A multi-scale information fusion framework with interaction-aware global attention for industrial vision anomaly detection and localization
Zhuo Li, Yifei Ge, Lin Meng · Information Fusion · 2025
This paper presents a practical industrial visual anomaly detection framework that can detect and locate anomalous regions accurately. Specifically, we design an interaction-aware global attention framework (IGAF) comprising a pre-trained convolutional neural network as well as spatial and channel interaction-aware global attention module. The pre-trained convolutional neural network is employed for multi-scale feature extraction. The spatial interaction-aware global attention module helps the model focus on specific information by strengthening the correlation between certain channels. Meanwhile, the channel interaction-aware global attention module effectively captures critical regions in the input image and promotes an understanding of global spatial relationships. We evaluate IGAF on standard industrial visual anomaly detection datasets. The proposed IGAF achieves an outstanding result of 99.13% pixel-level AUROC with WideResNet50 as the backbone of the popular used MVTec AD dataset. IGAF achieves an impressive result of 97.51% pixel-level AUROC, even using a smaller backbone network such as ResNet18. Furthermore, IGAF achieves a remarkable result of 98.75% on the widely used VisA dataset. Additionally, IGAF demonstrates an excellent result of 92.40% on the real-world industrial anomaly detection dataset BTAD. These results indicate the effectiveness of IGAF for the detection and location of industrial vision anomalies.