PCB Defect Detection

A production-ready system for automated detection and classification of defects in printed circuit board (PCB) images using deep learning and scalable APIs. This repository contains trained YOLOv8 model weights, an end-to-end inference pipeline, comprehensive analysis notebooks, and containerized deployment for local and cloud demonstration.

---

Overview

• Objective: Accurate, automated detection and localization of PCB manufacturing defects.
• Model: YOLOv8 Nano, trained on a custom multi-class defect dataset.
• API: FastAPI-based REST endpoint supporting image upload and robust error handling.
• Deployment: Docker container available; public cloud deployment for demonstration (Render).
• Notebooks: Includes EDA and training analysis workflows.

---

Model Details and Evaluation Metrics

• Architecture: YOLOv8 Nano (Ultralytics)
• Training Data: ~600 annotated PCB images (train, validation, test splits combined)
• Classes: 7 distinct PCB defect types
• Performance:
  • mAP@0.50: 99.49%
  • mAP@0.50-0.95: 83.08%
  • Precision: 99.33%
  • Recall: 99.99%
• Model Size: 6.2 MB

---

API & Cloud Demo

A robust FastAPI service wraps the detector with a clear REST API.

• Live API:
  https://pcb-defect-detection-6moo.onrender.com
  • Interactive docs available at: https://pcb-defect-detection-6moo.onrender.com/docs

Core Endpoints

• POST /predict
  Upload an image and receive detected defect locations and labels as a JSON response.
• GET /health
  Service health and model status.
• GET /docs
  Interactive Swagger (OpenAPI) documentation.

How to Test the API Using Swagger UI

1. Navigate to the interactive documentation:
   https://pcb-defect-detection-6moo.onrender.com/docs

2. Locate the POST /predict endpoint and click to expand it.

3. Click "Try it out" button in the top-right corner of the endpoint section.

4. Upload your PCB image:

   • Click "Choose File" next to the file parameter
   • Select a PCB image from your local machine
   • The filename will appear once selected

5. Click "Execute" to run inference.

   • Note: Inference typically takes 20-30 seconds on the free-tier Render deployment

6. View Results:

   • Scroll down to the "Response body" section
   • JSON response includes:
     • num_detections: Total number of defects found
     • detections: Array of detected defects with:
       • class_name: Defect type classification
       • confidence: Detection confidence score
       • bounding_box: Spatial coordinates of the defect

Example Response Format

{
  "filename": "pcb_sample.jpg",
  "image_size": {"width": 640, "height": 480},
  "num_detections": 2,
  "confidence_threshold": 0.5,
  "detections": [
    {
      "class_id": 0,
      "class_name": "defect_1",
      "confidence": 0.94,
      "bounding_box": {"x1": 120, "y1": 85, "x2": 195, "y2": 140}
    },
    {
      "class_id": 3,
      "class_name": "defect_4",
      "confidence": 0.89,
      "bounding_box": {"x1": 310, "y1": 220, "x2": 380, "y2": 275}
    }
  ]
}

Example: Inference via Python

import requests
url = "https://pcb-defect-detection-6moo.onrender.com/predict"
files = {"file": open("pcb_image.png", "rb")}
response = requests.post(url, files=files)
print(response.json())

---

Local Setup & Usage

1. Prerequisites: Python 3.11 or later.
2. Install dependencies:

   pip install -r requirements.txt
   

3. Run API server locally:

   uvicorn api.main:app --host 0.0.0.0 --port 8000
   

   Access the API at http://localhost:8000/docs

Docker Deployment

1. Build and run with Docker:

   docker build -f docker/Dockerfile -t pcb-defect-api .
   docker run -p 8000:8000 pcb-defect-api
   

---

Notebooks

The repository includes comprehensive Jupyter notebooks for analysis and training:

• 01_eda.ipynb: Exploratory data analysis and dataset statistics
• 02_training_analysis.ipynb: Training metrics, loss curves, and model evaluation

Located in the notebooks/ directory.

---

Customization and Retraining

• End-to-end notebooks provided under notebooks/ for EDA, data preparation, and training analysis.
• To support additional classes or datasets, update the training notebooks and data configuration.
• Model weights and configuration are easily replaced for further fine-tuning or evaluation.

---

License and Acknowledgments

MIT License.
Key technologies: YOLOv8 (Ultralytics), FastAPI, Docker, Render.