CCSFuse: Collaborative Compensation and Selective Fusion for UAV-Based RGB-IR Object Detection

Visible-Infrared (RGB-IR) object detection plays a crucial role in UAV-based vision tasks. However, existing methods still suffer from learning bias caused by imbalanced information distribution and inaccurate fusion due to modal conflicts. Inspired by the human multisensory information processing mechanism, we propose a novel “CompensationFusion” progressive detection framework, CCSFuse, to fully exploit the complementary relationship between modalities while eliminating conflict interference. Specifically, we design a cross-modal feature compensation module, which establishes inter-modal information interaction to achieve mutual complementarity and enhancement during feature extraction, effectively mitigating the issue of imbalanced modal information distribution. Additionally, we introduce an adaptive feature-selection fusion module to address modal conflicts. We employ a cross-modal channel attention to calibrate channel features of different modalities and utilizes a selective fusion strategy to dynamically assess modal importance, thereby achieving adaptive modal fusion. Finally, we validate the effectiveness of CCSFuse on the DroneVehicle and LLVIP datasets. The results confirm that CCSFuse significantly improves the efficiency of feature optimization and integration. In UAV-based object detection scenarios, CCSFuse outperforms state-of-the-art methods in both qualitative and quantitative comparisons, particularly for small objects and low-quality modalities. The code is available at https://github.com/ZhangT-xxl/CCSFuse.