Spatial-spectral attention-enhanced U-shaped blind spot network for hyperspectral anomaly detection

In recent years, hyperspectral anomaly detection (HAD) algorithms based on blind-spot networks have achieved good detection accuracy. However, existing methods still have deficiencies in capturing global information and often ignore the multi-class attribute of the background, which further limits their detection accuracy. To address these issues, this paper proposes a spatial-spectral attention enhancement U-shaped blind spot network for hyperspectral anomaly detection (SubNet). The overall architecture of the network is in the form of U-Net, and the mask convolution and downsampling operations are used to eliminate the interference of abnormal targets on subsequent feature extraction. First, to effectively capture global information, a spatial-spectral self-attention module (SSAB) is embedded in the network. Specifically, for the spatial self-attention module, to prevent blind-spot information from interfering with subsequent feature extraction, a mask is added to the attention map to meet the blind-spot requirements. In addition, to perceive the multi-class attribute of the background, a token dictionary-based attention module (TDAB) is designed in the latent space. This module perceives the multi-class attribute of the background through an additional dictionary and further classifies the input features. On each class, a multi-head self-attention module is used to learn its similarity. Thus, the quality of background is effectively enhanced, and the detection accuracy is further improved. Finally, when the network converges, the detection results are obtained through reconstruction error. Through comparisons with eight algorithms on four public hyperspectral datasets, the experimental results show that the proposed SubNet outperforms the other algorithms in terms of the AUC_(D,F) metric and has better detection accuracy. The code of this work is released at: https://github.com/xautzhaozhe/SubNet.