Answers are grounded in this paper and cite their sources.
Multi-modal image fusion (MMIF) integrates information from multiple imaging modalities to produce images with comprehensive information and detailed textures. In recent years, deep learning-based MMIF methods have made significant advancements. However, existing methods frequently encounter high-frequency detail loss and insufficient texture representation, resulting in inadequate scene detail representation. Moreover, current image fusion methods lack sufficient interaction to balance complementary information, leading to the loss of modality-specific information in the fused results. To overcome these challenges, we propose an invertible feature enhancement and cross-modal information balance network called ICFusion, which aims to enhance texture details and strengthen the interaction between multi-modal features. Specifically, we present the invertible Mamba module (IMM), which leverages an invertible structure for feature lossless transmission and extracts global multi-modal features to inject long-range pixel dependencies, thereby enhancing the representation of high-frequency details in global information modeling. Additionally, we introduce the cross-modal feature balance module (CFBM), which maps cross-modal features into space state models (SSMs) for interaction and adaptively fuses the interacted features with individual modality features to enhance the unique information of each modality, achieving the balance of complementary information across modalities. Extensive experiments on infrared-visible image fusion and medical image fusion tasks show that our proposed method yields remarkable results. Furthermore, we demonstrate that ICFusion can enhance the performance of downstream semantic segmentation in a unified benchmark.
No takes yet. Share an insight, caveat, or question.
Xu et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: