A novel low-rank blind spot reconstruction network based on multi-scale mask convolution for hyperspectral anomaly detection

With the emergence of blind spot reconstruction networks in hyperspectral anomaly detection, a series of new challenges have arisen. The most significant issues include local anomaly identity mapping and inaccurate detection of large anomaly targets in deep learning networks. Therefore, we propose a busy point reconstruction network for hyperspectral anomaly detection based on multi-scale mask convolution. We creatively design a multi-scale mask convolution module that uses an inner window mask to isolate the center block and edge background for reconstruction. To further enhance space-spectral information interactivity, we introduce a joint space-spectral module. Additionally, to increase the network's focus on global low-rank backgrounds, we incorporate a block low-rank penalty to guide global low-rank attention. Finally, by considering the scale difference in detection performance between different mask convolutions, we develop a dynamic learnable fusion module that performs adaptive fusion of two mask convolution results for optimal detection performance on two real data sets.