Deviation capture networks for anomaly detection

Feng Yan, Jiawei Cheng, Jian Feng, Yang Lu, Wenjie Zhang, Xiaoheng Jiang, Ming Liang Xu

Abstract: Anomaly detection is a key part of industrial manufacturing. Current anomaly detection methods relying on simulated anomaly data generally consider anomalies of different types as a whole during learning. This ignores the intrinsic differences of anomalies, such as texture and structure anomalies, which is adverse to the network learning. In this paper, we propose a Deviation Capture Network (DevCNet), which introduces a deviation capture module (DCM) to learn the differences between normal and anomalous samples and the differences among different types of anomalies. Specifically, DCM contains three deviation capture heads that identify anomaly deviation, texture anomaly deviation, and structural anomaly deviation, respectively. These deviation capture heads are jointly learned during training. In DevCNet, we adopt a memory bank to store features of normal samples and obtain deviation features of input samples. The deviation capture heads take the feature streams corresponding to their respective anomaly identification tasks as input and jointly identify the anomalies. In addition, we introduce an anomaly simulation module (ASM) to synthesize texture and structure anomalies randomly. Experiments conducted on the MVTec AD and BeanTech AD industrial datasets demonstrate the superiority of our method.

External IDs:dblp:journals/aei/YanCFLZJX26