• Proposes ED framework via intra- and inter-image discrepancies in reconstruction. • Achieves SOTA performance on three industrial benchmarks. • Open source implementation facilitates practical deployment in industrial inspection. Anomaly detection plays a crucial role in industrial inspection. Reconstruction-based anomaly detection methods face the well-known “learning shortcut” challenge, where models fail to reproduce normal patterns but instead resort to identity mapping or artificial denoising strategies. This paper proposes a novel self-supervised anomaly detection framework, Emphasize Discrepancies (ED), which addresses the prevalent learning shortcut problem in existing reconstruction-based methods through fusion of intra- and inter-image discrepancies. The advantage of this method lies in its simple transformer architecture that simultaneously processes both intra- and inter-image correlations, focusing features through these two types of differences. To maximally preserve this information, we designed a dual-path discrepancy integration approach that effectively combines intra- and inter-image difference signals, enhancing sensitivity to anomalous regions. Additionally, we introduced a channel-wise attention-based feed-forward mechanism to address the slow convergence problem. Extensive experiments on multiple real-world unsupervised benchmarks demonstrate that our proposed method achieves state-of-the-art performance. The implementation code is available at https://github.com/Vzoooong/ED .
Wei et al. (Thu,) studied this question.