Analysis of Generating Attention Map by Layer-wise External Attention for Deep Anomaly Detection

Abstract

Visual attention mechanisms aim to improve the classification performance of convolutional neural networks. The External Attention mechanism is a new visual attention mechanism for anomaly detection. It integrates an anomaly map from an output of networks trained with normal images only, into supervised classification models as an attention map. The External Attention mechanism improves the classification performance of existing anomaly detectors. However, it has not been clarified what kind of anomaly maps efficiently improve the performance of anomaly detection. In this study, we supposed a positional feature on anomalies is important for External Attention. Then, we used an extremely sparse anomaly map. Using MVTec AD, a benchmark for image anomaly detection, we analyzed the conversion process from an extremely sparse anomaly map to an attention map in External Attention. As a result, we found that an extremely sparse anomaly map efficiently improved the anomaly detection performance in most categories.