Abstract
Group sparsity has shown great potential in various low-level vision tasks such as image denoising, deblurring, and inpainting. This paper proposes a new prior model for image denoising via Group Sparsity Residual Constraint (GSRC). To enhance performance, the concept of group sparsity residual is introduced, translating the denoising problem into one of residual reduction. By obtaining a high-quality first-pass estimation of the group sparse coefficients from the noisy image and centralizing the noisy coefficients toward this estimation, the method achieves superior results. Experimental results demonstrate that GSRC not only outperforms state-of-the-art methods like BM3D and WNNM in quality but also offers faster execution speeds.
Introduction: The Challenge of High-Noise Denoising
Image denoising is a fundamental ill-posed problem in computer vision, where the goal is to recover a clean image $X$ from a noisy observation $Y = X + V$. Most modern techniques rely on Nonlocal Self-Similarity (NSS)—the observation that small patches in an image tend to repeat. Methods like BM3D group these similar patches and apply group sparsity to filter noise.
However, when an image is heavily corrupted, the noise itself obscures the "true" sparse structures. Extracting priors directly from a dirty source often leads to blurred details or artifacts. The GSRC framework addresses this by shifting the focus from finding the sparse representation of the clean image to minimizing the residual between noisy and estimated clean sparse coefficients.
Methodology: Group Sparsity Residual Constraint
The authors propose a two-stage approach centered on the Group Sparsity Residual.
- Defining the Residual: Instead of seeking absolute sparsity, the method defines a residual $R$ as the difference between the sparse coefficients of the noisy image and the unknown clean image.
- First-Pass Estimation: Since the true clean coefficients are unknown, a "first-pass" denoising step (using the BM3D algorithm) provides an initial estimate of the clean image to anchor the sparsity targets.
- Laplacian Modeling: Statistical analysis of the residual $R$ reveals that it closely follows a Laplacian distribution. This justifies the use of the $\ell_1$-norm in the regularization term, which is mathematically efficient for promoting sparsity.
- Iterative Refinement: The process is repeated, using the output of one iteration to improve the noise variance estimation and patch similarity for the next, progressively closing the gap between the noisy and clean representations.
Experimental Setup
The study utilizes 12 standard grayscale test images common in the image processing community, including classic portraits (Lena, Barbara), structures (House, Boat), and textures (Monarch, Leaves).
These images were artificially corrupted with additive white Gaussian noise at various standard deviations ($\sigma = 30, 40, 50, 100$) to benchmark performance against state-of-the-art competitors.
Key Results and Discussion
- Quantitative Gains: GSRC consistently outperformed benchmarks like BM3D, NCSR, and WNNM. In extreme noise cases ($\sigma = 100$), GSRC improved the Peak Signal-to-Noise Ratio (PSNR) by as much as 1.24dB to 1.37dB over BM3D.
- Visual Fidelity: Qualitatively, the method preserves sharp edges and fine textures (such as the threads in Barbara or the bricks in House) that other methods tend to blur or distort with "ringing" artifacts.
- Efficiency: Despite its iterative nature, GSRC is significantly faster than other high-end models like NCSR and WNNM. While it remains slower than the highly optimized C++ implementation of BM3D, it is very competitive for its complexity class.
Conclusion
The GSRC framework demonstrates that modeling the difference between sparse representations is more stable and effective than modeling the image structure directly from noise. By leveraging a "good enough" first-pass estimate, the method achieves superior reconstruction. Future work will explore extending GSRC to color images and video sequences, where cross-channel and temporal redundancies can further enhance performance.
Summary updated