SDRFPT-Net: A Spectral Dual-Stream Recursive Fusion Network for Multispectral Object Detection

Multispectral object detection combining visible and infrared imaging has emerged as a crucial technology for all-day and all-weather surveillance systems. However, effectively integrating complementary information from different spectral domains remains challenging. This paper proposes SDRFPT-Net (Spectral Dual-stream Recursive Fusion Perception Target Network), a novel architecture for multispectral object detection that addresses these challenges through three innovative modules. First, we introduce a Spectral Hierarchical Perception Architecture (SHPA) based on YOLOv10, which employs a dual-stream structure to extract domain-specific features from visible and infrared modalities. Second, a Spectral Recursive Fusion Module (SRFM) facilitates deep cross-modal feature interaction through a hybrid attention mechanism that integrates self-attention, cross-modal attention, and channel attention, coupled with a parameter-efficient recursive progressive fusion strategy. Third, a Spectral Target Perception Enhancement Module (STPEM) improves target region representation and suppresses background interference using lightweight mask prediction. Extensive experiments on the FLIR-aligned and LLVIP datasets demonstrate SDRFPT-Net's superior performance, achieving state-of-the-art results with 0.785 mAP50 and 0.426 mAP50:95 on FLIR-aligned, and 0.963 mAP50 and 0.706 mAP50:95 on LLVIP. Comprehensive ablation studies validate the effectiveness of each proposed component. The findings suggest that SDRFPT-Net offers a promising solution for reliable multispectral object detection in challenging environments, making it valuable for applications in autonomous driving, security surveillance, and remote sensing.