Different Input Resolutions and Arbitrary Output Resolution: A Meta Learning-Based Deep Framework for Infrared and...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Infrared and visible image fusion is crucial for various vision applications.
• Existing fusion methods face limitations with input/output spatial resolutions, hindering practical use.

Purpose of the Study:

• To propose a flexible meta-learning deep framework for infrared and visible image fusion.
• To overcome resolution limitations of current fusion techniques.

Main Methods:

• A meta-learning deep framework utilizing convolutional networks for feature extraction.
• A meta-upscale module for arbitrary resolution upscaling.
• A dual attention mechanism for feature fusion.
• A residual compensation module for enhanced detail extraction.
• Multi-task learning loss function for simultaneous fusion and super-resolution.
• A novel contrast loss for improved fused image contrast.

Main Results:

• The proposed framework successfully fuses infrared and visible images with different resolutions.
• It generates fused images at arbitrary resolutions using a single learned model.
• Experiments demonstrate superior effectiveness and performance compared to existing methods.

Conclusions:

• The meta-learning framework offers a flexible and effective solution for infrared and visible image fusion.
• The method addresses the critical challenge of varying spatial resolutions in image fusion.
• The approach shows significant potential for practical vision-based applications.