Surface Defect Detection Algorithm for Printed Circuit Boards Based on SRG-DETR

Abstract

Defect detection in printed circuit boards (PCBs) presents significant challenges due to the small size of defect targets, high false detection rates, and difficulties in model deployment. We propose an advanced defect detection method based on SRG-DETR model to address these issues. This method first introduces a star operation into the backbone network of the model, thereby significantly improving the model's ability to capture global information from defect images and substantially enhancing the inference speed. Secondly, an explicit attenuation mechanism and two-dimensional spatial prior knowledge are integrated into the neck network, enhancing the model's capacity to capture fine details and semantic information of PCB surface defects. Finally, GSConv is employed to improve network efficiency and reduce its size, facilitating easier deployment. Comparative experiments on a public PCB dataset demonstrate that our algorithm increases mAP by 3.8% and precision by 2.9%. Moreover, the computation amount is reduced by 59.5%, resulting in a compact size of 14.4 MB and achieving 95.1 FPS. These improvements meet the demands for real-time detection and deployment in industrial cameras.