Kashf Journal of Multidisciplinary Research

Vol: 02 - Issue 1 (2025)

P-ISSN: 3007-1992 E-ISSN: 3007-200X
https://kjmr.com.pk
A NEW TOTAL VARIATION REGULARIZATION-BASED MODEL FOR
ADDITIVE NOISE REMOVAL USING GLOBAL MESHLESS COLLOCATION
SCHEME
Abdul Kabir
University of Engineering and Technology Mardan.
Mushtaq Ahmad Khan
University of Engineering and Technology Mardan.
Muhammad Atif
University of Engineering and Technology Mardan.
Fazal Amin
University of Engineering and Technology Mardan.

*
Corresponding author: Mushtaq Ahmad Khan ([email protected])
DOI: https://doi.org/10.71146/kjmr223

Article Info Abstract

In this work, we proposed a new Total Variation (TV) regularization-based model

for additive noise removal problems using the Global Meshless Collocation
Scheme (GMCS). This new approach not only solves the associated Partial
Differential Equation (PDE) connected to the proposed model for the smooth
solution regarding image restoration, and preservation of edge but also for
minimization of the staircase effect due to which the image looks blocky. The
experimental result demonstrates that the proposed model and meshless scheme
seek to improve computational efficiency and noise removal accuracy in terms of
visual efficiency and Peak-Signal-to-Noise-Ratio (PSNR) values compared to
This article is an open other traditional-based methods.
access article distributed
under the terms and
conditions of the
Creative Commons
Attribution (CC BY)
licensehttps://creativecomm
ons.org/licenses/by/4.0

Keywords:
Total Variation (TV), Image Denoising, Euler-Lagrange PDE, Multi-quadratic
Radial Basis Function (MQ-RBF), Additive noise, Peak-signal-to-noise-ratio
(PSNR).

pg. 146
KJMR VOL.02 NO. 01 (2025) A NEW TOTAL VARIATION...

1. Introduction
One of the most significant features of computer vision and image processing is image denoising. The
focus of this study is additive noise reduction. Image denoising aims to restore the original image by
eliminating noise from a noisy image. However, because texture, edge, and noise are high-frequency
components, it isn't easy to detect them during the denoising process, and the denoised images may
inevitably lose some details. One of the most critical issues these days is recovering crucial information
from noisy images during the noise-removal process to produce high-quality images. [1].
The additive noise model is given by

𝓌0 = 𝓌 + 𝜁. (1)
Where 𝓌: Ω ⊂ ℜ → ℜ2 represents the given true image, 𝓌0 is the noisy image with additive noise 𝜁. In
literature, various nonlinear approaches have been utilized to tackle this problem, such as wavelet
approaches [2, 3], adaptive smoothing [4, 5], stochastic approaches [6, 7], and anisotropic diffusion [8, 9].
Recently Variational approaches have also been utilized to solve such problems, for instance, see [10, 11].
Rudin et al. (ROF) [10] provided the first TV regularization-based model for image restoration with
additive noise. In this model, the TV regularization term is essential for image denoising and edge
preservation. This approach produces good outcomes in the removal of image noise while maintaining the
quality of edges, see [12, 13]. It also presents certain undesirable features such as the staircase effect,
reduced image contrast, and increased computing time due to its nonlinearity and non-differentiability
[10, 14]. The main drawback of this model is that the image restoration outcome is satisfactory in terms
of quality, but it may not be visually apparent. The authors in [10] presented an artificial time-depending
technique for solving the associated EL-PDE of the ROF model. The efficiency of this strategy is limited
by its strict stability constraints in the time intervals. Furthermore, the time-dependent approach calculates
an approximation of the solution rather than the true solution. In recent times, various methodologies have
been employed to address this challenge, resulting in favorable outcomes, for instance [16–18]. But there
is still room for development.
To overcome the abovementioned issue we will propose a new TV regularization-based model.
We incorporate Weber’s law in the regularization term of the ROF model which will examine the image
visually and solve the arising non-linear associated EL-PDE of the proposed model by the Global
Meshless Collocation Scheme (GMCS). This suggested approach will assist minimize the staircase effect,
preserve textures, and preserve fine features throughout the restoration process in addition to aiding with
image denoising and edge preservation.
In recent years, radial basis function (RBF) approaches have gained popularity in both
approximation theory and the numerical solution of PDEs. The RBF collocation approach developed by
Kansa [19,20], also called the Kansa method, is the most widely used RBF strategy for the latter class of
problems. The meshless applications of the Kansa technique, which require only a set of points to
discretize the continuous difficulty, ultimately make it so prevalent. This addition to the method's
implementation is mainly simple, especially for issues involving g two or more dimensions and
complex shapes. As a result, the nonlinear problem solution yields a suitable value for the shape parameter
in addition to the coefficients in the RBF approximation [21]. The Kansa technique has demonstrated
more effectiveness in comparison to FDM [21, 22], a pseudo-spectral method [23], and FEM [24]. For
further information on RBF collocation schemes, see [21], [25, 26].
The RBF Global Meshless Collocation Scheme (GMCS) will be employed to solve the nonlinear
PDE that arises in the suggested model. The primary goals of the suggested meshless scheme for image
restoration are to minimize the staircase effect and preserve texture while maintaining the edges and fine
details in the images through the use of the RBF interpolation process. The smoothness property and the
absence of a mesh or integration procedure will yield the best restoration performance.
pg. 147
KJMR VOL.02 NO. 01 (2025) A NEW TOTAL VARIATION...

The remainder of the paper is organized as follows: Particular details regarding TV regularization, which
is employed in the ROF model [10] for image restoration, are given in Section 2. Information about RBFs
used to solve PDEs is also included in this section. Section 3 presents the suggested additive noise removal
model. The related EL-PDEs for the suggested model are also included in this section. This section also
includes the application of the Global Meshless Collocation Method (GMCM) for the numerical solution
of the suggested model. In terms of CPU time, iterations, and the quality of the restored images, as shown
by the PSNR, Section 4 gives the experimental results and a discussion of the suggested model and
meshless approach in comparison to mesh-based ROF models and other meshless methods. Conclusions
are included in Section 5.
2. Literature Review
2.1. TV regularization-based ROF Model for Additive Noise Removal (M1).
In their crucial work, Rudin, Osher, and Fatemi (ROF) presented an edge-preserving image de-noising
model with desired mathematical features [10]. The approach, which is based on total variation, was
created specifically to remove noise and other undesired fine-scale detail from images while maintaining
crisp discontinuities, or edges. The most basic convex Variational model is the ROF. The minimization
of this approach is as follows,
𝜆
𝓌̂ = 𝑎𝑟𝑔𝑚𝑖𝑛𝓌 𝔼(𝓌) = ∫Ω |∇𝓌| 𝑑𝔵𝑑𝔶 + 2 ‖𝓌 − 𝓌0 ‖22 , (2)

where |𝓌| = √𝓌𝔵2 + 𝓌𝔶2 .

The first term represents the TV regularization of 𝓌

̂ , whereas the second term represents the data fitting
component. The symbol for the regularization parameter is λ. The parameter λ balances the denoising and
smoothing of the denoised image, which are often affected by the amount of noise present. The EL-PDE
connected with the ROF Model is as follows:
∇𝓌
−∇ [|∇𝓌|2 ] + 𝜆(𝓌 − 𝓌0 ) = 0, (3)
+𝜀

in Ω 𝑓𝑜𝑟 𝜖 > 0, (𝔵, 𝔶) ∈ ℜ,

or

𝜕 𝓌𝔵 𝜕 𝓌𝔶
− 𝜕𝔵 ( ) + 𝜕𝔶 ( ) + 𝜆(𝓌 − 𝓌0 ) = 0 𝑖𝑛 Ω, (4)
√𝓌𝔵2 +𝓌𝔶2 √𝓌𝔵2 +𝓌𝔶2

𝜕𝓌
with = 0 on the boundary of Ω = 𝜕Ω. The time-dependent EL-PDE of equation (4) is given as follows:
𝜕𝑛

∇𝓌
𝓌𝑡 = ∇ [|∇𝓌|2 +𝜀] + 𝜆(𝓌 − 𝓌0 ), (5)

𝑖𝑛 Ω 𝑓𝑜𝑟 𝑡 > 0, (𝔵, 𝔶) ∈ ℜ,

or

𝜕𝓌 𝜕 𝓌𝔵 𝜕 𝓌𝔶
= 𝜕𝔵 ( ) + 𝜕𝔶 ( ) + 𝜆(𝓌 − 𝓌0 ) = 0, (6)
𝜕𝑡
√𝓌𝔵2 +𝓌𝔶2 √𝓌𝔵2 +𝓌𝔶2

pg. 148
KJMR VOL.02 NO. 01 (2025) A NEW TOTAL VARIATION...

in Ω 𝑓𝑜𝑟 𝑡 > 0, (𝔵, 𝔶) ∈ ℜ.

𝜕𝓌
For the given 𝓌(𝔵, 𝔶, 0) and also = 0 on 𝜕Ω. For further information, see [10].
𝜕𝑛

2.2. Radial Basis Function (RBF)

RBF interpolation of a continuous multivariate function, 𝒻(𝔵), 𝔵 ∈ Ω ⊆ 𝑅 𝑛 , where Ω is the bounded
domain. For 𝑁 interpolation function values {𝔶𝑗 }𝑁 𝑖=1 ∈ 𝑅 at the data location (which are traditionally
𝑁 𝑛
called centers in the RBF concept){𝔵𝑗 }𝑖=1 ∈ 𝑅 , then 𝒻(𝔵) can be approximated by a linear combination
of RBFs, namely,
𝑆(𝓌) = ∑𝑁 𝑖=1 Γ𝑖 Π(‖𝔵 − 𝔶𝑐𝑖 ‖),
(7)

where 𝜏𝑖 are unidentified coefficients that require determination. The following can be created by applying
the collocation technique:
𝑆(𝔵𝑖 ) = ∑𝑁𝑖=1 Γ𝑖 Π(‖𝔵𝑗 − 𝔶𝑖 ‖) = 𝒻(𝔵𝑖 ),
(8)

𝑓𝑜𝑟 𝑖, 𝑗 = 1,2,3,4 … . 𝑁.
Π(‖𝔵1 − 𝔶1 ‖) Π(‖𝔵1 − 𝔶2 ‖) ⋯ Π(‖𝔵1 − 𝔶𝑁 ‖) Γ1 𝒻(𝔵1 )

Π(‖𝔵2 − 𝔶1 ‖) Π(‖𝔵2 − 𝔶2 ‖) … Π(‖𝔵2 − 𝔶𝑁 ‖) Γ2 𝒻(𝔵2 )

= .
⋮ ⋮ ⋮ ⋮ ⋮ ⋮

[Π(‖𝔵𝑁 − 𝔶1 ‖)Π(‖𝔵𝑁 − 𝔶2 ‖) … Π(‖𝔵𝑁 − 𝔶𝑁 ‖)] [Γ𝑁 ] [𝒻(𝔵𝑁 )]

(9)

The following matrix can represent an 𝑁 × 𝑁 linear system form of the aforementioned system of linear
equations.
𝒜Γ = 𝒻, (10)

Γ = 𝒜−1 𝒻, (11)

Where Γ = (Γ1 , Γ2 , Γ3 … . . Γ𝑁 )𝑡 , 𝒻 = (𝒻1 , 𝒻2 , 𝒻3 , … … . 𝒻𝑁 )𝑡 and 𝒜 = [Π𝑖𝑗 ] = (𝒜𝑖𝑗 ) 𝜖 ℜ𝑁×𝑁

𝒜𝑖𝑗 = [Π𝑖𝑗 ] = Π(‖𝔵𝑗 − 𝔶𝑖 ‖), 𝑖, 𝑗 = 1,2,3, … 𝑁. With Π𝑖𝑗 = Π𝑗𝑖 .

Put equation (11) in equation (8) the interpolation matric is given as:

𝓌 = 𝐷𝒻. (12)
For further information, see [48].
3. Proposed Model
The Rudin-Osher-Fatemi (ROF) model [10] is a TV regularization-based model for the removal of
additive noise problems. The model is defined as follows:
𝜆
𝓌
̂ = 𝑚𝑖𝑛𝓌 𝐸(𝓌) = ℛ(𝓌) + ‖𝓌 − 𝓌0 ‖22 ,
2
(13)

pg. 149
KJMR VOL.02 NO. 01 (2025) A NEW TOTAL VARIATION...

where

ℛ(𝓌) = ∫𝛺 |𝛻𝓌| 𝑑𝔵𝑑𝔶, where |∇𝓌| = √𝓌𝔵2 + 𝓌𝔶2 .

The equation (22) contains the TV regularization term denoted by ℛ(𝓌), and the second term represents
the data fitting term. We incorporate the Weber,s law in the regularization term of the ROF model due to
examine the image visually. The minimization functional of the new TV regularization-based model for
additive noise removal is given below:
|∇𝓌| 𝜆
̂ = 𝑚𝑖𝑛𝓌 𝔼(𝓌) = ℓ1 ∫Ω|∇𝓌|𝑑𝔵𝑑𝔶 + ℓ2 ∫Ω
𝓌 𝑑𝔵𝑑𝔶 + 2 ‖𝓌 − 𝓌0 ‖22 . (14)
𝓌

In order to visually evaluate the image, equation (14) uses Weber's law in the regularization term. The TV
regularization component is addressed in the first part of the equation (14) and the data fitting term is
indicated in the second part.
This model contains three regularization parameters indicated as ℓ1 , ℓ2 , and 𝜆. To show an ideal balance
between the restoration and the degree of smoothness displayed by the restored image, the regularization
parameters are used. This new model's primary goal is to reduce noise and enhance visual efficiency by
getting rid of additive noise. The resulting EL-PDE of equation (14) is given below:
∇𝓌
∇ (|∇𝓌|) + 𝜆̃(𝓌0 − 𝓌) = 0, (15)

𝜆𝓌
Where, 𝜆̃ = (𝓌ℓ +ℓ ).
1 2

Equation (15) can be re-written as:

𝜕 𝓌𝔵 𝜕 𝓌𝔶
( )+ ( ) + 𝜆̃(𝓌0 − 𝓌) = 0, (16)
𝜕𝔵 𝜕𝔶
√𝓌𝔵2 +𝓌𝔶2 √𝓌𝔵2 +𝓌𝔶2

𝑖𝑛 Ω 𝑓𝑜𝑟 𝑡 > 0 (𝔵, 𝔶)𝜖 ℜ.

𝜕𝓌
For the given 𝓌(𝔵, 𝔶, 0), and also = 0 in 𝜕Ω.
𝜕𝑛

The primary benefit of the suggested model is that it uses TV regularization and the Weberized rule to
visually aspect the restored images, especially those with additive noise. Additive noise removal
difficulties are addressed by this suggested model. When a proper algorithm is used, this method reduces
the staircase effect and preserves the textures and edges.
3.1. Proposed Meshless Method (M2)
This subsection is concerned with the numerical solution of the EL-PDE equation (27), using the
global meshless scheme (GMCM). In this approach, the functional of the Weberized law-based TV
regularization-based functional is used in combination with MQ-RBF interpolation (25). The suggested
mesh-free method intends to preserve sharp edges and textures in addition to efficiently restoring the
image (both visually and denoising-wise) while reducing the staircase effect. Therefore, there is a
consistent improvement in PSNR values and calculation time when using the proposed mesh-free
technique.

pg. 150
KJMR VOL.02 NO. 01 (2025) A NEW TOTAL VARIATION...

Assume that {𝔵}𝑁 2

𝑖=1 be 𝑁𝑐 selected center point in a closed domain Ω ⊆ ℜ with RBF equation Π(𝑤) =
𝑁
‖𝑤‖2 in ℜ2 , i.e 𝑤 = (𝔵, 𝔶). Consequently the given known {𝔶𝑐𝑗 } 𝑐 𝑁𝑐 selected center points, the RBF
𝑗=1
interpolation is given by:

𝒮(𝔵) = ∑𝑁
𝑖=1 Γ𝑖 Π(‖𝔵 − 𝔶𝑖 ‖2 ). (17)

By using the following interpolation condition, the coefficient of Γ𝑖 in equation (17) can be found.

𝒮(𝔵𝑖 ) = 𝒻, (18)

Having, a group of points converging in the middle, 𝑁𝑐 . The RBF interpolation is illustrated using the
overdetermined interpolation form in the𝑁𝑐 , data center.

𝒞Γ = 𝒻,
(19)

This is employed to unravel the coefficient Γ and generate, 𝑁𝑐 × 𝑁𝑐 , a system of linear equations. Where,
Γ = (Γ1 , Γ2 , Γ3 , … … . Γ𝑁𝑐 )𝑡 and 𝒻 = (𝒻1 , 𝒻2 , 𝒻3 , … … . 𝒻𝑁𝑐 )𝑡 describe 𝑁𝑐 × 1 order matrices.
In equation (19), 𝒞 is denoted as the 𝑁𝑐 × 𝑁𝑐 system matrix or square interpolation and is expressed in
the following manner:

𝒞 = [Π𝑖𝑗 ] = [Π (‖𝔵𝑐𝑖 − 𝔶𝑐𝑗 ‖2 )] .

1≤𝑖,𝑗≤𝑁𝑐
(20)
Additionally, the matrix in equation (19) is convertible since it is positive definite which a necessary
condition for a unique solution. Therefore

Γ = 𝒞 −1 𝒻, (21)

Furthermore, at {𝔵𝑖 }𝑁
𝑖=1 be 𝑁 evaluation data point, the RBF interpolation by using equation (17) given
𝑁𝑐 × 𝑁 matrix 𝒟 which is written as follows:

𝒟 = [Π𝑖𝑗 ] = [Π (‖𝔵𝑖 − 𝔶𝑐𝑗 ‖2 )] 𝑓𝑜𝑟 𝑖 = 1,2,3, … … 𝑁, 𝑗 = 1,2,3 … … . . 𝑁𝑐 .

(22)
There are 𝑁 data points as well to approximate the over-determined condition of interpolation, the matrix-
vector product is utilized to generate 𝓌, as shown below:
𝓌 = 𝒟Γ. (23)
Merging equation (21) and equation (23) results in the given equation

𝓌 = 𝒟𝐶 −1 𝒻, (24)
or

𝓌 = 𝒮𝒻 𝑤ℎ𝑒𝑟𝑒 𝒮 = 𝒟𝒞 −1 . (25)

Which represents the evaluated solution at any point in Ω. where 𝓌 is 𝑁 × 1 order matrix.
Since equation (16) is

pg. 151
KJMR VOL.02 NO. 01 (2025) A NEW TOTAL VARIATION...

𝜕 𝓌𝔵 𝜕 𝓌𝔶
( ) + 𝜕𝔶 ( ) + 𝜆̃(𝓌0 − 𝓌) = 0, (26)
𝜕𝔵
√𝓌𝔵2 +𝓌𝔶2 √𝓌𝔵2 +𝓌𝔶2

∂𝓌
𝑖𝑛 Ω, = 0 𝑜𝑛 𝑡ℎ𝑒 Ω.
∂n

The time marching EL-PDE of equation (26) is given as the following equation:

𝑑𝓌 𝜕 𝓌𝔵 𝜕 𝓌𝔶
= 𝜕𝔵 ( ) + 𝜕𝔶 ( ) + 𝜆̃(𝓌0 − 𝓌),
𝑑𝑡
√𝓌𝔵2 +𝓌𝔶2 √𝓌𝔵2 +𝓌𝔶2

(27)

𝑖𝑛 Ω 𝑓𝑜𝑟 𝑡 > 0 (𝔵, 𝔶) 𝜖 ℜ,

𝜕𝓌
for the given 𝓌(𝔵, 𝔶, 0) with = 0 on 𝜕Ω. The equation (27) is re-written as
𝜕𝑛

(𝓌𝔵2 +𝓌𝔶2 )(𝓌𝔵𝔵 +𝓌𝔶𝔶 )−

( ).
𝑑𝓌 (𝓌𝔵𝔵 𝓌𝔵 +2𝓌𝔵 𝓌𝔶 (𝓌𝔵 𝓌𝔶 +𝓌𝔶 𝓌𝔵 )+𝓌𝔶𝔶 𝓌𝔶 )
= 3 + 𝜆̃(𝓌0 − 𝓌). (28)
𝑑𝑡
(𝓌𝔵2 +𝓌𝔶2 )2
The implicit gradient decent scheme is applied to equation (28) we get
(𝑛) (𝑛) (𝑛)
((𝓌𝔵2 )(𝑛) +(𝓌𝔶2 ) )(𝓌𝔵𝔵 +𝓌𝔶𝔶 )−
(𝑛) (𝑛) (𝑛) (𝑛) (𝑛) (𝑛)
𝓌 (𝑛+1) −𝓌 (𝑛) (2𝓌𝔵 𝓌𝔶 (𝓌𝔵 𝓌𝔶 +𝓌𝔵 𝓌𝔶 )𝓌𝔵 𝓌𝔵𝔵 +𝓌𝔶 𝓌𝔶𝔶 ) (0)
= 3 + 𝜆̃(𝓌0 − 𝓌 (𝑛) ) (29)
𝑑𝑡
((𝓌𝔵2 )(𝑛) +(𝓌𝔶2 )(𝑛) )2

( )
Equation (29) and equation (25) together produce a nonlinear system of equations, which the GMCS is
used to determine. The given equation must be taken into consideration in order to solve the nonlinear
restoration system of equations.

(𝑛) (𝑛) (𝑛) (𝑛)

𝐺(𝓌 (𝑛) )𝓌 (𝑛+1) = 𝐺(𝓌 (𝑛) )𝓌 (𝑛) + 𝑑𝑡 [((𝓌𝔵2 )(𝑛) + (𝓌𝔶2 ) ) (2𝓌𝔵 𝓌𝔶 (𝓌𝔵 𝓌𝔶 +
(𝑛) (𝑛) (𝑛) (0)
𝓌𝔵 𝓌𝔶 ) + 𝓌𝔵 𝓌𝔵𝔵 + 𝓌𝔶 𝓌𝔶𝔶 )] + 𝐺(𝓌 (𝑛) )𝑑𝑡 [𝜆̃(𝓌0 − 𝓌 (𝑛) )], (30)
3
𝜕𝓌
where 𝐺(𝓌) = (𝓌𝔵2 + 𝓌𝔶2 )2 , 𝓌𝔵 = 𝒮𝔵 𝒻, 𝓌𝔶 = 𝒮𝔶 𝒻, 𝓌𝔵𝔵 = 𝒮𝔵𝔵 𝒻, 𝓌𝔶𝔶 = 𝒮𝔶𝔶 𝒻, = 𝓌𝑛 = 𝒮𝑛 𝒻,
𝜕𝑛
𝑎𝑛𝑑 𝒻 0 = 𝒻.
The MQ-RBF is selected for better restoration performance as a basis function in this case. The derivation
has been done in a similar way as done in [27-29].
4. Result and Discussion
This section presents numerical results that illustrate the effectiveness of our suggested technique M2.
The outcomes of technique M1 and the acquired results are contrasted. Figure 1 displays the test images:
"Real 1 and Artificial 1."
In this study, it is expected that N = Nc = the size of the image, for our scheme M2, for the sake of
comparison with scheme M1. Here, the suggested approach M2 makes use of the Multi-quadric Radial

pg. 152
KJMR VOL.02 NO. 01 (2025) A NEW TOTAL VARIATION...

Basis Function (MQ-RBF). The peak signal-to-noise ratio (PSNR) is taken into consideration in order to
quantify the denoised image. The following formula can calculate it

𝑀×𝑁𝑚𝑎𝑥{𝓌̂ }2
𝑃𝑆𝑁𝑅 = 10 × 𝑙𝑜𝑔10 ( ). (31)
‖𝓌̂ −𝓌‖2
Where 𝓌
̂ is the given image, 𝓌 is the restored image and 𝑀 × 𝑁 is the size of an image.

Fig-1

Figure 1: Test Images

Figure 2: Left to right: (a) is true image; (b) Noisy image connected with additive Gaussian noise
of 𝑳𝟏 = 𝟑𝟕%; (c) Restored image with M1; (d) Restored image with M2. Also zoomed-in images
for staircase effect of true, noisy, restored by M1 and M2 are given in (e), (f), (g), and (h).

Figure 3: Left to right; (f) True image; (g) Noisy image connected with salt and pepper noise of
𝑳𝟐 = 𝟒𝟎%; (h) Obtained by M1; (i) Obtained by M2.

pg. 153
KJMR VOL.02 NO. 01 (2025) A NEW TOTAL VARIATION...

Table: 1 PSNR values, number of iterations required for convergence, and CUP time
comparison of Algorithm M1 and M3.

Image Size M1 M3
PSNR Iters Time(s) PSNR Iters Time(s)
Real1 2562 28.22 29 13.59 28.67 18 9.38

Artifical1 2562 23.93 34 16.70 24.30 27 10.79

Test Problem:
In this investigation, Real1 and Artifical1 images have been taken for image restoration having additive
Gaussian and salt and pepper noise with noise levels 𝐿1 = 37% and 𝐿2 = 40%, respectively. The true
images are given in Figures 2 and 3 as (a), (e) and (i) respectively while the noisy images are represented
by (b), (f), and (j), respectively. The mesh-based method M1 used on the ROF model produces a good
restoration result but produces the staircase effect which is the main drawback of the M1 and ROF model.
Also, the restoration results are also affected a bit due to the nonlinearity and non-differentiability of the
associated EL-PDE. All these images are shown in Figures 2 and 3 as (c), (g) and (k), respectively. But
the restoration results i.e., image denoising, preservation of edges and reduction of staircase effects
obtained by the proposed model and meshless scheme M2 are far better than M1 due to the application
of the new model and meshless method with applications of adaptive nature, and MQ-RBF applied on
them [20]. Similarity, due to the additional application of Weberized law involved in the proposed model
makes results effective in visual representation compared to the model M1 applied on the ROF model.
These obtained images are shown in Figures 2 and 3 as (d), (h) and (l), respectively. It can also be noticed
from Table 1 that the PSNR values of M2 are greater than M1 which shows the best restoration
performances of M2 over M1. Also, Table 1 indicates that the number of iterations and time required for
convergence of M2 are less than M1 which shows the quick restoration performance of M2 due to its
meshless applications and easy implementation in comparison to M1.

5. Conclusion
In this paper, a novel TV-based approach for removing additive noise from noisy image data is presented.
The GMCS solves the associated EL-PDE with a new model for a smooth solution that uses the MQ-RBF
as the basis function. This solution is adaptable and computationally easy because of the meshless and
MQ-RBF features. In comparison to the ROF Model, the outcomes are determined based on visual quality
and PSNR value, staircase effect minimization, and texture and edge preservation. The new model and
corresponding meshless technique are tested on both artificial and actual images, and the outcomes are
contrasted with those of conventional techniques and ROF models. The experimental results showed that
the suggested model and meshless approach are significantly better at image restoration PSNR values
(image denoising and visual efficiency due to the proposed model) and other related aspects, such as the
preservation of blocky effects, texture, and sharp edges, the number of iterations needed for convergence,
and the CPU time (proposed method).

pg. 154
KJMR VOL.02 NO. 01 (2025) A NEW TOTAL VARIATION...

References
1. R. C. Gonzalez and R. E. Woods, Digital Image Processing, 3rd ed. Englewood Cliffs, NJ, USA:
Prentice-Hall, 2008.
2. Li, X., Li, L., & Wang, Q. H. (2018). Wavelet based iterative perfect reconstruction in
computational integral imaging. JOSA A, 35(7), 1212-1220.
3. Naveed, K., Shaukat, B., Ehsan, S., Mcdonald Maier, K. D., & Ur Rehman, N. (2019). Multiscale
image denoising usinggoodness-of-fit test based on EDF statistics. PLoS One, 14(5), e0216197.
4. Kervrann, C. (2004). An adaptive window approach for image smoothing and structures
preserving. In Computer Vision-ECCV 2004: 8th European Conference on Computer Vision,
Prague, Czech Republic, May 11-14, 2004. Proceedings, Part III 8 (pp. 132-144). Springer
Berlin Heidelberg.
5. Polzehl, J., & Tabelow, K. (2007). Adaptive smoothing of digital images: The R package adimpro.
6. Ramadhan, A., Mahmood, F., & Elci, A. (2017). Image denoising by median filter in wavelet
domain. arXiv preprint arXiv:1703.06499.
7. Luo, L., Zhao, Z. Q., Li, X. P., & Feng, X. C. (2019). A stochastic image denoising method based
on adaptive patch size. Multidimensional Systems and Signal Processing, 30, 705-725.
8. Bai, J., & Feng, X. C. (2018). Image denoising using generalized anisotropic diffusion. Journal of
Mathematical Imaging and Vision, 60, 994-1007.
9. Liu, F., & Liu, J. (2012). Anisotropic diffusion for image denoising based on diffusion
tensors. Journal of Visual Communication and Image Representation, 23(3), 516-521.
10. Rudin, L. I., Osher, S., & Fatemi, E. (1992). Nonlinear total variation based noise removal
algorithms. Physica D: nonlinear phenomena, 60(1-4), 259-268.
11. Russ, J. C., & Russ, J. C.(2017). Introduction to image processing and analysis. CRC
press.
12. Osher, S., Burger, M., Goldfarb, D., Xu, J., & Yin, W. (2005). An iterative regularization
method for total variation based image restoration. Multiscale Modeling & Simulation, 4(2), 460-
489.
13. Chan, T., Esedoglu, S., Park, F., & Yip, A. (2006). Total variation image restoration: Overview
and recent developments. Handbook of mathematical models in computer vision, 17-31.
14. Chang, Q., Tai, X. C., & Xing, L. (2009). A compound algorithm of denoising using second-order
and fourth-order partial differential equations. Numer. Math. Theory Methods Appl, 2, 353-376.
15. Thanh, D. N., Prasath, V. S., Hieu, L. M., & Dvoenko, S. (2020). An adaptive method for image
restoration based on high-order total variation and inverse gradient. Signal, Image and Video
Processing, 14(6), 1189-1197.
16. Jiang, D. H., Tan, X., Liang, Y. Q., & Fang, S. (2015). A new nonlocal variational bi regularized
image restoration model via split Bregman method. EURASIP Journal on Image and Video
Processing, 2015, 1-10.
17. Xu, H., Sun, Q., Luo, N., Cao, G., & Xia, D. (2013). Iterative nonlocal total variation regularization
method for image restoration. PloS one, 8(6), e65865.
18. Guo, J., & Chen, Q. (2021). Image denoising based on nonconvex anisotropic total
variation regularization. Signal Processing, 186, 108124.
19. Kansa, E. J. (1990). Multiquadrics—A scattered data approximation scheme with
applications to computational fluid dynamics—I surface approximations and partial derivative
estimates. Computers & Mathematics with applications, 19(8 9), 127-145.
20. Kansa, E. J. (1999). Motivation for using radial basis functions to solve PDEs. RN, 64(1), 1.
21. Jankowska, M. A., Karageorghis, A., & Chen, C. S. (2018). Kansa RBF method for nonlinear
problems. Boundary Elements and Other Mesh Reduction Methods, 39.

pg. 155
KJMR VOL.02 NO. 01 (2025) A NEW TOTAL VARIATION...

22. Zerroukat, M., Power, H., & Chen, C. (1998). A numerical method for heat transfer problems using
collocation and radial basis functions. International Journal for numerical methods in
Engineering, 42(7), 1263-1278.
23. Larsson, E., & Fornberg, B. (2003). A numerical study of some radial basis function based solution
methods for elliptic PDEs. Computers & Mathematics with Applications, 46(5-6), 891-
902.
24. Li, J., Cheng, A. H. D., & Chen, C. S. (2003). A comparison of efficiency and error convergence
of multiquadric collocation method and finite element method. Engineering Analysis with
Boundary Elements, 27(3), 251-257.
25. Parand, K., & Rad, J. A. (2012). Numerical solution of nonlinear Volterra Fredholm-
Hammerstein integral equations via collocation method based on radial basis
functions. Applied Mathematics and Computation, 218(9), 5292-5309.
26. Doha, E. H., Baleanu, D., Bhrawy, A. H., & Abdelkawy, M. A. (2013). A Jacobi Collocation
Method for Solving Nonlinear Burgers‐Type Equations. In Abstract and Applied Analysis (Vol.
2013, No. 1, p. 760542). Hindawi Publishing Corporation.
27. Khan, M. A., Chen, W., Ullah, A., & Fu, Z. (2017). A mesh-free algorithm for ROF
model. EURASIP Journal on Advances in Signal Processing, 2017, 1-16.
28. Ahmad Khan, M., Altamimi, A., Hussain Khan, Z., Shehzad Khattak, K., Khan, S., Ullah, A., &
Ali, M. (2021). Multiquadric radial basis function approximation scheme for solution of total
variation based multiplicative noise removal model. Computer Modeling in Engineering &
Sciences, 126(1), 55-88.
29. Khan, M. A., Altamimi, A. B., Khan, Z. H., Khattak, K. S., Ali, M., Ullah, A., ... & Abrar, M. F.
(2020). Total Variation Filter via Multiquadric Radial Basis Function Approximation Scheme for
Additive Noise Removal. IEEE Access, 8, 88241-88258.

pg. 156