MsRAN: a multi-scale residual attention network for multi-model image fusion.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Image fusion is crucial for image processing tasks.
• Deep learning methods are increasingly used for information fusion.
• Existing methods struggle to highlight salient regions and retain sufficient information.

Purpose of the Study:

• To propose a multi-scale residual attention network (MsRAN) for enhanced image feature exploitation.
• To address limitations in highlighting typical regions and retaining useful information in source images.

Main Methods:

• Developed a generator network with two information refinement networks and one information integration network.
• Information refinement networks extract features at different scales using varying convolution kernel sizes.
• Information integration network incorporates merging and attention blocks to focus on salient regions and prevent information underutilization.

Main Results:

• The proposed MsRAN method demonstrates superior visual results compared to existing techniques.
• Experiments show that MsRAN retains more detailed information from multi-modal source images.
• Qualitative and quantitative evaluations on public datasets validate the method's effectiveness.

Conclusions:

• MsRAN effectively exploits image features for improved image fusion.
• The network successfully highlights salient regions and preserves detailed information.
• The dual adversarial structure and information loss function enhance detail capture during training.