Visible–infrared remote sensing image fusion for object detection based on hybrid-level dual-input Mamba fusion

Existing visible–infrared fusion detectors often struggle to balance cross modal complementarity, redundancy, and conflict. Most multimodal Mamba methods pre-fuse the two modalities before state space modeling, causing complementary and conflicting cues to be jointly encoded and obscuring their sources and reliability during state evolution. To address this issue, we propose hybrid-level dual-input Mamba fusion framework (HDMF) for UAV based vehicle detection. At the feature level, dual-branch SimAM–wavelet attention (DSWA) identifies regions with significant cross modal differences and performs complementary or modality specific enhancement to suppress noise amplification and spurious responses. Binary fusion Mamba (BFMamba) then treats visible, infrared, and explicit interaction features as independent inputs to the state equations, enabling cross modal fusion to directly guide hidden state evolution while modeling long range dependencies with linear complexity. At the decision level, the tri-branch aware decision fusion head (TDFH) combines illumination priors and cross branch consistency to fuse predictions from visible, infrared, and fused branches, mitigating residual modal mismatch and prediction inconsistency. HDMF achieves COCO style mAP@0.5:0.95 scores of 56.62%, 69.82%, and 48.11% on DroneVehicle, LLVIP, and FLIR, respectively.