DB-YOLO: An Improved YOLOV8 Model Based on The Second Backbone for Small Defects on PCB Surface

Abstract

This study introduces the DB-YOLO method, targeting the existing PCB defect detection algorithms, which struggle to identify minor flaws in intricate, small-scale layouts with comparable backdrops, to enhance detection precision. Initially, the deepening of the network layer will lead to a large number of loss of detailed features of the detection target. Although the significant features of the target can be retained, this loss will make the information of the small target incomplete in the small target detection task. Therefore, we combined CBLinear and CBFuse modules in SOTAYOLOv9 to design the second backbone. By redistributing information at different levels, the micro features in the original information are strengthened, and the communication between channels is increased at the same time, which effectively improves the richness of feature information.; next, in order to shorten the forward propagation process and retain the original information to extract the features of small defects in the structure, we use partial convolution to design a new module named CSPC, which reduces the waste of computer resources caused by feature map redundancy. Furthermore, a refined version of the inverted residual multiscale attention module (iREMA) was introduced to augment the representation of features. In conclusion, our comparison of the DB-YOLO model with other current models reveals through experiments that our suggested model surpasses them, enhancing the mAP in the validation set by 3.6% relative to YOLOv8n.