Random Walk Graph Laplacian-Based Smoothness Prior for Soft Decoding of JPEG Images

Area of Science:

• Digital Image Processing
• Signal Processing
• Computer Vision

Background:

• Joint Photographic Experts Group (JPEG) compression is widely used, necessitating effective image reconstruction methods.
• Traditional hard decoding in JPEG reconstruction is limited; soft decoding offers better performance by utilizing signal priors.
• Defining and applying suitable priors remains a key challenge in soft decoding.

Purpose of the Study:

• To develop an efficient JPEG soft decoding algorithm that optimizes image reconstruction.
• To integrate multiple image priors for improved accuracy and artifact reduction.
• To introduce and validate a novel graph-signal smoothness prior (LERaG).

Main Methods:

• Combined Laplacian prior for DCT coefficients, sparsity prior, and a new LERaG graph-signal smoothness prior.
• Utilized Laplacian prior for initial minimum mean square error solutions.
• Developed LERaG prior based on random walk graph Laplacian eigenvectors for improved high-frequency recovery.

Main Results:

• The proposed soft decoding algorithm integrates three signal priors with optimized weights.
• LERaG prior demonstrates effective image filtering properties and low computational overhead.
• Experimental results show significant improvements over existing state-of-the-art soft decoding methods in objective and subjective evaluations.

Conclusions:

• The combined use of Laplacian, sparsity, and LERaG priors leads to superior JPEG image reconstruction.
• The LERaG prior effectively addresses the limitations of sparsity priors in recovering high DCT frequencies.
• The developed algorithm represents a significant advancement in optimizing image reconstruction from JPEG compressed formats.