Improving Image Anomaly Localization: A Multi-branch and Skip Connection Framework

Image anomaly detection plays an important role in various fields, such as industrial defect detection and medical image analysis. Although the use of image restoration has led to significant progress in image anomaly detection, challenges persist in insufficient image feature extraction and inadequate integration of low-level and high-level features. In this study, a framework that combines reconstruction methods with knowledge distillation is designed. To address the challenge of insufficient image convolution feature extraction, a multi-branch feature extraction module is introduced, extracting both local and global features using convolution and self-attention modules, respectively. Additionally, to overcome the inadequate integration of low-level and high-level features in the skip connection part, a straightforward skip connection is designed to fuse semantic information from high-level features with finer details from low-level features through element-wise multiplication. Extensive experiments on industrial image datasets, such as Mvtec AD and BTAD, validate the effectiveness of the proposed method in improving anomaly detection localization performance. The Mvtec AD dataset yielded impressive results, with image-level and pixel-level ROC_AUC achieving average accuracy values of 99.16% and 98.09%, respectively. Similarly, on the BTAD dataset, our method exhibited remarkable performance, achieving average accuracy values of 93% and 96% for image-level and pixel-level ROC_AUC, respectively. These notable improvements across both datasets underscore the effectiveness of our proposed method.