Cross-Modal Image Fusion via Structure Preservation and Detail Enhancement Optimization

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Existing infrared-visible fusion methods often suffer from edge blurring, texture smoothing, and reduced target-to-background contrast.
• These limitations hinder the effective utilization of fused images in various applications.

Purpose of the Study:

• To propose a novel Structure-Detail Constrained Fusion (SDC-Fusion) method to overcome the limitations of existing fusion techniques.
• To enhance infrared target boundaries and visible textures while preserving background structure and luminance.

Main Methods:

• A frequency-decoupled framework utilizing the Haar wavelet transform to separate images into structural (low-frequency) and detail (high-frequency) components.
• A gated module applying Rectified Flow specifically to high-frequency wavelet subbands for enhanced target guidance and texture preservation.
• Integration of adaptive weighting, Mamba scanning, and depthwise convolutions in the low-frequency branch to maintain global and local image characteristics.

Main Results:

• SDC-Fusion achieved state-of-the-art performance, ranking first in multiple metrics (SSIM, VIF, Q_abf, SF, PSNR) on both MSRS and M^3FD datasets.
• Significant improvements were observed, with the largest gains reaching 10.44% in SF on MSRS and 5.72% in VIF on M^3FD compared to leading methods.
• The proposed model is computationally efficient with 0.535 M parameters and 67.5 G FLOPs.

Conclusions:

• SDC-Fusion effectively addresses the challenges of edge blurring and texture smoothing in infrared-visible image fusion.
• The frequency-decoupled approach with targeted Rectified Flow application offers superior preservation of structural and detailed information.
• The method demonstrates robust performance and significant advantages over existing fusion techniques.