Abstract: To address background blurring, insufficient feature extraction, and image distortion in image fusion, this paper proposes a multi-scale infrared and visible image fusion algorithm. The gradient operator is integrated into the residual feature extraction module to effectively extract edge and detail information by capturing gradient changes in different directions. The residual connection mechanism enhances local texture extraction and mitigates gradient disappearance while retaining underlying information. The multi-scale feature extraction module combines convolutional kernels and receptive fields at different scales to achieve multi-scale perception and richer semantic features. To ensure structural similarity of fused images, the Ghostconv module and CBAM attention mechanism are introduced for efficient feature optimization, enabling the model to focus on key features while reducing parameters. Evaluations on the TNO and MSRS datasets show that the fused images exhibit clearer textures and sharper contours, with significant visual quality advantages.