An Self-supervised Learning & Self-attention Based Method for Defects Classification on PCB Surface Images

Lijie Zhou,Xufeng Ling,Shan Zhu,Zheng Sun,Jie Yang

DOI: https://doi.org/10.1109/cecit53797.2021.00047

2021-01-01

Abstract:Automatic detection of PCB defects become a difficult work in electronics industry, with the rapid development of Integrated Circuit. A good detection method can effectively improve production efficiency and reduce the costs of quality inspection. The traditional image processing methods have the backwards of complicated system structure, difficult image segmentation, and low detection accuracy, etc. In order to solve the above problems, this paper proposes a method of extracting features by self-supervised learning with supervised learning to complete the classification task. This method uses a deep neural network model based on self-supervised learning technology, uses multi-head self-attention mechanism; automatically recognize single or multiple defect areas in PCB, to extract defects features of PCB, and uses full connection classifier to classify features. The open-source (DeepPCB) is selected as the dataset, including 6 defect categories and non-defect categories. In the training process, data enhancement technology is used to increase the training samples, train the model and make it converge, and realize low error recognition on the verification dataset. The experimental results show that the features extracted by the self-supervised learning method are richer than those extracted by the supervised learning. The composite model formed by combining the self-attention mechanism has high accuracy and good robustness, reaches the classification accuracy of 94.33%, and has high accuracy and good generalization ability.