MVF: A Novel Infrared and Visible Image Fusion Approach Based on the Morphing Convolutional Structure and the Light-Weight Visual State Space Block

Image/feature decomposition plays a crucial role in image fusion tasks. However, in deep learning-based fusion methods, existing decomposition operations tend to be either overly simplistic or excessively intricate to learn effectively. In this paper, a novel Fusion network (MVF) based on the Morphing convolutional (MConv) structure and the light-weight Visual state space block (LWVB) is proposed, which is used to extract richer deep features and enhance the quality of the fused images. The proposed structure incorporates convolutional kernels of different shapes to extract detail features from various directions (horizontal, vertical, and diagonal). These kernels consist of fixed shapes with learnable values, ensuring a simplified learning process while retaining the advantages of non-deep-learning filters. Moreover, to efficiently capture global spatial information without incurring significant storage overhead, we modify the Visual State Space (VSS) Block of VMamba and propose the light-weight VSS block which can compress and extract global information at both high and low frequencies, thereby enhancing the spatial awareness capabilities of proposed module. Comparison experiments show that the proposed fusion framework, utilizing the morphing convolutional structure and LWVB, achieves superior fusion performance compared to the state-of-the-art fusion methods.