Morphological-based Attention in Graph Networks for Hyperspectral Anomaly Detection

Kurzfassung

Characterizing spectral-spatial features at both local and global scales, and discriminating anomalies from background, are key objectives in hyperspectral anomaly detection. To address these issues, we propose a dual branch network extracting complementary local and global spectral-spatial features. A cross-attention mechanism is applied between the input patch and opening and closing morphological profiles, which convey information about anomalies and background, respectively, to enhance anomaly and background separation. In the first branch, consecutive convolutional layers with skip connections capture multi-level spectral-spatial semantic features. In the second branch, relationships between each spectrum and cluster centroids derived from the full hyperspectral image are modeled as a graph and processed by a graph convolutional network to extract global features. Features from both branches are fused to estimate the background, and the reconstruction residual is used as the anomaly score. Comparisons with several state-of-the-art methods on benchmark hyperspectral datasets demonstrate the proposed method's effectiveness and competitive performance, while maintaining a reasonable computational cost. An ablation study further assesses the contribution of each network component.