Image anomaly detection with a unified transformer model guided by dual-feature

Unsupervised industrial image anomaly detection has been extensively investigated by reconstruction-based frameworks. While demonstrating promising performance benchmarks, existing reconstruction networks frequently degenerate into identity mapping behavior: they are prone to directly copying inputs as outputs, which fails to truly learn the deep structural information and statistical distribution of the data, and often require training distinct models for individual object categories (one-class-one-model). In this paper, we propose a powerful multi-class unified model based on the Dual-Feature Guided Reconstruction Network (DFGR) for multi-class anomaly detection. One of DFGR strengthens the low-level feature to realize the guidance function of the model to reconstruct important normal features and significantly reducing the reliance on prior knowledge, and the other multi-layer fusion feature provides rich semantic features of the image. We utilize the main structural features to guide the reconstruction, realizes the interaction between the global information of the main structural features and the local information of the reconstructed feature map. Our method better balances the contribution of the low-level spatial structure information to the overall reconstruction process, and also effectively reduces the sharp response of the reconstructed network to small background noises. Experimental results on the MVTec dataset demonstrate an image-level area under the receiver operating characteristic (AUROC) of 98.0% and a pixel-level AUROC of 97.1%, and further validations on the DAGM2007, Rollei, and red and blue datasets confirm the feasibility of the dual-feature structure. The code: https://github.com/wyanb/DFGR.