[Paper Review] Image denoising via group sparsity residual constraint

Zhiyuan Zha, Xin Liu+8 more|arXiv (Cornell University)|Sep 12, 2016

Image and Signal Denoising Methods25 references3 citations

TL;DR

This paper proposes Group Sparsity Residual Constraint (GSRC), a novel image denoising method that reduces group sparsity residual by refining noisy image's group sparse coefficients using a first-pass clean image estimate. The method outperforms state-of-the-art techniques like BM3D and WNNM in PSNR while maintaining competitive computational speed.

ABSTRACT

Group sparsity has shown great potential in various low-level vision tasks (e.g, image denoising, deblurring and inpainting). In this paper, we propose a new prior model for image denoising via group sparsity residual constraint (GSRC). To enhance the performance of group sparse-based image denoising, the concept of group sparsity residual is proposed, and thus, the problem of image denoising is translated into one that reduces the group sparsity residual. To reduce the residual, we first obtain some good estimation of the group sparse coefficients of the original image by the first-pass estimation of noisy image, and then centralize the group sparse coefficients of noisy image to the estimation. Experimental results have demonstrated that the proposed method not only outperforms many state-of-the-art denoising methods such as BM3D and WNNM, but results in a faster speed.

Motivation & Objective

• Address the limitation of existing group sparsity-based denoising methods that rely solely on noisy image's nonlocal self-similarity, which can lead to suboptimal reconstruction.
• Improve denoising performance by introducing a new prior model that leverages both the noisy image's structure and an estimated clean image's group sparse coefficients.
• Reduce the group sparsity residual between the noisy image and the clean image to enhance reconstruction accuracy and detail preservation.
• Achieve a balance between high denoising quality and computational efficiency, outperforming methods like WNNM and AST-NLS in speed while maintaining or exceeding their PSNR performance.

Proposed method

• Introduce the concept of group sparsity residual as a measure of deviation between the group sparse representation of the noisy image and the true clean image.
• Perform a first-pass estimation of the clean image using a standard denoising method to obtain an initial estimate of the group sparse coefficients of the original image.
• Centralize the group sparse coefficients of the noisy image toward the estimated coefficients of the clean image to minimize the group sparsity residual.
• Utilize nonlocal similar patches from the noisy image to form 3D groupings, applying collaborative filtering in the transform domain to enhance sparsity and reduce noise.
• Apply a regularization framework that enforces group sparsity on the coefficient matrix derived from similar image patches, with parameters tuned per noise level.
• Implement the algorithm in MATLAB with a patch size adaptive to noise variance, and use a fixed set of parameters (c, γ, δ) for different noise levels.

Experimental results

Research questions

• RQ1 Can reducing the group sparsity residual between the noisy and clean image representations lead to improved denoising performance compared to existing methods?
• RQ2 Does incorporating an initial clean image estimate into the group sparse coefficient refinement process enhance reconstruction accuracy and detail preservation?
• RQ3 Can the proposed method achieve higher PSNR than state-of-the-art methods like BM3D and WNNM while maintaining competitive runtime?
• RQ4 How does the proposed GSRC method perform under high noise levels (σ > 50), where traditional methods often fail to preserve fine textures and edges?

Key findings

• GSRC achieves an average PSNR improvement of 0.39–0.56 dB over BM3D, 0.46–0.76 dB over NCSR, and 0.61–0.66 dB over MSEPLL across all tested noise levels.
• At σ = 100, GSRC achieves a PSNR gain of up to 1.24 dB on the Foreman image and 1.37 dB on the House image compared to BM3D, demonstrating strong performance under high noise.
• Visual results show that GSRC better preserves sharp edges and suppresses artifacts compared to BM3D, WNNM, and AST-NLS, especially in texture-rich and edge-dense regions.
• GSRC runs significantly faster than WNNM, NCSR, and AST-NLS, with a median runtime of 55–115 seconds on test images, compared to over 200 seconds for WNNM and NCSR.
• The method maintains competitive speed despite being implemented in MATLAB, outperforming all methods except BM3D, which uses optimized C++ mex-functions.
• The proposed method achieves state-of-the-art PSNR results on 12 standard test images across all noise levels, with consistent gains in both low and high noise regimes.