SEPDNet: simple and effective PCB surface defect detection method.

Area of Science:

• Computer Vision
• Manufacturing Automation
• Artificial Intelligence

Background:

• Manual inspection of Printed Circuit Boards (PCBs) is time-consuming and costly.
• Existing automated defect detection methods often use models like YOLO and Faster R-CNN.
• A significant challenge lies in detecting small targets and concentrated size distributions in PCB datasets.

Purpose of the Study:

• To develop an efficient and accurate defect detection algorithm tailored for PCB datasets with specific characteristics.
• To design a simplified network architecture that reduces parameter count while enhancing detection performance.
• To demonstrate the effectiveness of customizing model design based on dataset attributes.

Main Methods:

• A novel network, SEPDNet (Simple and Effective PCB Defect Detection Network), was designed.
• SEPDNet utilizes a single detection head to minimize parameters and boost performance.
• RepConv (Re-parameterizable Convolution) was employed in the backbone for enhanced representation, and FPN (Feature Pyramid Network) in the neck for model simplification.

Main Results:

• SEPDNet achieved superior performance compared to YOLOv9u-s despite having less than 30% of its parameters.
• Improvements were noted in F1 score (+0.025), mAP50 (+2.7%), and mAP50:95 (+3.8%).
• The customized approach demonstrated strong performance with a simplified structure.

Conclusions:

• Customizing model design to dataset characteristics is a viable strategy for improving PCB defect detection.
• SEPDNet offers a promising solution for efficient and accurate automated PCB inspection.
• The study highlights the potential of tailored deep learning models in industrial applications.