Interactive Global-Local Anomaly Fusion Network for Hyperspectral Anomaly Detection

The core goal of hyperspectral anomaly detection (HAD) is to identify anomalous targets that significantly differ from the background. However, most existing methods typically perform anomaly detection (AD) based on a single background modeling approach. Due to the intricate and complex distribution of background features, this often leads to insufficient background reconstruction accuracy and inadequate anomaly feature representation. To address these issues, this paper proposes an interactive global-local anomaly fusion network (IGFNet) to simultaneously perform feature analysis on the global and local content in hyperspectral image. The method first constructs background prototypes and integrates a background enhancement module to reconstruct a pure background for global content analysis. To avoid missing local content details, sparse mamba is introduced to directly extract local spatial-spectral feature patterns. Finally, the fusion of global and local perspective content achieves accurate detection results. Experimental results on multiple public datasets and comparisons with other state-of-the-art methods demonstrate the effectiveness and superiority of our approach.