Poisson image restoration using a novel directional regularization
Introduction
Poisson noise is inevitable in various applications such as electronic microscopy [11] and astronomical imaging [22]. So it is very important to reduce its influence on images through image restoration. The degree of Poisson noise significantly depends on the peak value of pixel intensity values. Specifically, the smaller the peak value of an original image is, the higher the intensity of Poisson noise is. Moreover, the Poisson noise magnitude in an image increases with the pixel intensity of the region of this image. These facts make image restoration under Poisson noise very challenging. Since Poisson noise is different from the Gaussian noise, the restoration models proposed for Gaussian noise are not effective in removing Poisson noise [43]. Based on the statistical property of the Poisson distribution and maximum likelihood estimation [23], the proposed Poisson restoration methods often prefer the following data fidelity term [2], [14], [15], [27]:where is a bounded open domain in , is a bounded linear operator representing the convolution kernel, and represent an observed image and the corresponding restored image, respectively.
Due to the lack of some prior information, the inverse problem of Poisson image restoration is ill-posed. Therefore, many regularization techniques have been widely investigated to deal with this problem in the past two decades, such as total variation (TV) regularization [30], high-order TV regularization [26], [42], [44], [46], total generalized variation (TGV) regularization [5], fractional-order total variation [9], [10] and directional TV regularization [3]. In particular, the TV-based Poisson restoration model [25], [33], [45] can be written aswhere is a regularization parameter. A number of fast algorithms have been proposed to solve this model, such as the EM-TV methods [31], [32], the augmented Lagrangian method [15], the split Bregman method [33] and the primal-dual algorithms [36]. These algorithms have also been successfully applied to remove noisy images corrupted by Gaussian, impulse, and multiplicative noises [4], [37], [38]. If , (1.2) becomes a denoising model, where is the known noisy image. In [12], for modeling the restoration of images corrupted by Poisson noise, the authors proposed to minimize a generalized Kullback–Leiblerdivergence term plus a smoothed version of TV.
For the Poisson restoration problem, the TV regularization performs very well on piecewise constant images for preserving edges while removing noise. But it often causes staircase effects in flat regions. In particular, this phenomenon is more obvious for processing piecewise smooth images. In this case, some high-order total variation (HOTV) regularizations have been introduced [18], [19], [24], [43]. These regularizations have better restoration performance than the TV regularization for preserving the smooth regions. One example of the HOTV regularizations can be described aswhere
However, while using this model for Poisson image restoration, the image edges will be blurred. To preserve edges while suppressing staircase effects, the variational model with TGV regularization was presented for dealing with Poisson image restoration problem [16]. Recently, researchers have proposed a directional TGV regularization for Poisson image restoration [13].
In some special images, the textures show obvious directionality. Therefore, the proposed model should be built to describe the geometric features of images. However, in the numerical calculation, the finite difference scheme of gradient for the TV-based models only depends on the horizontal and vertical directions. In this case, these models based on this scheme cannot couple with local structures of images efficiently. Fortunately, the directional total variation (DTV) method could overcome this shortcoming by adding a rotation matrix to rotate the gradient operator [3], [21], [29], [41]. Besides, this method adds a weighted matrix to enhance the diffusion of the corresponding Euler–Lagrange equation. Specifically, the DTV-based Poisson restoration model can be formulated as follows:where the matrix and the rotation matrix are defined byHere is a weight parameter and is the affine angle. As reported in Bayram and Kamasak [3], for restoring images with a dominant direction, the DTV-based model has better performance than the classical TV-based models. However, the DTV-based model does not perform very well when restoring images with complex structures [29]. In other words, when the image has several dominant directions, the DTV-based model will not achieve the ideal restoration effect. This fact motivates us to study some new schemes. In a recent work, Pang et al. proposed an adaptive weighted TV regularization for Gaussian noise removal [29]. Inspired by the model proposed in Pang et al. [29], in this paper, we propose a Poisson restoration model based on an adaptive weighted directional regularization (AWDTV) to describe the local structures of the image. The major contributions of the paper are three-fold:
- •
To the best of our knowledge, this is the first time that AWDTV is considered for Poisson noise. Owing to the combination of the rotation matrix, the adaptive weighted matrix and the -quasinorm, the proposed model has more robust adaptivity and more stronger restoration abilities.
- •
Since the proposed model is non-smooth and non-Lipschitz, an ADMM is presented for solving it, where the convergence of the relevant subproblem is guaranteed. The complexity analysis of the proposed method is also given.
- •
Experiments including natural images and synthetic images are conducted under different blurs and noise levels to demonstrate that our AWDTV has superior performance over the state-of-the-art methods in Poisson image restoration.
This paper is organized as follows. In Section 2, the AWDTV-based Poisson restoration model is proposed. Section 3 presents the alternating direction method of multipliers for solving this model. In Section 4, experiments are carried out to show the effectiveness of our method, compared with other three models for Poisson image restoration. Finally, Section 5 concludes the paper.
Section snippets
Our proposed model
In this section, we first introduce the definition of adaptive weighted DTV -quasinorm (AWDTV). Then we propose a Poisson restoration model based on the AWDTV.
Generally speaking, different regions of an image have different structures. To describe them, it is very important to keep the diffusion of the corresponding Euler–Lagrange equation along with the tangential direction of the edge. Consequently, we need to adaptively rotate the gradient to fit in with the tangential direction. To this
Algorithm
In this section, we will present an ADMM for solving the proposed model (2.3). The solution of the original problem is transformed into alternating calculation of several related subproblems. Firstly, we need to focus on how to solve the problem, since one of the subproblems has this form.
Numerical experiments
In this section, we give some experimental results to show the performance of AWDTV. Compared with the TV model [33], the HOTV model [43] and the DTV model (1.5) in Poisson image restoration, numerical results and images will prove the superiority of our AWDTV model. In fact, the DTV model is also our new model which has some advantages on restoring images with a dominate direction. All the numerical experiments are conducted in Matlab environment on a PC with 2.30 GHz Intel(R) Core(TM) i5
Conclusion
In this paper, we proposed a novel Poisson restoration model based on an adaptive weighted directional regularization. Furthermore, we presented an efficient ADMM to solve the proposed model. One of the subproblems was solved by the half-quadratic algorithm with guaranteed convergence. Experimental results demonstrated the superior performance of the proposed approach over other three competing ones. In the future, we will explore some acceleration techniques to further reduce the
CRediT authorship contribution statement
Jun Zhang: Conceptualization, Methodology, Software, Formal analysis, Writing – review & editing. Pengcheng Li: Software, Investigation, Data curation, Writing – original draft. Junci Yang: Visualization. Mingxi Ma: Validation, Supervision. Chengzhi Deng: Writing – review & editing.
Declaration of Competing Interest
Authors declare that they have no conflict of interest.
Acknowledgments
This work was supported by the Science Foundation for Post Doctorate of China (2020M672484), the Natural Science Foundation of Jiangxi Province (20192BAB211005), and the NNSF of China (61865012), the Guangxi Natural Science Foundation Program (2018GXNSFAA138056).
References (46)
- et al.
ACQUIRE: an inexact iteratively reweighted norm approach for TV-based Poisson image restoration
Appl. Math. Comput.
(2020) - et al.
High-order total variation-based Poissonian image deconvolution with spatially adapted regularization parameter
Appl. Math. Model.
(2017) - et al.
Deblurring Poisson noisy images by total variation with overlapping group sparsity
Appl. Math. Comput.
(2016) - et al.
Nonlinear total variation based noise removal algorithms
Phys. D
(1992) - et al.
Deblurring Poissonian images by split Bregman techniques
J. Vis. Commun. Image Represent.
(2010) - et al.
Nonconvex and nonsmooth total generalized variation model for image restoration
Signal Process.
(2018) - et al.
High-order total bounded variation model and its fast algorithm for Poissonian image restoration
Math. Probl. Eng.
(2019) - et al.
Nonlinear multigrid method for solving the LLT model
Appl. Math. Comput.
(2013) - et al.
Proximal alternating minimization and projection methods for nonconvex problems: an approach based on the kurdyka-lojasiewicz inequality
Math. Oper. Res.
(2010) - et al.
Variance stabilization in Poisson image deblurring
IEEE 14th International Symposium on Biomedical Imaging (ISBI 2017)
(2017)