[Paper Review] Deconvolution of Poissonian Images Using Variable Splitting and Augmented Lagrangian Optimization

Mário A. T. Figueiredo, José M. Bioucas‐Dias|ArXiv.org|Apr 30, 2009

Sparse and Compressive Sensing Techniques19 references5 citations

TL;DR

This paper proposes a novel deconvolution method for Poissonian images using variable splitting and augmented Lagrangian optimization to efficiently solve the non-smooth, non-quadratic regularization problem arising from combining Poisson log-likelihood with total variation (TV) regularization. The approach transforms the original problem into a constrained optimization task, enabling fast, convergent algorithms via alternating minimization, and achieves state-of-the-art performance with faster convergence than existing methods.

ABSTRACT

Although much research has been devoted to the problem of restoring Poissonian images, namely in the fields of medical and astronomical imaging, applying the state of the art regularizers (such as those based on wavelets or total variation) to this class of images is still an open research front. This paper proposes a new image deconvolution approach for images with Poisson statistical models, with the following building blocks: (a) a standard regularization/MAP criterion, combining the Poisson log-likelihood with a regularizer (log-prior) is adopted; (b) the resulting optimization problem (which is difficult, since it involves a non-quadratic and non-separable term plus a non-smooth term) is transformed into an equivalent constrained problem, via a variable splitting procedure; (c) this constrained problem is addressed using an augmented Lagrangian framework. The effectiveness of the resulting algorithm is illustrated in comparison with current state-of-the-art methods.

Motivation & Objective

• To address the challenge of deconvolving Poisson-distributed images corrupted by blur, where standard Gaussian-based methods fail due to non-Gaussian noise models.
• To overcome the limitations of existing algorithms in handling the non-Lipschitz gradient of the Poisson log-likelihood and the non-smooth TV regularizer.
• To develop a robust, convergent optimization framework for Poissonian image deconvolution that enables efficient computation using variable splitting and augmented Lagrangian methods.
• To demonstrate superior performance in terms of reconstruction quality and computational efficiency compared to state-of-the-art approaches.

Proposed method

• Transform the original non-convex, non-smooth optimization problem into an equivalent constrained problem via variable splitting, introducing auxiliary variables to decouple complex terms.
• Apply the augmented Lagrangian framework to solve the constrained problem, ensuring convergence and enabling alternating minimization over variables.
• Solve the subproblem in the x-variable via a linear system involving the convolution operator and identity matrix, efficiently computed using the FFT for block Toeplitz/circulant structures.
• Address the z-variable subproblem via a closed-form solution derived from a quadratic equation, ensuring non-negativity automatically.
• Perform TV-based denoising via the Moreau proximity operator using Chambolle’s fast algorithm to enforce sparsity in image gradients.
• Iteratively update dual variables and penalty parameters, with convergence guaranteed under standard conditions for the augmented Lagrangian method.

Experimental results

Research questions

• RQ1 Can variable splitting and augmented Lagrangian optimization be effectively applied to Poisson image deconvolution with non-smooth TV regularization?
• RQ2 How does the proposed method compare in convergence speed and reconstruction accuracy to existing state-of-the-art algorithms for Poissonian images?
• RQ3 Can the algorithm handle both high and low signal-to-noise ratio scenarios effectively, especially when the Poisson log-likelihood lacks a Lipschitz-continuous gradient?
• RQ4 Does the use of a fixed penalty parameter in the augmented Lagrangian framework still ensure convergence and good performance in practice?
• RQ5 Can the method be extended to other regularizers beyond total variation, such as frame-based sparse representations?

Key findings

• The proposed PIDAL algorithm achieved an average improvement in signal-to-noise ratio (ISNR) of 6.96 dB in high-SNR experiments, outperforming the state-of-the-art method from [13], which reported 6.61 dB.
• In low-SNR experiments, PIDAL achieved lower mean absolute error (MAE) across all tested intensity levels (5, 30, 100, 255), with values of 0.37, 1.34, 3.99, and 8.65 respectively, compared to 0.44, 1.44, 4.69, and 10.40 for the method in [9].
• Each run of the algorithm from [9] took approximately ten times longer than PIDAL, demonstrating a significant speed advantage of the proposed method.
• The algorithm automatically satisfies the non-negativity constraint on the image estimate due to the closed-form solution for the z-variable subproblem.
• The method is robust across varying signal-to-noise ratios and maintains high reconstruction fidelity even at low intensities.
• The use of TV regularization with a simpler structure yielded better results than more complex locally adaptive techniques in [13], highlighting the effectiveness of the proposed optimization framework.