LR-SMAN: Low-Rank Symmetrical Mutual Attention Network for multispectral object detection

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Multispectral object detection, especially via fusing RGB and infrared (IR) imagery, has attracted growing attention due to its robustness under challenging conditions such as low illumination, nighttime, and adverse weather. Nevertheless, effectively integrating these heterogeneous modalities remains challenging. Existing attention-based fusion methods often adopt asymmetric cross-attention designs and learn cross-modal affinities in a purely data-driven manner without explicit structural constraints, which may yield suboptimal fusion, particularly in complex scenes. To address these issues, we propose a Low-Rank Symmetrical Mutual Attention Network (LR-SMAN) for multispectral object detection. The core of LR-SMAN is twofold: (i) a Symmetrical Mutual Attention (SMA) mechanism that constructs a shared consensus topology and enables balanced, bidirectional cross-modal interactions; and (ii) Tensor Nuclear Norm Regularization (TNNR) imposed on the paired modality-specific affinity tensor to explicitly promote low-rank structure. The proposed regularization suppresses noise and redundancy, encouraging compact, robust, and discriminative shared semantic representations. Extensive experiments show that LR-SMAN achieves competitive performance on several widely used multispectral benchmarks, including VEDAI, M 3 FD, and LLVIP, with particularly notable improvements in challenging settings involving low illumination, small objects, and cluttered backgrounds.