Topological Methods in Nonlinear Analysis

Journal of the Juliusz Schauder Center

Volume 31, 2008, 305–313

APPLICATION OF HOMOTOPY PERTURBATION METHOD

TO REGULARIZATION OF SCALAR IMAGES

Qin Ma — Ru-Yi Xing — Shu-Li Mei

Abstract. The homotopy perturbation method is implemented to solve

nonlinear equations. Based on this method, a multi-step scheme is con-
structed for a kind of Hamilton–Jacobi formulations by assuming the ho-
motopy parameter is a linear function of time. Using this multi-step scheme,
a minimal surface regularization equation is solved, which designates a reg-
ularization process that doesn’t smooth the image with the same weight
in all the spatial directions. Some image denoisying examples illustrate its
effectiveness and convenience.

1. Introduction

For several years, regularization algorithms have attracted a growing interest

in the computer vision community. Isotropic linear regularization is a natural
way to smooth and simplify data and has consequently been reached by several
mathematical formulations: from the restoration scheme to the classical linear
filtering of images, all these methods lead to the same regularization behavior,
e.g. the signal is blurred little by little in an isotropic way during the PDE evolu-
tion. To overcome the limitations of linear methods leading to isotropic smooth-
ing, several nonlinear extensions of the heat equations are proposed, which can

2000 Mathematics Subject Classification. Primary 35F25; Secondary 68U10.

Key words and phrases. Hamilton–Jacobi formulation, homotopy perturbation method,
precise integration method, image regularization, variational iteration method.
The third named author (corresponding author) was supported by NSF of China Grant
#60772038 and National High Technology Research and Development Program of China Grant
#2006AA10Z235.

c 2008 Juliusz Schauder Center for Nonlinear Studies

305
306 Q. Ma — R.-Y. Xing — S.-L. Mei

be classified as the Hamilton–Jacobi formulation. Comparing with other image

processing methods, the PDE based method becomes one of the major tools in
computer vision and image processing [4]. In recent years, many excellent algo-
rithms were developed such as TVD schemes, ENO and WENO schemes [14],
but almost all of these methods were based on the finite difference method. The
usual finite difference method can be executed with a narrow band-width, but
have to face the problems of numerical stability and precision [25].
The homotopy perturbation method (HPM) proposed by He [9], [11] is con-
stantly being developed and applied to solve various nonlinear problems by Ji-
Huan He [7], [17] and by others [1], [2], [5], [6], [8], [10], [12], [15], [16], [18]–[22].
Unlike other analytical perturbation methods, HPM does not depend on small
parameter which is difficult to be found. The variational iteration method was
another simple and effective method for nonlinear equations proposed by He [8],
[12], which can provide analytical approximations to a rather wide class of nonlin-
ear equations [24], [25], [3] without linearization, perturbation, or discretization
which can result in massive numerical computation.
In fact, the Hamilton–Jacobi formulation could be transformed to matrix
differential equations derived by the semi-analytical method from the PDEs. In
this paper, we try use the coupling technique of He’s VIM and HPM to solve
the matrix differential equation. The corresponding numerical result could be
obtained by the precise integration method (PIM) proposed by Zhong [25]. In
contrast to the traditional finite difference approximation, the numerical results
of PIM for a set of simultaneous linear time-invariant ODEs have computer
precision and also are free from the stiff problem.

2. Fundamental theory of coupling technique of VIM and HPM

Consider the nonlinear matrix differential equation as follows

(2.1) L(V̇, V, t) + N (V̇, V, t) = G(t)

where L is a linear operator, N a nonlinear operator and G an inhomogeneous

term, V is an n-dimensional unknown vector, the dot stands for the differential
with respect to the time t. For convenience, (2.1) can be rewritten as

(2.2) V̇ − HV − F(V̇, V, t) = 0

Here H is a given n × n constant matrix, and F(V̇, V, t) is a n-dimensional non-

linear external force vector. According to the variational iteration method [24],
we can write down a correction functional as follows
Z t
Vn+1 (t) = Vn (t) + e˙ n , V
λ[V̇n (τ ) − HVn (τ ) − F(V e n , τ )] dτ
0
 Appication of Homotopy Perturbation Method 307

where λ is a general Lagrange vector multiplier [24] which can be identified opti-
mally via the variational theory. The subscript n denotes the nth approximation
and Ve n is considered as a restricted variation, i.e. δ V
e n = 0.
Taking the exact analytic solution of V̇ − HV = 0 as the initial approxima-
tion, we have:
Z t
Vn+1 (t) = Vn (t) + e˙ n , V
eH (t−τ ) F(V e n , τ ) dτ
0

Ht
where the exponential matrix e can be calculated accurately by PIM. A linear
homotopy function for (2.2) can be constructed as

(2.3) V̇ − HV − F(V̇, V, t) + ε[f0 − F(V̇, V, t)] = 0

where f0 is a known initial value vector, and ε ∈ [0, 1] is the homotopy parameter.
According to the perturbation theory, the solution of (2.3) can be expressed as
the power series expansion of ε

(2.4) V = V0 + εV1 + ε2 V2 + . . .

Substituting equation (2.4) into (2.3), we have:

(2.5) ε0 : V̇0 = HV0 + f0 ,

(2.6) ε1 : V̇1 = HV1 − f0 + f1 ,
(2.7) ε2 : V̇2 = HV2 + f2

where f1 and f2 are the terms containing ε1 and ε2 , respectively in the expansion
of εF(V̇, V, t). According to VIM, taking the iteration once and setting the
initial values of V1 and V2 as 0, the general solutions of (2.5)–(2.7) could be
obtained exactly. Substituting solutions of (2.5)–(2.7) into (2.4), and assuming
ε = 1, the approximate analytic solution of (2.1) can be obtained subsequently.
As the exponential matrix eHt can be calculated accurately by PIM [25], the
numerical solution of (2.1) can be obtained lastly.

3. The application of HPM in regularization

of scalar images with PDEs

A scalar image can be defined as

(
Ω ⊂ Rp → R,
I:
x → I(x),

where x= x when p = 1, x= (x, y) when p = 2, Ω ⊂ Rp is a closed spatial

domain of dimension p, Ω is the defined domain of images. The derivative of
the image I with respect to the variable a is written as Ia = ∂I/∂a, and the
308 Q. Ma — R.-Y. Xing — S.-L. Mei

second derivative of a scalar image I with respect to a then to b is denoted by

Iab = ∂ 2 I/(∂a∂b).
Consider the diffusion PDE as follows

 It=0 = Inoisy ,
 0
(3.1)

 ∂I = div φ (k∇Ik) ∇I ,
∂t k∇Ik
√
where φ = 2 1 + s2 −2, which called Hyper-surface schemes. Of course, different
choices of function φ lead to different regularization methods.
According definitions of the image gradient ∇I and the image gradient norms
k∇Ik, (3.1) can be rewritten as

It=0 = Inoisy ,
(
(3.2) ∂I
= a(Ix , Iy )Ixx + b(Ix , Iy )Ixy + c(Ix , Iy )Iyy ,
∂t

where
 1 + Iy 2

 a(I x , I y ) = 2 q ,
2 2 2 2

(I + I ) I + I

x y x y




4Ix Iy Ixy



(3.3) b(Ix , Iy ) = − q ,

 (Ix + Iy 2 ) Ix 2 + Iy 2
2


1 + Ix 2



c(I , I ) = 2 .

x y

 q
(Ix 2 + Iy 2 ) Ix 2 + Iy 2


It is easy to see that a(Ix , Iy ), b(Ix , Iy ) and c(Ix , Iy ) are decreasing functions
vanishing on edges in order to stop the diffusion. So, we can linearize these
functions with respect to the time t approximately, and the influence of this
approximation on the regularization of scalar images should be small.
Using the central difference scheme, we have
 I(xi+1 , yj ) − I(xi−1 , yj )
 Ix = ,
2h





 I(xi , yj+1 ) − I(xi , yj−1 )
Iy = ,


2k




I(xi+1 , yj ) − 2I(xi , yj ) + I(xi−1 , yj )

(3.4) Ixx = ,

 h2
I(xi , yj+1 ) − 2I(xi , yj ) + I(xi , yj−1 )




 Iyy = ,
k2




I(xi+1 , yj+1 ) − I(xi−1 , yj+1 ) − I(xi+1 , yj−1 ) + I(xi−1 , yj−1 )


Ixy = ,

4hk
 Appication of Homotopy Perturbation Method 309

where xi = x0 + ih, yj = y0 + jk. Substituting (3.4) into (3.2), a system of

nonlinear ODEs can be obtained as follows:

dI(xi , yj , t) Ii+1,j − 2Ii,j + Ii−1,j

(3.5) = ai,j
dt h2
Ii+1,j+1 − Ii−1,j+1 − Ii+1,j−1 + Ii−1,j−1 Ii,j+1 − 2Ii,j + Ii,j−1
+ bi,j + ci,j
4hk k2

where t ∈ [t0 , t1 ], Ii,j denotes I(xi , yj , t0 ); ai,j , bi,j , ci,j denote a(Ix (xi , yj , t),
Iy (xi , yj , t)), b(Ix (xi , yj , t), Iy (xi , yj , t)) and c(Ix (xi , yj , t), Iy (xi , yj , t)), respec-
tively. Let

I = ((I(x0 , y0 ), . . . , I(xn , y0 )), . . . ,

(I(x0 , yj ), . . . , I(xi , yj ), . . . , I(xn , yj )), . . . , I(xn , ym ))T ,
A = diag((a0,0 , a1,0 , . . . , an,0 ), (a0,1 , a1,1 , . . . , an,1 ), . . . , (a0,m , a1,m , . . . , an,m )),
B = diag((b0,0 , b1,0 , . . . , bn,0 ), (b0,1 , b1,1 , . . . , bn,1 ), . . . , (b0,m , b1,m , . . . , bn,m )),
C = diag((c0,0 , c1,0 , . . . , cn,0 ), (c0,1 , c1,1 , . . . , cn,1 ), . . . , (c0,m , c1,m , . . . , cn,m )),

 −2 1 0 0 ... 0 

 1 −2 1 0 ... 0 

1  0 1 −2 1 ... 0 
M1 = 2 
 
h  0 ... 0 1 −2 1


 .. .. .. .. . . .. 
 . . . . . . 
0 0 ... 0 1 −2 (n×m)×(n×m)

 n−1 n 
z }| { z }| {
0 −1 0 ... 0 0 ... 0 −1 0 1 0 ... 0
1 
1 0 −1 0 ... 0 ... 0 0 −1 0 1

0 ...
M2 =
 
4hk
 0 1 0 −1 0 0 ... 0 0 0 −1 0 1 0
.
.. . . . . . . . . . ... .. . . .. .. .. .. . . . . .. 
. . . . . . . . .
0 ... 0 1 0 1 0 ... 0 0 0 ... −1 0 (n×m)×(n×m)

 n n 
z }| {z }| {
1 0 ... 0 −2 0 ... 0 1 0 ... 0
1  0 −2 0 ...

M3 = 2
 0 1 0 ... 0 1 0 ...
 .
k  .. . . .. . . .. . . .. . . .. . . .. .. 
. . . . . . . . . . . .
0 ... 0 1 0 ... 0 −2 0 ... 0 1 (n×m)×(n×m)

And then (3.5) can be rewritten as the matrix differential equation:

d
I = (AM1 + BM2 + CM3 )I.
dt
310 Q. Ma — R.-Y. Xing — S.-L. Mei

According to HPM, we can construct the linear homotopy function as follows

 
dI
(3.6) − (A0 M1 + B0 M2 + C0 M3 )I
dt
 
dI
+ε − ((A0 − A)M1 + (B0 − B)M2 + (C0 − C)M3 )I = 0
dt
where A0 , B0 , C0 are the known initial values, ε ∈ [0, 1] is the homotopy pa-
rameter. According to the homotopy perturbation theory, the solution of (3.6)
can be expressed as the power series expansion of ε:

(3.7) I = I0 + εI1 + ε2 I2 + . . .

Substituting (3.7) into (3.6), we have:

dI0
(3.8) ε0 : = (A0 M1 + B0 M2 + C0 M3 )I0 ,
dt
dI1
(3.9) ε1 : = [(A − A0 )M1 + (B − B0 )M2 + (C − C0 )M3 ] I1 .
dt
Assuming ε = t/τ , t ∈ [0, τ ], A in (3.9) can be identified as [A(I0 )](n×m)×(n×m)
by the definition of Taylor series. So, the solution can be obtained by multi-
step scheme, that is, we can calculate I0 from (3.8) firstly, and then obtaine the
matrix A in (3.9). Secondly, we solve (3.9) and obtain I1 by PIM. Letting ε = 1,
and substituting I0 and I1 into (3.7), the regularization image I can be obtained.
The comparison of HPM with the common difference method is shown in
Figure 1. As this system has no exact analytic solution, the solution obtained by
the Adams–Bashforth–Moulton method (ABM) which is built into Matlab 7.0
is taken as the exact solution. In the options of ABM, the scalar relative error
tolerance ’RelTol’ (1e-3 by default) is taken as 1e-6, and the vector of absolute
error tolerances ’AbsTol’ (all components 1e-6 by default) is taken as 1e-12.
Solutions obtained by ABM at specific times 3, 6, 9 are shown in Figure 1
(c1)–(c3), respectively. The corresponding solutions obtained by HPM and the
difference method (DM) are shown in Figure 1 (d1)–(d3) and Figure 1 (e1)–(e3),
respectively. It should be noted that the time step τ in HPM is the same as
in DM. Obviously, regularization results of a noisy image obtained by HPM and
ABM are similar. But in the results with the difference method, a large amount
of noise doesn’t vanish and even is enhanced. This indicates that the difference
method is invalid if a larger time step is taken.

4. Conclusion

The multi-step scheme based on HPM developed in this paper can solve non-
linear diffusion differential equations successfully. Comparison of scalar images
regularization with the hyper-surfaces schemes reveals that HPM can obtain
higher accuracy than the common difference method. Otherwise, HPM is not
 Appication of Homotopy Perturbation Method 311

(a) (b)

(c1) (c2) (c3)

(d1) (d2) (d3)

(e1) (e2) (e3)

Figure 1. HPM and the difference method applied in regularization of

a noisy scalar image. (a) Scalar image, (b) noisy image, (c1) t = 30, (c2) t =
60, (c3) t = 90, (d1) HPM, τ = 30, 1 iteration, (d2) τ = 30, 2 iterations,
(d3) τ = 30, 3 iterations, (e1) DM, τ = 30, 1 iteration, (e2) τ = 30,
2 iterations, (e3) τ = 30, 3 iterations.
312 Q. Ma — R.-Y. Xing — S.-L. Mei

sensitive to the time step and possesses better stability. Actually, HPM for ma-
trix differential equations has the uniform analytic solution format, and so it can
be easily used to solve various nonlinear problems.

References
[1] P. D. Ariel, T. Hayat and S. Asghar, Homotopy perturbation method and axisym-
metric flow over a stretching sheet, Internat. J. Nonlinear Sci. 7 (2006), 399–406.
[2] A. Belendez, A. Hernandez and T. Belendez, Application of He’s homotopy per-
turbation method to the Duffing-harmonic oscillator, Internat. J. Nonlinear Sci. 8 (2007),
79–88.
[3] N. Bildik and A. Konuralp, The use of variational iteration method, differential
transform method and Adomian decomposition method for solving different types of non-
linear partial differential equations, Internat. J. Nonlinear Sci. 7 (2006), 65–70.
[4] R. P. Fedkiw, Shock capturing, level sets, and PDE based methods in computer vision
and image processing: a review of Osher’s contributions, J. Comput. Phys. 185 (2003),
309–341.
[5] D. D. Ganji and A. Sadighi, Application of He’s homotopy-perturbation method to
nonlinear coupled systems of reaction-diffusion equations, Internat. J. Nonlinear Sci. 7
(2006), 411–418.
[6] A. Ghorbani and J. Saberi-Nadjafi, He’s homotopy perturbation method for calcu-
lating adomian polynomials, Internat. J. Nonlinear Sci. 8 (2007), 229–232.
[7] Q. K. Ghori, M. Ahmed and A. M. Siddiqui, Application of homotopy perturbation
method to squeezing flow of a Newtonian fluid, Internat. J. Nonlinear Sci. 8 (2007),
179–184.
[8] J. H. He, Variational iteration method — a kind of non-linear analytical technique:
some examples, Internat. J. Non-Linear Mech. 34 (1999), 699–708.
[9] , A coupling method of a homotopy technique and a perturbation technique for
non-linear problems, Internat. J. Non-Linear Mech. 35 (2000), 37–43.
[10] , Homotopy perturbation method for bifurcation of nonlinear problems, Internat.
J. Nonlinear Sci. 7 (2005), 207–208.
[11] , Some asymptotic methods for strongly nonlinear equations, Internat. J. Mod.
Phys. B 20 (2006), 1141–1199.
[12] , Variational iteration method — some recent results and new interpretations,
J. Comput. Appl. Math. 207 (2007), 3–17.
[13] Z. M. Odibat and S. Momani, Application of variational iteration method to nonlinear
differential equations of fractional order, Internat. J. Nonlinear Sci. 7 (2006), 27–34.
[14] S. Osher and J. A. Sethian, Fronts propagating with curvature dependent speed:
algorithms based on Hamilton–Jacobi formulations, J. Comput. Phys. 79 (1988), 12–49.
[15] T. Ozis and A. Yildirim, Traveling wave solution of Korteweg–de Vries equation using
He’s homotopy perturbation method, Internat. J. Nonlinear Sci. 8 (2007), 239–242.
[16] , A comparative study of He’s homotopy perturbation method for determining
frequency-amplitude relation of a nonlinear oscillator with discontinuities, Internat. J.
Nonlinear Sci. 8 (2007), 243–248.
[17] M. A. Rana, A. M. Siddiqui and Q. K. Ghori, Application of He’s homotopy per-
turbation method to Sumudu transform, Internat. J. Nonlinear Sci. 8 (2007), 203–210.
[18] M. Rafei and D. D. Ganji, Explicit solutions of Helmholtz equation and fifth-order
KdV equation using homotopy perturbation method, Internat. J. Nonlinear Sci. 7 (2006),
321–328.
 Appication of Homotopy Perturbation Method 313

[19] D. H. Shou and J. H. He, Application of parameter-expanding method to strongly

nonlinear oscillators, Int. J. Nonlinear Sci. 8 (2007), 121–124.
[20] A. M. Siddiqui, R. Mahmood and Q. K. Ghori, Thin film flow of a third grade fluid
on a moving belt by He’s homotopy perturbation method, Internat. J. Nonlinear Sci. 7
(2006), 7–14.
[21] , TCouette and Poiseuille flows for non-Newtonian fluids, Internat. J. Nonlinear
Sci. 7 (2006), 15–26.
[22] H. Tari, D. D. Ganji and M. Rostamian, Approximate solutions of K(2, 2), KdV and
modified KdV equations by variational iteration method, homotopy perturbation method
and homotopy analysis method, Internat. J. Nonlinear Sci. 8 (2007), 185–190.
[23] D. Tschumperle, PDE’s Based Regularization of Multivalued Images and Applications,
Ph.D. thesis, Universite De Noce, Sophia Antipolis, 2002.
[24] E. Yusufoglu, Variational iteration method for construction of some compact and
noncompact structures of Klein–Gordon equations, Internat. J. Nonlinear Sci. 8 (2007),
153–158.
[25] W. X. Zhong, On precise integration method, J. Comput. Appl. Math. 163 (2004),
59–78.

Manuscript received September 17, 2007

Qin Ma and Shu-Li Mei

College of Information and Electrical Engineering
Agricultural University
Beijing 100083, P. R. CHINA
E-mail address: [email protected], [email protected]

Ru-Yi Xing
Center of Computer
Hebei University of Engineering
Handan, Heibei 056038, P. R. CHINA
E-mail address: [email protected]

TMNA : Volume 31 – 2008 – No 2