Scd-Net: An Efficient Detection Network for Pcb Defects Based on Spatial and Channel Reconstruction Convolution and Deformable Convolutions

Abstract

The detection of defects on printed circuit boards (PCBs) is crucial for guaranteeing the quality and performance of electronic products. However, due to the size of defects is quite small in high-resolution industrial acquisition images, existing detection models are unable to identify them accurately and effectively. This study introduces a new detection model, SCD-Net, which improves the accuracy, effectiveness and robustness of PCB defect detection systems. By combining the proposed modules CNISC and DCCN with the YOLOv8s network structure, the model can reduce computational costs and improve detection accuracy. Furthermore, replacing the original loss function with a new SIoU function helps the model converge faster to improve performance. In comparison to other PCB detection models, this model achieves a PCB defect detection accuracy of up to 98.0%, a model size of 19.2M, and an detection speed of 79.1 FPS, providing a feasible solution for PCB defect detection.

Keywords: Deformable Convolutions；Defect detection；Printed circuit board (PCB)； Spatial and Channel Reconstruction Convolution；Tiny object detection； YOLOv8