PCB micro-defect detection algorithm based on receptive field attention

Printed circuit board (PCB) are critical components in electronics manufacturing, yet micro-defects caused by materials, processes, equipment, environmental, and human factors remain challenging. Current deep learning based methods for detecting PCB defects encounter difficulties in achieving a balance among accuracy, speed, and parameter efficiency. To overcome these drawbacks, this paper puts forward a high precision algorithm for detecting micro-defects on PCB, which makes use of a receptive field attention mechanism. Firstly, a receptive field attention mechanism (RFAConv) module is integrated into the C2f module of the YOLOv8 backbone network. This integration serves to boost the network's capacity to extract features related to small defects. Secondly, the C2f module in the YOLOv8 neck network is substituted with the CSPStage module. Through a segmentation and merging strategy, gradient information is propagated along distinct paths. so that the gradient combination is richer. An improved YOLOv8 detection algorithm model is proposed for PCB microdefect localization. In the training process, This paper aims to improve the accuracy of boundary box regression and proposes a boundary box similarity comparison loss function regression model based on the minimum point distance(MPDIoU). Experimental results show that the improved algorithm model achieves 97.8% accuracy and 98.9% recall, only 31MB parameters, and has great advantages on mAP50.