SEPDNet: simple and effective PCB surface defect detection method
Du Lang, Zhenzhen Lv
Abstract
Replacing time-consuming and costly manual inspections on production lines with efficient and accurate defect detection algorithms for Printed Circuit Boards (PCBs) remains a significant challenge. Current PCB defect detection methods are typically optimized using existing models such as YOLO and Faster R-CNN to enhance detection accuracy. In this study, we analyse a PCB defect dataset characterized by small targets and a concentrated size distribution. SEPDNet (Simple and Effective PCB Defect Detection Network) is designed for the characteristics of the dataset, only one detection head is used, which reduces the number of parameters and improves the detection performance at the same time. SEPDNet uses RepConv (Re-parameterizable Convolution) to improve the backbone representation ability, and FPN (Feature Pyramid Network) is used in the neck part to simplify the model. SEPDNet has fewer than 30% of the parameters of YOLOv9u-s, yet achieves an improvement of 0.025 in the F1 score, 2.7% in mAP50, and 3.8% in mAP50:95 compared to YOLOv9u-s. We propose the method of designing the model according to the characteristics of the dataset. Our experiments show that customizing the model design according to dataset characteristics can achieve strong performance with a simplified structure and fewer parameters.