Unsupervised Image Anomaly Detection and Localization using Autoencoder and U-Net Architectures

Abstract

Image anomaly detection and localization holds a significant im portanceinvariousfieldsincludingindustrialmanufacturing, med ical care, and information processing. Among the existing methods for image anomaly detection and localization the AnoViT model framework combines a vision transformer and encoder-decoder ar chitectures. The encoder-decoder architecture detects anomalous images and the ViT localizes the anomalous regions in the anoma lous images. Due to exploitation of long term dependencies and f ine grained details at global level, AnoViT performance suffers in the case of less amount of training data. In this thesis, the proposed approach combines the strengths of autoencoder and U-Net mod els. Using the autoencoder, we achieve anomalous image detection in an unsupervised way. Wecapture the fine-grained spatial details at the neighboring pixel level and the local spatial context using the U-Net model. The experimental evaluation shows that our ap proach attains an average enhancement of 15.13% for anomaly de tection and 11.48% for localization across widely-accepted bench mark the MVTecAD dataset in comparison to AnoViT. The pro posed approach is computationally efficient and demands minimal training data for robust performance