EqMambaFuse: An Equivariance-Constrained Mamba Network for Infrared and Visible Image Fusion

Abstract

Infrared and visible image fusion aims to integrate thermal target responses from infrared images with texture and structural information from visible images, and is important for nighttime surveillance, object perception, and intelligent vision systems. Existing methods often suffer from weakened infrared targets, loss of visible details, and insufficient global structural modeling in complex scenes. To address these issues, this article proposes an equivariance-constrained Mamba network for infrared and visible image fusion, termed EqMambaFuse. The proposed method adopts a dual-branch lightweight state-space encoding architecture to extract multi-level infrared and visible features separately, thereby enhancing long-range spatial dependency modeling while maintaining low computational complexity. To improve cross-modal information selection, CBAM attention and quality-aware adaptive fusion modules are introduced at each level, enabling the network to dynamically assign infrared and visible feature weights according to regional information characteristics. Residual dense blocks are further employed to strengthen the propagation of shallow textures and edge information, while a lightweight Transformer-based global context module is applied to the highest-level fused feature to supplement semantic correlations. Multi-level feature aggregation is then used for image reconstruction. To correct local modality contribution deviations in the initial reconstruction, an improved EM-inspired dynamic weight refinement strategy is designed, which iteratively optimizes infrared and visible information using pixel discrepancies and edge priors. During training, a joint constraint consisting of basic fusion loss, perceptual imaging loss, equivariance consistency loss, and mutual information, spatial correlation difference, and Qabf surrogate losses is constructed to guide optimization from the perspectives of information preservation, spatial structural consistency, and edge quality. Experimental results demonstrate that EqMambaFuse effectively enhances infrared salient targets while preserving visible background textures and structural details, producing fused images with improved target clarity, edge preservation, structural consistency, and visual naturalness.