EAFF-Net: Efficient attention feature fusion network for dual-modality pedestrian detection

The pedestrian detection network utilizing a combination of infrared and visible image pairs can improve detection accuracy by fusing their complementary information, especially in challenging illumination conditions. However, most existing dual-modality methods only focus on the effectiveness of feature maps between different modalities while neglecting the issue of redundant information in the modalities. This oversight often affects the detection performance in low illumination conditions. This paper proposes an efficient attention feature fusion network (EAFF-Net), which suppresses redundant information and enhances the fusion of features from dual-modality images. Firstly, we design a dual-backbone network based on CSPDarknet53 and combine with an efficient partial spatial pyramid pooling module (EPSPPM), improving the efficiency of feature extraction in different modalities. Secondly, a feature attention fusion module (FAFM) is built to adaptively weaken modal redundant information to improve the fusion effect of features. Finally, a deep attention pyramid module (DAPM) is proposed to cascade multi-scale feature information and obtain more detailed features of small targets. The effectiveness of EAFF-Net in pedestrian detection has been demonstrated through experiments conducted on two public datasets.