Abstract
Most existing image denoising algorithms are designed to handle a single type of noise. However, real-world images often suffer from mixed noise—typically a combination of Gaussian and impulse noise—acquired during capture and transmission. This paper proposes a new variational algorithm that simultaneously exploits image local consistency and nonlocal consistency. Local consistency is enforced via a hyper-Laplace prior to maintain smoothness and sharp edges, while nonlocal consistency is captured through the three-dimensional sparsity of similar image patches to preserve repetitive textures. Using a Split-Bregman optimization technique, the proposed method achieves significant performance improvements over state-of-the-art schemes.
Introduction
In practical scenarios, images are rarely corrupted by a single noise source. Instead, they often exhibit Gaussian noise (graininess from sensor heat) and impulse noise (sharp "salt-and-pepper" pixels from transmission errors). Traditional denoising methods struggle with this combination: Gaussian filters often leave impulse speckles behind, while impulse-focused filters like median filters tend to blur the underlying image structure.
Existing "two-phase" approaches attempt to detect impulse noise first, but they frequently over-smooth the image and lose fine textures because they fail to fully exploit the statistical properties of natural images. There is a clear need for a unified framework that can address both noise types without sacrificing image detail.
Methodology
The proposed algorithm utilizes two powerful image priors within a unified variational framework to reconstruct clean images from mixed-noise observations:
- Image Local Consistency (Hyper-Laplace Prior): Rather than using a standard Laplacian distribution (Total Variation), the model employs a "heavy-tailed" hyper-Laplace prior ($||\nabla x||^{2/3}$). This more accurately models the gradient distribution of natural images, which are characterized by large smooth regions punctuated by sharp edges. This prevents the over-smoothing common in older variational methods.
- Image Nonlocal Consistency (3D Sparsity): This exploits the inherent self-similarity of natural images. The algorithm identifies similar patches across the image, stacks them into 3D arrays, and enforces sparsity in the transform domain. This allows the model to recover repetitive patterns that might be completely obscured by noise.
To solve the resulting complex optimization problem efficiently, the researchers implemented the Split-Bregman technique. This approach decomposes the global problem into smaller, iterative sub-problems, allowing for a progressive refinement of the image and the "map" of corrupted pixels.
Experimental Results
The algorithm was rigorously tested on benchmark images including Lena, Boat, and Barbara. The researchers simulated mixed noise by adding Gaussian noise ($\sigma = 10$) and varying levels of impulse noise (10% to 50%).
Quantitatively, the proposed method consistently outperformed standard techniques like Total Variation (TV) and IFASDA, providing an average improvement of 2 dB in Peak Signal-to-Noise Ratio (PSNR). The visual results are equally compelling, showing sharper edges and fewer artifacts than competing methods.
One of the most significant results was observed in the Barbara dataset, which is known for its complex textures. In scenarios with 50% impulse noise, the proposed method achieved a 3.6 dB gain over IFASDA. As shown in the figure below, the algorithm successfully recovered the fine stripes on the clothing, which other methods reduced to a gray blur.
Conclusion
By combining local and nonlocal image consistencies, the proposed variational algorithm offers a robust solution for the challenging problem of mixed Gaussian-impulse noise removal. The use of a hyper-Laplace prior and 3D patch sparsity allows for superior edge preservation and texture recovery. Future work may involve optimizing the computational search for nonlocal patches and extending the framework to address other multi-degradation problems, such as simultaneous deblurring and denoising.
Summary updated