GLMFuse: Multi-Scale Image Fusion Based on Multimodal Features

In the process of infrared and visible light image fusion, a multi-scale image fusion method based on multimodal features (GLMFuse) is proposed to address the problems of reduced fusion information fidelity and blurred structural information. First, a cross-spatial learning approach is designed to obtain multi-scale images; second, a fusion block that preserves context-aware features is designed to obtain holographic feature images; and finally, a lightweight decoder is designed to reconstruct images. Conducting fusion framework evaluation experiments on the public domain datasets TNO and Roadscene, the experiments show that objectively, EN, MI, SF, SD, and PSNR are improved by an average of 6.7% and 3.4% compared to the sub-optimal models. Subjectively, the fused image is rich in detail texture, with better fusion performance and stronger generalization.