DSAFusion: Detail-semantic-aware network for infrared and low-light visible image fusion

It is important to simultaneously preserve detail and semantic information in both infrared and visible (especially low-light) images for the pursuit of high-quality fusion maps. Unfortunately, the existing fusion methods fails to balance them, resulting in the fusion results are over-smoothed, low-contrast, and sensitive to application scenarios. To address these problems, this paper develops a detail-semantic-aware network for low-light infared and visible image fusion, termed as DSAFusion. Our DSAFusion mainly includes the following key parts: Firstly, a dual-branch encoder is employed to extract the detail and semantic features in infrared and visible images. Then, the features from the two typical modes are respectively concatenated and fused by detail and semantic information fusion networks (respectively named as DFNet and SFNet). Finally, the fused features contribute to reconstruct the final fusion map by a decoder to decode them. Experimental results in both quantitation and qualification show that our DSAFusion is effective and performs better than the existing SOTA fusion methods on the preservation of textures and semantic information.