Dual-Net: Dual Visual Spectral Affinity Monitoring Network for Hyperspectral Anomaly Detection

Hyperspectral anomaly detection (HAD) refers to the identification of anomalies that significantly differ from the background without relying on anomalous spectral priors. By utilizing effective deep learning networks, reconstruction-based methods have become one of the primary solutions for HAD tasks. However, most existing reconstruction-based methods fail to fully leverage the local affinity between a central pixel and its surrounding pixels. To address this issue, a novel Dual Visual Spectral Affinity Monitoring Network (Dual-Net) is proposed. Specifically, the dual network separately extracts the unique features of each pixel and its spectral similarity with surrounding pixels. The Spectral Affinity Monitoring Strategy (SAMS) uses this similarity as an affinity matrix to monitor the impact of anomalous pixels on the network reconstruction process. To further enhance the effectiveness of SAMS, a Visual Monitoring Focused Attention (VMFA) mechanism inspired by the human visual system is introduced. This mechanism effectively captures anomalous features at both the pixel and semantic levels while suppressing background information. Extensive experiments on nine public datasets validate the superiority and effectiveness of the Dual-Net.