Abstract
This paper introduces a 'low-rank prior' for small oriented noise-free image patches: considering an oriented patch as a matrix, a low-rank matrix approximation is enough to preserve the texture details in the properly oriented patch. Based on this prior, we propose a single-patch method within a generalized joint low-rank and sparse matrix recovery framework to simultaneously detect and remove non-pointwise random-valued impulse noise (e.g., very small blobs). A weighting matrix is incorporated in the framework to encode an initial estimate of the spatial noise distribution. An accelerated proximal gradient method is adapted to estimate the optimal noise-free image patches. Experiments show the effectiveness of our framework in removing non-pointwise random-valued impulse noise.
Introduction
Impulse noise—often seen as "salt-and-pepper" or random-valued pixels—is a common artifact in digital imaging. While single-pixel noise is well-studied, "non-pointwise" impulse noise, where noise forms small, irregular blobs, is significantly harder to remove. This type of noise is prevalent in specific environments, such as underwater "marine snow" or particles in medical hydrocolonoscopy videos.
Existing methods generally fall into two categories: two-stage filters and multi-patch methods. Two-stage filters detect noisy pixels and interpolate them, but they often fail when noise is structured in blobs. Multi-patch methods group similar patches to find a low-rank representation; however, as noise blobs increase in size, finding identical large patches becomes mathematically difficult, often resulting in blurring or poor noise suppression. This research proposes a single-patch approach that utilizes larger patch sizes (e.g., 40x40) to encompass and remove significant noise structures without requiring external similarity searches.
Methodology
The core of the proposed method is the observation that a small, noise-free image patch is approximately low-rank when treated as a matrix, provided it is oriented correctly.
1. Characteristic Orientation
An edge or texture pattern may have a high rank if it is diagonal relative to the pixel grid. However, by rotating the patch so the dominant structure aligns with the matrix axes, the rank drops significantly while texture details are preserved. The authors find this "characteristic orientation" by minimizing the error of a rank-1 approximation.
2. Weighted Low-Rank and Sparse Recovery
Denoising is formulated as an optimization problem based on Robust Principal Component Analysis (RPCA). A noisy patch $P$ is decomposed into a low-rank component $L$ (the clean signal) and a sparse component $S$ (the impulse noise). To handle blobs effectively, a weighting matrix $W$ is introduced as a "soft" noise detector. By assigning lower weights to suspected noisy pixels, the framework reduces the influence of noise during the reconstruction of the low-rank signal.
3. Optimization
The problem is solved using an Accelerated Proximal Gradient (APG) method. This convex optimization technique is adapted to handle entry-wise weighting and the specific single-patch model structure.
Experimental Results
The authors validated the low-rank hypothesis using approximately 450,000 patches from the Caltech256 and SceneCategory15 datasets. Quantitative evaluations were performed on standard benchmark images like 'Barbara' and 'Lena' with synthetic blobs (1x1 to 4x4 pixels) added.
Key findings include:
- Superior Performance on Blobs: On the 'Barbara' image with 3x3 noise blobs, the method achieved a PSNR of 26.40, significantly outperforming the ROLD-EPR (18.99) and multi-patch MPLR (24.06) methods.
- Edge Preservation: The method maintains sharp edges where standard low-rank or median filters typically cause blurring.
- Real-World Application: In hydrocolonoscopy videos, the method successfully removed large suspended particles that multi-patch methods either blurred or failed to detect.
Conclusion
This single-patch low-rank prior offers a robust solution for removing structured impulse noise. While the method is computationally intensive—taking several minutes for standard resolution images in Matlab—it demonstrates superior edge preservation and noise suppression for non-pointwise artifacts. Future work may involve hardware acceleration and extending the framework to video denoising to leverage temporal consistency.
Summary updated