MDAFusion: Multi-scale Dual Attention Feature Fusion for Multispectral Object Detection

In fields like autonomous driving and security surveillance, multispectral images provide a richer and more comprehensive representation for object detection tasks compared to single-spectrum images. This advantage helps mitigate detection accuracy degradation in complex environments, thereby improving the reliability and robustness of object detection networks in real-world applications. However, efficiently fusing different spectral information remains a challenge in multispectral object detection. To address this, we propose the Multi-scale Dual Attention Feature Fusion Network (MDAFusion), which aggregates multi-scale information from multispectral images and optimizes the fusion of visible and infrared features through a dual attention mechanism. The channel spatial attention mechanism focuses on key features and their precise locations, while the Transformer-based self-attention mechanism captures long-range dependencies in multispectral feature, creating a global receptive field that compensates for the limitations of traditional convolutional neural networks. Extensive experiments on the FLIR, LLVIP, and M3FD datasets demonstrate that MDAFusion outperforms existing methods, achieving state-of-the-art performance in multispectral object detection.