Denoiser Learning for Infrared and Visible Image Fusion

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Infrared (IR) and Visible Image (VI) fusion enhances information content and visual quality.
• Current fusion methods rely on manual operators (e.g., intensity, gradient), limiting complete information extraction.
• Existing techniques struggle to accurately describe and fuse information, hindering overall performance.

Purpose of the Study:

• To propose a novel information measurement method for improved IR and VI image fusion.
• To guide a generator network using denoisers for more accurate feature representation.
• To develop a semantic adaptive loss function for adaptive fusion of semantic information.

Main Methods:

• A generator network is guided by learning from denoisers that restore images corrupted by noise.
• A mutual competition between denoisers helps the generator explore source image data specificity.
• A semantic adaptive measurement loss function is introduced to fuse semantic information based on density.

Main Results:

• The proposed method demonstrates superior information fusion quality.
• Experimental results show a faster fusion speed compared to advanced methods.
• Quantitative and qualitative evaluations on three public datasets validate the effectiveness.

Conclusions:

• The novel denoiser-guided approach significantly enhances IR and VI image fusion.
• The semantic adaptive loss function improves the fusion of complex semantic details.
• The method offers a promising advancement in high-quality, efficient image fusion techniques.