An Anomaly Detection and Localization Method Based on Feature Fusion and Attention

In the context of anomaly detection and localization, the small proportion of anomaly data and its unknown nature increase the difficulty of the model in learning the anomaly information. At the same time, the current mainstream detection methods all need help with the problems of limited feature learning ability and weak model generalization ability. Therefore, this paper proposes a multi-module combination of anomaly detection and localization method, which enhances the model’s differential learning of normal and abnormal through anomaly simulation strategy and memory module, and fuses multi-scale semantic information through multi-scale feature fusion to improve the accuracy of the judgment of the model on the different scales of abnormality. In addition, the low-cost attention modules acquire global information from multiple perspectives and augment essential features to enhance the feature learning capability. The superiority of the method in this paper is demonstrated in the MVTec anomaly detection dataset.