Self-Supervised Masking for Unsupervised Anomaly Detection and Localization

Recently, anomaly detection and localization in multimedia data have received significant attention among the machine learning community. In real-world applications such as medical diagnosis industrial defect detection, anomalies only present a fraction of images. To extend reconstruction-based architecture to localized anomalies, we propose self-supervised approach through random masking then restoring, named Self-Supervised Masking (SSM) for unsupervised localization. SSM is able not enhance training inpainting network but also lead great improvement efficiency mask prediction at inference. Through masking, each image augmented into diverse set triplets, thus enabling autoencoder learn reconstruct with masks various sizes shapes during training. improve effectiveness inference, novel progressive refinement that progressively uncovers normal regions finally locates anomalous regions. The proposed method outperforms several state-of-the-arts both localization, achieving 98.3% AUC on Retinal-OCT 93.9% MVTec AD, respectively.