Improved THT Solder Joint in PCB Defect Detection Model Based on YOLOv8

Detection of solder joint defects on printed circuit boards is an important means of improving the production quality of printed circuit boards. In this study, improvements are made to the YOLOv8 model by restructuring the neck network architecture of YOLOv8s through the BiFPN network. This modification reduces the number of parameters of the model while improving its multi-scale feature detection capability. To address the issue of sample imbalance in the target detection dataset for THT solder joint defects. Equalization Loss is introduced to adjust the weights of the samples, thereby improving the detection capability for defect categories. Finally, the improved model is compared with the yolov8s model, an improvement of 2.2% in average precision mAP@0.5:0.95, with a 33.78% reduction in the number of parameters.