Abstract: To address the partial loss of depth details in traditional infrared and visible image fusion methods, this study proposes a framework based on a multi-scale convolution squeeze–excitation fusion network for infrared and visible image fusion. First, a squeeze-and-excitation residual dense block was constructed to extract the deep features of the source images, which can fully utilize the reuse of features in each convolutional layer. Subsequently, a dual-channel multi-scale convolution attention fusion network was designed in the fusion layer to fuse the extracted features of the source images. The corresponding loss functions were designed based on the characteristics of this fusion network. Finally, a decoder network that matched the encoder was designed to achieve feature reconstruction. Experiments using public datasets revealed that the proposed fusion network had a better comprehensive performance in subjective and objective evaluations and preserved more deep-detail features compared to typical relevant methods from recent years.