Teacher-student Network As An Attention Module With Memory for Defect Localization

The goal of defect detection is to detect and locate defects on product surface. Since the complexity of defect types in actual industrial production and it is difficult to obtain a large number of samples for supervised training. Current defect detection methods are thought to be mainly unsupervised training methods. However, the current unsupervised methods all learn the normal sample feature representation from the whole and lack direct defect location information. In this paper, A teacherstudent network is used as the attention mechanism to provide the network with more direct information on the location of defects. At the same time, to increase the network’s ability to characterize normal samples, a memory module is used to enhance the reconstruction ability of the network. The experimental results show that our approach obtains the best results compared to similar teacher-student networks and attention mechanism networks.

I. Introduction

The task of anomaly detection is a very important part of the industrial manufacturing process and is aimed at detecting defects in the product in advance. The design of automated anomaly detection methods has been the focus of industrial and theoretical attention. In the traditional anomaly detection field [1], [2], methods are used to manually design feature extraction modules for specific defects. Such methods often only solve specific defect detection problems and require complex feature design modules, which are susceptible to be affected surrounding scenarios and cannot be used on a large scale.