Dual-modality adaptive feature fusion for object detection guided by illumination awareness

To address the significant differences in representational capabilities between infrared and visible images under varying illumination conditions, and the limitations of existing fusion strategies in fully exploiting the complementary advantages of dual modalities, this paper proposes an adaptive dual-modality feature fusion method guided by illumination awareness. Specifically, we design a dual-stream feature extraction network to separately extract features from infrared and visible images. An Illumination-Aware Guidance Module (IAGM) is embedded into the visible branch to perceive the illumination level of the input scene. The predicted illumination value is then used to guide the Adaptive Feature Fusion Module (AFFM), which dynamically adjusts channel-wise weights between the two modalities. Experiments conducted on the dual-modality DroneVehicle dataset demonstrate that the proposed method outperforms seven state-of-the-art deep learning-based detectors. Furthermore, transfer experiments on the LLVIP pedestrian detection dataset confirm the strong cross-scene generalization capability of our approach.