EvaNet: Toward More Efficient and Consistent Infrared and Visible Image Fusion Assessment.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Existing image fusion evaluation metrics are often adapted from other vision tasks, lacking specificity and computational efficiency.
• Traditional metrics struggle to accurately capture the true quality of fused images due to complex transformations.

Purpose of the Study:

• To develop a unified, efficient, and accurate evaluation framework tailored for image fusion.
• To address the limitations of traditional metrics in assessing fusion quality and computational cost.

Main Methods:

• Propose a lightweight neural network to approximate existing image fusion metrics efficiently.
• Implement a divide-and-conquer strategy by decomposing fused images into infrared and visible components for evaluation.
• Utilize contrastive learning and large language model-based perceptual scene assessment during model training.
• Introduce a consistency evaluation framework to align fusion metrics with human visual perception.

Main Results:

• The proposed learning-based evaluation paradigm demonstrates superior efficiency, achieving up to 1,000 times speedup.
• The framework shows greater consistency across various standard image fusion benchmarks.
• The method effectively disentangles the evaluation process by assessing information preservation in separated image components.

Conclusions:

• The developed unified evaluation framework offers a significant advancement in image fusion research.
• The approach provides a more accurate, efficient, and human-aligned method for evaluating image fusion quality.
• This work paves the way for more reliable benchmarking and development in the field of image fusion.