Image Quality Assessment Driven Self-Supervised Anomaly Detection

Abstract

Anomaly detection is a challenging task due to the bottleneck of anomalous sample collection. Recently, self-supervised learning has shown great potential for anomaly detection. However, these methods use task-related attributes as selfsupervised signals and are suitable for specific application scenarios. In this paper, we propose a more general selfsupervised method to detect anomaly which regards image quality as a supervised signal. Specially, we convert anomaly detection task into an image quality restoration task which contains distortion and restoration processing. Firstly, we introduce content distortion which erases hardness blocks in normal image based on no reference quality assessment method. Meanwhile, we introduce resolution distortion which down-samples normal images to low-resolution images preserving the essential appearance information. Then, a restoration network is trained to repair such distortions that encourages the model to learn detail information and the content of hardness regions in normal image. Extensive experiments on public benchmark datasets verify the effectiveness and generality of the proposed method.