HCMambaAD: anomaly detection and localization on industrial textured surface with hybrid context clustering and Mamba reconstruction

Abstract In the field of industrial textured image anomaly detection and localization, methods based on convolutional neural network (CNN) and Transformer architecture have achieved remarkable results. However, CNN is difficult to capture global information, and Transformer has high computational complexity. In this paper, we propose a texture image anomaly detection and localization method HCMambaAD, which reconstructs an image based on the developed Mamba architecture with hybrid context clustering. HCMambaAD overcomes the problem of ignoring local information in the fixed global scanning method of Mamba architecture by introducing a hybrid feature extraction strategy of global scanning and context clustering, which makes the model able to combine global and local information, and can dynamically capture image contextual information. The experiments were extensively conducted on MVTec-AD-Texture, ISP‐AD, KolektorSDD2 and Texture-AD datasets, and the results demonstrate that the proposed method achieves state-of-the-art performance in comparison with existing methods, such as DRAEM, EfficientAD, VitaLnet, DeSTSeg, MambaAD, SuperSimpleNet, ViTAD, Dinomaly, and INPFormer.