Self-supervised learning for textured surface anomaly detection and localization

Abstract

Aiming at the problems of insufficient labeled samples and high missed detection rate in common textured surface anomaly detection, the paper designs a self-supervised learning model based on masked Autoencoder, which can realize accurate detection and location of anomalies without providing mass anomaly samples. Autoencoder is widely used, but it is difficult to detect and locate anomalies by reconstruction error due to its strong generalization ability reconstructed anomalies with small errors. Then, masked reconstruction method is proposed to reduce the generalization performance. First, each input image is masked to obtain multiple masked input images which are sequentially reconstruct by the Autoencoder. Second, these reconstructed images are complementarily masked and recombined to obtain the final reconstructed image. Finally, anomaly detection and localization are achieved by evaluating the reconstruction error between the input and reconstructed image. The experiment results indicate that the anomaly detection rate of this method is 95.09 % and the anomaly location rate is 93.32% under the anomaly detection standard metric，and the performance can be significantly improved.