[Paper Review] A note on patch-based low-rank minimization for fast image denoising

Haijuan Hu, Jacques Froment+1 more|arXiv (Cornell University)|Jun 28, 2015

Image and Signal Denoising Methods15 references3 citations

TL;DR

This paper proposes a fast patch-based low-rank minimization (PLR) method for image denoising that groups similar image patches into similarity matrices and applies hard thresholding via PCA or SVD to denoise them in a single pass. The method achieves state-of-the-art denoising performance—particularly in preserving textures—while being significantly faster than iterative methods like K-SVD, BM3D, and WNNM.

ABSTRACT

Patch-based low-rank minimization for image processing attracts much attention in recent years. The minimization of the matrix rank coupled with the Frobenius norm data fidelity can be solved by the hard thresholding filter with principle component analysis (PCA) or singular value decomposition (SVD). Based on this idea, we propose a patch-based low-rank minimization method for image denoising. The main denoising process is stated in three equivalent way: PCA, SVD and low-rank minimization. Compared to recent patch-based sparse representation methods, experiments demonstrate that the proposed method is rather rapid, and it is effective for a variety of natural grayscale images and color images, especially for texture parts in images. Further improvements of this method are also given. In addition, due to the simplicity of this method, we could provide an explanation of the choice of the threshold parameter, estimation of PSNR values, and give other insights into this method.

Motivation & Objective

• To develop a computationally efficient image denoising method that avoids iterative optimization.
• To leverage non-local self-similarity in natural images by grouping similar patches into similarity matrices.
• To improve denoising performance, especially for texture-rich regions, compared to existing sparse representation and low-rank methods.
• To achieve high denoising quality with a single-pass processing pipeline, reducing runtime compared to iterative methods.

Proposed method

• Divides the noisy image into overlapping d×d patches and identifies similar patches within an n×n neighborhood around each reference patch.
• Constructs similarity matrices of size d²×m by stacking the m most similar patches (including the reference) as column vectors.
• Applies a hard thresholding filter using the principal component basis (via PCA) or left singular vectors (via SVD) to the similarity matrix to minimize the rank under Frobenius norm fidelity.
• Assumes zero mean for the similarity matrix patches to simplify the low-rank approximation process.
• Restores the denoised image by averaging the denoised patches back into their original positions, using symmetric boundary handling.
• Uses a single pass per similarity matrix, avoiding iterative refinement and enabling high computational efficiency.

Experimental results

Research questions

• RQ1 Can a non-iterative, patch-based low-rank minimization method achieve competitive denoising performance with significantly reduced computational cost?
• RQ2 How does the proposed PLR method compare to state-of-the-art methods like K-SVD, BM3D, and WNNM in terms of PSNR and visual quality?
• RQ3 To what extent does the method preserve fine textures and structural details in natural images compared to sparse representation-based approaches?
• RQ4 Can hard thresholding via PCA/SVD provide an effective and efficient solution to the low-rank matrix approximation problem in image denoising?

Key findings

• The proposed PLR method achieves PSNR values of 35.90 dB on Lena and 35.50 dB on Barbara at σ=10, outperforming K-SVD, LPG-PCA, and PLOW, and matching or exceeding BM3D and WNNM in several cases.
• At σ=20, the method achieves 33.03 dB on Lena and 32.12 dB on Barbara, outperforming K-SVD, LPG-PCA, and PLOW, and matching or surpassing BM3D and WNNM in key metrics.
• The method is exceptionally fast, with a runtime of only 2 seconds on 256×256 images, significantly outperforming K-SVD (210 s), LPG-PCA (138 s), and WNNM (134 s).
• Visual results in Figure 1 show that PLR preserves textures in Lena and Barbara better than BM3D, WNNM, and other methods, particularly in complex regions.
• The method’s performance is robust across diverse images, including Peppers, House, Cameraman, Boats, Bridge, and Lena, with consistent gains in PSNR and visual fidelity.
• The use of compact dictionaries via PCA/SVD and single-pass hard thresholding enables high efficiency without sacrificing denoising quality, making it suitable for real-time applications.