Graph-Based Non-Convex Low-Rank Regularization for Image Compression Artifact Reduction.

Area of Science:

• Image processing
• Computer vision
• Signal processing

Background:

• Block transform coding, like Discrete Cosine Transform (DCT), often introduces artifacts in compressed images at low bitrates due to independent coefficient quantization.
• Image prior models are crucial for effective compressed image reconstruction.
• Natural image patches exhibit sub-manifold structures in high-dimensional spaces, offering valuable prior information.

Purpose of the Study:

• To develop an advanced image reconstruction method that mitigates artifacts in block transform coded images.
• To leverage the sub-manifold structure of natural image patches as a prior for improved reconstruction.
• To enhance both objective and perceptual quality of reconstructed images.

Main Methods:

• Characterizing sub-manifold structure at the patch level using graph Laplacian regularization.
• Estimating patch distributions by exploiting similar patches as samples.
• Employing graph-domain distance for patch similarity measurement, instead of Euclidean distance.
• Applying low-rank regularization on similar-patch groups with a non-convex lp penalty.
• Utilizing an alternating minimization strategy to solve the non-convex optimization problem.

Main Results:

• The proposed method demonstrates superior reconstruction accuracy compared to existing state-of-the-art techniques.
• Both objective image quality metrics and perceptual assessments confirm the method's effectiveness.
• The graph-based approach successfully models patch distributions and reduces compression artifacts.

Conclusions:

• The developed technique effectively reconstructs compressed images by exploiting intrinsic image patch structures.
• Graph Laplacian regularization and graph-domain distance provide a robust way to model image priors.
• The method offers significant improvements in image reconstruction quality, outperforming current approaches.