Dual-window transformer framework with pyramid structure and constrained self-attention for hyperspectral anomaly detection

Hyperspectral anomaly detection (HAD) is widely used in various fields including military, agriculture, mining, and food safety inspection. However, the absence of prior information on targets poses significant challenges to feature extraction and anomaly identification. To address this issue, this paper proposes a novel dual-window transformer framework integrated with a pyramid structure and constrained self-attention mechanism, which effectively leverages both local and global spectral information for anomaly detection. The dual-window transformer is designed to extract deep spectral features by capturing discriminative patterns between central pixels and their surrounding background. Simultaneously, the constrained self-attention module incorporates global contextual information into the feature representation. Furthermore, a stepwise downsampling pyramid architecture is introduced to reduce the sensitivity of the model to dual-window size selection while facilitating the propagation of global information from higher to lower layers. Extensive hyperparameter analysis and comparative experiments demonstrate the robustness and superiority of the proposed framework. The source code is publicly available at: <ns0:ext-link ext-link-type="uri" ns4:href="https://github.com/aosilu/DWT-P-CSA-HAD">https://github.com/aosilu/DWT-P-CSA-HAD</ns0:ext-link>.