)

Capstone deployment overview

This Hugging Face Space is the online deployment artifact for the CSUF master's capstone project on automated PCB defect inspection. It demonstrates that the trained detector can be exported and served through a browser-based workflow suitable for manufacturing triage demos.

Student: Aditya Dhayapulay
Institution: California State University, Fullerton
Contact: aditdhayapulay@gmail.com
Runtime model: YOLO11s (best.pt, PyTorch)
Inference device: CPU

What the demo does

Accept a clean uploaded PCB inspection image or bundled raw PCB sample.
Run the exported detector with adjustable confidence and NMS thresholds.
Return an annotated image, per-detection table, class breakdown, inspection verdict, and downloadable results.

Reference offline metrics

These numbers come from the saved YOLO_PCB unified evaluation used in the capstone report. Space latency will differ from the offline V100 GPU benchmark.

Metric	Value
Dataset	YOLO_PCB (5,551 / 1,016 / 1,016 train / val / test)
Test mAP50	0.902
Test mAP50-95	0.502
Test recall	0.866
V100 batch-1 inference latency	12.8 ms/image

Defect classes

Class	Typical appearance
Missing hole	Drill hole absent from pad region
Mouse bite	Small edge erosion on copper trace
Open circuit	Broken or interrupted conductor
Short	Unintended bridge between traces
Spur	Thin protrusion from a trace
Spurious copper	Extra copper not in the design

What this demo claims

The project trained a YOLO-family detector for six PCB defect classes.
The best checkpoint was exported to PyTorch (best.pt) and ONNX (best.onnx) for deployment evidence.
The online demo satisfies the web deployment requirement for the capstone presentation.

What this demo does not claim

Embedded Jetson / TensorRT production latency (hardware access was unavailable).
Full factory-line integration or live camera streaming.
State-of-the-art benchmark leadership across all public PCB datasets.

Reproducibility

Artifacts, metrics tables, and training logs are archived in the project repository and Kaggle/Nautilus export bundles used to build the final ESCS paper tables.

For additional manual tests, use clean raw PCB images from the public PCB-DATASET image folders. Avoid annotated mosaics, screenshots, or files with text overlays.