Answers are grounded in this paper and cite their sources.
Deep learning-based hyperspectral anomaly detection (HAD) has drawn increasing interest owing to its competitive performance. However, most existing methods struggle to achieve a favorable balance between false-alarm suppression and anomaly identification, which limits their overall performance. To this end, we develop a novel dual-domain background modeling network (DDBMNet) with anomaly suppression for HAD, which extracts complementary frequency- and spatial-domain background features while suppressing anomaly reconstruction to enhance detection performance. Specifically, a corner-visible frequency attention (CVFA) module is developed to model frequency-domain background characteristics, where a corner-visible convolutional (CVC) block is integrated to suppress anomaly identity mapping by aligning potential anomalies with the background distribution. To capture spatial-domain background features, we develop an attention-aware spatial Mamba (AASM) module. With an attention-aware scanning (AAS) block it can further weaken the learning of anomaly features by adaptively assigning lower weights to anomaly features during the scanning process. Moreover, a frequency-spatial interaction fusion (FSIF) module is devised to enhance background representations via a frequency-spatial inter-attention mechanism that exploits complementary information in frequency- and spatial-domains. Comprehensive experiments on six hyperspectral datasets demonstrate the effectiveness of our proposed modules and verify the overall superiority of DDBMNet over state-of-the-art HAD approaches.
No takes yet. Share an insight, caveat, or question.
Huo et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: