Inferring normality from noised samples: Enhanced deep autoencoder with image denoising for anomaly detection

Abstract

The goal of image anomaly detection is to detect and localize the patterns that do not meet the normal expectations in an image. The widely used reconstruction-based method typically trains an autoencoder just with normal images and detects abnormal images based on the reconstruction error. It is expected that abnormal images cannot be reconstructed well because they are not included in the training process. However, due to its strong generalization ability, the autoencoder often reconstruct abnormal images quite well, thus reducing the performance of anomaly detection. To address this problem, this paper proposes a novel anomaly detection method based on image reconstruction and image denoising. Specifically, this method enhances the autoencoder to learn robust normality by adding noise to the input images. Besides, we proposed a metric to guide the model training based on the spatial scale of abnormal regions. The experimental results demonstrate that our method achieves competitive performance comparing to the state-of-the-art methods. Moreover, based on its concise architecture, our method achieves high efficiency.