ampragatish/yolov8n-pcb-defects-detection

Model Card: PCB Defect Detection Model

Model Details

• Model Name: PCB Defect Detection Model
• Version: 1.0
• Date: December 30, 2025
• Architecture: Object Detection (YOLO format compatible)
• Framework: Compatible with YOLO-family architectures (YOLOv5, YOLOv8, etc.)
• Input Format: PCB images with XML annotations
• Output Format: YOLO-style bounding boxes with class labels

Intended Use

• Primary Use: Detection of defects on Printed Circuit Boards (PCBs)
• Target Audience: Electronics manufacturers, quality control engineers, PCB inspection systems
• Use Cases:
  • Automated PCB quality inspection
  • Manufacturing defect detection
  • Quality assurance in electronics production
  • Research and development in computer vision for electronics

Defect Classes

The model detects 6 types of PCB defects:

1. Missing_hole (Class ID: 0) - Absence of required holes in PCB
2. Mouse_bite (Class ID: 1) - Irregular edge imperfections
3. Open_circuit (Class ID: 2) - Broken or incomplete circuit connections
4. Short (Class ID: 3) - Unintended connections between circuits
5. Spur (Class ID: 4) - Unwanted metal traces
6. Spurious_copper (Class ID: 5) - Excess copper on PCB surface

Data

• Source Dataset: "PCB Defects" dataset from Kaggle (akhatova/pcb-defects)
• Dataset Size: Approximately 1.88 GB
• Image Format: Standard image formats with XML annotations
• Annotation Format: VOC XML converted to YOLO format
• Data Processing:
  • XML to YOLO format conversion
  • Normalized bounding box coordinates (center-x, center-y, width, height)
  • Case sensitivity handling for class names

Training Data (Assumed)

• Train/Val/Test Split: Standard object detection split (typically 70/20/10 or 80/10/10)
• Augmentation: Potential rotation augmentation available in dataset
• Class Distribution: Balanced across 6 defect types

Performance Metrics

Note: Actual performance metrics would depend on specific training implementation

• Expected Metrics (when trained):
  • mAP@0.5 (Mean Average Precision at IoU 0.5)
  • Precision and Recall per class
  • F1-score
  • Inference speed (FPS)

Limitations

• Domain Specific: Only trained/tested on PCB defect detection
• Image Conditions: Performance may vary with different lighting, camera angles, or PCB types
• Defect Types: Limited to 6 specific defect classes
• Scale Sensitivity: May not detect very small defects on high-resolution PCBs
• Data Bias: Performance depends on dataset representation

Ethical Considerations

• Safety Critical: PCB defects can lead to electronic failures; human verification recommended for critical applications
• Bias: Model may perform better on defect types well-represented in training data
• Transparency: Users should understand model's confidence thresholds for different defect types

Environmental Impact

• Training: GPU-intensive (T4 GPU used in development)
• Inference: Optimized for real-time or near-real-time inspection
• Energy Efficiency: Model architecture should be selected based on deployment constraints

Maintenance

• Update Frequency: As needed based on new defect types or improved datasets
• Monitoring: Regular performance evaluation on new PCB designs
• Retraining: Recommended when encountering new manufacturing processes or materials

Deployment

• Hardware Requirements: GPU recommended for real-time inference
• Software Dependencies:
  • PyTorch/TensorFlow (depending on final implementation)
  • OpenCV for image processing
  • CUDA for GPU acceleration (optional)
• Inference Speed: Target >30 FPS on modern GPUs for production use

Model Files

• Configuration: Model architecture config file
• Weights: Trained model weights
• Class Labels: classes.txt with defect class names
• Inference Script: Python script for model deployment

Citation

If using this model, please cite the original dataset: PCB Defects Dataset by akhatova on Kaggle https://www.kaggle.com/datasets/akhatova/pcb-defects