An Efficient Inexact Gauss–Seidel-Based Algorithm for Image Restoration with Mixed Noise

Tingting Wu,Yue Min,Chaoyan Huang,Zhi Li,Zhongming Wu,Tieyong Zeng

DOI: https://doi.org/10.1007/s10915-024-02510-8

2024-04-17

Journal of Scientific Computing

Abstract:A challenge in image restoration is to recover a clear image from the blurry observation in the presence of different types of noise. There are few works addressing image deblurring under mixed noise. To handle this issue, we propose a general model based on classical wavelet tight frame regularization. We utilize a convexity-preserving term to obtain a component-wise convex model under a mild condition. Indeed, to reduce the cost of solving subproblems, the inexact Gauss–Seidel-based majorized semi-proximal alternating direction method of multipliers (sGS-imsPADMM) with relative error control is developed. Besides, the global convergence of sGS-imsPADMM is demonstrated. Numerical results for the image restoration problems show that the proposed model and solving approach are superior to some state-of-the-art methods both in numerical analysis and visual quality.

mathematics, applied

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to restore blurred images in the presence of mixed noise (additive Gaussian noise and multiplicative Gamma noise). Specifically, the authors propose a general model based on the classical wavelet tight - frame regularization and use convexity - preserving terms to obtain component - convex models under mild conditions. To reduce the cost of solving sub - problems, they develop an inexact symmetric Gauss - Seidel - based semi - proximal alternating direction method of multipliers (sGS - imsPADMM) with relative error control. In addition, the global convergence of sGS - imsPADMM is also proved. Numerical results show that the proposed model and solution method are superior to some existing state - of - the - art methods in terms of numerical analysis and visual quality.

The mathematical expressions mentioned in the paper are as follows:

- The mathematical expression for image degradation is:
\[
f=(Hw + b)\eta
\]
where \(f\) is the degraded image, \(w\) is the original image, \(H\) is a linear operator, \(b\) is additive Gaussian noise, and \(\eta\) is multiplicative Gamma noise.

- The proposed model is:
\[
\min_{u > 0, w}\left\langle\log u+\frac{f}{u}, 1\right\rangle+\alpha\left\|\sqrt{u}-\sqrt{f}\right\|_{2}^{2}+\frac{\|Hw - u\|_{2}^{2}}{2\sigma^{2}}+\mu\|Ww\|_{1}
\]
where \(\alpha\) and \(\mu\) are positive parameters, and \(W\) is a frame transformation matrix satisfying \(W^{T}W = I\).

- By introducing an auxiliary variable \(x\), the minimization problem (7) is transformed into an equivalent form:
\[
\min_{u, w, x}\left\langle\log u+\frac{f}{u}, 1\right\rangle+\alpha\left\|\sqrt{u}-\sqrt{f}\right\|_{2}^{2}+\frac{\|Hw - u\|_{2}^{2}}{2\sigma^{2}}+\mu\|x\|_{1}
\]
subject to the constraint:
\[
x - Ww=0
\]

- To solve the above model, an inexact sGS - imsPADMM algorithm with relative error control is proposed, which can reduce the cost of solving sub - problems while ensuring theoretical convergence.

Through these methods, the paper demonstrates the superior performance of the proposed model and algorithm in handling mixed - noise and blur - removal tasks.