Upload
shu-mei-guo
View
214
Download
0
Embed Size (px)
Citation preview
An intelligent image agent based on soft-computing techniques
for color image processing
Shu-Mei Guoa, Chang-Shing Leeb,*, Chin-Yuan Hsua
aDepartment of Computer Science and Information Engineering, National Cheng Kung University, Tainan, 701, Taiwan, ROCbDepartment of Information Management, Chang Jung Christian University, Tainan, 711, Taiwan, ROC
Abstract
An intelligent image agent based on soft-computing techniques for color image processing is proposed in this paper. The intelligent image
agent consists of a parallel fuzzy composition mechanism, a fuzzy mean related matrix process and a fuzzy adjustment process to remove
impulse noise from highly corrupted images. The fuzzy mechanism embedded in the filter aims at removing impulse noise without destroying
fine details and textures. A learning method based on the genetic algorithm is adopted to adjust the parameters of the filter from a set of
training data. By the experimental results, the intelligent image agent achieves better performance than the state-of-the-art filters based on the
criteria of Peak-Signal-to-Noise-Ratio (PSNR) and Mean-Absolute-Error (MAE). On the subjective evaluation of those filtered images, the
intelligent image agent also results in a higher quality of global restoration.
q 2005 Elsevier Ltd. All rights reserved.
Keywords: Impulse noise; Image filtering; Fuzzy inference; Genetic algorithm
1. Introduction
Nowadays, the techniques of image processing have been
well developed, but there are still some bottlenecks that are
not solved. For example, many image processing algorithms
cannot work well in a noisy environment, so the image filter
is adopted as a preprocessing module. The process of image
transmission could be corrupted by impulse noise and the
corrupted image is different from the original image. A
number of approaches have been developed for impulse
noise removal. For example, a median filter (Arakawa, 1996)
is the most used method, but it will not work efficiently when
the noise rate is above 0.5. Abreu and Mitra (1995) proposed
an efficient nonlinear algorithm to suppress impulse noise
from highly corrupted images while preserving details and
features. The algorithm is based on detection–estimation
strategy, called Signal-Dependent Rank Ordered Mean (SD-
ROM) filter. SD-ROM filter can achieve an excellent tradeoff
between noise suppression and detail preservation, and
0957-4174/$ - see front matter q 2005 Elsevier Ltd. All rights reserved.
doi:10.1016/j.eswa.2004.12.010
* Corresponding author. Tel.: C886 6278 5123x2059; fax: C886 6278
5657.
E-mail addresses: [email protected] (C.-S. Lee), [email protected].
ncku.edu.tw (C.-S. Lee).
outperform a number of well-known techniques for highly
corrupted images. Weighted Fuzzy Mean (WFM) filter (Lee,
Kuo, & Yu, 1997) has a better ability for removing high
impulse noise. Especially, when the noise rate is above 0.5,
WFM filter still maintains a steady result. Adaptive
Weighted Fuzzy Mean (AWFM) filter (Kuo, Lee, & Chen,
2000) can improve the WFM filter’s incapability in a low
noisy environment, and still retains its capability of
processing in the heavily noisy environment. Russo (1999,
2000) presented the hybrid neuro-fuzzy filters for images,
which are highly corrupted by impulse noise. The network
structure of the filter is specifically designed to detect
different patterns of noisy pixels typically occurring in highly
corrupted data. The proposed filters are able to yield a very
effective noise cancellation and to perform significantly
better than the other approaches. Wang, Liu, and Lin (2002)
presented a histogram-based fuzzy filter (HFF) to the
restoration of noise-corrupted images, which is particularly
effective at removing highly impulsive noise while preser-
ving image details. Lukac (2003) proposed an adaptive
vector median filter for impulse noise suppression and
outliers rejection in multichannel images. Pok, Liu, and Nair
(2003) proposed a decision-based, signal-adaptive median
filtering algorithm for removal of impulse noise. Chang and
Expert Systems with Applications 28 (2005) 483–494
www.elsevier.com/locate/eswa
Fig. 1. The structure of intelligent image agent.
S.-M. Guo et al. / Expert Systems with Applications 28 (2005) 483–494484
Chen (2004) proposed a classifier-augmented median filter
for impulse noise removal from images. Liu (2002) presented
a representation of digital image by fuzzy neural network, by
which the predetermined fuzzy system can be constructed to
express a given two-dimensional (2D) digital image. Tsai
and Yu (1999, 2000) proposed adaptive fuzzy hybrid
multichannel filters for removal of impulsive noise from
color images. Lin and Hsueh (2000) proposed a multichannel
filtering by gradient information. Barni, Buti, Bartolini, and
Cappellini (2000) proposed a quasi-Euclidean norm to speed
up vector median filtering. Vardavoulia, Andreadis, and
Tsalides (2001) proposed a new vector median filter for color
image processing.
Intelligent agents are a new paradigm of modern
Artificial Intelligence (AI) research in computer science.
An agent is a physical or virtual entity, which is capable of
acting in an environment and communicating directly with
other agents (Ferber, 1999). Soft computing differs from
conventional (hard) computing in that, unlike hard comput-
ing, it is tolerant of imprecision, uncertainty and partial
truth. Neural network theory, fuzzy logic, probabilistic
reasoning, genetic algorithms, chaos theory and parts of
learning theory all are in soft computing. Fuzzy inference is
the process of formulating the mapping from a given input
to an output using fuzzy logic. The mapping then provides a
basis from which decisions can be made, or patterns
discerned. The process of fuzzy inference involves member-
ship functions, fuzzy logic operators, and if–then rules.
In this paper, we propose an intelligent image agent to
remove impulse noise from highly corrupted images. The
proposed filter consists of a parallel fuzzy composition
mechanism, a fuzzy mean related matrix process, and a fuzzy
adjustment process. The genetic learning approach proposed
by Cord’on, Herrera, and Villar (2001) and Lee and Pan
(2004) is applied to tune the parameters of the membership
functions. The intelligent image agent performs better than
our previous AWFM operator (Kuo, Lee, & Chen, 2000) and
is able to largely outmatch state-of-the art methods in the
literature. The rest of this paper is organized as follows.
In Section 2, we briefly introduce the knowledge base of the
image agent. Section 3 describes the novel structure of the
intelligent image agent. Section 4 focuses on parameter
encoding and genetic learning. The experimental results for
intelligent image agent are described in Section 5. Finally, we
make the conclusion in Section 6.
Fig. 2. The luminance fuzzy variable with five linguistic terms.
2. Knowledge base construction for intelligent
image agent
An intelligent image agent is a special fuzzy system
having an image knowledge base and a fuzzy inference
mechanism. Fig. 1 shows the structure of the intelligent
image agent.
In this system, the RGB color space is adopted to
represent color images. X(i,j) denotes the color image that
may be corrupted by impulse noise, and Y(i,j) is the output
image after filtering. R($), G($) and B($) are the functions to
produce the projections of X(i,j) in the red axis XR(i,j), green
axis XG(i,j), and blue axis XB(i,j), respectively, i.e. the
functions can be represented as following formulas:
Xði; jÞ Z ðXRði; jÞ;XGði; jÞ;XBði; jÞÞ (1)
RðXði; jÞÞ Z XRði; jÞ (2)
GðXði; jÞÞ Z XGði; jÞ (3)
BðXði; jÞÞ Z XBði; jÞ (4)
After filtering in individual color channel, the function
T($) aggregates the partial results to construct the filtered
color image Y(i,j), that is,
Yði; jÞ Z TðFFðXRði; jÞÞ;FFðXGði; jÞÞ;FFðXBði; jÞÞÞ (5)
In this paper, we propose a new construction algorithm of
image knowledge base (IKB), where the trapezoidal
function is adopted to be the membership function of
fuzzy sets. Eq. (6) denotes the membership function fA(x) of
fuzzy set A.
fAðxÞ Z
0 x!aA
ðx KaAÞ=ðbA KaAÞ aA%x!bA
1 bA%x!cA
ðdA KxÞ=ðdA KcAÞ cA%x!dA
0 xRdA
8>>>>>>><>>>>>>>:
(6)
The trapezoidal membership function of fuzzy set A is
denoted by the parameter set AZ[aA,bA,cA,dA]. Fig. 2
illustrates an example for luminance fuzzy variable with five
linguistic terms. The membership degree is usually a value
S.-M. Guo et al. / Expert Systems with Applications 28 (2005) 483–494 485
in the range [0, 1], where ‘1’ denotes a full membership and
‘0’ denotes no membership.
The image knowledge base consists of the parameters
of the membership functions. In this paper, we define
five fuzzy sets for an image including very dark (VDK),
dark (DK), medium (MD), bright (BR) and very bright
(VBR) shown in Fig. 2. The membership functions of
fuzzy sets VDK, DK, MD, BR and VBR for color
image are denoted as VDKz Z ½azVDK ; b
zVDK ; c
zVDK ; d
zVDK�,
DKzZ ½azDK ; b
zDK ; c
zDK ; d
zDK�, MDzZ ½az
MD; bzMD; c
zMD; d
zMD�,
BRz Z ½azBR; b
zBR; c
zBR; d
zBR� and VBRz Z ½az
VBR; bzVBR; c
zVBR;
dzVBR� where zZ[R, G, B] means the three axes of color
image, respectively. The fuzzy sets describing the
intensity feature of a noise-free image can be derived
from the histogram of the source image. Now we describe
the construction algorithm for the image knowledge base
as follows:
Construction Algorithm for the Knowledge Base of
Intelligent Image Agent:
Input: The histogram of sample image or noise-free
image.
Output: The parameter of the membership functions.
Method:
Step 1:
Decide the overlap range of the fuzzy sets,respectively.
Step 1.1: Set czVDK of z axis be the first sz
k such that
gzk O0, az
DK )czVDK .
Step 1.2: Set bzVBR of z axis be the last sz
k such that
gzk O00, dz
BR )bzVBR.
Step 1.3: Set
rangez )ðbz
VBR KczVDKÞ
2,Nzf K3
$ %
where Nzf is the number of fuzzy sets of z
axis.
Step 1.4: Set azVDK )0, bz
VDK )0.
Step 1.5: Set czVBR )0, dz
VBR )255.
Step 2:
Decide the parameter values of the membershipfunction f zVDK of fuzzy set VDK in z axis:
dzVDK )cz
VDK Crangez.
Step 3:
Decide the parameter values of the membershipfunction f zDK of fuzzy set DK in z axis by the
following sub-steps:
Step 3.1: Set bzDK )dz
VDK .
Step 3.2: Set czDK )rangez Cbz
DK .
Step 3.3: Set dzDK )rangezCcz
DK .
Step 4:
Decide the parameter values of the membershipfunction f zMD of fuzzy set MD in z axis by the
following sub-steps:
Step 4.1: Set azMD )cz
DK .
Step 4.2: Set b2MD )d2
DK .
Step 4.3: Set czMD )rangez Cbz
MD.
Step 4.4: Set dzMD )rangezCcz
MD.
Step 5:
Decide the parameter values of the membershipfunction f zBR of fuzzy set BR in z axis by the
following sub-steps:
Step 5.1: Set azBR )cz
MD.
Step 5.2: Set bzBR )dz
MD.
Step 5.3: Set czBR )bz
BR Crangez.
Step 6:
Decide the parameter values of the membershipfunction f zVBR of fuzzy set VBR in z axis:
azVBR)cz
BR.
Step 7:
Stop.Then we can apply the construction algorithm to perform
the red channel, green channel and blue channel in color
image, respectively.
3. The structure of the intelligent image agent
In this section, we describe the structure of the intelligent
image agent. The proposed agent operates on a 3$3
neighborhood in order to restore image data highly
corrupted by impulse noise. Fig. 3 shows the architecture
of the intelligent image agent for the impulse noise removal.
The intelligent image agent consists of a parallel fuzzy
inference mechanism, a fuzzy mean related matrix process
and a fuzzy adjustment process. Now, we describe them as
follows.
3.1. Parallel fuzzy inference mechanism
The architecture of parallel fuzzy inference mechanism is
shown in Fig. 3. In Fig. 3, the structure consists of five
layers. Now, we will describe each layer in details.
Layer 1 (Input linguistic layer): The nodes in the first
layer just directly transmit input values to the next layer. If
the input vector is ðxz1; x
z2;.; xz
9Þ, where xzi is denoted as
input value of the ith pixel. Then, the output for this layer
will be
m1;zij Z xz
ij; i Z 1;.; 9; j Z 1;.; 5; (7)
where xzij is input value of the jth linguistic term for the ith
pixel from z axis.
Layer 2 (Input term layer): Each fuzzy variable of the
second layer appearing in the premise part is represented with
a condition node. This layer performs the first inference step
to compute matching degrees. If the input vector of this layer
is ððxz11; x
z12;.; xz
15Þ; ðxz21; x
z22;.; xz
25Þ;.; ðxz91; x
z92;.; xz
95ÞÞ,
then the output vector will be
m2;zij Z f z
AjðxzijÞ; i Z 1;.; 9; j Z 1;.; 5; (8)
where f zAjðx
zijÞ is the membership degree of the jth term for the
ith pixel.
Layer 3 (Rule layer): The third layer is called the rule
layer, where each node is a rule to represent a fuzzy rule.
The links in this layer are used to perform precondition
Fig. 3. The Structure of the intelligent image agent.
S.-M. Guo et al. / Expert Systems with Applications 28 (2005) 483–494486
matching of fuzzy logical rules. If the input vector of this
layer is ððf zA1ðx
z11Þ; f
zA2ðx
z12Þ;.; f z
A5ðxz15ÞÞ; ðf
zA1ðx
z21Þ; f
zA2ðx
z22Þ;.
; f zA5ðx
z25ÞÞ;.; ðf z
A1ðxz91Þ; f
zA2ðx
z92Þ;.; f z
A5ðxz95ÞÞÞ then the output
vector will be
m3;zij ZMINff z
AjðxzijÞ; f
zAjðy
zmeanÞg; i Z1;.;9; j Z1;.;5 (9)
Layer 4 (Subrulebase layer): Let m4;zi be the output of the
ith node (iZ1,.,9). The node function is defined by
MAXjZ1;.;5
fm3;zij g; i Z1;.;9 (10)
Layer 5 (Output Linguistic Layer): The final output v is
evaluated by means of the following relation:
v Zm5;z ZX9
iZ1
ðm4;zi !xz
i Þ (11)
3.2. Fuzzy mean related matrix process
The fuzzy mean related matrix process performs the
fuzzy mean of input variables. Eq. (12) denotes the
computing process with fuzzy interval F_mean from z
axis for fuzzy mean related matrix process.
yzmean Z
P9iZ1 f z
F_meanðxzi Þ,xz
iP9iZ1 f z
F_meanðxzi Þ
; ifP9
iZ1 f zF_meanðx
zi Þ
O0
0; otherwise
8<:
(12)
where f zF_mean Z ½0;az;bz; 255�. Then, we set the fuzzy
mean related matrix f zAjðy
zmeanÞ, jZ1,.,5. The fuzzy mean
related matrix is used to evaluate input variables and
perform weighted input variables.
3.3. Fuzzy adjustment process
There are four computation functions including f zKð,Þ,
f z,1ð,Þ, f z
,2ð,Þ, f z
sumð,Þ and two membership functions
including f zsmall and f z
large utilized in fuzzy adjustment process.
Now we briefly describe them as follows:
f zKðm5;z; xz
5Þ Z jm5;z Kxz5j (13)
f z,1ðv; f z
largeðfz
Kð,ÞÞÞ Z m5;z !f zlargeðf
zKð,ÞÞ (14)
f z,2ðxz
5; fzsmallðf
zKð,ÞÞÞ Z xz
5 !f zsmallðf
zKð,ÞÞ (15)
Fig. 4. The architecture of genetic learning process of the intelligent image agent.
S.-M. Guo et al. / Expert Systems with Applications 28 (2005) 483–494 487
f zsumðf
z,1ð,Þ; f z
,2ð,ÞÞ Z f z
,1ð,ÞC f z
,2ð,Þ (16)
The fuzzy sets of small and large denote smallZ[0, 0,
sz, lz] and largeZ[sz, lz, 255, 255] for the fuzzy adjustment
process. The parameters s and l for fuzzy sets small and
large are defined as follows:
sz Z lz,f zF_meanðx
z5Þ (17)
The final output y of fuzzy decision process is the
computing result of f zsumð,Þ. The membership functions f z
large
and f zsmall define the detail preserving process of the filter. It
basically executes full correction of large amplitude noise
pulses, partial correction of median amplitude noise
pulses, and no correction of small amplitude noise pulses.
In fact, the quantity f zKð,Þ can be interpreted as measure
of the modification process by previous layers. If this
measure is large, a full correction is allowed. If this measure
is small, on the contrary, the correction is further reduced in
order to better preserve the quality of fine details and
textures.
Fig. 5. (a) Encoding of fuzzy sets and parameters. (b) Encoding of the
linguistic modifiers of the linguistic terms.
4. Parameter encoding and genetic learning
This section introduces the genetic learning for intelli-
gent image agent. As previously mentioned, we adopt a
supervised learning method based on the genetic learning
for the fuzzy filtering system shown in Fig. 4.
The important questions when using the genetic learning
are how to encode each solution, how to evaluate these
solutions and how to create new solutions from existing ones.
In order to adopt a genetic learning method we encode the set
of fuzzy sets and the linguistic modifiers of the linguistic
terms. Here, we apply the learning approach proposed by
Cord’on et al. (2001) to learning image knowledge base
containing image DB and image RB for the next behavior
learning.
The three components of image knowledge base to be
encoded are the membership functions of the fuzzy
variables and the linguistic modifiers of the linguistic
terms. This chromosome is composed of the following sub-
parts CSza and CSz
b shown in Fig. 5.
1.
CSza is that a 23-gene which encodes the fuzzy setparameters.
2.
CSzb is that a 5-gene which encodes the linguisticmodifiers of the linguistic terms.
Next, a linguistic modifier used in IKB is a function with
the parameter d that lets us alter the membership functions.
Two of the most well known modifiers are the erosion
linguistic modifier ‘very’ (dZ2) and the dilation linguistic
modifier ‘more-or-less’ (dZ0.5) (Lee & Pan, 2004).
Eqs. (18) and (19) denote the functions of the two modifiers
used in this paper:
mveryðxÞ Z ðmðxÞÞ2 (18)
mmore�or�lessðxÞ Z ðmðxÞÞ0:5 (19)
Fig. 6. The experimental website for intelligent image agent.
Fig. 7. (a) Original ‘Lena’ image, (b) noise image corrupted by impulse noise (
Fig. 8. (a) The fuzzy sets of ‘Lena’ image constructed by t
S.-M. Guo et al. / Expert Systems with Applications 28 (2005) 483–494488
The factors of the luminance considered here are
the fuzzy variables VDK, DK, MD, BR, and
VBR, represented as ½azVDK ; b
zVDK ; c
zVDK ; d
zVDK�, ½az
DK ; bzDK ;
czDK ; d
zDK�, ½az
MD; bzMD; c
zMD; d
zMD�, ½az
BR; bzBR; c
zBR; d
zBR� and
½azVBR; b
zVBR; c
zVBR; d
zVBR�. The simple genetic algorithm
(Lee & Pan, 2004) operates as follows. The method
starts with a randomly generated population of individ-
uals and produces the subsequent populations by means
of reproduction, crossover, and mutation operators. The
individuals having the best fitness have more chances to
be reproduced. The object function F, which
measures the fitness of each individual, is based on the
mean-absolute error (MAE) between the processed and
prob. 0.4), and (c) result yielded by genetic learning after 50 generations.
he construction algorithm. (b) The tuned fuzzy sets.
Table 1
The parameters of fuzzy sets for ‘Lena’ image constructed by the intelligent image agent
Axis Terms Before tuning After tuning
[a, b, c, d] d l a b [a, b, c, d] d l a b
Red VDK [0, 0, 53, 81] 1 72 28 224 [0, 13, 28, 46] 0.5 40 3 234
DK [53, 81, 109, 137] 1 [28, 62, 84, 112] 0.5
MD [109, 137, 165, 193] 1 [84, 112, 112, 168] 0.5
BR [165, 193, 221, 249] 1 [133, 168, 168, 218] 0.5
VBR [221, 249, 255, 255] 1 [171, 224, 255, 255] 0.5
Blue VDK [0, 0, 1, 35] 1 72 28 224 [0, 1, 3, 69] 0.5 72 7 236
DK [1, 35, 69, 103] 1 [35, 79, 81, 112] 0.5
MD [69, 103, 137, 171] 1 [83, 116, 140, 168] 1
BR [137, 171, 205, 239] 1 [140, 168, 196, 224] 0.5
VBR [205, 239, 255, 255] 1 [219, 224, 255, 255] 1
Green VDK [0, 0, 8, 39] 1 72 28 224 [1, 1, 2, 66] 2 7 72 193
DK [8, 39, 70, 101] 1 [46, 67, 75, 118] 1
MD [70, 101, 132, 163] 1 [84, 121, 136, 182] 1
BR [132, 163, 194, 225] 1 [140, 183, 188, 221] 0.5
VBR [184, 225, 255, 255] 1 [198, 240, 249, 254] 1
Fig. 9. Values of fitness obtained during the learning process and effects of different choices of genetic parameters for ‘Lena’ image.
S.-M. Guo et al. / Expert Systems with Applications 28 (2005) 483–494 489
Fig. 10. MAE curves of the proposed method and others on the color images corrupted by salt-and-pepper impulse noises with the noise corruption rate p,
where pZ0.1–0.8.
Fig. 11. PSNR curves of the proposed method and others on the color images corrupted by salt-and-pepper impulse noises with the noise corruption rate p,
where pZ0.1–0.8.
S.-M. Guo et al. / Expert Systems with Applications 28 (2005) 483–494490
the original noise-free image:
F Z
P256iZ1
P256jZ1 jy
Rði; jÞKsRði; jÞj
256!256
C
P256iZ1
P256jZ1 jy
Gði; jÞKsGði; jÞj
256!256
C
P256iZ1
P256jZ1 jy
Bði; jÞKsBði; jÞj
256!256
! 3
Table 2
PSNR values of the compared approaches for salt-and-pepper impulse noisy “Le
Filters pZ0.1 pZ0.2 pZ0.3 pZ0.4
Russo 48.09 42.81 38.90 34.04
AWFM 29.86 28.68 27.37 26.13
Median 30.22 29.89 29.37 28.08
Lin 30.73 26.42 23.46 20.62
Proposed 39.37 37.82 36.35 34.62
The learning process stops when an assigned number
of generations have been evolved or when a satisfactory
value of fitness has been obtained.
5. Experimental results
There are many different methods for removing impulse
noise from corrupted images. In this paper, we compare our
approach with other famous filters including Russo’s filter,
na” image with the corruption rate p, where pZ0.1–0.8
pZ0.5 pZ0.6 pZ0.7 pZ0.8
29.94 25.33 21.23 18.15
25.04 23.59 21.36 19.50
25.79 22.69 18.74 15.82
18.04 16.00 14.07 12.83
33.25 31.90 30.17 27.46
Table 3
PSNR values of the compared approaches for salt-and-pepper impulse noisy ‘House’ image with the corruption rate p, where pZ0.1–0.8
Filters pZ0.1 pZ0.2 pZ0.3 pZ0.4 pZ0.5 pZ0.6 pZ0.7 pZ0.8
Russo 51.96 44.09 40.37 36.54 31.49 26.92 22.84 19.72
AWFM 33.79 31.44 29.35 27.67 26.14 24.13 21.44 19.45
Median 33.30 32.65 31.42 29.61 26.76 23.00 19.04 16.05
Lin 31.27 27.29 24.00 21.07 18.18 15.59 13.83 13.02
Proposed 50.28 45.11 42.48 39.75 37.21 35.62 33.22 30.25
S.-M. Guo et al. / Expert Systems with Applications 28 (2005) 483–494 491
AWFM, Median and Lin filter to test the performance of the
intelligent image agent. In the noise model for experiments,
the noise-free image is corrupted by additive identical
independent distribution (i.i.d.) impulse noise with the
corruption rate p, and the impulses take on positive and
negative values with an equal p/2, i.e. the x is a Bernoulli
random variable (Kuo, Lee, & Chen, 2000), as follows:
xði; jÞ Z
sði; jÞCnði; jÞ with corruption rate p=2
sði; jÞKnði; jÞ with corruption rate p=2
sði; jÞ with probability ð1 KpÞ
8><>: (21)
where s(i,j) is the gray level of the noise-free pixel on
location (i,j), n(i,j) is the noise amplitude corrupted on
location (i,j), and x(i,j) is the gray level of the noisy pixel for
s(i,j). In the beginning, we analyze the properties of the
intelligent image agent, then verify the noise removal
capability of the intelligent image agent by comparing with
the other filters. To decide the parameter set of the
intelligent image agent for the experiment, we adopt the
well-known 256!256 ‘Lena’ color image to be the sample
image to construct image knowledge base. In addition, we
also produce a salt-and-pepper noisy ‘Lena’ color image
with a corruption rate 0.4 for the intelligent image agent. We
have chosen a small population of 20 individuals and set the
parameters of genetic learning as follows: crossover
probability 1.0, mutation rate 0.005 and 50 generations.
We have implemented an experimental website to test
Fig. 12. Results of color image ‘Lena’ with pZ0.8 impulse noise.
Table 4
Runtime (in s) consumed at various noise densities p using the intelligent
image agent and other filters based on ‘Lena’ image
Filters pZ0.1 pZ0.3 pZ0.5 pZ0.7
Russo (3!3) 11.87 11.80 11.75 11.90
AWFM (3!3) 1.67 1.63 1.54 1.60
Median (3!3) 1.32 1.36 1.30 1.33
Lin (3!3) 1.26 1.23 1.24 1.25
Proposed (3!3)
(filtering time) (s)
6.62 6.33 6.40 6.64
Proposed (3!3)
(tuning time) (min)
149.2 152.7 150.6 150.3
Fig. 13. Results of color image ‘House’ with pZ0.8 impulse noise.
Fig. 14. Results of color image ‘Lena’ with pZ0.8 impulse noise.
S.-M. Guo et al. / Expert Systems with Applications 28 (2005) 483–494492
the performance of the proposed approach. Fig. 6 shows the
experimental website.
Fig. 7(a)–(c) shows the noise-free ‘Lena’ image, noise
‘Lena’ image with probability 0.4 and result image by the
intelligent image agent, respectively.
Fig. 8(a) illustrates the fuzzy sets of ‘Lena’ color image
constructed by the construction algorithm. The tuned fuzzy
sets are shown in Fig. 8(b).
Table 1 shows the parameters of fuzzy sets for ‘Lena’
color image constructed by the intelligent image agent.
In order to analyze the behavior of the intelligent image
agent, we choose the well-known ’Lena’ color image to test
the convergence for the intelligent image agent. In addition,
we also choose a small population of 20 individuals and 50
generations and set the crossover probability 0.9, 0.6 or 1.0,
the mutation rate 0.05, 0.1, or 0.005. Fig. 9 shows the fitness
curves of the intelligent image agent with various
parameters for ‘Lena’ color image.
By this experimental result, we can see that genetic
learning is robust for various parameters and images. Next,
we analyze the filtering capability of the intelligent image
agent. We compare the noise removal capability of Russo,
AWFM, Median, Lin and the intelligent image agent in
Fig. 15. Results of color image ‘House’ with pZ0.8 impulse noise.
S.-M. Guo et al. / Expert Systems with Applications 28 (2005) 483–494 493
the following experiments. The parameters of Russo’s
method are setting as follows: M1Z4, M2Z2, crossover
probability 1.0, mutation rate 0.005, population size with 40
individuals and 50 generations. Figs. 10 and 11 show the
MAE and PSNR curves of all compared approaches for
‘Lena’, and ’House‘ images, respectively.
The extrapolated PSNR value of ‘Lena’ image and
‘House’ color image resulted from using various filters at
different noise densities, ranging from 0.1 to 0.8, are shown
in Tables 2 and 3, respectively.
The runtime analysis of the intelligent image agent and
other concerned filters were conducted for ‘Lena’ image
using Pentium IV 2.4 GHz Personal Computer and docu-
mented in Table 4.
Figs. 12 and 13 show the salt-and-pepper noisy ‘Lena’
and ‘House’ images with a corruption rate 0.8, the results of
the intelligent image agent and other filter, respectively.
Figs. 14 and 15 show the subjective evaluation on edge
detection results of Figs. 12 and 13, respectively. By the
results, we can see that the intelligent image agent can
preserve the fine details and textures better than the other
approaches.
6. Conclusions
In this paper, we present an intelligent image agent
including a parallel fuzzy composition mechanism, a fuzzy
mean related matrix process and a fuzzy adjustment process
to remove impulse noise from highly corrupted color
images. The intelligent image agent will receive sample
images or the noise-free color image, then construct image
knowledge base for the filter. It will also adjust the
parameters of fuzzy sets for getting the optimal image
knowledge base. From the experimental results, we observe
that the PSNR and MAE curves of the intelligent image
agent achieve the most efficient results than other
approaches including Russo’s method, AWFM, median
and Lin for removing heavily corrupted additive impulse
noise. Subjective evaluation of the intelligent image agent
also shows a higher quality of global restoration.
Acknowledgements
This work was partially supported by the National
Science Council of TAIWAN (ROC), under Grant NSC
90-2213-E-309-007 and NSC 92-2213-E-309-005.
References
Abreu, E., & Mitra, S. K. (1995). A signal-dependent rank ordered mean
(SD-ROM) filter. Proceedings of IEEE International Conference
on Acoustics, Speech and Signal Processing, ICASSP-95, Detroit ,
2371–2374.
Arakawa, K. (1996). Median filter based on fuzzy rules and its application
to image restoration. Fuzzy Sets and Systems, 77, 3–13.
Barni, M., Buti, F., Bartolini, F., & Cappellini, V. (2000). A quasi-
Euclidean norm to speed up vector median filtering. IEEE Transactions
on Image Processing, 9, 1704–1709.
Chang, J. Y., & Chen, J. L. (2004). Classified-augmented median filters for
image restoration. IEEE Transactions on Instrumentation and
Measurement, 53, 351–356.
Cord’on, O., Herrera, F., & Villar, P. (2001). Generating the knowledge
base of a fuzzy rule-based system by the genetic learning of the data
base. IEEE Transactions on Fuzzy System, 9, 667–674.
Ferber, J. (1999). Multi-agent systems. New York: Addison-Wesly.
Kuo, Y. H., Lee, C. S., & Chen, C. L. (2000). High-stability AWFM filter
for signal restoration and its hardware design. Fuzzy Sets and Systems,
114, 185–202.
S.-M. Guo et al. / Expert Systems with Applications 28 (2005) 483–494494
Lee, C. S., Kuo, Y. H., & Yu, P. T. (1997). Weighted fuzzy mean filters for
image processing. Fuzzy Sets and Systems, 89, 157–180.
Lee, C. S., & Pan, C. Y. (2004). An intelligent fuzzy agent for
meeting scheduling decision support system. Fuzzy Sets and Systems,
142, 467–488.
Lin, R. S., & Hsueh, Y. C. (2000). Multichannel filtering by gradient
information. Signal Processing, 80, 279–293.
Liu, P. (2002). Representation of digital image by fuzzy neural network.
Fuzzy Sets and Systems, 130, 109–123.
Lukac, R. (2003). Adaptive vector median filtering. Pattern Recognition
Letters, 24, 1889–1899.
Pok, G., Liu, J. C., & Nair, A. S. (2003). Selective removal of impulse noise
based on homogeneity level information. IEEE Transactions on Image
Processing, 12, 85–92.
Russo, F. (1999). Hybrid neuro-fuzzy filter for impulse noise removal.
Pattern Recognition, 32, 1843–1855.
Russo, F. (2000). Noise removal from image data using recursive
neurofuzzy filters. IEEE Transactions on Instrumentation and
Measurement, 49, 307–314.
Tsai, H. H., & Yu, P. T. (1999). Adaptive fuzzy hybrid multichannel filters
for removal of impulsive noise from color images. Signal Processing,
74, 127–151.
Tsai, H. H., & Yu, P. T. (2000). Genetic-based fuzzy hybrid
multichannel filters for color image restoration. Fuzzy Sets and
Systems, 114, 203–224.
Vardavoulia, M. I., Andreadis, I., & Tsalides, Ph. (2001). A new vector
median filter for colour image processing. Pattern Recognition Letters,
22, 675–689.
Wang, J. H., Liu, W. J., & Lin, L. D. (2002). Histogram-based fuzzy filter
for image restoration. IEEE Transactions on Systems, Man and
Cybernetics, Part B, 32, 230–238.