PCB Defect Detection

Description

Automated detection and classification system for Printed Circuit Board (PCB) defects using YOLO11. The project combines Kaggle training with a local graphical interface for testing trained models.

Main Features:

• Automatic detection of 6 types of PCB defects
• Modern GUI interface with ttkbootstrap theme
• Fast inference with PyTorch (GPU accelerated)
• Detection precision of 97.4%
• Optimized training pipeline (50% faster)

GUI Interface Demo

Modern interface with ttkbootstrap theme featuring:

• Control panel with model and image loading
• Large image display area with zoom controls
• Color-coded detection results panel
• Real-time defect detection and visualization

Technologies Used

Component	Technology	Version	Usage
Deep Learning	YOLO11 (Ultralytics)	8.3.0+	Object detection
Framework	PyTorch	2.2+	Model training and inference
GUI Interface	ttkbootstrap	1.10+	Modern user interface
Image Processing	OpenCV, Pillow	4.9+, 10.2+	Image manipulation
Training Platform	Kaggle	-	Free GPU T4

Training Platform

This project was trained on Kaggle Notebooks with the following configuration:

• GPU: NVIDIA Tesla T4 × 2 (15 GB VRAM each)
• RAM: 30 GB
• Storage: 299.8 GB available
• Training Duration: 42 minutes
• Kaggle Notebook: View Training Results

Performance Results

Dataset: PCB Defects - Akhatova

• Total Images: 693 annotated images
• Split: 79.9% training (554 images) / 20.1% validation (139 images)
• Model: YOLO11m (20M parameters)
• Training: 100 epochs on GPU T4 × 2 (Kaggle)
• Training Time: 42 minutes
• Training Date: January 30, 2026

Global Metrics

Metric	Score	Description
Detection Precision	96.4%	Mean Average Precision at IoU 0.5
Strict Precision	53.8%	Mean Average Precision at multiple IoU thresholds
Mean Precision	97.0%	Precision across all classes
Mean Recall	92.5%	Detection rate
F1-Score	94.7%	Precision-recall balance

Training Configuration

Parameter	Value
GPU	NVIDIA Tesla T4 × 2
Epochs	100
Batch Size	16
Learning Rate	0.001
Image Size	640×640
Optimizer	auto (AdamW)

Training Visualization

Key Observations:

• Rapid convergence in first 20 epochs
• Stable training with no overfitting
• Validation errors near zero after epoch 10
• Learning rate smoothly decays from 0.01 to 0.0001

Key Observations:

• Rapid convergence in first 40 epochs
• Stable training with no overfitting
• Validation errors follow training errors closely
• Learning rate smoothly decays from 0.01 to 0.0001
• Training completed in 42 minutes on GPU T4 × 2

Sample Detection Results

The model successfully detects various PCB defects with high confidence and accurate bounding boxes across different defect types.

Defect Classes

ID	Class	Description
0	missing_hole	Missing drill hole
1	mouse_bite	Irregular edge
2	open_circuit	Broken trace
3	short	Short circuit
4	spur	Copper protrusion
5	spurious_copper	Unwanted copper

Training on Kaggle

1. Create Kaggle notebook with GPU enabled
2. Add dataset: akhatova/pcb-defects
3. Run training code:

!pip install ultralytics -q
!wget -q https://github.com/alainpaluku/pcb-defect-detection/archive/main.zip
!unzip -q main.zip
!mv pcb-defect-detection-main pcb-defect-detector
%cd pcb-defect-detector
!python run_kaggle.py

4. Download trained model:

from IPython.display import FileLink
FileLink('/kaggle/working/pcb_model.pt')

Local GUI Testing

git clone https://github.com/alainpaluku/pcb-defect-detection.git
cd pcb-defect-detection
pip install -r requirements.txt

1. Place downloaded model in models/ directory:

   • pcb_model.pt

2. Place test images in images/ directory

3. Run GUI: python -m gui_test.app

Project Architecture

pcb-defect-detection/
├── src/                   # Main Python modules
│   ├── model.py          # YOLO11 model and export
│   ├── trainer.py        # Training pipeline
│   ├── detector.py       # Detection interface
│   └── config.py         # Configuration
├── gui_test/             # Graphical interface
│   ├── app.py           # Main application
│   ├── main_window.py   # Main window
│   └── model_loader.py  # Model loading
├── models/               # Trained models (.pt)
├── images/               # Test images
├── run_kaggle.py         # Kaggle training script
└── requirements.txt      # Dependencies

---

Author: Alain Paluku - @alainpaluku