DSPFusion: Image Fusion via Degradation and Semantic Dual-Prior Guidance

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Existing infrared-visible image fusion methods often fail with degraded images.
• Current degradation-aware methods typically address specific degradations or require text prompts, limiting practical use.

Purpose of the Study:

• To develop a unified framework for infrared-visible image fusion that handles diverse image degradations without auxiliary prompts.
• To improve the practicality and efficiency of image fusion in real-world scenarios.

Main Methods:

• Introduced DSPFusion, a dual-guided framework utilizing degradation and semantic priors.
• Employed a semantic prior diffusion model for high-quality semantic restoration in a latent space.
• Developed an enhancement and fusion network guided by restored semantic and degradation priors.

Main Results:

• DSPFusion effectively handles diverse image degradations while preserving complementary information.
• Achieved over 30x inference speedup compared to mainstream diffusion model-based fusion schemes.
• Demonstrated competitive performance and broadened the scope of practical image fusion applications.

Conclusions:

• DSPFusion offers a robust and efficient solution for degraded infrared-visible image fusion.
• The framework's ability to jointly restore and fuse images without prompts enhances its applicability.
• The method provides a significant advancement in automatic image fusion technology.