DBCM: a dual-branch cross-modal fusion object detection network

Visible and infrared (IR) multimodal fusion object detection can leverage the complementary advantages of the two modalities. However, although visible images contain rich texture details, they are prone to redundant background interference. IR images can provide target thermal radiation information, but they face issues of thermal noise and insufficient texture details, with the difficulty of small target detection further exacerbated. To fully exploit the complementary advantages of IR and visible images and improve the performance of object detection after multimodal fusion, We propose DBCM, a dual-branch cross-modal collaborative detection network for RGB-IR small object detection. On the basis of existing mature technologies, customized optimizations are performed according to the specific requirements of RGB-IR detection. This framework adopts a two-stage strategy of modal purification prior to bidirectional fusion. Specifically, the HSF-PC module firstly suppresses redundant information and enhances critical features in the visible light branch. Subsequently, the hierarchical dilated context enhancement module dynamically compensates texture details and restrains thermal noise for features of the IR branch. After feature enhancement of the two individual branches, the adaptive bidirectional feature fusion network module realizes cross-level feature interaction and fusion, strengthens the transmission of multi-scale features, and effectively boosts object detection performance. Experimental results on M3FD and FLIR datasets demonstrate that the proposed framework achieves substantial improvements in detection accuracy and robustness, offering a feasible solution for multimodal object detection tasks.