Anomaly detection via end-to-end learning of mask generator and denoising autoencoder

Abstract

Anomaly detection is a task that distinguishes whether incoming data is normal or abnormal. To give a network the ability to detect anomaly samples, we propose a method that deliberately limiting and distorting information of the data and then restoring original data from such corrupted data by using denoising and inpainting. As most of anomaly detection algorithm does, the reconstruction error would be the measure of abnormality. The main idea behind our proposed method is that the restored data distribution inevitably follows the normal sample distribution if only limited information of the data is provided. Experimental results show that the proposed method is superior to the existing generative model-based abnormal detection method in both quantitative and qualitative aspects.