Detection and segmentation of image anomalies based on unsupervised defect reparation

Anomaly detection is a challenging task in the field of data analysis, especially when it comes to unsupervised pixel-level segmentation of anomalies in images. In this paper, we present a novel multi-stage image resynthesis framework for detecting and segmenting image anomalies. In contrast to existing reconstruction-based approaches, our method is based on repairing suspicious regions of defective images so that the defects can be localized in the residual map between inputs and the repaired outputs. To avoid the reconstruction artifacts caused by defects, we propose to generate each pixel of the image by its context in the first coarse reconstruction stage. Then, while excluding all safe pixels, our method repairs suspicious regions that have large deviations to the original input image in subsequent stages. After several iterations, the defects will be detected in the final residual map. The experimental results show that we achieved better performances than the state-of-the-art benchmarks using the publicly available MVTec dataset as well as a real-world equipment surface dataset. In addition, the method also demonstrates an excellent capability of repairing defects in abnormal samples.