Group Sparsity Mixture Model and Its Application on Image Denoising

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Prior learning is crucial for image processing tasks.
• Existing methods struggle to effectively model image patch group priors.

Purpose of the Study:

• To propose a novel prior model for image patch groups.
• To develop an efficient framework for image denoising using learned priors.

Main Methods:

• Introduced the Group Sparsity Mixture Model (GSMM) using bilateral matrix multiplication.
• Developed a plug-and-play framework for patch group-based image denoising.
• Implemented two GSMM-based denoising methods.

Main Results:

• GSMM effectively models local patch features and non-local patch relationships, capturing inherent sparsity.
• GSMM-based methods outperform competing models like FoE and GMM.
• The improved GSMM method achieves state-of-the-art performance comparable to WNNM, with an 8x speedup.

Conclusions:

• The Group Sparsity Mixture Model (GSMM) offers a powerful approach for learning image patch priors.
• The proposed denoising framework is flexible and enhances image denoising efficiency and effectiveness.
• GSMM-based methods represent a significant advancement in image denoising technology.