Self-Adaptive Background-Attentive Network With Anomaly Suppression for Hyperspectral Anomaly Detection

In recent years, deep learning has emerged as one of the prominent technologies in hyperspectral anomaly detection (HAD) due to its powerful feature extraction capability. Deep learning-based methods aim to reconstruct the background image and use the reconstruction error as a criterion for anomaly detection. However, existing methods rarely exploit the spatial and spectral differences between background and anomalies to suppress anomaly reconstruction. To overcome this limitation, we propose a self-adaptive background-attentive network with anomaly suppression (BASNet). Specifically, a pyramid similarity downsampling module (PSDM) is employed to reduce the proportion of anomalies in the input image and improve multiscale background representation. To enhance the network's attention to background information, a multiscale attention module (MAM) is designed, and an inverse attention strategy is adopted to guide the network to focus on background regions. In addition, a depthwise separable blind-block convolution module (DBCM) is constructed for pixel reconstruction. This module applies weighted multiscale blind-block convolutions according to the anomaly score of each pixel, allowing the network to effectively handle anomalies at different scales. Due to the stochastic nature of deep learning, the optimal stopping epoch for network training is not fixed. Therefore, reconstructed image difference (RID) is developed to dynamically determine the stopping epoch. Extensive experiments on eight datasets demonstrate that the proposed BASNet achieves satisfactory detection performance.