Low-Rankness Guided Group Sparse Representation for Image Restoration

Area of Science:

• Computer Vision
• Image Processing
• Signal Processing

Background:

• Group Sparse Representation (GSR) is a powerful nonlocal image model for restoration.
• Existing GSR methods leverage Nonlocal Self-Similarity (NSS) but neglect other priors like low-rankness (LR).
• Ignoring LR can degrade image restoration quality by failing to preserve structural information.

Purpose of the Study:

• To propose a novel Low-Rankness guided Group Sparse Representation (LGSR) model.
• To effectively integrate both sparsity and low-rankness priors within a unified framework for image restoration.
• To enhance the preservation of both texture and structure information in natural images.

Main Methods:

• Developed the LGSR model that jointly utilizes sparsity and LR priors for image patches.
• Employed an alternating minimization algorithm to solve the proposed LGSR model.
• Incorporated an adaptive parameter adjustment scheme within the optimization process.

Main Results:

• LGSR demonstrated superior performance across various image restoration tasks.
• Achieved state-of-the-art results in image denoising, inpainting, and compressive sensing (CS).
• The joint use of sparsity and LR priors proved effective in preserving image details.

Conclusions:

• The proposed LGSR model offers a significant advancement in image restoration.
• Combining complementary priors (sparsity and LR) enhances the preservation of image texture and structure.
• LGSR provides a robust and effective framework for diverse image restoration applications.