Cross-Modal Fusion and Detection Network for RGB-IR Vehicle

With the rapid development of UAV remote sensing technology, RGB-IR object detection has demonstrated significant value in the aviation field. However, existing fusion methods struggle to effectively address the imbalance between feature redundancy and complementarity across modalities, while their detection performance remains limited in complex aerial scenarios. There is an urgent need to develop more robust multi-modal fusion detection algorithms. Therefore, we proposed a Cross-Modal Fusion and Detection Network (CMFD-Net) for RGB-IR vehicle detection. First, a Complementary-Shared Feature Extraction (CSFE) module is designed to dynamically suppress redundant information through a cross-modal channel calibration mechanism, while extracting both modality-specific and shared features via a dual-path architecture. Second, a Cross-Modal Feature Refinement (CMFR) Network is introduced, leveraging a spatial attention-guided multi-scale fusion strategy to adaptively integrate edge-contour features from the infrared modality and texture-detail features from the RGB modality. Furthermore, a mutual information loss is incorporated to enable synergistic optimization between the fusion strategy and detection task. Experiments on the DroneVehicle dataset show that CMFDNet achieves a remarkable mAP of 75.3%, significantly outperforming 6 comparative methods.