Disentangle to Fuse: Toward Content Preservation and Cross-Modality Consistency for Multi-Modality Image Fusion.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Multi-modal image fusion (MMIF) integrates data from different sensors.
• Cross-modality discrepancies often degrade fused image quality and semantic representation.
• Existing MMIF methods struggle with preserving content and ensuring consistency.

Purpose of the Study:

• To propose C²MFuse, a novel framework for MMIF that preserves content and ensures cross-modality consistency.
• To explicitly disentangle style and content representations across modalities for superior image fusion.
• To enhance semantic consistency and reduce distributional uncertainty in fused images.

Main Methods:

• C²MFuse employs a content-preserving style normalization to suppress modality-specific variations.
• Features are progressively aggregated to enhance fine-grained details and content completeness.
• A novel adaptive consistency loss with learnable transformation enforces global consistency across heterogeneous inputs.

Main Results:

• C²MFuse achieves efficient and high-quality fusion across five datasets and three MMIF tasks.
• The framework surpasses existing state-of-the-art methods in fusion performance.
• Experimental results demonstrate effective generalization to downstream visual applications.

Conclusions:

• C²MFuse offers a significant advancement in multi-modal image fusion by preserving content and ensuring cross-modality consistency.
• The proposed method effectively handles cross-modality discrepancies, leading to improved semantic representation.
• C²MFuse provides a robust and generalizable solution for various MMIF applications.