Anomaly Detection of Hyperspectral Images Based on Transformer With Spatial–Spectral Dual-Window Mask

Anomaly detection has become one of the crucial tasks in hyperspectral images processing. However, most deep learning-based anomaly methods often suffer from incapability utilizing spatial–spectral information, which decreases accuracy. To address this problem, we propose a novel method with dual-window mask transformer, termed as S2DWMTrans, can fully extract features global and local perspectives, suppress reconstruction targets adaptively. Specifically, transformer aggregates background information entire image perspective to neutralize anomalies, uses neighboring pixels reconstruction. An adaptive-weighted loss function is designed further adaptively during network training process. According our investigation, first work apply detection. Comparative experiments ablation studies demonstrate that proposed S2DWMTrans achieves competitive performance.