UAV-based RGB-Infrared Object Detection via Multi-Domain Feature Interaction

Unmanned aerial vehicles (UAVs) are playing an increasingly important role in the field of remote sensing due to their flexibility and efficiency. Meanwhile, UAV-based RGB-infrared image pairs possess stronger environmental adaptability and can provide rich object details, making them worthy of research. In this paper, we propose a multi-domain interaction detector (MDID) to detect various objects in extreme environments through the multi-domain effective integration of RGB-infrared images. The main feature interaction blocks consist of the dual wavelet enhancement block (DWEB) and the dual inception fusion block (DIFB). Specifically, DWEB can acquire valuable shared features between RGB-infrared image pairs in the frequency domain to mitigate modality feature conflicts during the feature extraction process. Then, DIFB can represent the object fusion information of varying scales and enhance the channel attention of the fused feature maps, which achieves the fusion of RGB-infrared image pairs in the spatial and channel domains. The experimental results with a challenging dataset confirm that our MDID performs better than the SOTA methods.