Coupling local and nonlocal diffusion equations for image denoising
Introduction
In this paper, we investigate the properties of the evolution equation and discuss its applications in image restoration under Gaussian noise. Here () is a bounded domain with smooth boundary , is the unit outward normal on , , , is a constant and . The kernel is a nonnegative continuous radial function with compact support, , and . Given a noisy image , the solution is the restored image with scale variable .
Eq. (1) is coupled with a weighted -Laplacian equation and a weighted nonlocal -Laplacian equation. The -Laplacian equation has been widely investigated for image processing. If , this is the heat equation, whose solution is given as the convolution of the initial value with a Gaussian Kernel [1]. That is to say, image denoising by the heat equation is equivalent to the Gaussian filter and always causes over-smoothness in the restored images. For edge-preserving, it is straightforward to consider nonlinear equations. The total variation (TV) flow ( in Eq. (2)) [2], which comes from the gradient descent scheme of the famous ROF model [3], has been thoroughly studied in the past two decades. The main advantage of the TV flow for image denoising is the powerful ability to restore edges and homogeneous regions. However, it utilizes only the image’s local information, which means that it fails to preserve textured structures and repeated patterns in images.
The nonlocal diffusion equation was proposed as a remedy [4], [5], [6]. It is often referred to as the nonlocal -Laplacian equation, since the solutions of Eq. (3) converge strongly in to the solution of Eq. (2) with homogeneous Neumann boundary condition if the kernel is rescaled appropriately [7]. In image denoising problems, the kernel represents the nonlocal weight that measures the similarity between two patches [8]. Consequently, Eq. (3) avoids the drawbacks of Eq. (2) and restores textures and details in images effectively. It is noticed that a main difference between the usual -Laplacian equation (2) and the nonlocal -Laplacian equation (3) is the lack of regularizing effect [9] of the latter. As a result, Eq. (3) causes residual noise in homogeneous regions of the restored images.
Inspired by the local method and the nonlocal method mentioned above, we propose Eq. (1) to reduce artifacts of each method for image denoising. The local -Laplacian equation and the nonlocal -Laplacian equation in (1) are coupled through a space varying function . In regions with rich textures and repeated structures, we let , which leads to nonlocal diffusion In regions with cartoon structures, we let , which leads to If for any (recall that has a compact support), then the nonlocal term in the above disappears and it becomes local diffusion. At last, shall be used for complex regions.
In the nonlocal part of Eq. (1), the diffusion coefficient is constructed as . The authors of [10] suggested using as the weight for nonlocal equations. As we shall see in Theorem 9, both of them are well-defined in the sense that the nonlocal problem converges to the local problem if the kernel is rescaled appropriately. Note that our construction has the nice symmetric property Then we could define weak solutions for Eq. (1) straightforwardly. Besides, our construction is a more appropriate approach for digital images, since is not necessary on the grid.
The weighted -Laplacian equation in Eq. (1) has a smoothing effect. It can be replaced by other weighted nonlinear diffusion equations for image denoising. In this paper, we also consider the evolution equation which is coupled with the weighted regularized Perona–Malik equation [11] and the weighted linear nonlocal equation. Notice that the regularized Perona–Malik equation has a better smoothing effect than the -Laplacian equation and the nonlocal -Laplacian equation (3) performs the best for image denoising when . Eq. (5) outperforms Eq. (1) for image denoising.
To finish the introduction, we briefly review some related works. Coupling local and nonlocal methods for image denoising have been proposed in [12], [13]. The variational model combines the total variation regularization term and a nonlocal data fidelity term to reduce the undesirable artifacts caused by each method. The model can be viewed as the adaptive regularization of the nonlocal means filter [8]. Recently, qualitative properties for the gradient flow of has been studied in [14], where . Notice that Eq. (1) is the gradient flow of the energy functional If and , then it follows from (4) that functional (8) coincides with (7).
The rest of this paper is organized as follows. In Section 2 we define solution spaces and state main results. In Section 3 we prove the existence and uniqueness of weak solutions for Eq. (1). We also prove that solutions of Eq. (1) converge to the solution of the usual -Laplacian equation if the kernel is rescaled appropriately. Section 4 is devoted to the applications of Eqs. (1), (5) for image denoising. We conclude the paper in Section 5.
Section snippets
Preliminaries and main results
The weights and are allowed to vanish in sets of positive measure. This causes difficulties in analyzing the properties of Eq. (1). In this paper, we consider a specific class of coefficients called the Muckenhoupt weight that has been discussed for image processing in [15]. More precisely, we assume that where , are two smooth subdomains of .
Proofs of main results
For the proof of Theorem 7, let us consider the auxiliary equation where is a fixed constant and .
Lemma 10 Eq. (15) admits a uniqueness weak solution , such that and holds for any and any . Besides,
Numerical demonstrations
In this section, we present the finite difference schemes for Eqs. (1), (5) and show the image denoising results of them. As we have mentioned in the introduction, the nonlocal -Laplacian equation (3) performs the best for image denoising when . Consequently, the linear nonlocal equation is utilized in Eq. (1). The radial kernel in both Eqs. (1), (5) should be replaced by a symmetric nonlocal weight with and .
Conclusion
A local and nonlocal coupling diffusion equation has been proposed to reduce the artifacts created by the local method and the nonlocal method. The existence and uniqueness of weak solutions for the equation have been discussed in a space involving the Muckenhoupt weight. Numerical experiments have been presented comparing the performance of the proposed equation with that of TV flow, nonlocal TV flow, and a hybrid method.
Acknowledgments
The author would like to thank the referees for the valuable suggestions and comments. This work was supported by the National Natural Science Foundation of China (12001509) and the Natural Science Foundation of Zhejiang Province, China (LQ21A010010).
References (20)
- et al.
Nonlinear total variation based noise removal algorithms
Phys. D
(1992) - et al.
A nonlocal -Laplacian evolution equation with Neumann boundary conditions
J. Math. Pures Appl.
(2008) - et al.
Asymptotic behavior for nonlocal diffusion equations
J. Math. Pures Appl.
(2006) - et al.
Cauchy problem of a nonlocal -Laplacian evolution equation with nonlocal convection
Nonlinear Anal. TMA
(2014) Fast noise variance estimation
Comput. Vis. Image Underst.
(1996)- et al.
Mathematical Problems in Image Processing: Partial Differential Equations and the Calculus of Variations, Vol. 147
(2006) - et al.
Minimizing total variation flow
Differential Integral Equations
(2001) - et al.
Deblurring and denoising of images by nonlocal functionals
Multiscale Model. Simul.
(2005) - et al.
Nonlocal operators with applications to image processing
Multiscale Model. Simul.
(2009) - et al.
Nonlocal linear image regularization and supervised segmentation
Multiscale Model. Simul.
(2007)
Cited by (5)
Hybrid variable exponent model for image denoising: A nonstandard high-order PDE approach with local and nonlocal coupling
2024, Journal of Mathematical Analysis and ApplicationsCOUPLING LOCAL AND NONLOCAL FOURTH-ORDER EVOLUTION EQUATIONS FOR IMAGE DENOISING
2023, Inverse Problems and ImagingInitial state reconstruction on graphs
2023, Contemporary MathematicsInitial State Reconstruction on Graphs
2022, arXiv