An innovative optimization strategy based on Mamba and generative adversarial networks for efficient and high-performance multimodal image fusion

Multimodal image fusion (MIF) integrates multisource data into a single high-quality image with minimal redundancy. While deep learning has advanced MIF by improving fusion quality, convolutional neural networks (CNNs) struggle with long-range dependencies, and Transformers incur high computational costs. Additionally, preserving fine textures, suppressing noise, and achieving high efficiency remain challenges, particularly for infrared and visible image fusion (IVIF). This paper proposes MMGFuse, a novel MIF framework based on a Multi-Parallel Vision Mamba Generative Adversarial Network. MMGFuse leverages the Mamba model's efficiency and generative adversarial networks' realism, introducing a residual parallel vision Mamba (ResPViM4) module to enhance texture and detail preservation and a multi-parallel vision Mamba (MPViM) module to capture both global and local features across scales. A dual-modality image discriminator further optimizes visual quality. Experiments show that MMGFuse outperforms state-of-the-art methods in subjective visual quality and objective metrics for IVIF and medical image fusion, demonstrating its effectiveness, efficiency, and broad applicability in advancing image fusion. The codes are available at https://github.com/sunyichen1994/MMGFuse.