A Unified Dual Attention-Guided Reverse Distillation Framework for Anomaly Detection

One of the great challenges is how to learn a unified model for multi-class anomaly detection. The methods based on reverse distillation gained impressing detection results, but they face a problem that the anomaly information is possibly introduced in the following inference process. To address the above issue, we employ two strategies to reduce the introduction of anomalous information. Firstly, channel attention is utilized to assign distinct weights to each channel, thereby emphasizing the importance of specific weight information. Secondly, spatial attention is introduced to selectively focus on critical regions within the image, thus disregarding extraneous background information. This approach ensures that the student network precisely learns the normal representation of essential features, thereby significantly reducing the possibility of incorrect reconstructions by the student network during the inference process. Moreover, to enhance anomaly localization accuracy, image features are independently introduced into spatial and channel attention mechanisms, resulting in the acquisition of features distinguished with varied channel and spatial weights via convolutional fusion. This process contributes to the enhancement of feature discriminability across diverse categories, thereby diminishing interference between categories.