DCSAFNet: Dual-Channel and Spatial Attention Fusion Network for Multispectral Object Detection in Low-Light Conditions

In recent years, object detection technology has made significant progress, yet it still faces substantial challenges in low-light conditions. Existing methods often address this issue by introducing thermal cameras to assist RGB cameras. However, most of these methods primarily focus on the shared features across different modalities, often neglecting the diversity of modality-specific characteristics. To address this limitation, we propose a novel Dual-Channel and Spatial Attention Fusion Module (DCSAFM), which constructs unique attention feature maps for each modality along the channel dimension and achieves deep fusion through self-attention in the spatial dimension. This module simultaneously benefits from the distinctive information within each modality and their complementary relationships, thereby significantly enhancing the accuracy and robustness of multispectral object detection. By embedding DCSAFM at multiple feature levels, we propose a multispectral object detection network named Dual-Channel and Spatial Attention Fusion Network (DCSAFNet), which achieves state-of-the-art performance on the FILR dataset with an mAP of 83.7%. Moreover, extensive experiments on the KAIST and LLVIP datasets demonstrate the strong generalization capability of our method, achieving a miss rate of 7.7% and an mAP of 97.7%, respectively. The code is available at https://github.com/wliu318/DCSAFNet1 .