ASL-YOLO-Based Detection of PCB Solder Joint Defects

Abstract

With the advancement of electronic manufacturing towards high density and miniaturization, PCB soldering defect detection has become a critical component of industrial quality control. Traditional human visual inspection suffers from low efficiency, while existing deep learning algorithms generally face challenges such as feature loss during downsampling and complex background interference when dealing with tiny defects. To address these issues, this paper proposes an improved model named ASL-YOLO based on YOLOv8, designs an EOC dual-branch pooling feature extraction module to effectively preserve edge details of tiny targets, and introduces an AGCA attention mechanism combining asymmetric convolution and depthwise convolution to highlight directional structures and suppress background interference. This work is based on a real SMT production line PCB dataset. Experimental results demonstrate that ASL-YOLO achieves high detection accuracy on both self-built and public datasets while maintaining favorable inference speed, verifying its industrial application value. The self-built dataset and key code have been open-sourced at https://github.com/Lssppssl/ Dataset-and-key-code-for-PCB-soldering-defect-detection-based-on-ASL-YOLO.