Rev-RetinaNet: PCB defect detection algorithm based on improved RetinaNet

Abstract

PCBs are important electronic components and are used in a wide range of industries. The efficient detection of defects in PCBs can improve the yield of PCB production, and this research direction has a wide range of prospects for development. Nowadays, the increase in hardware computing power and rapid advances in imaging technology continue to facilitate the development of deep learning, and the use of convolutional neural networks for defect detection has become possible. Defect detection faces the following challenges: small target size, more difficult feature extraction, and low detection accuracy using generic target detectors. To address these challenges, this paper proposes a network called Rev-RetinaNet, which uses the well-known single-stage network RetinaNet as the baseline network with improvements and ConvNext as the backbone network part, effectively improving the adequacy of the network for feature extraction. In addition, to avoid overfitting, this paper uses the Drop Path method to randomly "remove" the sub-paths of the multi-branch structure in the deep learning model, and adopt the YOLOXHSVRandomAug random luminance adjustment strategy to simulate the situation of PCB boards under different lighting conditions, so that the network can adapt to different situations of PCB detection, thus improving the accuracy of defect detection. This enables the network to adapt to different situations of PCB inspection, thus improving the accuracy of defect detection. This paper conducted detailed ablation experiments on a publicly available PCB dataset from Peking University. The results show that the Rev-RetinaNet achieves a mAP of 89.74% compared to the baseline network, an improvement of 3.1%, demonstrating the effectiveness of the proposed network.

Topics & Concepts