PCB Defect Detection using YOLOv5

Internship Certificate from iNeuron

Demo Video

github

Industry Context

In modern electronics manufacturing, ensuring Printed Circuit Board (PCB) quality is crucial for preventing malfunctioning devices and costly recalls. Manual inspection is time-consuming and error-prone, hence automated defect detection using deep learning significantly improves speed, accuracy, and reliability in quality assurance pipelines.

Problem Definition

The goal is to build an object detection model that accurately identifies and classifies six types of PCB defects using YOLOv5.

About the Data

The dataset contains annotated images of PCB boards with six defect types:

• missing_hole
• mouse_bite
• open_circuit
• short
• spur
• spurious_copper

Each image is paired with a .txt file in YOLO format indicating bounding boxes and corresponding class labels. The dataset was split into:

• Training Images: 92
• Validation Images: 23

To prepare for YOLOv5 training, the annotations were converted into YOLO format and directories were structured accordingly using a custom to_v5_directories() utility function.

Data Preprocessing

• Verified and converted all annotations to YOLOv5-compatible format.
• Ensured the structure: images/train, images/val, labels/train, labels/val.
• Verified class balance and number of annotations per class.

Actionable Insights using EDA

• The dataset is imbalanced with varying instances per class.
• Defects like missing_hole and short are more frequent than spur or spurious_copper.
• Most images contain one or two defects; very few have multiple overlapping defects.
• Bounding boxes are relatively small, requiring high-resolution training.
• The defect types are visually distinct, making object detection feasible with the right model.

Data Preparation

• All images resized to 416x416 pixels.
• Real-time augmentations using Albumentations were applied:
  • Horizontal Flip
  • CLAHE
  • Blur
  • Grayscale
• YOLOv5's built-in caching was used to speed up data loading during training.

Model Selection

Chose YOLOv5s (small variant of YOLOv5) due to its balance between speed and accuracy, making it suitable for real-time defect detection on edge devices.

Model Training

• Framework: PyTorch with YOLOv5
• Pre-trained weights: yolov5s.pt
• Custom training on 6 defect classes
• Trained for 100 epochs using default hyperparameters
• Loss Function: Combined objectness, classification, and bounding box regression losses
• Optimizer: SGD (used by default in YOLOv5)

Model Evaluation

• Accuracy: 93.7%
• mAP@0.5: High across classes
• Precision: Up to 99.8% for specific classes
• Overall Performance: 47.9% across all metrics
• Detection samples were visualized and confirmed using bounding boxes.

Feature Importance

Feature importance in object detection is reflected in how YOLO learns spatial and class-related features. The most dominant classes had better localization and confidence scores. Visual inspection showed that open_circuit and missing_hole were more accurately localized.

Key Results

• Successfully trained YOLOv5 model to detect six PCB defect types.
• Achieved strong performance with high accuracy and precision.
• Validated through visual inspection and inference scripts.
• Model saved in .pt format for deployment.

Business Impact

• Automated PCB inspection can reduce manual labor by over 70%.
• Improves defect detection consistency and reliability.
• Saves time in production pipelines, boosting throughput.
• Reduces return rates and product failures due to faulty PCBs.

Deployment

CI/CD Pipeline with GitHub Actions and AWS

1. Environment Setup

   • Created a Conda environment with Python 3.9.
   • Installed dependencies using pip install -r requirements.txt.

2. Dockerization

   • Docker image built from source code including app.py, models, and utility components.
   • Dockerfile includes all setup steps to run the inference API.

3. AWS Services Used

   • Amazon ECR: Container image registry to store Docker images.
   • Amazon EC2: Used as a host to run the containerized application.

4. GitHub Actions Workflow

   • CI/CD pipeline defined in .github/workflows/main.yml.
   • On every push to main branch:
     • Builds the Docker image
     • Authenticates and pushes it to ECR
     • SSH into EC2
     • Pulls and runs the latest Docker image

5. Self-Hosted Runner

   • Configured EC2 instance as a self-hosted GitHub runner.
   • Ensures secure and direct deployment without manual intervention.

6. Secrets Used

   • AWS credentials, region, ECR URI, and repository name are securely stored as GitHub secrets.

Future Scope

• Expand dataset size for better generalization.
• Fine-tune hyperparameters using YOLOv5’s hyperparameter evolution.
• Integrate defect severity classification to aid manufacturing decisions.
• Deploy model on an edge device (e.g., Jetson Nano) for real-time inspection.

Tools Used

• Python
• YOLOv5
• PyTorch
• Albumentations
• OpenCV
• Pandas
• NumPy
• Matplotlib
• Conda
• Docker
• GitHub Actions
• AWS EC2
• AWS ECR
• Git