Guided Deep Generative Model-Based Spatial Regularization for Multiband Imaging Inverse Problems.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Solving inverse problems in multiband imaging necessitates spatial and spectral regularizations.
• Existing methods often rely on direct spectral information extraction and conventional spatial penalizations like total variation.
• These conventional methods may not fully capture the complex spatial features required for optimal image reconstruction.

Purpose of the Study:

• To propose a generic framework for deriving data-driven spatial regularizations in multiband imaging.
• To leverage auxiliary high-resolution acquisitions for improved spatial regularization.
• To demonstrate the framework's versatility in multiband image fusion and inpainting.

Main Methods:

• A model-based formulation for inverse problems in multiband imaging.
• Utilizing deep learning, specifically a deep generative network, to encode spatial semantic features from auxiliary high-resolution images.
• Instantiating the framework for multiband image fusion and multiband image inpainting tasks.

Main Results:

• The proposed framework successfully derives tailored, data-driven spatial regularizations.
• Experimental results show significant benefits of these informed regularizations compared to conventional methods.
• The approach demonstrated effectiveness in both multiband image fusion and inpainting.

Conclusions:

• The developed framework offers a powerful method for creating data-driven spatial regularizations in multiband imaging.
• Deep learning enables the extraction of high-level spatial semantic features for enhanced image reconstruction.
• This approach represents a significant advancement over traditional regularization techniques in specific multiband imaging applications.