Pixel-Weighted Background Reconstruction Network with Dual-Windows Blind Block For Hyperspectral Target Detection

Target detection is a pivotal task in hyperspectral remote sensing imagery. However, the inherent target sparsity and limited prior information have precipitated a bottleneck in algorithmic advancement, thereby imposing signiffcant challenges on deep learning methodologies. To address these issues, a pixel-weighted background reconstruction network with dual-windows blind block (DWBRN) is proposed for hyperspectral target detection. Firstly, a pre-detection-guided background reconstruction network architecture is built to alleviate the sparsity of target prior information by learning the background information of the data. Secondly, the design of dual-windows blind block (DWBB) and surrounding pixel-weighted module (SPW) minimizes the interference from targets during background generation, while simultaneously mitigating the impact of the quality of pre-detection results on network learning. Finally, the detection-reconstruction aggregation module is designed for result fusion, further suppressing the background while retaining the ability of the detection method to highlight targets. Experiments conducted on four datasets illustrate that DWBRN outperforms eight comparison algorithms.