Robust Anomaly Detection under Data Contamination for Pixel-Level Defective Region Segmentation
TL;DR
This work proposes an unsupervised robust anomaly detection framework under data contamination for pixel-level defective region segmentation, built on an autoencoder structure with a memory module and holds two mild assumptions: across-sample low-rank, and sparse anomalies.
Abstract
Industrial image-based anomaly detection plays a significant role in quality control, where pixel-level defective region segmentation is an important task to locate anomalies for various applications. Due to the data-abundance and label-scarcity nature of industrial products, the unsupervised learning approach has attracted great attention because such a method does not require a labeling process. However, it is often based on a normality assumption that all training data is normal, which can be broken in practice. Thus, a contaminated dataset, i.e., a mixture of normal and anomalous samples without label, should be considered. To deal with this challenge, we propose an unsupervised robust anomaly detection framework under data contamination for pixel-level defective region segmentation. Specifically, unlike matrix-decomposition-based methods that need certain explicit assumptions on the normal background, our proposed model, built on an autoencoder structure with a memory module, allows flexible patterns presented in the background and holds two mild assumptions: (i) across-sample low-rank, and (ii) sparse anomalies. Additionally, the proposed model will be trained and pixel-wisely locate anomalies by decomposing each input image into a normal background and an anomaly mask in a one-stop fashion. The effectiveness is demonstrated in the case study on a public anomaly detection dataset.