A Lightweight Infrared and Visible Image Fusion Based on Semantic Guidance and End-to-End Residual Network

This paper proposes a lightweight infrared and visible image fusion framework based on semantic guidance and an end-to-end residual network. To start with, on the original end-to-end fusion framework, by introducing the Segment Anything Model (SAM), a cross-modal semantic attention module is constructed to achieve feature alignment guided by semantic priors, effectively enhancing the fusion quality of key regions (such as human bodies and vehicles). In addition, a lightweight multiscale feature extraction network is designed. The structure of replacing ordinary convolutions with Depthwise Separable Convolutions (DSC) is adopted, which reduces the number of parameters while improving the ability to preserve edge details. Experiments show that the proposed method achieves good performance on the TNO dataset. It not only preserves the source details well but also integrates semantic information, and greatly improves the computational efficiency.