PCB Defect Detection Method Based on Improved RetinaNet

Abstract

AbstractIn the electronic industry product quality control, PCB defect detection is a crucial part, which has the characteristics of small defect size and high similarity. The existing defect detection methods are still not good enough for detecting small target defects; therefore, the algorithm in this paper proposes an improved algorithm for PCB defect detection based on the RetinaNet model. The ResNet-D residual structure and efficient channel focus module are introduced in the model backbone network to enhance its feature extraction capability and achieve the purpose of improving the detection accuracy. At the same time, the method replaces the original multi-step learning decay strategy with a cosine annealing scheduling learning strategy, which optimizes the training process of the model. Finally, the performance of the method is verified on the publicly available PCB defect dataset from the Open Laboratory of Intelligent Robotics, Peking University. The experimental results show that the algorithm improves the mAP value by 3.2% compared with the original algorithm, while the fastest detection speed reaches 36.9 FPS, which can effectively improve the defect detection performance of PCB.KeywordsDefect detectionRetinaNetChannel attention