A NOVEL NONLINEAR COUPLED NONLOCAL REACTION–DIFFUSION

SYSTEM FOR IMAGE RESTORATION

YUHANG LI, ZHICHANG GUO, JINGFENG SHAO, BOYING WU

Abstract. A novel nonlinear coupled nonlocal reaction-diffusion system is proposed for image restora-
tion, characterized by the advantages of preserving low gray level features and textures. The gray level
indicator in the proposed model is regularized using a new method based on porous media type equa-
tions, which is suitable for recovering noisy blurred images. The well-posedness, regularity, and other
properties of the model are investigated. Numerical experiments conducted on texture and satellite
images demonstrate the effectiveness of the proposed model in denoising and deblurring tasks.

1. Introduction
In many image processing applications, the acquired images are often blurred and corrupted by noise.
The degradation model of an image can typically be formulated as
f = Ku + n,
where K is some linear operator, especially convolution operator, n is noise, f is the observed image,
and u is the original image. In this paper, we mainly focus on the non-blind image restoration problem,
which aims to recover a clean image u from a noisy blurred image f when K is known exactly. This can
be achieved by minimizing the following energy functional:
λ
E(u) = Φ(u) + kKu − f k22 , (1.1)
2
where kKu − f k2 is called the fidelity term and Φ(u) is called the regularization term, λ ⩾ 0 is a
parameter to balance the fidelity term and regularization term. Φ(u) is generally chosen as G(|∇u|2 ) in
earlier methods, yielding the gradient flow corresponding to (1.1) as
∂u    
= div g |∇u| ∇u − λK ′ (Ku − f )
2
(1.2)
∂t
is a reaction-diffusion equation, where K ′ is the adjoint operator of K, g(s2 ) = G′ (s2 ). When λ = 0, (1.2)
reduces to a pure diffusion equation, which can be employed for image denoising. Many models have been
1
proposed for image restoration. In the well-known ROF model [1, 2], g(s2 ) = |s| , which results in total
variation (TV) diffusion in (1.2). The use of TV can effectively preserve edges but easily cause a staircase
effect. Various improved versions [3–5] have been proposed to eliminate this effect. Another popular
class of choices is to use anisotropic diffusivity g(s2 ) = 1+k11 s2 , leading to Perona-Malik diffusion [6, 7] in
(1.2). Perona-Malik diffusion is a forward-backward diffusion process which is known to sharpen edges
effectively but cause staircase instability. To reduce the noise sensitivity of Perona-Malik diffusion [8],
various regularization methods [9–13] can be utilized to enhance its stability. Higher order regularization
terms and PDEs have been proposed to overcome the staircase effect of edge-preserving second-order

Key words and phrases. Nonlinear nonlocal parabolic equations, Nonlinear nonlocal parabolic equations, Fractional deriva-
tives, Image restoration, Maximal regularity.
1
2 YUHANG LI, ZHICHANG GUO, JINGFENG SHAO, BOYING WU

PDEs. Some early works [14, 15] indicated that methods based on fourth-order PDEs can effectively
reduce the staircase effect but encourage piecewise planar solutions, with modification can be found in
[16–19].
In natural images, there are often more complex structures such as textures and other repetitive fea-
tures, which impose higher demands on image restoration models. Nonlocal operators have been defined
in [20, 21] to extend the nonlocal method to the variational framework. The nonlocal total variation
(NLTV) regularization [21, 22] remains one of the most popular deblurring methods for recovering tex-
ture to date. Other image restoration methods based on nonlocal functionals or PDEs [23–26] have also
achieved success in preserving textures. Fractional derivative-based variational and PDE models were
proposed in the past two decades for image processing due to the nonlocal property of the fractional
derivative. Bai et al. [27] proposed a fractional diffusion equation formulated as
∂u    
2
= divα g |∇α u| ∇α u , (1.3)
∂t
where 1 < α < 2, ∇α denotes the fractional gradient operator and − divα denotes the adjoint operator
of ∇α . This model was proposed for image denoising and can be viewed as generalizations of second-
order and fourth-order anisotropic diffusion equations. It has been demonstrated to effectively eliminate
the staircase effect. Zhang et al. [28] established a comprehensive total α-order variation framework
for image restoration, where α can take any positive value. This framework along with other fractional
derivative-based models [29–31] demonstrates the texture preservation capability in image restoration.
In [32], a doubly degenerate (DD) diffusion model was proposed as a novel image restoration framework,
originally utilized for multiplicative noise removal. The framework was described as
∂u
= div (c (|∇u| , u) ∇u) − λh(f, u), (1.4)
∂t
where c(|∇u| , u) = a(u)b(|∇u|), a(u) and b(|∇u|) are referred to as the gray level indicator and edge
detector, respectively, while h(u, f ) represents certain source terms derived from variational models. The
gray level indicator is used to control the diffusion speed at different gray levels and the edge detector
is employed to preserve image edges; various options are available for selecting both. Numerous PDE-
based image restoration models can be incorporated into this framework. Yao et al. generalized the DD
model to fractional order, initially applied it for multiplicative noise removal [33], and later extended its
application to deblurring tasks [34]. The model takes the form
∂u
= divα (c (|∇α u| , u) ∇α u) − λK ′ (Ku − f ), (1.5)
∂t
where 0 < α < 2, c(|∇α u| , u) = a(u)b(|∇α u|). The gray level indicator is chosen as
 γ
|u|
a(u) = , (1.6)
M
1
where M = sup u(x), γ > 0. On the other hand, b(|∇α u|) = 1+k1 |∇α u|β
can be seen as a texture detection
x∈Ω
function, where k1 > 0, β > 0. The design of the diffusion coefficient in this model emphasizes the
preservation of low gray level features and texture information. Experimental results have demonstrated
the effectiveness of the model in restoring texture-rich images. However, the model (1.5) still faces two
challenges: Firstly, since the observed image is always noisy or noisy blurred, the gray level indicator (1.6)
may result in a large difference in diffusion speeds at neighboring pixels in homogeneous regions, thereby
reducing the visual quality of the restored image. Secondly, the presence of fractional-order divergence
makes the model violates the Fick’s law, which complicates its physical interpretation and increases
 A NOVEL NONLINEAR COUPLED NONLOCAL REACTION–DIFFUSION SYSTEM FOR IMAGE RESTORATION 3

computational costs. In [35], a model based on fractional Fick’s law was proposed for multiplicative noise
removal, attempting to address the challenges mentioned above. The model takes on the form
∂u
= div (c (|∇α u| , uσ ) ∇α u) , (1.7)
∂t
where 0 < α ⩽ 3, c(|∇u| , uσ ) = a(uσ )b(|∇α u|). The authors chose the gray level indicator as
 γ
|uσ |
a(uσ ) = , (1.8)
Mσ
where uσ = Gσ ∗ u, Gσ is a Gaussian convolution kernel, Mσ = sup uσ (x), σ, γ > 0. Gaussian filters
x∈Ω
can reduce the interference of noise on image features, enabling the gray level indicator to perform more
effectively in controlling the diffusion speed. Similar operations have also been observed in previous
works [36–38]. The selection of the texture detection function is similar to that in (1.5). This model
possesses has a more comprehensive physical background and experimental results have demonstrated
that it achieves superior image restoration performance and fast computational speeds.
The motivation of this paper is to propose an anisotropic diffusion model from a clear physical back-
ground, with the aim of preserving low gray level image features and texture details in denoising and
deblurring tasks on both texture-rich images and natural images. A simple idea is to consider the equation
∂u
= div (c (|∇α u| , uσ ) ∇u) − λK ′ (Ku − f ), (1.9)
∂t
and select the same gray level indicator as in (1.7). However, this is not suitable for restoring noisy
blurred images because the Gaussian filter makes the already blurred image over smooth (uσ belongs to
the C ∞ class), which would result in a rough gray level indication and decrease the visual quality of the
restored image. This requires us to adopt a ”gentler” approach to regularize u in (1.6). The gray level
indicator in (1.6) was initially inspired by the gamma correction [32, 39], but we now provide another
interpretation. In DD model, the gray level indicator ensures that the diffusion speed is slow when u
is small and fast when u is large, which is the characteristic of porous medium equations (PMEs), also
known as Newtonian filtration slow diffusion [40]. In fact, if we set a(u) = |u|
γ

M γ , b(·) ≡ 1, h(f, u) = 0 in
the DD model, and taking M as a positive constant, such as M = sup f (x), (1.4) becomes a PME:
x∈Ω
−γ −γ
∂u M M
= div (|u|γ ∇u) = ∆|u|γ+1 . (1.10)
∂t γ+1 γ+1
This type of equation appears in different contexts, such as in the flow of gases through porous media
[41, 42], in the heat conduction at high temperatures [43] and in groundwater flow [44]. For the case
where b(·) is chosen as a general edge detection operator, the equation
 γ 
∂u |u|
= div b(|∇u|)∇u
∂t Mγ
can be viewed as an anisotropic PME.
Inspired by the analysis above, we consider introducing a variable v into the gray value indicator,
designing it to satisfy a type of slow diffusion equation and ensuring it has the same initial data as u,
so that it gradually becomes more regular during the diffusion process, rather than always remaining
smooth like uσ in (1.8). The diffusion equation for denoising and deblurring is written as
!
∂u |v|γ 1
= div ∇u − λK ′ (Ku − f ). (1.11)
∂t M γ 1 + k1 |∇α u|β
4 YUHANG LI, ZHICHANG GUO, JINGFENG SHAO, BOYING WU

To emphasize mutual transfer of information between the two equations, we design the diffusion equation
for v as
 µ   γ 
∂v |u| |v|
= λ1 div ∇v + (1 − λ 1 ) div ∇v , (1.12)
∂t Mµ Mγ
which exhibits characteristics of a PME, where 0 < λ1 < 1 and µ > 0, the diffusion speed is simultaneously
controlled by both u and v. As u becomes clearer during the restoration process governed by (1.11), it
facilitates positive mutual transfer of information within the coupled reaction-diffusion system formed by
(1.11) and (1.12). Besides, following the recommendation in [7], we adopt periodic boundary conditions
for this system.
Taking everything into account, we propose the following reaction diffusion system as an image restora-
tion model for some fixed positive T0 :

 ∂u

 = div (c (|∇α u| , v) ∇u) − λK ′ (Ku − f ), in (0, T0 ) × Ω,

 ∂t
 ∂v
= div (a(u, v)∇v) , in (0, T0 ) × Ω, (1.13)

 ∂t

 u, v periodic on Ω for (0, T0 ),


u(0, x) = v(0, x) = f (x), in Ω,

where
|v|γ 1
c (|∇α u| , v) =γ
b(|∇α u|), b(|∇α u|) = β
,
M 1 + k1 |∇α u|
|u|µ |v|γ
a (u, v) = λ1 µ + (1 − λ1 ) γ ,
M M


Ω ⊂ RN is an N -dimensional cube for N ⩾ 1, K ∈ L L1 (Ω), L2 (Ω) , M = sup f (x), and 0 < α < 1,
x∈Ω
β > 0, γ > 0, µ > 0, λ > 0, 0 < λ1 < 1, k1 > 0 are some given constants. The proposed model maintains
the characteristics similar to those of existing models: v can be regarded as a regularized version of the
observed image, when v is small, the diffusion speed at low gray value regions will be slow. The diffusion
speed is also controlled by the texture detection function b(|∇α u|), which becomes small when |∇α u|
is large, leading to the protection of the textures. Our model also possesses the following advantages:
the gentler regularization of u in the gray level indicator aligns the model more closely with the task of
recovering noisy blurred images; the model is based on a more classical diffusion framework, providing a
clearer physical background.
The main contribution of this paper lies in the theoretical analysis of the proposed model. Due to the
low regularity of the diffusion coefficient in system (1.13), it is challenging to establish its well-posedness
using classical fixed-point methods as done in [36–38]. Therefore, our work also serves as a theoretical
complement to equations of the type in [32–34]. In this paper, we employ the Maximal Regularity Theory
developed by Amann et al. to investigate the well-posedness of (1.13). The regularity and other properties
of solutions are also concerned. Additionally, we present a semi-implicit finite difference scheme for the
proposed model and validate its effectiveness in image restoration tasks using both texture and satellite
images.
The rest of this paper is organized as follows. Some mathematical preliminaries and our main result
are stated in Section 2. The main result and some properties of weak solutions are proven in Section
3. Regularity results for the proposed model are provided in Section 4. In Section 5, some numerical
examples are presented to demonstrate the effectiveness of our model.
 A NOVEL NONLINEAR COUPLED NONLOCAL REACTION–DIFFUSION SYSTEM FOR IMAGE RESTORATION 5

2. Mathematical Preliminaries and main result

In this section, we state some necessary preliminaries of the fractional gradient, fractional order Sobolev
spaces and maximal Lp -regularity which will be used below. One of the definitions of the fractional
gradient is based on the Fourier transform. The Fourier transform of a function u that is periodic on Ω
is given by Z
(Fu)(k) = e−2πik·x u(x)dx, k ∈ ZN ,
Ω
and for α ∈ (0, 1), the fractional gradient of u is defined by
 
∇α u = F −1 diag ((2π ik)α )k∈ZN Fu,
where diag [(mk )k∈ZN ] denotes the multiplier given by the sequence (mk )k∈ZN . The j-th component of
the fractional-order gradient ∇α u is defined as the partial derivative with order α of u with respect to
xj , denoted as ∂xj u.
In the following we will make essential use of two types of periodic Sobolev spaces of fractional order.
We refer to [45, Chapter 1] as well as [46, Section 5] and the references therein for more basic results of
general fractional order Sobolev spaces. Let U be an open set in RN and p ∈ [1, +∞). Denote by
Z Z  p1
|f (y) − f (x)|p
[f ]W s,p (U ) = dxdy
U U |x − y|
N +sp

the Gagliardo seminorm of a measurable function f in U for s ∈ (0, 1). Then for s ∈ (0, +∞)\N, the
periodic Slobodeckij spaces are the Banach spaces defined by
n o
Wπs,p (Ω) = u ∈ W ⌊s⌋,p (Ω) : u is periodic on Ω, [u]W s,p (Ω) < +∞

equipped with the norm

X
kukWπs,p (Ω) = kukLpπ (Ω) + [∂ α u]W s−⌊s⌋,p (Ω) .
|α|=⌊s⌋

The subscript π indicates that we are in a periodic function space. For p ∈ (1, +∞) and s ∈ (0, +∞), the
Bessel potential spaces are the Banach spaces defined by

Hπs,p (Ω) = u ∈ Lpπ (Ω) : kukHπs,p (Ω) < +∞ ,
where  
−1

2s 
kuk s,p
Hπ (Ω) = F diag 1 + |k| 2
Fu .
k∈ZN Lp
π (Ω)

We will apply Maximal Regularity Theory to the study of problem (1.13). For more details about this
d
theory, see [46–50] and the references therein. Let E0 and E1 be Banach spaces such that E1 ,− → E0 , i.e.
E1 is densely embedded in E0 . Suppose that 1 < p < +∞. Set E1− p1 = (E1 , E0 )1− p1 ,p , (·, ·)θ,p denoting
the standard real interpolation functor. Given T0 ∈ (0, +∞), for any T ∈ (0, T0 ], put
W 1,p (0, T ; (E1 , E0 )) = Lp (0, T ; E1 ) ∩ W 1,p (0, T ; E0 ).
An operator B ∈ L∞ (0, T ; L(E1 , E0 )) is said to possess the property of maximal Lp -maximal regularity
on [0, T ) with respect to (E1 , E0 ) if the map
 
du
W (0, T ; (E1 , E0 )) → L (0, T ; E0 ) × E1− p1 , u 7→
1,p p
+ Bu, u(0)
dt
6 YUHANG LI, ZHICHANG GUO, JINGFENG SHAO, BOYING WU

is a bounded isomorphism. The set of all such operators B is denoted by MRp (0, T ; (E1 , E0 )). Denote
by MRp (E1 , E0 ) the set of all C ∈ L(E1 , E0 ) such that the map [0, T ) → L(E1 , E0 ), t 7→ C belongs
0,1
to MRp (0, T ; (E1 , E0 )). Let X and Y be metric spaces. Denote by CVolt (X, Y ) the space of all maps
f : X → Y which satisfies the following conditions:
(i) f is bounded on bounded sets and uniformly Lipschitz continuous on such sets,
(ii) for each T ∈ (0, T0 ] and each u, v ∈ X with u|[0,T ) = v|[0,T ) , it follows that f (u)|[0,T ) = f (v)|[0,T ) .
0,1
Let Y1 and Y0 be Banach spaces such that Y1 ,→ Y0 . Denote by CVolt (X; Y1 , Y0 ) the set of all maps
0,1
g : X → Y0 such that g − g(0) ∈ CVolt (X, Y1 ).
Consider a quasilinear abstract Cauchy problem

 du
+ A(u)u = F (u), t ∈ (0, T0 ),
dt (2.1)
 u(0) = u .
0
1,p
A function u ∈ Wloc (0, T ; (E1 , E0 )) is said to be a solution of (2.1) on [0, T ) if it satisfies (2.1) in the
a.e. sense on (0, T ). A solution is said to be maximal if it cannot be extended to a solution on a strictly
larger interval.
The following existence and uniqueness result is applied to the study of problem (1.13).
Theorem 2.1. (See [47, Theorem 2.1]) Suppose that
0,1

(i) A ∈ CVolt W 1,p (0, T ; (E1 , E0 )), MRp (0, T ; (E1 , E0 )) ,
(ii) there exists q ∈ (p, +∞] such that
0,1

F ∈ CVolt W 1,p (0, T ; (E1 , E0 )), Lq (0, T ; (E1 , E0 )), Lp (0, T ; (E1 , E0 )) ,
(iii) u0 ∈ E1− p1 .
Then there exist a maximal Tmax ∈ (0, T0 ] and a unique solution u of (2.1) on [0, Tmax ).
Reformulating problem (1.13) as a weak Lp -formulation, the weak solution of (1.13) can be defined as
follows.
1− 2 ,p
Definition 2.2. Given f ∈ Wπ p (Ω), a couple of functions (u, v) is called a weak solution of problem
(1.13) on [0, T ), if it satisfies the following conditions:



(i) u, v ∈ Lploc 0, T ; Wπ1,p (Ω) ∩ Wloc1,p
0, T ; Wπ−1,p (Ω) ;
(ii) u(0, ·) = v(0, ·) = f ;
′
(iii) For any ϕ, ψ ∈ Wπ1,p (Ω), the following integral equalities hold:
Z Z Z
∂u
ϕdx + c (|∇α u| , v) ∇u · ∇ϕdx + λ (Ku − f )Kϕdx = 0,
Ω ∂t
ZΩ Z Ω
(2.2)
∂v
ψdx + a (u, v) ∇v · ∇ψdx = 0,
Ω ∂t Ω
for almost all t ∈ (0, T ).
Now we state the main result as follows.
  1− 2 ,p
Theorem 2.3. Assume that p ∈ 2+N 1−α , ∞ and β ⩾ 1. For any given 0 < f ∈ Wπ p (Ω), there exist a
unique maximal Tmax ∈ (0, T0 ] and a unique weak solution (u, v) of problem (1.13) on [0, Tmax ).
Utilizing the Maximum Regularity Theory, we prove Theorem 2.3 by incorporating the techniques
from [10, 18, 51]. The proof is given in the next section.
 A NOVEL NONLINEAR COUPLED NONLOCAL REACTION–DIFFUSION SYSTEM FOR IMAGE RESTORATION 7

3. Local well-posedness and some properties of weak solutions

In this section, we first establish the local well-posedness of problem (1.13) via maximum regularity.
For all ε > 0, let v̄ε = max{v, ε}. We consider the following auxiliary problem:

 ∂u

 = div (c (|∇α u| , v̄ε ) ∇u) − λK ′ (Ku − f ), in (0, T0 ) × Ω,

 ∂t
 ∂v
= div (a (u, v̄ε ) ∇v) , in (0, T0 ) × Ω, (3.1)

 ∂t

 u, v periodic on Ω for (0, T0 ),


u(0, x) = v(0, x) = f (x), in Ω,
Similar to Definition 2.2, we define the weak solution of the auxiliary problem (3.1).
1− 2 ,p
Definition 3.1. Given f ∈ Wπ p (Ω), a couple of functions (u, v) is called a weak solution of auxiliary
problem (3.1) on [0, T ), if it satisfies the following conditions:



(i) u, v ∈ Lploc 0, T ; Wπ1,p (Ω) ∩ Wloc 1,p
0, T ; Wπ−1,p (Ω) ;
(ii) u(0, ·) = v(0, ·) = f ;
′
(iii) For any ϕ, ψ ∈ Wπ1,p (Ω), the following integral equalities hold:
Z Z Z
∂u
ϕdx + c (|∇ u| , v̄ε ) ∇u · ∇ϕdx + λ (Ku − f )Kϕdx = 0,
α
(3.2a)
Ω ∂t
ZΩ Z Ω
∂v
ψdx + a (u, v̄ε ) ∇v · ∇ψdx = 0, (3.2b)
Ω ∂t Ω

for almost all t ∈ (0, T ).

Proof of Theorem  2.3. Firstly, we reformulate problem (3.1) as a quasilinear abstract Cauchy problem.
Since p ∈ 1−α , ∞ , by the almost reiteration property [48, Theorem V.1.5.3] and a classical embedding
2+N
 
theorem [45, Theorem 4.6.1(e)], for any δ ∈ 0, 1 − 2+N
p , we get
2
1− p ,p 2+N
Wπ (Ω) ,→ C 0,1−δ− p (Ω),
which implies that inf f > 0. Henceforth, we always set ε = inf f and suppose that 0 < T ⩽ T0 . In
x∈Ω x∈Ω
this section, we always choose
E0 = Wπ−1,p (Ω)2 , E1 = Wπ1,p (Ω)2 .
d
→ E0 . We denote by
For this choice it is well-known that E1 ,−
2
1− p ,p
E1− p1 = (E0 , E1 )1− 1 ,p = Wπ (Ω)2
p

the respective real interpolation space. To simplify the notation, we set

2
1− p ,p
= Wπ−1,p (Ω),
1/2 1/2 1/2
E0 E1 = Wπ1,p (Ω), E1− 1 = Wπ (Ω).
p

For fixed w0 = (u0 , v0 )T ∈ W 1,p (0, T ; (E1 , E0 )), let a bilinear form a11 [w0 ](·, ·) be given by
Z
′
a11 [w0 ](u, ϕ) := c (|∇α u0 | , v0ε ) ∇u · ∇ϕdx, u ∈ Wπ1,p (Ω), ϕ ∈ Wπ1,p (Ω).
Ω
8 YUHANG LI, ZHICHANG GUO, JINGFENG SHAO, BOYING WU

Then a linear differential operator A11 (w0 ) is naturally induced by a11 [w0 ], i.e. for almost all t ∈ (0, T ),
′
hA11 (w0 )u, ϕi(W −1,p (Ω),W 1,p′ (Ω)) = a11 [w0 ](u, ϕ), ∀ϕ ∈ Wπ1,p (Ω), .
π π

Actually, we have already defined an operator

  
A11 : W 1,p (0, T ; (E1 , E0 )) → L∞ 0, T ; L E1 , E0
1/2 1/2
, w 7→ A11 (w).

In the same way, for fixed w0 = (u0 , v0 )T ∈ W 1,p (0, T ; (E1 , E0 )), let a22 [w0 ](·, ·) be given by
Z
′
a22 [w0 ](v, ψ) := a (u, v̄ε ) ∇v · ∇ψdx, v ∈ Wπ1,p (Ω), ψ ∈ Wπ1,p (Ω),
Ω

then we can define A22 (w0 ) by

′
hA22 (w0 )v, ψi(W −1,p (Ω),W 1,p′ (Ω)) = a22 [w0 ](v, ψ), ∀ψ ∈ Wπ1,p (Ω)
π π

for almost all t ∈ (0, T ), and introduce another operator

  
A22 : W 1,p (0, T ; (E1 , E0 )) → L∞ 0, T ; L E1 , E0
1/2 1/2
, w 7→ A22 (w).

Now we define the operator matrix A given by

 
∞ A11 (w) 0
A:W 1,p
(0, T ; (E1 , E0 )) → L (0, T ; L (E1 , E0 )) , w 7→ .
0 A22 (w)
For fixed w = (u, v)T ∈ W 1,p (0, T ; (E1 , E0 )), we define
 
−λK ′ (Ku − f )
F :W 1,p
(0, T ; (E1 , E0 )) → L (0, T ; E0 ) ,
p
w 7→ . (3.3)
0
Then problem (3.1) can be written as a quasilinear abstract Cauchy problem

 dw
+ A(w)w = F (w), t ∈ (0, T0 ),
dt (3.4)
 w(0) = (f, f )T ,

Let us now see that the linearized operator matrix has the property of maximal Lp -regularity. It is
well known that for any s ∈ (α, 1),


∂xαj ∈ L Hπs,p (Ω), Hπs−α,p (Ω) , j = 1, 2, · · · , N.
Thealmost reiteration
 property and classical embedding theorem imply that when (s − α)p > N , for any
δ ∈ 0, s − α − p ,
N
 
p −δ
s−α− N
∂xj ∈ L Wπ (Ω), Cπ
α s,p
(Ω) , j = 1, 2, · · · , N. (3.5)

Taking s = 1 − 2
in (3.5) and using embedding theorem [48, Theorem III.4.10.2]
p
    
1/2 1/2 1/2
W 1,p 0, T ; E1 , E0 ,→ C [0, T ], E1− 1 (3.6)
p
  
1/2 1/2
we know that for any u ∈ W 1,p 0, T ; E1 , E0 , there exists ρ = 1 − α − Np+2 − δ ∈ (0, 1) such that


|∇α u| ∈ C [0, T ], Cπ0,ρ (Ω) .
 A NOVEL NONLINEAR COUPLED NONLOCAL REACTION–DIFFUSION SYSTEM FOR IMAGE RESTORATION 9

Since k1 > 0 and β ⩾ 1, the function b ∈ BU C ∞ [0, ∞). It follows from the property of Nemytskii
operators [52, Theorem 3.1] that
g 7→ b(i) (g), C 0,ρ0 (Ω) → C 0,ρ0 (Ω), i = 0, 1
is bounded and uniformly Lipschitz continuous on bounded sets for any ρ0 ∈ (0, 1). Therefore,


b (|∇α u|) ∈ C [0, T ], Cπ0,ρ (Ω) . (3.7)
  
1/2 1/2
For any v ∈ W 1,p 0, T ; E1 , E0 , by (3.6) again we know that
 
v̄εγ 0,γ (1−δ− 2+N
p )
∈ C [0, T ], C π (Ω) . (3.8)
Mγ
n  o
Let ρ1 = min ρ, min{1, γ} 1 − δ − 2+N p , then ρ1 ∈ (0, 1). Combining (3.7) with (3.8), it holds that

v̄εγ

c (|∇α u| , v ε ) =
γ
b(|∇α u|) ∈ C [0, T ], Cπ0,ρ1 (Ω) (3.9)
M
since Hölder spaces are algebras. It follows that
  
1/2 1/2
[t 7→ A11 (w(t))] ∈ C [0, T ], L E1 , E0 (3.10)

for any w = (u, v)T ∈ W 1,p (0, T ; (E1 , E0 )). On the other hand, there exists a constant C1 > 1 such that
for any x ∈ Ω, t ∈ (0, T ) and |ξ| = 1,
1
0< < c (|∇α u| , v ε ) |ξ|2 < C1 . (3.11)
C1
It follows from (3.10) and (3.11) and the weak generation theorem [46, Theorem 8.5] that for any t ∈ (0, T ),
1/2
−A11 (w(t)) is sectorial in E0 . To obtain the maximal Lp -regularity of A11 (w(t)), as pointed out in
[49, Section 13], the only requirement in the end is sufficient regularity of the coefficients, with Hölder
continuity being adequate. Thus, for any t ∈ (0, T ),
 
1/2 1/2
A11 (w(t)) ∈ MRp E1 , E0 . (3.12)
    
1/2 1/2
Similarly, let ρ2 = min{γ, µ, 1} 1 − δ − 2+Np , then ρ 2 ∈ (0, 1). For any u ∈ W 1,p
0, T ; E 1 , E 0 ,
using (3.6) again we deduce that
|u|µ v̄εγ

a (u, v̄ε ) = λ1 + (1 − λ 1 ) ∈ C [0, T ], Cπ0,ρ2 (Ω) (3.13)
Mµ Mγ
and that   
1/2 1/2
[t 7→ A22 (w(t))] ∈ C [0, T ], L E1 , E0 (3.14)
for any w = (u, v)T ∈ W 1,p (0, T ; (E1 , E0 )). There also exists a constant C2 > 1 such that for any x ∈ Ω,
t ∈ (0, T ) and |ξ| = 1,
1
0< < a (u, v̄ε ) |ξ|2 < C2 , (3.15)
C2
then (3.13) and (3.15) imply that
 
1/2 1/2
A22 (w(t)) ∈ MRp E1 , E0 (3.16)
10 YUHANG LI, ZHICHANG GUO, JINGFENG SHAO, BOYING WU

for any t ∈ (0, T ). It follows from (3.12) and (3.16) that for all t ∈ (0, T ) and each G ∈ Lp (0, T ), the
linear abstract Cauchy problem

 dz
+ A(w(t))z = G(τ ), τ ∈ (0, T ),
dτ
 z(0) = (0, 0)T

has exactly one strong Lp solution on (0, T ), that is

A(w(t)) ∈ MRp (E1 , E0 ) . (3.17)
Using [50, Theorem 7.1] in combination with (3.10), (3.14) and (3.17) we obtain that
A(w) ∈ MRp (0, T ; (E1 , E0 ))
for any w ∈ W 1,p (0, T ; (E1 , E0 )).
Next, we establish Lipschitz estimates of the operator matrix. For R > 0, take w1 = (u1 , v1 )T , w2 =
(u2 , v2 )T ∈ BW 1,p (0,T ;(E1 ,E0 )) (0, R), then
kA(w1 ) − A(w2 )kL∞ (0,T ;L(E1 ,E0 ))
= sup sup kA(w1 (t))w − A(w2 (t))wkE0
t∈[0,T ] w=(u,v)T
∥w∥E1 =1

= sup sup kA11 (w1 (t))u − A11 (w2 (t))ukE 1/2

t∈[0,T ] ∥u∥ 1/2 =1
0
E1

+ sup sup kA22 (w1 (t))v − A22 (w2 (t))vkE 1/2

t∈[0,T ] ∥v∥ 1/2 =1
0
E1
Z
= sup sup sup (c (v1ε (t), |∇α u1 (t)|) − c (v2ε (t), |∇α u2 (t)|)) ∇u · ∇ϕdx
t∈[0,T ] ∥u∥ 1,p =1 ∥φ∥ 1,p′ =1 Ω
Wπ (Ω) Wπ (Ω)
Z
+ sup sup sup (a (u1 (t), v1ε (t)) − a (u2 (t), v2ε (t))) ∇v · ∇ψdx
t∈[0,T ] ∥v∥ 1,p =1 ∥ψ∥ 1,p′ =1 Ω
Wπ (Ω) Wπ (Ω)

1
⩽ sup γ
kv1ε (t)γ b (|∇α u1 (t)|) − v2ε (t)γ b (|∇α u2 (t)|)k∞
t∈[0,T ] M
λ1 µ µ 1 − λ1
+ sup k|u1 (t)| − |u2 (t)| k∞ + sup kv1ε (t)γ − v2ε (t)γ k∞
M µ t∈[0,T ] M γ t∈[0,T ]
:= I1 + I2 + I3 .
We estimate I1 , I2 and I3 separately. From now on C = C(R) denotes a positive constant which can take
different value in different places.
I1 ⩽ C sup kv1ε (t)γ k∞ kb (|∇α u1 (t)|) − b (|∇α u2 (t)|)k∞
t∈[0,T ]

+ C sup kb (|∇α u2 (t)|)k∞ kv1ε (t)γ − v2ε (t)γ k∞

t∈[0,T ]

⩽ C sup k|∇α u1 (t)| − |∇α u2 (t)|k∞ + C sup kv1ε (t) − v2ε (t)k∞
t∈[0,T ] t∈[0,T ]

X
N
⩽ C sup ∂xαj (u1 (t) − u2 (t)) 0,ρ
+ C sup kv1 (t) − v2 (t)kCπ0,ρ (Ω)
t∈[0,T ] j=1 Cπ (Ω) t∈[0,T ]
 A NOVEL NONLINEAR COUPLED NONLOCAL REACTION–DIFFUSION SYSTEM FOR IMAGE RESTORATION 11

X
N
⩽C ∂xαj ( ) sup ku1 (t) − u2 (t)k
1− 2 −δ,p 1−α− 2 −δ,p 1− 2 ,p
p p p
j=1 L Hπ (Ω),Hπ (Ω) t∈[0,T ] Wπ (Ω)

+ C sup kv1 (t) − v2 (t)k 1− 2 ,p

p
t∈[0,T ] Wπ (Ω)

⩽ C ku1 − u2 k ( ) + C kv1 − v2 k ( ) .
1− 2 ,p 1− 2 ,p
p p
C [0,T ],Wπ (Ω) C [0,T ],Wπ (Ω)

Similarly, we have
I2 ⩽ C ku1 − u2 k ( ) , I3 ⩽ C kv1 − v2 k ( ) .
1− 2 ,p 1− 2 ,p
p p
C [0,T ],Wπ (Ω) C [0,T ],Wπ (Ω)

Combining the estimates of I1 , I2 and I3 , we obtain that

 
 ( ) ( )
kA(w1 ) − A(w2 )kL∞ (0,T ;L(E1 ,E0 )) ⩽ C ku1 − u2 k + kv1 − v2 k 
1/2 1/2
C [0,T ],E 1 C [0,T ],E 1
1− 1−
p p

= C kw1 − w2 k ( )
C [0,T ],E1− 1
p

⩽ C kw1 − w2 kW 1,p (0,T ;(E1 ,E0 )) ,

from which it follows that
0,1

A ∈ CVolt W 1,p (0, T ; (E1 , E0 )), MRp (0, T ; (E1 , E0 )) . (3.18)

Now we check the Lipschitz property of the source term. It is clear that F (0) = (λK ′ f, 0) ∈
T

L (0, T ; E0 ). Set F̃ = F − F (0) = (−λK ′ K, 0) . For R > 0, take w1 = (u1 , v1 )T , w2 = (u2 , v2 )T ∈

∞ T

BW 1,p (0,T ;(E1 ,E0 )) (0, R), we have that

F̃ (w1 ) − F̃ (w2 ) ⩽ sup F̃ (w1 (t)) − F̃ (w2 (t))

L∞ (0,T ;E0 ) t∈[0,T ] E0
′
⩽ λ sup kK K (u1 (t) − u2 (t))k∞
t∈[0,T ]

⩽ C sup ku1 (t) − u2 (t)k 1− 2 ,p

p
t∈[0,T ] Wπ (Ω)

⩽ C kw1 − w2 k ( )
C [0,T ],E1− 1
p

⩽ C kw1 − w2 kW 1,p (0,T ;(E1 ,E0 )) ,

which implies that


0,1
F ∈ CVolt W 1,p (0, T ; (E1 , E0 )); L∞ (0, T ; E0 ) , Lp (0, T ; E0 ) . (3.19)
It follows from (3.18), (3.19) and Theorem 2.1 that there exist a maximal Tmax ∈ (0, T0 ] and a unique
1,p
solution w = (u, v)T ∈ Wloc (0, Tmax ; (E1 , E0 )) of the quasilinear abstract Cauchy problem (3.4). Then
(u, v) is the unique weak solution of auxiliary problem (3.1) on [0, Tmax ).
It is an easy consequence of the minimum principle that v satisfies
v(x, t) ⩾ inf f = ε a.e. x ∈ Ω,
x∈Ω
12 YUHANG LI, ZHICHANG GUO, JINGFENG SHAO, BOYING WU

for every t ∈ (0, Tmax ), which implies that v = v̄ε . Hence, (u, v) is also the unique weak solution of
problem (1.13), which concludes the proof. □

A general continuity result for quasilinear parabolic problems implies the continuous dependence of
the maximal existence time of solutions on both the initial data and the right-hand side. Utilizing this
continuity result, we can derive the following proposition of problem (1.13).
Proposition 3.2. Let the assumptions of Theorem 2.3 hold. Given τ ∈ (0, T0 ] and c > 0, there exists
r = r(τ ) > 0 such that for each f ∈ B 1− p2 ,p (c, r), the weak solution of problem (1.13) exists on [0, τ ].
Wπ (Ω)

Proof. The constant functions (u, v) ≡ (c, c) are clearly stationary solutions of problem (1.13). Set
w0 = (c, c)T , then for any τ ∈ (0, T0 ], there exists R > 0 such that kw0 kW 1,p (0,τ ;(E1 ,E0 )) < R. It
follows from [47, Theorem 3.1] that there exists r > 0 such that for any (f, f )T ∈ BE1− 1 (w0 , 2r), the
p
corresponding maximal interval of existence contains [0, τ ]. This fact concludes the proof. □

For a general K, it is challenging to employ the Stampacchia’s truncation method to establish the
maximum and minimum principles satisfied by u. However, Theorem 2.3 implies the local solution u is
actually bounded. Using the Moser’s method, we establish a specific L∞ estimate for u.
Proposition 3.3. Let the assumptions of Theorem 2.3 hold. Let (u, v) be the weak solution of problem
(1.13) on [0, T ) with the initial data f , then for every t ∈ (0, T ),
ku(t, ·)k∞ ⩽ e2λt (kf k∞ + kK ′ f k∞ ) . (3.20)
Proof. Since u(t) is Hölder continuous, we can take ϕ = u2r−1 (t) in (2.2) with r ⩾ 1. Then we have
Z Z Z Z
1 d
u2r (t)dx + (2r − 1) c (|∇α u| , v) u2(r−1) |∇u|2 + λ K ′ Ku · u2r−1 dx − λ K ′ f · u2r−1 dx = 0.
2r dt Ω Ω Ω Ω

Applying Young’s and Hölder’s inequalities, we obtain that

Z Z
1 d λ ′ 2r − 1
ku(t, ·)kL2r (Ω) ⩽
2r
|K Ku| dx +
2r
λ |u|2r dx
2r dt 2r Ω 2r
Z Z Ω
λ ′ 2r 2r − 1
+ |K f | dx + λ |u|2r dx
2r Ω 2r Ω
λ 4r − 2 + kKk2 |Ω|
⩽ kK ′ f k2rL2r (Ω) + λku(t, ·)k2rL2r (Ω) .
2r 2r
It follows from Grönwall’s inequality that
 
′
ku(t, ·)k2r
2r
L (Ω) ⩽ exp{(4r − 2 + kKk 2
|Ω|)λt} ku(0, ·)k 2r
2r
L (Ω) + λtkK f k 2r
L (Ω) .
2r

Using the monotonicity of `p norms, we arrive the estimate

   
kKk2 |Ω| − 2
ku(0, ·)kL2r (Ω) + (λt) 2r kK ′ f kL2r (Ω)
1
ku(t, ·)kL2r (Ω) ⩽ exp 2+ λt
2r
which yields, as r → +∞, the estimate (3.20) holds. □

Now we apply the well-known generalized principle of linearized stability to conclude the following
stability result for the equilibria of (1.13). This result pertains to the case of λ = 0, which leads to a
pure diffusion system.
 A NOVEL NONLINEAR COUPLED NONLOCAL REACTION–DIFFUSION SYSTEM FOR IMAGE RESTORATION 13

Theorem 3.4. Let the assumptions of Theorem 2.3 hold. For any given c > 0, there exists r > 0 such
that the unique weak solution (u, v) of problem (1.13) without the source term and with initial value
2
1− p ,p
f ∈B 1− 2 ,p (c, r) exists globally and converges exponentially in the topology of Wπ (Ω)2 to some
Wπ p (Ω)
equilibrium (u∞ , v∞ ) of (1.13) as t → ∞.
1− 2 ,p
Proof. It is sufficient to prove that the equilibrium (c, c)T is normally stable in Wπ p (Ω)2 when λ = 0.
For a definition of normally stable, see [53]. Let E denote the set of equilibrium of (3.4), which means
′
that w = (u, v)T ∈ E if and only if w ∈ E1 and A(w)w = 0 = F (w), i.e. for any ϕ, ψ ∈ Wπ1,p (Ω),
Z Z
c (|∇α u| , v) ∇u · ∇ϕdx = a (u, v) ∇v · ∇ψdx = 0. (3.21)
Ω Ω
It is clear that E0 := R2 ⊂ E. At each w∗ = (c, c)T ∈ E0 , the linearization of A is given by
 
∗ d ∗ −c11 ∆ 0
A0 := A(w ) + [A(w)w ] =
dw w=w∗
0 −c22 ∆
where c11 = c11 (c), c22 = c22 (c) > 0, ∆ : Wπ1,p (Ω) ⊂ Wπ−1,p (Ω) → Wπ−1,p (Ω). Since the spectrum of −∆
is discrete and consists only of non-negative eigenvalues, we know that
σ(A0 )\{0} ⊂ C+ = {z ∈ C : Rez > 0}
,→ Wπ−1,p (Ω) that A0 has compact resolvent, which
and {0} is isolated in σ(A0 ). It follows from Wπ1,p (Ω) ,−
implies that A0 generates an eventually compact semigroup since −∆ is sectorial. By [54, Corollary 5.3.2],
we can conclude that 0 is a pole of R(·, A0 ).
To show 0 is a semi-simple eigenvalue, we will prove that N (A0 ) = N (A20 ). Clearly N (A0 ) = E0 , taking
z ∈ N (A20 ), then there exists w ∈ E0 such that A0 z = w. The periodic condition ensures that w = 0, and
it follows that z ∈ N (A0 ). Therefore, [55, Remark A.2.4] yields that 0 is semi-simple. As E0 is a subspace
1− 2 ,p
of Wπ p (Ω)2 , the tangent space at w∗ is isomorphic to N (A0 ).
In [53, Remark 2.2] it is shown that all equilibria close to w∗ are contained in a manifold M of
dimension dim (N (A0 )) = 2. Since the dimension of E0 is also 2, there exists an open neighborhood
1− 2 ,p
U0 ⊂ Wπ p (Ω)2 of w∗ such that M ∩ U0 = E0 ∩ U0 . Thus, U0 contains no other equilibria than the
elements of E0 , i.e. E0 ∩ U0 = E ∩ U0 .
So all assumptions of the generalized principle of linearized stability [53, Theorem 2.1] are satisfied.
This principle concludes the proof. □
Remark 3.5. Let the assumptions of Theorem 2.3 hold and (u, v) be the weak solution of problem
(1.13) on [0, T ) with the initial data f . The weak solution of problem (1.13) without a source term has
additionally the following properties:
(a) (Extremum principle) For every t ∈ (0, T ),
inf f ⩽ u(x, t), v(x, t) ⩽ sup f, a.e. x ∈ Ω.
x∈Ω x∈Ω

(b) (Average invariance) For every t ∈ (0, T ), it holds that

Z Z Z
u(t, x)dx = v(t, x)dx = f (x)dx.
Ω Ω Ω
(c) (Lyapunov functional) The energy functional
Z
1
E= (u2 + v 2 )dx
2 Ω
14 YUHANG LI, ZHICHANG GUO, JINGFENG SHAO, BOYING WU

defines a Lyapunov functional for (1.13).

These properties correspond to the case of λ = 0 in problem (1.13) and their proofs are straightforward.
In addition, if λ 6= 0 but K = I, properties (a) and (b) still hold.

4. Regularity results for the proposed model

In this section, we establish regularity results for the solutions of problem (1.13). Firstly, when the
regularity of the initial data is improved, the local existence and uniqueness of strong solutions to problem
(1.13) is obtained via Maximum Regularity Theory. The proof is similar to that of Theorem 2.3.
 
Theorem 4.1. Assume that p ∈ 2+N 1−α , ∞ , β ∈ {1} ∪ [2, ∞), µ > 1 and γ ⩾ 1. For any given
2− 2 ,p
0 < f ∈ Wπ p (Ω), there exist a unique maximal Tmax ∈ (0, T0 ] such that problem (1.13) possess a
unique strong solution (u, v) on [0, Tmax ) satisfying


u, v ∈ Lploc 0, Tmax ; Wπ2,p (Ω) ∩ Wloc
1,p
(0, Tmax ; Lpπ (Ω)) .
Proof. The method for proving the existence of strong solutions is similar to the proof of Theorem 2.3,
with the main differences lying in establishing the maximum Lp -regularity of the linearized operator ma-
trix and the Lipschitz estimate between the operator matrix and the source term. Firstly, we reformulate
an auxiliary problem

 ∂u

 = c (|∇α u| , v̄ε ) ∆u + ∇c (|∇α u| , v) · ∇u − λK ′ (Ku − f ), in (0, T0 ) × Ω,

 ∂t
 ∂v
= a (u, v̄ε ) ∆v + ∇a (u, v) · ∇v, in (0, T0 ) × Ω, (4.1)

 ∂t

 u, v periodic on Ω for (0, T0 ),


u(0, x) = v(0, x) = f (x), in Ω,
as a quasilinear abstract Cauchy problem. In this section, we always choose
E0 = Lpπ (Ω)2 , E1 = Wπ2,p (Ω)2 ,
d
→ E0 . Then
which is well-known to satisfy E1 ,−
2 2
1/2 1/2 2− p ,p 1/2 2− p ,p
E0 = Lpπ (Ω), E1 = Wπ2,p (Ω), E1− p1 = Wπ (Ω)2 , E1− 1 = Wπ (Ω).
p

For fixed w0 = (u0 , v0 )T ∈ W 1,p (0, T ; (E1 , E0 )), we define linear differential operators A11 (w0 ) and
A22 (w0 ) as follows:
A11 (w0 )u = −c (|∇α u0 | , v0ε ) ∆u − ∇c (|∇α u0 | , v0 ) · ∇u, u ∈ Wπ1,p (Ω),
A22 (w0 )v = −a (u0 , v0ε ) ∆v − ∇a (u0 , v0 ) · ∇v, v ∈ Wπ1,p (Ω).
An operator matrix A can then be given by
 
A11 (w) 0
A : W 1,p (0, T ; (E1 , E0 )) → L∞ (0, T ; L (E1 , E0 )) , w 7→ .
0 A22 (w)
Defining F as in (3.3), then problem (4.1) can be rewritten as a quasilinear abstract Cauchy problem
with the same form as (3.4):

 dw
+ A(w)w = F (w), t ∈ (0, T0 ),
dt (4.2)
 w(0) = (f, f )T ,
 A NOVEL NONLINEAR COUPLED NONLOCAL REACTION–DIFFUSION SYSTEM FOR IMAGE RESTORATION 15

Next, we check the maximal Lp -regularity of the linearized operator

 matrix. The  almost reiteration
theorem and classical embedding theorem imply that for any δ ∈ 0, 1 − α − p , 2+N

 
2− 2 ,p 1,1−α− Np+2 −δ
∂xαj ∈ L Wπ p (Ω), Cπ (Ω) , j = 1, 2, · · · , N.

It follows from Amann’s embedding theorem (3.6) that there exists ρ = 1 − α − N +2

− δ ∈ (0, 1) such
   p
1/2 1/2
that for any u ∈ W 1,p 0, T ; E1 , E0 ,


|∇α u| ∈ C [0, T ], Cπ1,ρ (Ω) .
The property of Nemytskii operators implies that


b (|∇α u|) , b′ (|∇α u|) ∈ C [0, T ], Cπ0,ρ (Ω) .
Set ρ0 = 1 − α − Np+2 − δ, we choose ρ1 = min{ρ, ρ0 } if γ = 1 and ρ1 = min {ρ, min{1, γ − 1}ρ0 } if γ > 1,
  
1/2 1/2
then for any u, v ∈ W 1,p 0, T ; E1 , E0 ,


c (|∇α u| , v̄ε ) , ∂xj c (|∇α u| , v) ∈ C [0, T ], Cπ0,ρ1 (Ω) (4.3)
since Hölder spaces are algebras. It follows that
  
1/2 1/2
[t 7→ A11 (w(t))] ∈ C [0, T ], L E1 , E0 (4.4)

for any w = (u, v)T ∈ W 1,p (0, T ; (E1 , E0 )). On the other hand, for any t ∈ (0, T ), the principle symbol
of −A11 (w(t)) := A11 is
A11,π (x, ξ) = c (|∇α u(t)| , v̄ε (t)) |ξ|2 , x ∈ Ω, ξ ∈ RN \{0}.
It follows from (4.3) that there exists C1 > 0 such that for any x ∈ Ω and |ξ| = 1,
1
0< < c (|∇α u(t)| , v ε (t)) |ξ|2 < C1 ,
C1
n o

which implies that λ ∈ C : Reλ < C11 ⊂ ρ(A11,π ). Aπ is real, so there exists θ ∈ 0, π2 such that
 
1
σ(A11,π ) ⊂ λ ∈ C : Reλ ⩾ ∩ R ⊂ {z ∈ C \ {0} : |arg(z)| < θ} .
C1
Thus, −A11 (w(t)) is uniformly (C1 , θ)-elliptic for all t ∈ (0, T ). From [56, Corollary 9.5], it follows that
A11 (w(t)) is a sectorial operator on L2 (Ω), i.e. A11 (w(t)) is the generator of an analytic semigroup T2 (t)
on L2 (Ω). [57, Theorem 9.4.2] implies that T2 (t) can be represented by a kernel satisfying a Gaussian
bound. It follows from [58, Theorem 3.1] that for any t ∈ (0, T ),
 
1/2 1/2
A11 (w(t)) ∈ MRp E1 , E0 . (4.5)

Set ρ3 = 1 − N +2
− δ, we choose ρ2 = min{1, µ − 1}ρ3 if γ = 1 and ρ2 = min{1, γ − 1, µ − 1}ρ3 if γ > 1,
p   
1/2 1/2
it is easy to see that for any u, v ∈ W 1,p 0, T ; E1 , E0 ,


a (u, v̄ε ) , ∂xj a (u, v) ∈ C [0, T ], Cπ0,ρ2 (Ω) ,
from which it follows that
  
1/2 1/2
[t 7→ A22 (w(t))] ∈ C [0, T ], L E1 , E0 (4.6)
16 YUHANG LI, ZHICHANG GUO, JINGFENG SHAO, BOYING WU

for any w = (u, v)T ∈ W 1,p (0, T ; (E1 , E0 )). The process of obtaining (4.5) reveals that the maximal Lp -
regularity of the linearized operator ultimately boils down to sufficient regularity of the coefficients. A
similar argument yields that for any t ∈ (0, T ),
 
1/2 1/2
A22 (w(t)) ∈ MRp E1 , E0 . (4.7)
From (4.5) and (4.7), it follows that for all t ∈ (0, T ) and each G ∈ Lp (0, T ), the linear abstract Cauchy
problem 
 dz
+ A(w(t))z = G(τ ), τ ∈ (0, T ),
dτ
 z(0) = (0, 0)T
has a unique strong Lp solution on (0, T ), i.e.
A(w(t)) ∈ MRp (E1 , E0 ) . (4.8)
By (4.4), (4.6), (4.8) and [50, Theorem 7.1], we can conclude that for any w ∈ W 1,p
(0, T ; (E1 , E0 )),
A(w) ∈ MR (0, T ; (E1 , E0 )).
p

Now we establish Lipschitz estimates of the operator matrix. For R > 0, take w1 = (u1 , v1 )T , w2 =
(u2 , v2 )T ∈ BW 1,p (0,T ;(E1 ,E0 )) (0, R), then
kA(w1 ) − A(w2 )kL∞ (0,T ;L(E1 ,E0 )) = sup sup kA(w1 (t))w − A(w2 (t))wkE0
t∈[0,T ] w=(u,v)T
∥w∥E1 =1

= sup sup kA11 (w1 (t))u − A11 (w2 (t))ukE 1/2

t∈[0,T ] ∥u∥ 1/2 =1
0
E1

+ sup sup kA22 (w1 (t))v − A22 (w2 (t))vkE 1/2

t∈[0,T ] ∥v∥ 1/2 =1
0
E1

:= I1 + I2 .
We estimate I1 and I2 separately. From now on C = C(R) denotes a positive constant which can take
different value in different places. Through a process similar to the Lipschitz estimate of A11 in the proof
of Theorem 2.3, we obtain the following estimate by applying the trick of adding and subtracting the
same term repeatedly.
I1 ⩽ sup sup k(c (|∇α u1 | , v1ε ) − c (|∇α u2 | , v2ε )) ∆ukLpπ (Ω)
t∈[0,T ] ∥u∥ 2,p =1
Wπ (Ω)

+ sup sup k(∇c (|∇α u1 | , v1 ) − ∇c (|∇α u2 | , v2 )) · ∇ukLpπ (Ω)

t∈[0,T ] ∥u∥ 2,p =1
Wπ (Ω)

⩽ C sup kv1ε (t)γ b (|∇α u1 (t)|) − v2ε (t)γ b (|∇α u2 (t)|)k∞

t∈[0,T ]

X
N
+C sup kv1 (t)γ b (|∇α u1 (t)|)k∞ ∂xj (v1 (t) − v2 (t)) ∞
j=1 t∈[0,T ]

X
N
+C sup ∂xj v2 (t) ∞
v1 (t)γ−1 b (|∇α u1 (t)|) − v2 (t)γ−1 b (|∇α u2 (t)|) ∞
j=1 t∈[0,T ]

X
N
+C sup kv1 (t)γ k∞ ∂xj (b (|∇α u1 (t)|) − b (|∇α u2 (t)|)) ∞
j=1 t∈[0,T ]
 A NOVEL NONLINEAR COUPLED NONLOCAL REACTION–DIFFUSION SYSTEM FOR IMAGE RESTORATION 17

X
N
+C sup ∂xj b (|∇α u2 (t)|) ∞
kv1 (t)γ − v2 (t)γ k∞
j=1 t∈[0,T ]
⩽ C sup ku1 (t) − u2 (t)k 2− 2 ,p
p
+ C sup kv1 (t) − v2 (t)k 2− 2 ,p
p
t∈[0,T ] Wπ (Ω) t∈[0,T ] Wπ (Ω)

X
N
+C kb′ (|∇α u1 (t)|)k∞ ∂xj (|∇α u1 (t)| − |∇α u2 (t)|) ∞
j=1

X
N
+C ∂xj (|∇α u2 (t)|) ∞
kb′ (|∇α u1 (t)|) − b′ (|∇α u2 (t)|)k∞
j=1
⩽ C ku1 − u2 k ( ) + C kv1 − v2 k ( )
2− 2 ,p 2− 2 ,p
p p
C [0,T ],Wπ (Ω) C [0,T ],Wπ (Ω)

+ C sup k|∇α u1 (t)| − |∇α u2 (t)|kC 1,ρ1 (Ω)

t∈[0,T ]

⩽ C kw1 − w2 kW 1,p (0,T ;(E1 ,E0 )) .

In the same way, we have
I2 ⩽ C kw1 − w2 kW 1,p (0,T ;(E1 ,E0 ))
and furthermore
kA(w1 ) − A(w2 )kL∞ (0,T ;L(E1 ,E0 )) ⩽ C kw1 − w2 kW 1,p (0,T ;(E1 ,E0 )) ,
which implies that

0,1
A ∈ CVolt W 1,p (0, T ; (E1 , E0 )), MRp (0, T ; (E1 , E0 )) . (4.9)
The Lipschitz estimate of the source term is similar to that done in the proof of Theorem 2.3. We have


0,1
F ∈ CVolt W 1,p (0, T ; (E1 , E0 )); L∞ (0, T ; E0 ) , Lp (0, T ; E0 ) . (4.10)
It follows from (4.9), (4.10) and Theorem 2.1 that there exist a maximal Tmax ∈ (0, T0 ] and a unique
1,p
solution w = (u, v)T ∈ Wloc (0, Tmax ; (E1 , E0 )) of the quasilinear abstract Cauchy problem (4.2). Then
(u, v) is the unique strong solution of auxiliary problem (4.1) on [0, Tmax ). The minimum principle v ⩾ ε
implies that v = v̄ε . Therefore, (u, v) is also the unique strong solution of problem (1.13), which concludes
the proof. □
Now we present a regularity result for the strong solutions of problem (1.13). Denote QT = Ω × (0, T )
for T ∈ (0, T0 ].
Theorem 4.2. Let the assumptions of Theorem 4.1 hold. Let (u, v) be the strong solution of problem (1.13)
ρ
on its maximal interval of existence [0, Tmax ). Then there exists ρ ∈ (0, 1) such that u, v ∈ C 2+ρ,1+ 2 (QT )
for every T ∈ (0, Tmax ).
Proof. We will accomplish the proof using the techniques provided in [46, Section 14]. Henceforth, T
can take any value in (0, Tmax ), C > 0 denotes a constant which can take different value in different
places. Theorem 4.1 and Amann’s embedding theorem imply that
 

2
2− p ,p
u, v ∈ L 0, T ; Wπ (Ω) ∩ W (0, T ; Lπ (Ω)) ∩ C [0, T ], Wπ
p 2,p 1,p p
(Ω) .

Utilizing Hölder’s inequality and the interpolation inequality of Slobodeckij spaces, it can be inferred that
ku(t) − u(s)k (2− 2 )θ,p ⩽ Cku(t) − u(s)k1−θ
Lp
π (Ω)
ku(t) − u(s)kθ 2− 2 ,p
p p
W π(Ω) W (Ω) π
18 YUHANG LI, ZHICHANG GUO, JINGFENG SHAO, BOYING WU

Z t  1−θ
p′
Z t  1−θ
p
p
⩽C 1dτ ku̇(τ )kLpπ (Ω) dτ
s s
(1−θ)(p−1)
⩽ C|t − s| p

holds for all 0 < s ⩽ t < T and 0 < θ < 1, from which it follows that
 
(1−θ)(p−1) 2θ− 2θ
p ,p
u∈C p [0, T ], Wπ (Ω) .

Using arguments similar to those employed in the proofs of Theorems 2.3 and 4.1, it can be deduced that
 
(1−θ)(p−1) 1,2θ−1− 2θ+N −δ
u∈C p [0, T ], Cπ p
(Ω)

and  
(1−θ)(p−1) 1,2θ−1− 2θ+N −α−δ
|∇ u| ∈ C
α p [0, T ], Cπ p
(Ω)
   
for θ ∈ N +(α+1)p
2(p−1) , 1 and δ ∈ 0, 2θ − 1 − 2θ+N
p − α . Similarly,
 
(1−θ)(p−1) 1,2θ−1− 2θ+N −δ
v∈C p [0, T ], Cπ p
(Ω) .

Set
(1 − θ)(p − 1) 2θ + N 2θ + N
ρ1 = , ρ2 = 2θ − 1 − − δ, ρ3 = 2θ − 1 − − α − δ,
p p p
we choose ρ = min{1, γ − 1} min{ρ1 , ρ3 } if γ = 1, and ρ = min{1, γ − 1, µ − 1} min ρi if γ > 1, then
i=1,2,3

′ ρ, ρ


∂xj c (|∇ u| , v) , ∂xj a(u, v), −λK (Ku − f ) ∈ C
α 2 QT
since Hölder spaces are algebras. Rewrite
c (|∇α u| , v) := a11 (x, t), a(u, v) := a22 (x, t), −λK ′ (Ku − f ) := f1 (x, t),
∇c (|∇α u| , v) := b1 (x, t), ∇a(u, v) := b2 (x, t),
then (u, v) is the unique strong solution of the linear parabolic system

 ∂ ũ

 = a11 (x, t)∆ũ + b1 (x, t) · ∇ũ − f1 (x, t), in (0, T0 ) × Ω,

 ∂t
 ∂ṽ
= a22 (x, t)∆ṽ + b2 (x, t) · ∇ṽ, in (0, T0 ) × Ω, (4.11)

 ∂t

 ũ, ṽ periodic on Ω for (0, T0 ),


ũ(0, x) = ṽ(0, x) = f (x), on Ω,
which has Hölder continuous coefficients and a Hölder continuous right-hand side. It follows from classical
ρ
results [59, Section 14.18] that (4.11) possesses a unique classical solution ũ, ṽ ∈ C 2+ρ,1+ 2 (QT ). It is
obvious that (ũ, ṽ) is also the unique strong solution of (4.11), which implies that u = ũ, v = ṽ. Therefore,
ρ
(u, v) is the unique classical solution of (1.13), and u, v ∈ C 2+ρ,1+ 2 (QT ). □

Corollary 4.3. Assume that β2 , µ2 , γ are integers, 0 < f ∈ Cπ∞ (Ω). Let (u, v) be the strong solution
of problem (1.13) on its maximal interval of existence [0, Tmax ). Then u, v ∈ C ∞ (QT ) for every T ∈
(0, Tmax ).
 A NOVEL NONLINEAR COUPLED NONLOCAL REACTION–DIFFUSION SYSTEM FOR IMAGE RESTORATION 19

Proof. The bootstrapping argument used in the proof of Theorem 4.2 can be iterated in order to see
2 k+ρ
that for k ⩾ 2, if f ∈ W k− p ,p (Ω), then there exists ρ ∈ (0, 1) such that u, v ∈ C k+ρ, 2 (QT ) by means
of classical results [59, Section 14.18], as repeated differentiations can be achieved under the assumption.
The arbitrariness of k concludes the proof. □

5. Numerical experiments
In this section, we present several numerical experiment examples to illustrate the effectiveness of the
proposed model in image restoration. Firstly, a numerical discretization of (1.13) is derived. Assume
that the discrete image to be I × I pixels, τ to be the time step size and h the space grid size. Then the
equidistant spatio-temporal grid is given by
Tτ,h = {(tn , xi , yj ) : tn = nτ, xi = ih, yj = jh, n = 0, 1, 2, · · · , i, j = 0, 1, · · · , I − 1} .
Denote by un = {uni,j }∀i,j the grid function at time tn , which approximates the values of u(tn , ·) at grid
points. Some other notations and assumptions are given for the following numerical scheme.
u0i,j = fi,j := f (xi , yj ),
uni,0 = uni,I , un0,j = unI,j , uni,−1 = uni,I−1 , un−1,j = unI−1,j ,
i±1,j −ui,j
un −un
n
un
∆x± uni,j := ± h , ∆y± uni,j := ± i,j±1h i,j ,
δx2 uni,j := uni+1,j − 2uni,j + uni−1,j , δy2 uni,j := uni,j+1 − 2uni,j + uni,j−1 .
Similar notations and assumptions are used for v and other functions. Denote by ûn the 2-D discrete
Fourier transform (DFT) of the grid function un and F−1 the 2-D inverse discrete Fourier transform
(IDFT) operator. To approximate |∇α un |, we use a central difference scheme provided by [27]. The
fractional-order difference can be defined as
h
α i
Dxα un = F−1 1 − e−iω1 h eiαω1 2 ûn (ω1 , ω2 ) ,
h

h
α i (5.1)
Dyα un = F−1 1 − e−iω2 h eiαω2 2 ûn (ω1 , ω2 ) ,
h

and the discrete fractional-order gradient ∇α α n α n

h = (Dx u , Dy u ). The numerical approximation of (1.13)
could be derived by local linearization method. Denote by an = a(un , v n ) the values at the grid points
of the diffusion coefficient a(·, ·). If we have a numerical solution (un , v n ), an explicit finite difference
scheme
n+1
vi,j − vi,j
n



= ∆x− ani,j ∆x+ vi,j
n
+ ∆y− ani,j ∆y+ vi,j
n
, (5.2)
τ
can be used to obtain v n+1 , then we can use a semi-implicit scheme

i,j − ui,j
un+1 n




= ∆x− cni,j ∆x+ uni,j + ∆y− cni,j ∆y+ uni,j − λ K′ ∗ K ∗ un+1 i,j + λ (K′ ∗ f )i,j , (5.3)
τ


to obtain un+1 . Here cn = c |∇α h u |, v
n n+1
, ∇α
h is the discrete fractional-order gradient, K = (ki,j )
is a discrete convolution kernel with adjoint K′ , which is obtained from the convolution kernel k, i.e.
ki,j = k(ih, jh). The function f here is only defined on the grid, but we do not denote it.
In the numerical implementation of the proposed model, the explicit scheme (5.2) can be directly
computed, and the semi-implicit scheme (5.3) can be solved using the DFT and IDFT. The image
restoration process based on the proposed model is summarized as Algorithm 1.
The following theorem concerns the stability of the difference scheme (5.3).
20 YUHANG LI, ZHICHANG GUO, JINGFENG SHAO, BOYING WU

Algorithm 1 The restoration algorithm based on the proposed model

Input: Original image f , τ , α, β, k1 , λ, K, γ, µ, λ1 , tol, maxIter
Initialize: u0 = v 0 = f , n = 0
Output: un
1: Compute M = max fi,j
i,j
2: for each image un , v n do
Compute an = λ1 |uM µ| + (1 − λ1 ) |vM γ|
n µ n γ
3:
4: Compute v n+1 by (5.2)
5: Compute Dxα un and Dyα un by (5.1)
|v n+1 |γ 1
6: Compute cn = M γ 1+k (D α un +D α un )β/2
1 x y
7: for i, j = 0, 1, · · · , I − 1 do



8: Compute dni,j := uni,j + τ ∆x− cni,j ∆x+ uni,j + ∆y− cni,j ∆y+ uni,j + τ λ (K′ ∗ f )i,j
9: end for h i
dˆn
10: Compute un+1 = F−1 I+τ ˆ λK̂′ ◦K̂
∥un+1 −f ∥22
11: if > tol and i < maxIter then
∥un+1 ∥22
12: Set n = n + 1
13: Repeat 3-10 steps for un and v n
14: end if
15: end for

Theorem 5.1. To ensure strict stability in the `2 norm of the scheme (5.3) for any λ > 0 and K, it is
required that
τ
1
2
max c |∇α h u |, v
n n+1
⩽ . (5.4)
h n,i,j 4
Proof. Freezing cni,j as a constant c, consider the following constant coefficient difference equation
τ
′


i,j = ui,j + 2 δx ui,j + δy ui,j − τ λ K ∗ K ∗ u
un+1 n 2 n 2 n n+1
i,j
, (5.5)
h
whose stability conditions can be obtained by von Neumann analysis. The values of the grid function un
at each grid point can be represented by its DFT:
X
I−1 X
I−1
2πp 2πq
uni,j = ûn (ω1 , ω2 )ei(ω1 xi +ω2 yj ) , ω1 = , ω2 = . (5.6)
p=0 q=0
I I

Noting that the DFT of K′ always equals the complex conjugate of the DFT of K, utilizing the convolution
property of the DFT, we have

X
I−1 X
I−1
K′ ∗ K ∗ un+1 i,j
= ûn+1 (ω1 , ω2 )|K̂(ω1 , ω2 )|2 ei(ω1 xi +ω2 yj ) . (5.7)
p=0 q=0

Substituting (5.6) and (5.7) into (5.5) and comparing coefficients, we obtain that
  
4τ c 2 ω1 h 2 ω2 h
û n+1
(ω1 , ω2 ) = 1 − 2 sin + sin ûn (ω1 , ω2 ) − τ λ|K̂(ω1 , ω2 )|2 ûn+1 (ω1 , ω2 ). (5.8)
h 2 2
 A NOVEL NONLINEAR COUPLED NONLOCAL REACTION–DIFFUSION SYSTEM FOR IMAGE RESTORATION 21

Then the amplification factor can be derived by simple calculation:

ûn+1 (ω1 , ω2 ) 1 − 4cτ
h2 sin
2 ω1 h
2 + sin
2 ω2 h
2
G(ω1 , ω2 ) = n = .
û (ω1 , ω2 ) 1 + τ λ|K̂(ω1 , ω2 )|2
To ensure strict von Neumann condition
|G(ω1 , ω2 )| ⩽ 1
holds for all λ > 0 and K, it is necessary and sufficient that τc
h2 ⩽ 14 . Therefore,
τ
1
2
max c |∇αh u |, v
n n+1
⩽
h n,i,j 4
becomes the necessary condition to ensure that the scheme (5.3) is strictly stable for all λ > 0 and K. □
In the deblurring case, using fully explicit schemes imposes stricter restriction on the time step size.
By employing similar techniques as in the proof of Theorem 5.1, it can be shown that the necessary
condition for the explicit scheme
i,j − ui,j
un+1 n



= ∆x− cni,j ∆x+ uni,j + ∆y− cni,j ∆y+ uni,j − λ (K′ ∗ (K ∗ un − f ))i,j (5.9)
τ
to be strictly stable for any K is

2
τ max c |∇α h u |, v
n n+1
⩽ .
n,i,j λ + h82
It is evident that when λ is large, strict stability of the explicit scheme (5.9) requires using very small
time step size, resulting in a considerable number of time steps for the deblurring process. To mitigate
computational costs, we opted for the semi-implicit scheme (5.3) in the following numerical experiments.
Numerical experiments indicate that even if τ slightly exceeds the constraint in (5.4), the semi-implicit
scheme (5.9) still performs well.
Remark 5.2. The explicit scheme (5.2) requires that the necessary condition for stability is
τ 1
max a (un , v n ) ⩽ .
h2 n,i,j 4
n
For the computation of v , although there exist semi-implicit schemes that are unconditionally stable
(see [60]), the benefits of such schemes are very limited due to the restriction (5.4). There is no necessity
to use semi-implicit schemes to solve v n in practice.
In experiments, three common types of blur kernels are considered: the motion blur kernel with angle π3
and length 20, disk kernel with a radius of 3 and 5×5 average kernel. We test several noisy blurred images,
distorted by Gaussian noise with a standard deviation σ = 3 and by blurring with one of the previously
mentioned kernels, to verify the effectiveness of the proposed model. For illustration, the result for the
texture1 (256 × 256 pixels), texture2 (256 × 256 pixels), hybrid (256 × 256 pixels), satellite1(512 × 512
pixels), satellite2(512 × 512 pixels) and satellite3(512 × 512 pixels) are presented, see the original test
images in Figure 1.
The proposed model is compared with fast total variation (FastTV) minimization [4], nonlocal to-
tal variation (NLTV) method [21, 22], nonlocal adaptive biharmonic (NLABH) regularization [26], a
nonlinear fractional diffusion (NFD) model [34] and a Perona-Malik type equation regularized by the
p-Laplacian (PLRPM) [12]. The free parameters for FastTV, NLTV, NLABH, NFD and PLRPM are
set as suggested in the reference papers. For the proposed model, while there may be more suitable
parameter choices for each image, in this paper, we set τ = 0.5, α = 0.9, γ = 1, µ = 0.4, k1 = 1, λ1 = 0.9,
22 YUHANG LI, ZHICHANG GUO, JINGFENG SHAO, BOYING WU

(a) Texture1 (b) Texture2 (c) Hybrid

(d) Satellite1 (e) Satellite2 (f) Satellite3

Figure 1. Test images.

maxIter = 500, and tol = 0.005 throughout all experiments. For the texture1, texture2, hybrid, satellite
1, satellite 2 and satellite 3 images, we select β = 0.96, 0.63, 1, 0.7, 0.9 and 0.8 respectively. The parameter
λ is chosen based on the specific image and the type of blur kernel.

Table 1. Comparison of PSNR and SSIM for different models in the experiments on
texture1, texture2 and hybrid image. Bold values indicate the best result.

Figure 2 Figure 3 Figure 4

PSNR SSIM PSNR SSIM PSNR SSIM
FastTV 15.17 0.5320 19.00 0.6887 21.00 0.7136
NLTV 16.35 0.6599 19.26 0.7427 21.14 0.7087
NLABH 16.11 0.6698 18.89 0.6710 20.21 0.4598
NFD 16.18 0.6738 19.06 0.6763 20.36 0.4723
PLRPM 16.30 0.6599 19.29 0.7068 21.17 0.6415
Ours 16.25 0.6745 19.39 0.6992 21.17 0.6281

In order to quantify the restoration effect, for the original image f and the compared image u, the
restoration performance is measured in terms of the peak signal noise ratio (PSNR)
 
Σi,j 2552
PSNR(u, f ) = 10 log10 ,
Σi,j (ui,j − fi,j )2
and the structural similarity index measure (SSIM)
(2µu µf + c1 )(2σuf + c2 )
SSIM(u, f ) = ,
µ2u + µ2f + c1 + σu2 + σf2 + c2
 A NOVEL NONLINEAR COUPLED NONLOCAL REACTION–DIFFUSION SYSTEM FOR IMAGE RESTORATION 23

Table 2. Comparison of PSNR and SSIM for different models in the experiments on
three satellite images. Bold values indicate the best result.

Figure 5 Figure 6 Figure 7

PSNR SSIM PSNR SSIM PSNR SSIM
FastTV 23.97 0.8888 24.91 0.9322 22.41 0.9229
NLTV 23.87 0.8838 25.46 0.9387 22.59 0.9291
NLABH 22.56 0.8475 24.82 0.9325 22.00 0.9049
NFD 22.67 0.8548 24.77 0.9301 22.06 0.9059
PLRPM 24.04 0.8949 25.28 0.9380 22.55 0.9205
Ours 24.44 0.9029 25.47 0.9395 22.55 0.9212

where c1 , c2 are two variables to stabilize the division with weak denominator, µu , µf , σu , σf and σuf are
the local means, standard deviations and cross-covariance for image u, f , respectively. The better quality
image will have higher values of PSNR and SSIM. For a comparison of the performance quantitatively,
we list the PSNR and SSIM values of the restored results in Table 1 and Table 2.
Now we report the numerical experiments of deblurring and denoising for the original test images in
Figure 1. The corresponding results are shown in Figures 2-7. Firstly, the restoration results for texture1
and texture2 are shown in Figures 2 and 3. Texture1 and texture2 images are blurred by disk kernel and
average kernel, respectively. For these two experiments, we select λ = 45 in the proposed model. We find
that in Figures 2c and 3c, there is minimal noise present, but the visual quality is over-smooth, with an
evident loss of texture information. On the other hand, in Figures 2e, 2f, 3e and 3f, texture details are
better preserved, however, there is an increase in the noise level of the images. Compared to Figures 2d,
3d, 2g and 3g, the restoration results of the proposed model exhibit slightly higher noise levels but retain
more texture information, resulting in a better visual quality, see Figures 2h and 3h.
For the restoration results of the hybrid image, some new phenomena emerge, as illustrated in Figure
4. The central area of the hybrid image is rich in texture, while the surrounding area in smooth, as
shown in Figure 4a. The hybrid image is blurred by a motion kernel, posing a challenge for all models,
see Figure 4b. For this experiment, we select λ = 15 in the proposed model. It is observed from Figure
4c that the restoration result of FastTV contains fewer noise and artifacts in smooth areas, but it fails to
preserve texture information. NLABH and NFD noticeably amplify noise in smooth regions and produce
obvious artifacts, as seen in Figures 4e and 4f. In the restoration results of the proposed model, noise
in smooth areas is slightly more pronounced compared to NLTV and PLRPM, but texture preservation
is better than PLRPM, and there are no artifacts in smooth regions as seen in the restoration results of
NLTV, see Figures 4d, 4g, and 4h.
Finally, the restoration results for three satellite images are shown in Figures 5-7. The three satellite
images are respectively blurred by motion kernel, disk kernel, and average kernel, as shown in Figures
5b, 6b, and 7b. We select λ = 15, 15, 10 in the proposed model respectively for these three experiments.
Similar to the previous experiments, the restoration results of FastTV and NLTV are relatively smooth,
and the noise levels in the restoration results of NLABH and NFD are increased. Overall, the methods
with better visual restoration results are NLTV, PLRPM, and the proposed method. Our method slightly
reduces visual quality on satellite3 due to increased noise in smooth regions, but performs best on satellite1
and satellite2, as shown in Figures 5h, 6h, and 7h.
24 YUHANG LI, ZHICHANG GUO, JINGFENG SHAO, BOYING WU

(a) Original (b) Blurred (c) FastTV (d) NLTV

(e) NLABH (f) NFD (g) PLRPM (h) Ours

Figure 2. Recovery results for the texture1 image with disk blur and corrupted by
the noise of standard deviation σ = 3. (a) original image. (b) noisy blurred image,
PSNR=12.77. (c)-(h) recovered images.

(a) Original (b) Blurred (c) FastTV (d) NLTV

(e) NLABH (f) NFD (g) PLRPM (h) Ours

Figure 3. Recovery results for the texture2 image with average blur and corrupted by
the noise of standard deviation σ = 3. (a) original image. (b) noisy blurred image,
PSNR=15.60. (c)-(h) recovered images.
A NOVEL NONLINEAR COUPLED NONLOCAL REACTION–DIFFUSION SYSTEM FOR IMAGE RESTORATION 25

(a) Original (b) Blurred (c) FastTV (d) NLTV

(e) NLABH (f) NFD (g) PLRPM (h) Ours

Figure 4. Recovery results for the hybrid image with motion blur and corrupted by
the noise of standard deviation σ = 3. (a) original image. (b) noisy blurred image,
PSNR=18.07. (c)-(h) recovered images.

(a) Original (b) Blurred (c) FastTV (d) NLTV

(e) NLABH (f) NFD (g) PLRPM (h) Ours

Figure 5. Recovery results for the satellite1 image with motion blur and corrupted by
the noise of standard deviation σ = 3. (a) original image. (b) noisy blurred image,
PSNR=18.64. (c)-(h) recovered images.
26 YUHANG LI, ZHICHANG GUO, JINGFENG SHAO, BOYING WU

(a) Original (b) Blurred (c) FastTV (d) NLTV

(e) NLABH (f) NFD (g) PLRPM (h) Ours

Figure 6. Recovery results for the satellite2 image with disk blur and corrupted by
the noise of standard deviation σ = 3. (a) original image. (b) noisy blurred image,
PSNR=21.68. (c)-(h) recovered images.

(a) Original (b) Blurred (c) FastTV (d) NLTV

(e) NLABH (f) NFD (g) PLRPM (h) Ours

Figure 7. Recovery results for the satellite3 image with average blur and corrupted
by the noise of standard deviation σ = 3. (a) original image. (b) noisy blurred image,
PSNR=19.20. (c)-(h) recovered images.
 A NOVEL NONLINEAR COUPLED NONLOCAL REACTION–DIFFUSION SYSTEM FOR IMAGE RESTORATION 27

References
[1] L. I. Rudin, S. Osher, E. Fatemi, Nonlinear total variation based noise removal algorithms, Physica
D: Nonlinear Phenomena 60 (1) (1992) 259–268. doi:10.1016/0167-2789(92)90242-F.
URL https://www.sciencedirect.com/science/article/pii/016727899290242F
[2] L. I. Rudin, S. Osher, Total variation based image restoration with free local constraints, in:
Proceedings of 1st International Conference on Image Processing, Vol. 1, 1994, pp. 31–35 vol.1.
doi:10.1109/ICIP.1994.413269.
URL https://ieeexplore.ieee.org/document/413269
[3] H. Takeda, S. Farsiu, P. Milanfar, Deblurring using regularized locally adaptive kernel regression,
IEEE Transactions on Image Processing 17 (4) (2008) 550–563. doi:10.1109/TIP.2007.918028.
URL https://ieeexplore.ieee.org/document/4459371
[4] Y. Huang, M. K. Ng, Y. W. Wen, A fast total variation minimization method for image restoration,
Multiscale Modeling & Simulation 7 (2) (2008) 774–795. doi:10.1137/070703533.
URL https://epubs.siam.org/doi/10.1137/070703533
[5] K. Bredies, K. Kunisch, T. Pock, Total generalized variation, SIAM Journal on Imaging Sciences
3 (3) (2010) 492–526. doi:10.1137/090769521.
URL https://epubs.siam.org/doi/10.1137/090769521
[6] P. Perona, J. Malik, Scale-space and edge detection using anisotropic diffusion, IEEE Transactions
on Pattern Analysis and Machine Intelligence 12 (7) (1990) 629–639. doi:10.1109/34.56205.
URL https://ieeexplore.ieee.org/document/56205
[7] M. Welk, D. Theis, T. Brox, J. Weickert, PDE-based deconvolution with forward-backward diffusiv-
ities and diffusion tensors, in: R. Kimmel, N. A. Sochen, J. Weickert (Eds.), Scale Space and PDE
Methods in Computer Vision, Springer Berlin Heidelberg, Berlin, Heidelberg, 2005, pp. 585–597.
doi:10.1007/11408031_50.
URL https://link.springer.com/chapter/10.1007/11408031_50
[8] J. Weickert, A review of nonlinear diffusion filtering, in: B. ter Haar Romeny, L. Florack, J. Koen-
derink, M. Viergever (Eds.), Scale-Space Theory in Computer Vision, Springer Berlin Heidelberg,
Berlin, Heidelberg, 1997, pp. 1–28. doi:10.1007/3-540-63167-4_37.
URL https://link.springer.com/chapter/10.1007/3-540-63167-4_37
[9] F. Catté, P. L. Lions, J. M. Morel, T. Coll, Image selective smoothing and edge detection by nonlinear
diffusion, SIAM Journal on Numerical Analysis 29 (1) (1992) 182–193. doi:10.1137/0729012.
URL https://epubs.siam.org/doi/10.1137/0729012
[10] P. Guidotti, A new nonlocal nonlinear diffusion of image processing, Journal of Differential Equations
246 (12) (2009) 4731–4742. doi:10.1016/j.jde.2009.03.017.
URL https://www.sciencedirect.com/science/article/pii/S0022039609001247
[11] P. Guidotti, A backward–forward regularization of the perona–malik equation, Journal of Differential
Equations 252 (4) (2012) 3226–3244. doi:10.1016/j.jde.2011.10.022.
URL https://www.sciencedirect.com/science/article/pii/S0022039611004608
[12] P. Guidotti, Y. Kim, J. Lambers, Image restoration with a new class of forward-backward-forward
diffusion equations of perona–malik type with applications to satellite image enhancement, SIAM
Journal on Imaging Sciences 6 (3) (2013) 1416–1444. doi:10.1137/120882895.
URL https://epubs.siam.org/doi/10.1137/120882895
[13] J. Shao, Z. Guo, W. Yao, D. Yan, B. Wu, A non-local diffusion equation for noise removal, Acta
Mathematica Scientia 42 (5) (2022) 1779–1808. doi:10.1007/s10473-022-0505-1.
URL https://link.springer.com/article/10.1007/s10473-022-0505-1
28 YUHANG LI, ZHICHANG GUO, JINGFENG SHAO, BOYING WU

[14] J. Tumblin, G. Turk, Lcis: A boundary hierarchy for detail-preserving contrast reduction, in: Pro-
ceedings of the 26th Annual Conference on Computer Graphics and Interactive Techniques, SIG-
GRAPH 1999, Association for Computing Machinery, Inc, 1999, pp. 83–90. doi:10.1145/311535.
311544.
URL https://dl.acm.org/doi/10.1145/311535.311544
[15] Y. L. You, M. Kaveh, Fourth-order partial differential equations for noise removal, IEEE Transactions
on Image Processing 9 (10) (2000) 1723–1730. doi:10.1109/83.869184.
URL https://ieeexplore.ieee.org/document/869184
[16] A. L. Bertozzi, J. B. Greer, Low-curvature image simplifiers: Global regularity of smooth solutions
and laplacian limiting schemes, Communications on Pure and Applied Mathematics 57 (6) (2004)
764–790. doi:10.1002/cpa.20019.
URL https://onlinelibrary.wiley.com/doi/abs/10.1002/cpa.20019
[17] M. R. Hajiaboli, A self-governing hybrid model for noise removal, in: T. Wada, F. Huang, S. Lin
(Eds.), Advances in Image and Video Technology, Springer Berlin Heidelberg, Berlin, Heidelberg,
2009, pp. 295–305.
[18] P. Guidotti, K. Longo, Well-posedness for a class of fourth order diffusions for image processing,
Nonlinear Differential Equations and Applications NoDEA 18 (4) (2011) 407–425. doi:10.1007/
s00030-011-0101-x.
URL https://link.springer.com/article/10.1007/s00030-011-0101-x
[19] Y. Wen, J. Sun, Z. Guo, A new anisotropic fourth-order diffusion equation model based on image
features for image denoising, Inverse Problems and Imaging 16 (4) (2022) 895–924. doi:10.3934/
ipi.2022004.
URL https://www.aimsciences.org/article/id/621ec7fd2d80b7479e4357d4
[20] G. Gilboa, S. Osher, Nonlocal linear image regularization and supervised segmentation, Multiscale
Modeling & Simulation 6 (2) (2007) 595–630. doi:10.1137/060669358.
URL https://epubs.siam.org/doi/10.1137/060669358
[21] G. Gilboa, S. Osher, Nonlocal operators with applications to image processing, Multiscale Modeling
& Simulation 7 (3) (2009) 1005–1028. doi:10.1137/070698592.
URL https://epubs.siam.org/doi/10.1137/070698592
[22] X. Zhang, M. Burger, X. Bresson, S. Osher, Bregmanized nonlocal regularization for deconvolution
and sparse reconstruction, SIAM Journal on Imaging Sciences 3 (3) (2010) 253–276. doi:10.1137/
090746379.
URL https://epubs.siam.org/doi/10.1137/090746379
[23] S. Kindermann, S. Osher, P. W. Jones, Deblurring and denoising of images by nonlocal functionals,
Multiscale Modeling & Simulation 4 (4) (2005) 1091–1115. doi:10.1137/050622249.
URL https://epubs.siam.org/doi/abs/10.1137/050622249
[24] F. Karami, K. Sadik, L. Ziad, A variable exponent nonlocal p(x)-laplacian equation for image restora-
tion, Computers & Mathematics with Applications 75 (2) (2018) 534–546. doi:10.1016/j.camwa.
2017.09.034.
URL https://www.sciencedirect.com/science/article/pii/S0898122117305977
[25] K. Shi, Image denoising by nonlinear nonlocal diffusion equations, Journal of Computational and
Applied Mathematics 395 (2021) 113605. doi:10.1016/j.cam.2021.113605.
URL https://www.sciencedirect.com/science/article/pii/S0377042721002259
[26] Y. Wen, L. A. Vese, K. Shi, Z. Guo, J. Sun, Nonlocal adaptive biharmonic regularizer for image
restoration, Journal of Mathematical Imaging and Vision 65 (3) (2023) 453–471. doi:10.1007/
s10851-022-01129-4.
 A NOVEL NONLINEAR COUPLED NONLOCAL REACTION–DIFFUSION SYSTEM FOR IMAGE RESTORATION 29

URL https://link.springer.com/article/10.1007/s10851-022-01129-4
[27] J. Bai, X. C. Feng, Fractional-order anisotropic diffusion for image denoising, IEEE Transactions on
Image Processing 16 (10) (2007) 2492–2502. doi:10.1109/TIP.2007.904971.
URL https://ieeexplore.ieee.org/document/4303134
[28] J. Zhang, K. Chen, A total fractional-order variation model for image restoration with nonhomo-
geneous boundary conditions and its numerical solution, SIAM Journal on Imaging Sciences 8 (4)
(2015) 2487–2518. doi:10.1137/14097121X.
URL https://epubs.siam.org/doi/10.1137/14097121X
[29] R. H. Chan, A. Lanza, S. Morigi, F. Sgallari, An adaptive strategy for the restoration of textured
images using fractional order regularization, Numerical Mathematics: Theory, Methods and Appli-
cations 6 (1) (2013) 276–296. doi:doi.org/10.4208/nmtma.2013.mssvm15.
URL http://global-sci.org/intro/article_detail/nmtma/5904.html
[30] M. Xu, J. Yang, D. Zhao, H. Zhao, An image-enhancement method based on variable-order
fractional differential operators, Bio-Medical Materials and Engineering 26 (s1) (2015) S1325–S1333.
doi:10.3233/BME-151430.
URL https://content.iospress.com/articles/bio-medical-materials-and-engineering/
bme1430
[31] X. Yin, S. Zhou, M. A. Siddique, Fractional nonlinear anisotropic diffusion with p-laplace variation
method for image restoration, Multimedia Tools and Applications 75 (8) (2016) 4505–4526. doi:
10.1007/s11042-015-2488-6.
URL https://link.springer.com/article/10.1007/s11042-015-2488-6
[32] Z. Zhou, Z. Guo, G. Dong, J. Sun, D. Zhang, B. Wu, A doubly degenerate diffusion model based
on the gray level indicator for multiplicative noise removal, IEEE Transactions on Image Processing
24 (1) (2015) 249–260. doi:10.1109/TIP.2014.2376185.
URL https://ieeexplore.ieee.org/document/6967828
[33] W. Yao, Z. Guo, J. Sun, B. Wu, H. Gao, Multiplicative noise removal for texture images based on
adaptive anisotropic fractional diffusion equations, SIAM Journal on Imaging Sciences 12 (2) (2019)
839–873. doi:10.1137/18M1187192.
URL https://epubs.siam.org/doi/10.1137/18M1187192
[34] Z. Guo, W. Yao, J. Sun, B. Wu, Nonlinear fractional diffusion model for deblurring images with
textures, Inverse Problems and Imaging 13 (6) (2019) 1161–1188. doi:10.3934/ipi.2019052.
URL https://www.aimsciences.org/article/id/aea138af-3399-4ba5-91b1-f75dfc529952
[35] X. Shan, J. Sun, Z. Guo, W. Yao, Z. Zhou, Fractional-order diffusion model for multiplicative noise
removal in texture-rich images and its fast explicit diffusion solving, BIT Numerical Mathematics
62 (4) (2022) 1319–1354. doi:10.1007/s10543-022-00913-3.
URL https://link.springer.com/article/10.1007/s10543-022-00913-3
[36] S. Majee, R. K. Ray, A. K. Majee, A gray level indicator-based regularized telegraph diffusion
model: Application to image despeckling, SIAM Journal on Imaging Sciences 13 (2) (2020) 844–870.
doi:10.1137/19M1283033.
URL https://epubs.siam.org/doi/10.1137/19M1283033
[37] X. Shan, J. Sun, Z. Guo, Multiplicative noise removal based on the smooth diffusion equation,
Journal of Mathematical Imaging and Vision 61 (6) (2019) 763–779.
[38] Z. Zhou, Z. Guo, D. Zhang, B. Wu, A nonlinear diffusion equation-based model for ultra-
sound speckle noise removal, Journal of Nonlinear Science 28 (2) (2018) 443–470. doi:10.1007/
s00332-017-9414-1.
URL https://link.springer.com/article/10.1007/s00332-017-9414-1
30 YUHANG LI, ZHICHANG GUO, JINGFENG SHAO, BOYING WU

[39] C. Poynton, Digital video and HD: Algorithms and Interfaces, second edition Edition, Morgan Kauf-
mann, 2012.
URL https://www.sciencedirect.com/book/9780123919267
[40] J. L. Vazquez, The Porous Medium Equation: Mathematical Theory, Oxford University Press, 2006.
doi:10.1093/acprof:oso/9780198569039.001.0001.
URL https://academic.oup.com/book/7922
[41] G. I. Barenblatt, V. M. Entov, V. M. Ryzhik, Theory of fluid flows through natural rocks, Vol. 395,
Springer, 1990.
URL https://link.springer.com/book/9780792301677
[42] M. Muskat, The flow of fluids through porous media, Journal of Applied Physics 8 (4) (1937) 274–
282. doi:10.1063/1.1710292.
URL https://pubs.aip.org/aip/jap/article-abstract/8/4/274/175338
[43] Y. B. Zel’Dovich, Y. P. Raizer, Physics of shock waves and high-temperature hydrodynamic phe-
nomena, Courier Corporation, 2002. doi:10.1016/B978-0-12-395672-9.X5001-2.
URL https://www.sciencedirect.com/book/9780123956729
[44] J. Boussinesq, Recherches théoriques sur l’écoulement des nappes d’eau infiltrées dans le sol et sur
le débit des sources, Journal de Mathématiques Pures et Appliquées 10 (1904) 5–78.
URL http://eudml.org/doc/235283
[45] H. Triebel, Interpolation Theory, Function Spaces, Differential Operators, North-Holland Mathe-
matical Library, 1978.
URL https://www.sciencedirect.com/bookseries/north-holland-mathematical-library/
vol/18
[46] H. Amann, Nonhomogeneous Linear and Quasilinear Elliptic and Parabolic Boundary Value Prob-
lems, Vieweg+Teubner Verlag, Wiesbaden, 1993, pp. 9–126. doi:10.1007/978-3-663-11336-2_1.
URL https://link.springer.com/chapter/10.1007/978-3-663-11336-2_1
[47] H. Amann, Quasilinear parabolic problems via maximal regularity, Advances in Differential
Equations 10 (10) (2005) 1081–1110. doi:10.57262/ade/1355867805.
URL https://projecteuclid.org/journals/advances-in-differential-equations/
volume-10/issue-10/Quasilinear-parabolic-problems-via-maximal-regularity/ade/
1355867805.short
[48] H. Amann, Linear and Quasilinear Parabolic Problems, Birkhäuser Basel, 1995. doi:10.1007/
978-3-0348-9221-6.
URL https://link.springer.com/book/10.1007/978-3-0348-9221-6
[49] P. C. Kunstmann, L. Weis, Maximal Lp -regularity for Parabolic Equations, Fourier Multiplier Theo-
rems and H ∞ -functional Calculus, Springer Berlin Heidelbergv, Berlin, Heidelberg, 2004, pp. 65–311.
doi:10.1007/978-3-540-44653-8_2.
URL https://link.springer.com/chapter/10.1007/978-3-540-44653-8_2
[50] H. Amann, Maximal regularity for nonautonomous evolution equations, Advanced Nonlinear Studies
4 (4) (2004) 417–430. doi:10.1515/ans-2004-0404.
URL https://www.degruyter.com/document/doi/10.1515/ans-2004-0404/html
[51] P. Guidotti, A new well-posed nonlinear nonlocal diffusion, Nonlinear Analysis: Theory, Methods &
Applications 72 (12) (2010) 4625–4637. doi:10.1016/j.na.2010.02.040.
URL https://www.sciencedirect.com/science/article/pii/S0362546X10001173
[52] R. Chiappinelli, R. Nugari, The nemitskii operator in hölder spaces: some necessary and sufficient
conditions, Journal of the London Mathematical Society 51 (2) (1995) 365–372. doi:10.1112/jlms/
51.2.365.
 A NOVEL NONLINEAR COUPLED NONLOCAL REACTION–DIFFUSION SYSTEM FOR IMAGE RESTORATION 31

URL https://londmathsoc.onlinelibrary.wiley.com/doi/abs/10.1112/jlms/51.2.365
[53] J. Prüss, G. Simonett, R. Zacher, On convergence of solutions to equilibria for quasilinear parabolic
problems, Journal of Differential Equations 246 (10) (2009) 3902–3931. doi:10.1016/j.jde.2008.
10.034.
URL https://www.sciencedirect.com/science/article/pii/S0022039608004403
[54] K. J. Engel, R. Nagel, S. Brendle, One-parameter semigroups for linear evolution equations, Springer
New York, 2000. doi:10.1007/b97696.
URL https://link.springer.com/book/10.1007/b97696
[55] A. Lunardi, Analytic semigroups and optimal regularity in parabolic problems, Birkhäuser Basel,
1995.
URL https://link.springer.com/book/9783034805568
[56] H. Amann, M. Hieber, G. Simonett, Bounded H∞ -calculus for elliptic operators, Differential and
Integral Equations 7 (3-4) (1994) 613 – 653. doi:10.57262/die/1370267697.
URL https://projecteuclid.org/journals/differential-and-integral-equations/
volume-7/issue-3-4/Bounded-H_infty-calculus-for-elliptic-operators/die/1370267697.
short
[57] A. Friedman, Partial Differential Equations of Parabolic Type, R.E. Krieger Publishing Company,
1983.
[58] M. G. Hieber, J. Prüss, Heat kernels and maximal Lp -Lq estimates for parabolic evolution equa-
tions, Communications in Partial Differential Equations 22 (9-10) (1997) 1647–1669. doi:10.1080/
03605309708821314.
URL https://www.tandfonline.com/doi/abs/10.1080/03605309708821314
[59] V. A. Solonnikov, Boundary value problems of mathematical physics. Part 3. On boundary value
problems for linear parabolic systems of differential equations of general form, Vol. 83, Russian
Academy of Sciences, Steklov Mathematical Institute of Russian, 1965.
URL http://mi.mathnet.ru/book1212
[60] J. Weickert, B. Romeny, M. Viergever, Efficient and reliable schemes for nonlinear diffusion filtering,
IEEE Transactions on Image Processing 7 (3) (1998) 398–410. doi:10.1109/83.661190.
URL https://ieeexplore.ieee.org/document/661190
Department of Mathematics, Harbin Institute of Technology, Harbin 150001, China
[email protected]