SRAD: A spatially-aware reconstruction network with anomaly suppression for multi-class anomaly detection

Abstract

Reconstruction-based methods have achieved remarkable outcomes in unsupervised image anomaly detection by training separate models for different categories. However, when it comes to a practical unified model, these methods often face the “identical shortcut” problem, where both normal and abnormal samples can be recovered well, leading to failure in anomaly detection. To address this problem, we propose a Spatially-aware Reconstruction network with anomaly suppression for multi-class Anomaly Detection (SRAD). Firstly, we propose a Spatially-aware Channel Convolutional (SCC) neural network, which replaces general convolution with channel convolution and incorporates a Spatial Information Fusion (SIF) block during encoding and decoding. The SIF block is proposed to allow the model to capture rich feature representations while avoiding overemphasizing details. Secondly, to prevent the model from learning identical reconstruction, we propose an Anomaly Suppression Feedback Learning (ASFL) strategy. The ASFL strategy encourages the model to accurately reconstruct normal samples while inhibiting the reconstruction of abnormal samples through a feedback mechanism. Experiments on MVTec AD, VisA and BTAD datasets demonstrate the clear superiority of SRAD compared to previous state-of-the-art unified models, e.g., achieving 98.2% I-AUROC and 97.6% P-AUROC on multi-class MVTec AD dataset. Furthermore, SRAD exhibits a high frame rate of 51 FPS on a 2080 Ti GPU, making it potential for practical applications.