Reconstruction by inpainting for visual anomaly detection

Abstract Visual anomaly detection addresses the problem of classification or localization of regions in an image that deviate from their normal appearance. A popular approach trains an auto-encoder on anomaly-free images and performs anomaly detection by calculating the difference between the input and the reconstructed image. This approach assumes that the auto-encoder will be unable to accurately reconstruct anomalous regions. But in practice neural networks generalize well even to anomalies and reconstruct them sufficiently well, thus reducing the detection capabilities. Accurate reconstruction is far less likely if the anomaly pixels were not visible to the auto-encoder. We thus cast anomaly detection as a self-supervised reconstruction-by-inpainting problem. Our approach (RIAD) randomly removes partial image regions and reconstructs the image from partial inpaintings, thus addressing the drawbacks of auto-enocoding methods. RIAD is extensively evaluated on several benchmarks and sets a new state-of-the art on a recent highly challenging anomaly detection benchmark.