MFIFusion: An infrared and visible image enhanced fusion network based on multi-level feature injection

Infrared and visible image fusion aims to integrate complementary information from both types of images. Existing deep learning-based fusion methods rely solely on the final output of the feature extraction network, which may overlook valuable information presented in the middle layers of the network, ultimately reducing the richness of the fusion results, i.e., detailed texture information might not be fully extracted and integrated into the fused image. This study proposes a multi-level feature injection method based on an image decomposition model for infrared and visible image fusion, termed as MFIFusion. On the one hand, we introduce an attention-guided multi-level feature injection module designed to mitigate information loss during the feature extraction stage of the image scale decomposition process. More specifically, the proposed method integrates multiple fusion branches in the encoder network and employs an attention mechanism to guide the feature fusion process. On the other hand, based on the characteristics that superficial features retain image detail information and profound features are more suitable for extracting semantic information from images, we use distinct fusion strategies in these two phases to adaptively control the intensity distribution of the salient targets and to preserve the texture information in the background region. Qualitative and quantitative results demonstrate that our proposed approach produces fused images that are more visually salient to the target and contain richly detailed textures.