MindX: Denoising Mixed Impulse Poisson-Gaussian Noise Using Proximal Algorithms

Mohamed Aly, Wolfgang HeidricharXiv 2016 · OpenReview

TL;DR

This paper introduces "MindX," a novel algorithm for denoising images corrupted by a challenging combination of impulse, Poisson, and Gaussian noise. By cleverly transforming the noise, applying proximal optimization techniques, and then reverting the transformation, MindX achieves state-of-the-art denoising performance across various noise conditions.

What Problem Does This Paper Address?

Digital imaging sensors and acquisition processes are susceptible to various types of noise. While Gaussian noise is common and well-studied, real-world scenarios often involve more complex noise mixtures. Specifically, this paper tackles the problem of images corrupted by:

  • Poisson noise: Arises from photon-limited situations (e.g., low light, medical imaging) and is signal-dependent, meaning its variance changes with the signal intensity.
  • Gaussian noise: A signal-independent noise common in electronic components.
  • Impulse noise: Characterized by sudden, extreme pixel value changes (salt-and-pepper or random-valued) often caused by sensor defects or transmission errors.

Denoising these mixed noise types simultaneously, especially when the locations of impulse noise are unknown (blind denoising), is a significant challenge. Existing methods often focus on only one or two of these noise types, leaving a gap for robust algorithms that can handle all three.

Methods

The MindX algorithm is a three-step process:

  1. Variance Stabilization Transform (VST): The first step uses the Generalized Anscombe Transform (GAT) to convert the mixed Poisson-Gaussian noise into white Gaussian noise. This is crucial because it simplifies the problem by making the noise signal-independent and having a stable variance, making it easier to handle with standard denoising techniques.
  2. Proximal Optimization for Mixed Noise Denoising: The core of the algorithm lies in denoising the transformed image, which now has mixed impulse and Gaussian noise. This is achieved by solving a combinatorial optimization problem using proximal algorithms. The objective function combines:
  3. A data fidelity term (ensuring the denoised image is close to the noisy one).
  4. An $l_0$ quasi-norm prior to model the sparsity of impulse noise (i.e., only a few pixels are corrupted).
  5. A spatial regularizer, specifically Total Variation (TV) and the BM3D (Block-Matching and 3D filtering) prior, to leverage image self-similarity and structural information. This step employs a two-step alternating minimization approach (z-step and x-step) that iteratively estimates the impulse noise locations and the denoised image.

  6. Inverse VST: Finally, the inverse GAT is applied to transform the denoised image back to its original pixel range, yielding the final clean image estimate.

The novelty lies in the seamless integration of these three noise types within a proximal optimization framework, particularly the use of GAT to handle Poisson noise and the combinatorial optimization for impulse noise alongside standard image priors.

Datasets

The paper evaluates MindX on standard, publicly available 512x512-pixel images: "lena," "cameraman," "barbara," and "pepper." These are common benchmarks in image processing research. No new datasets were introduced.

Key Insights and Ideas

  • The power of noise transformation: Using the Generalized Anscombe Transform is a critical step that effectively converts the complex, signal-dependent Poisson-Gaussian noise into a more manageable, signal-independent Gaussian noise, paving the way for subsequent denoising.
  • Synergy of priors: Combining different regularization priors (Total Variation for piecewise smoothness and BM3D for patch-based self-similarity) within the proximal optimization framework leads to superior denoising performance compared to using a single prior.
  • Robustness to impulse noise: The $l_0$ quasi-norm regularization, coupled with proximal algorithms, proves effective in handling impulse noise, even when its locations are unknown (blind denoising).
  • MindX's broad applicability: The algorithm demonstrates strong performance across a wide range of noise conditions, suggesting its potential for real-world applications where noise is often a mix of different types.

Key Results

MindX consistently outperforms state-of-the-art methods in denoising images corrupted by mixed impulse, Poisson, and Gaussian noise.

  • Superior Performance: In most tested scenarios, MindX achieved the best Peak Signal-to-Noise Ratio (PSNR) values. It was ranked first in all but six settings across the experiments.
  • Competitive Baselines: GAT-AOP and GAT-BM3D were identified as close runners-up, highlighting the effectiveness of their respective approaches.
  • Impact of Priors: Adding the BM3D prior to the TV prior (as in MindX-TV-BM3D) generally improved performance over methods using only TV (MindX-TV) or only BM3D (GAT-BM3D), underscoring the benefit of complementary priors.
  • Convergence: The paper proves that Algorithm 1 (MindX) converges to a local minimum. Figure 1 visually demonstrates the convergence of the algorithm and the effect of GAT on variance stabilization.

Limitations and Future Work

  • Local Minimum: While the algorithm is proven to converge to a local minimum, it does not guarantee a global optimum for the non-convex optimization problem.
  • Parameter Tuning: The performance of MindX, like many optimization-based methods, can be sensitive to parameter tuning (e.g., weights for different priors).
  • Future Work: The authors suggest extending their work to general deconvolution problems (both blind and non-blind) and integrating MindX into applications like fluorescence microscopy and low-dose X-ray tomography. They also note the potential for further investigation into the tradeoff between different priors for extreme noise scenarios.

Summary updated

Abstract

We present a novel algorithm for blind denoising of images corrupted by mixed impulse, Poisson, and Gaussian noises. The algorithm starts by applying the Anscombe variance-stabilizing transformation to convert the Poisson into white Gaussian noise. Then it applies a combinatorial optimization technique to denoise the mixed impulse Gaussian noise using proximal algorithms. The result is then processed by the inverse Anscombe transform. We compare our algorithm to state of the art methods on standard images, and show its superior performance in various noise conditions.
BibTeX
@misc{aly2016mindx,
  author = {Mohamed Aly and Wolfgang Heidrich},
  title = {{MindX: Denoising Mixed Impulse Poisson-Gaussian Noise Using Proximal Algorithms}},
  year = {2016},
  eprint = {1608.07802},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1608.07802},
}
Figure