Weighted Anisotropic-Isotropic Total Variation for Poisson Denoising

Kevin Bui, Yifei Lou, Fredrick Park, Jack Xin

TL;DR

This paper introduces a highly efficient denoising model for images corrupted by Poisson noise using a nonconvex regularizer called Weighted Anisotropic-Isotropic Total Variation (AITV). By leveraging a specialized proximal operator within an ADMM framework, the authors achieve superior image restoration quality and sharper edges while running nearly 10 times faster than current state-of-the-art methods.

What Problem Does This Paper Address?

In photon-limited imaging—common in fields like astronomy and medical imaging—images are often plagued by Poisson noise. Unlike the more familiar Gaussian noise, Poisson noise is signal-dependent: the noise level varies based on the intensity of the pixels. This makes denoising particularly tricky because a "one size fits all" smoothing approach often fails.

The classic solution is Total Variation (TV) regularization. While TV is great at removing noise, it suffers from "staircasing" artifacts (turning smooth gradients into flat blocks) and tends to blur out fine details and edges. More modern approaches like Non-local PCA or Convolutional Neural Networks (CNNs) offer better results but come with heavy computational costs, requiring either expensive patch-matching or massive training datasets and GPUs. The authors sought a "middle ground": a model that provides the edge-preserving benefits of nonconvex optimization without the massive computational overhead.

Methods

The core innovation is the application of Weighted Anisotropic-Isotropic Total Variation (AITV) to the Poisson denoising problem.

The AITV Regularizer

AITV is defined as the difference between the $\ell_1$ norm (anisotropic) and the $\ell_{2,1}$ norm (isotropic) of the image gradient: $$\|\nabla u\|_1 - \alpha\|\nabla u\|_{2,1}$$ where $\alpha \in [0, 1]$ is a weighting parameter. This nonconvex formulation allows the model to promote sparsity in the gradient more effectively than standard TV, which helps in recovering sharp, oblique edges that standard TV often smears.

The Algorithm: ADMM + Proximal Operator

To solve the resulting optimization problem, the authors developed a customized Alternating Direction Method of Multipliers (ADMM). The "secret sauce" for their efficiency is the use of a proximal operator for the $\ell_1 - \alpha\ell_2$ term.

  • They split the problem into subproblems for the image ($u$), an auxiliary variable for the Poisson log-likelihood ($v$), and the gradient ($w$).
  • The $u$-subproblem is solved efficiently in the frequency domain using Fast Fourier Transforms (FFT).
  • The $w$-subproblem, which involves the nonconvex AITV term, is solved using a closed-form proximal mapping. This avoids the slow, iterative "Difference-of-Convex" (DCA) algorithms typically used for such problems.

Datasets

The model was evaluated using five standard grayscale images from the Berkeley Segmentation Dataset (BSD): River, Butterfly, Starfish, Penguin, and Boat.

To simulate various noise conditions, the authors rescaled these images to three different peak intensity values (80, 55, and 30). Lower peak values represent "darker" conditions with significantly higher relative Poisson noise.

Key Insights and Ideas

  • Nonconvexity for Edge Preservation: The paper reinforces the idea that nonconvex regularizers can "outperform" convex ones in edge recovery. By subtracting a portion of the isotropic TV from the anisotropic TV, the model becomes less biased toward axis-aligned edges.
  • Speed via Closed-Form Solutions: Most nonconvex models are avoided in real-time applications because they are slow to converge. This paper demonstrates that if you can derive a closed-form proximal operator (as shown in Lemma 1), nonconvex models can actually be faster than their convex counterparts.
  • Signal-Dependent Adaptation: Because the model incorporates the Poisson log-likelihood directly into the data fidelity term, it naturally handles the varying noise levels across different intensity regions of the image.

Key Results

The authors compared AITV against standard TV, Non-local PCA (NL-PCA), and Fractional-order TV (FOTV).

  • Quantitative Performance: AITV consistently achieved the highest PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index) across almost all test images and noise levels. For example, at a peak value of 30 (high noise), AITV averaged a PSNR of 27.57 compared to TV's 26.72.
  • Qualitative Sharpness: Visual inspections and line profile plots (Fig 3 in the paper) show that AITV produces much cleaner results in flat areas (like the sky) while maintaining significantly sharper spikes at edge boundaries.
  • Computational Efficiency: This is the most striking result. AITV averaged a processing time of 1.89 seconds, while NL-PCA took 20.05s and standard TV took 16.59s. This represents a nearly 10x speedup over traditional ADMM-based TV solvers.

Limitations and Future Work

  • Convergence Guarantees: While the algorithm performed reliably in experiments, the authors note that global convergence for ADMM on nonconvex problems is not always guaranteed under the specific conditions of their gradient operator.
  • Parameter Tuning: The model requires tuning the fidelity parameter ($\lambda$) and the AITV weight ($\alpha$). While the authors provided ranges, these currently need to be optimized per image/noise level.
  • Future Directions: The authors plan to extend this model to color images and explore "algorithm unfolding"—a technique where the ADMM steps are used to design the architecture of a deep neural network, potentially combining the strengths of variational methods and deep learning.

Summary updated

Abstract

Poisson noise commonly occurs in images captured by photon-limited imaging systems such as in astronomy and medicine. As the distribution of Poisson noise depends on the pixel intensity value, noise levels vary from pixels to pixels. Hence, denoising a Poisson-corrupted image while preserving important details can be challenging. In this paper, we propose a Poisson denoising model by incorporating the weighted anisotropic-isotropic total variation (AITV) as a regularization. We then develop an alternating direction method of multipliers with a combination of a proximal operator for an efficient implementation. Lastly, numerical experiments demonstrate that our algorithm outperforms other Poisson denoising methods in terms of image quality and computational efficiency.
BibTeX
@inproceedings{bui2023weighted,
  author = {Kevin Bui and Yifei Lou and Fredrick Park and Jack Xin},
  title = {{Weighted Anisotropic-Isotropic Total Variation for Poisson Denoising}},
  booktitle = {ICIP},
  year = {2023},
  doi = {10.1109/icip49359.2023.10222230},
}

Loading related content...

Figure