A computational framework known as the parallel linearized alternating direction method of multipliers, or PLADMM, provides an efficient solution for solving large-scale regularized inverse problems in multichannel imaging without requiring burdensome matrix inversions. Developed by Chuan He, Wenshen Peng, Junwei Wang, Xiaowei Feng, and Licheng Jiao, the algorithm targets complex imaging tasks such as color image deblurring, missing pixel inpainting, and parallel magnetic resonance imaging reconstruction. Many inverse formulations in imaging rely on convex optimization models that combine a data-fidelity term with non-smooth regularization functions coupled through linear operators. Solving such formulations through standard operator splitting techniques often forces computational algorithms to execute explicit matrix inversion steps. In multichannel settings involving cross-channel blurs or complex forward operators, these linear inversion stages cannot be diagonalized straightforwardly by fast Fourier transforms, causing severe bottlenecks. The newly introduced linearized framework bypasses direct matrix inversions entirely while offering a parallel structure suited for simultaneous processing across distributed computing architectures.
Inverse imaging problems aim to recover an underlying high-quality scene from degraded, blurred, incomplete, or noisy measurements across several image channels. Mathematically, these setups are frequently posed as the minimization of an objective function composed of a convex term and a sum of non-smooth convex penalties acting on linearly transformed representations of the unknown image. While classical gradient descent methods cannot directly handle non-smooth penalties such as total variation or total generalized variation, operator splitting algorithms like the alternating direction method of multipliers break the overall problem into smaller, proximable subproblems. However, existing splitting variants are generally configured for problems featuring only two split functional terms. Applying traditional splitting to intricate multi-term, multichannel problems typically couples the auxiliary variables and linear operators together, necessitating the inversion of large-scale matrices at every iteration. This computational burden substantially degrades single-step execution speed and limits practical applicability in high-dimensional imaging domains.
To overcome this computational obstacle, the research team constructed a parallel linearized algorithm by introducing independent auxiliary variables for each linear operator term and formulating an augmented Lagrangian function. Distinct positive penalty parameters are assigned individually to each linear constraint, giving the system greater flexibility than standard ADMM frameworks where a single uniform parameter is shared. In the primal update stage of the classical scheme, computing the primary image estimate requires minimizing a quadratic objective that links all operators simultaneously, yielding a massive linear system. The authors linearize this quadratic component around the current iteration point, converting the subproblem into an isolated proximity operation governed by an adjustable step-size parameter. Because linearization decouples the operator products from the update inversion, every auxiliary variable and dual multiplier vector can be evaluated independently and concurrently at each iteration step.
Building upon the foundational full-splitting scheme, the researchers established a compact primal-dual representation that further streamlines practical computations. By rearranging the update sequence so that dual variables are computed before the primal image variable and deploying the Moreau decomposition identity, the explicit auxiliary variables can be eliminated from memory storage. This primal-dual formulation reduces the memory footprint required during distributed computing, as fewer high-dimensional arrays must be transferred or stored in cache memory. The resulting routine updates dual components via projection operators acting on scaled linear transforms, followed by a consolidated primal step that aggregates dual feedback through adjoint linear operators. Theoretical analysis using variational inequalities confirmed that the sequence generated by the algorithm is contractive, guaranteeing convergence from arbitrary initial starting points to an exact saddle point at a worst-case theoretical convergence rate of O(1/k).
The team demonstrated the practical efficacy of the scheme by applying it to multichannel inverse problems regularized by second-order total generalized variation, known as TGV. As an extension of classical total variation, second-order TGV balances first-order and second-order spatial differences, penalizing image discontinuities while simultaneously regularizing moderate slope regions. This structural balance enables the model to suppress the unnatural piecewise-constant staircasing artifacts commonly produced by standard total variation while retaining sharp boundaries. To implement TGV regularization within the proposed solver, the researchers decomposed the overall optimization into decoupled functional components subject to pixel value box constraints bounding intensity to the interval between 0 and 255 alongside discrepancy-based residual fidelity constraints. Closed-form evaluations were derived for each component, employing two-dimensional and four-dimensional projection operators alongside multi-dimensional shrinkage operations.
Performance evaluations for multichannel image deblurring were conducted in MATLAB on an Intel Core i5 personal computer equipped with 8 gigabytes of random-access memory. The experiments examined four standard color test images—Lena at 256 by 256 pixels, Peppers at 512 by 512 pixels, Monarch at 768 by 512 pixels, and Race at 768 by 512 pixels—subjected to synthetic blurs created from average, Gaussian, and motion kernels combined across color channels. Synthetic Gaussian noise was added across different degradation scenarios, with iterations capped at 300. The proposed TGV-regularized algorithm was evaluated against several established methods, including parallel ADMM with TGV, fast total variation deconvolution, and parallel primal-dual splitting. Visual inspection and quantitative metrics, including peak signal-to-noise ratio in decibels and structural similarity index measurements, indicated that the TGV-regularized approach restored sharper edges and fewer staircasing artifacts than competing total variation techniques across all tested image sizes and noise levels.
The deblurring trials demonstrated clear computational speed advantages when handling cross-channel blurs. In the first experimental problem, which contained exclusively within-channel blurs, the speed difference between linearized and standard parallel ADMM was relatively modest because block-diagonal components could be inverted reasonably fast. However, in problems involving cross-channel blurring, standard parallel ADMM required complex matrix inversion steps through Cramer’s rule that could not be diagonalized by fast Fourier transforms, causing its execution time to increase sharply. By contrast, the linearized algorithm avoided matrix inversions entirely, achieving significantly faster convergence per unit of central processing unit time. Although the linearized scheme exhibited slightly slower initial progress during the opening iterations due to inexact primal approximation steps, its superior single-step efficiency allowed it to rapidly overtake competing algorithms and achieve higher objective image quality scores in less overall processing time.
In a second experimental phase focusing on RGB image inpainting, the methodology was tested on the 768 by 512 pixel Girl test image to evaluate its capability in reconstructing missing image regions. Pixels were randomly removed at rates ranging from 0 percent up to 65 percent, and the degraded data were further corrupted with Gaussian noise possessing a standard deviation of 20. When the pixel loss level reached 55 percent, representing the destruction of the majority of image information, the algorithm successfully inpainted the missing structure across an iteration ceiling of 150 cycles. Compared against total variation implementations of linearized ADMM and fast deconvolution, the TGV-regularized variant delivered superior peak signal-to-noise ratio and structural similarity values across all missing-pixel fractions. Detailed close-up visualizations demonstrated that while total variation methods suffered from noticeable staircasing across smooth tonal transitions, the second-order TGV formulation preserved natural gradients and edge definition.
The framework was subsequently validated on parallel magnetic resonance imaging reconstruction using sensitivity encoding principles to reconstruct undersampled spatial data from multiple receiver coils. A 512 by 512 pixel brain scan acquired from a 3-Tesla scanner via a T2-weighted turbo spin echo sequence with eight surface receiver coils was tested alongside an undersampling radial mask having an acquisition rate of 0.147. Under additive Gaussian noise with standard deviations of 25 and 100 across 100 iterations, the proposed TGV-based linearized method was compared against TV-based formulations including FUSL, LADM, and ALADMM. The algorithm attained higher peak signal-to-noise ratios, improved structural similarity scores, and lower relative reconstruction errors compared to the ground truth image, while demanding less processing time than competing total variation methods despite managing a mathematically more complex regularization structure.
Collectively, the experimental investigations and theoretical convergence proofs confirm that parallel linearized ADMM provides a mathematically sound, highly scalable solver for multichannel imaging inverse problems. By replacing matrix inversion operations with decoupled proximal evaluations, the method resolves longstanding computational bottlenecks in multichannel deconvolution and parallel magnetic resonance reconstruction. The authors noted that while second-order total generalized variation effectively suppresses staircasing artifacts using analytical mathematical priors, future iterations of the algorithm could incorporate learned feature models to further improve restoration fidelity in specialized imaging domains.
Subject of Research: Parallel linearized alternating direction method of multipliers for imaging inverse problems
Article Title: Parallel linearized ADMM with application to multichannel image restoration and reconstruction
Article References: He, C., Peng, W., Wang, J., Feng, X., & Jiao, L. (2024). Parallel linearized ADMM with application to multichannel image restoration and reconstruction. EURASIP Journal on Image and Video Processing, 2024(1), Article 39. https://doi.org/10.1186/s13640-024-00654-z
Image Credits: AI Generated
DOI: 10.1186/s13640-024-00654-z
Keywords: ADMM, linearized ADMM, parallel computing, image restoration, image deblurring, image inpainting, parallel MRI, TGV regularization, total generalized variation, inverse problems, convex optimization, operator splitting, multichannel deconvolution, proximal operators, PSNR, SSIM, primal-dual algorithm
Cite Scienmag News
Scienmag. (November 21, 2026). Parallel Linearized ADMM Speeds Multichannel Image Restoration Without Matrix Inversion. https://scienmag.com/parallel-linearized-admm-speeds-multichannel-image-restoration-without-matrix-inversion/
Scienmag. "Parallel Linearized ADMM Speeds Multichannel Image Restoration Without Matrix Inversion." Scienmag, 21 November 2026, https://scienmag.com/parallel-linearized-admm-speeds-multichannel-image-restoration-without-matrix-inversion/. Accessed 21 November 2026.
Scienmag. "Parallel Linearized ADMM Speeds Multichannel Image Restoration Without Matrix Inversion." Scienmag. November 21, 2026. https://scienmag.com/parallel-linearized-admm-speeds-multichannel-image-restoration-without-matrix-inversion/








