A Method for Surface Defect Detection of Tempered Glass Based on Polarization Characteristics and Unsupervised Learning

Self-detonation of tempered glass due to defects has always been a major problem in its engineering applications, and in the actual defect detection, it is susceptible to the interference of dust on tempered glass. In this paper, combining the different polarization characteristics of defects and dust, we design a polarization system combining polarizer and 1/4 wave plate to collect images, combined with deep learning for detection and differentiation. However, in the actual detection process, the samples of tempered glass with defects are relatively rare, and the probability density of abnormalities is low, typically resulting in a long-tailed distribution of normal and abnormal data. The lack of abnormal sample data leads to a low detection rate of defects by supervised learning methods, which can't satisfy the actual industrial inspection needs. Therefore, this paper proposes an unsupervised learning method based on FastFlow to solve the problem of insufficient abnormal samples. Finally, the experiments show that the unsupervised learning method based on polarization characteristics and FastFlow proposed in this paper effectively detects defects and distinguishes dust and the recall rate of defects reaches 96.1%, and the false detection rate is only 3.1%, which realizes the intelligent detection of tempered glass defects and meets the requirements of the industrial inspection of tempered glass.