High-quality Image Restoration Using Low-Rank Patch Regularization and Global Structure Sparsity.

Area of Science:

• Computer Vision
• Image Processing
• Signal Processing

Background:

• Image restoration techniques often rely on either nonlocal self similarity or global structure regularization.
• Nonlocal self similarity leverages repeating image patches for reconstruction.
• Global structure regularization enforces sparsity in image structure to minimize artifacts.

Purpose of the Study:

• To develop a novel image restoration method integrating both nonlocal self similarity and global structure sparsity.
• To improve the efficiency and performance of image restoration algorithms.

Main Methods:

• A combined model utilizing nonlocal self similarity and global structure sparsity.
• Simultaneous reconstruction of similar patch groups using adaptive regularization (weighted nuclear norm).
• Image decomposition into smooth and sparse components, with the residual regularized by l1 norm.
• Optimization via the Alternating Direction Method of Multipliers (ADMM) algorithm.

Main Results:

• The proposed method effectively combines nonlocal self similarity and global structure sparsity.
• Demonstrated superior performance in image completion and super-resolution tasks compared to state-of-the-art methods.
• Achieved high-quality image recovery across various corruption types and levels.

Conclusions:

• The novel approach offers significant improvements in image restoration by synergistically employing two powerful priors.
• The method provides an efficient and effective solution for complex image restoration challenges.
• This work advances the field of image processing by presenting a unified framework for restoration.