694

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART B: CYBERNETICS, VOL. 35, NO. 4, AUGUST 2005

Genetic-Based Fuzzy Image Filter and Its Application

to Image Processing
Chang-Shing Lee, Shu-Mei Guo, and Chin-Yuan Hsu

AbstractIn this paper, we propose a Genetic-based Fuzzy

Image Filter (GFIF) to remove additive identical independent
distribution (i.i.d.) impulse noise from highly corrupted images.
The proposed filter consists of a fuzzy number construction process,
a fuzzy filtering process, a genetic learning process, and an image
knowledge base. First, the fuzzy number construction process receives sample images or the noise-free image and then constructs
an image knowledge base for the fuzzy filtering process. Second,
the fuzzy filtering process contains a parallel fuzzy inference
mechanism, a fuzzy mean process, and a fuzzy decision process to
perform the task of noise removal. Finally, based on the genetic
algorithm, the genetic learning process adjusts the parameters
of the image knowledge base. By the experimental results, GFIF
achieves a better performance than the state-of-the-art filters
based on the criteria of Peak-Signal-to-Noise-Ratio (PSNR),
Mean-Square-Error (MSE), and Mean-Absolute-Error (MAE).
On the subjective evaluation of those filtered images, GFIF also
results in a higher quality of global restoration.

Fig. 1. Structure of GFIF.

Index TermsFuzzy inference, fuzzy number, genetic algorithm, image processing, impulse noise.

I. INTRODUCTION

OWADAYS, the techniques of image processing have

been well developed, but there are still some bottlenecks
that have not been solved. For example, many image processing
algorithms cannot work well in a noisy environment; therefore,
the image filter is adopted as a preprocessing module. The
process of image transmission could be corrupted by impulse
noise, which causes the corrupted image to be different from
the original one. A number of approaches have been developed
for impulse noise removal. For example, Tukey [1], Astola et al.
[2], and Pitas et al. [3] introduce the median filter to eliminate
impulse noise. The median filter [4] can achieve reasonably
good performance for low corrupted images, but it will not
work efficiently when the noise rate is above 0.5. In addition,
the median filter has been intensively studied and extended
to promising approaches such as Weighted Median (WM) [5]
and Center Weighted Median (CWM) [6] filters. The WM
filter uses a set of weighting parameters to control the filtering
performance in order to preserve more signal details than that
which the median filter can accomplish. The CWM filter is a
Manuscript received May 15, 2004; revised August 17, 2004 and November
26, 2004. This work was supported in part by the National Science Council of
Taiwan, R.O.C., under Grants NSC 90-2213-E-309-007 and NSC 92-2213-E309-005. This paper was recommended by Associate Editor Sudeep Sarkar.
C.-S. Lee is with the Department of Information Management, Chang Jung
Christian University, Tainan 711, Taiwan, R.O.C.
S.-M. Guo and C.-Y. Hsu are with the Department of Computer Science and
Information Engineering, National Cheng Kung University, Tainan 701, Taiwan,
R.O.C. (e-mail: leecs@mail.cju.edu.tw; leecs@cad.csie.ncku.edu.tw).
Digital Object Identifier 10.1109/TSMCB.2005.845397

Fig. 2.

Luminance fuzzy variable with five linguistic terms.

Fig. 3. An 8-neighborhood.

special case of the WM filter, where only the center pixel of the
filtering window has a weighted factor [7]. Eng et al. [7] present
a novel switching-based median filter with incorporation of
the fuzzy-set concept called the Noise Adaptive Soft-switching
Median (NASM) filter to achieve a much-improved filtering

1083-4419/$20.00 2005 IEEE

LEE et al.: GENETIC-BASED FUZZY IMAGE FILTER AND ITS APPLICATION TO IMAGE PROCESSING

695

Fig. 4. Architecture of the fuzzy filtering process.

performance in terms of effectiveness in removing impulse

noise while preserving signal details and robustness in combating noise density variations. The proposed NASM filter
consists of two stages, including a soft-switching noise-detection scheme and three WM-based filters. Experimental
results show that the NASM filter impressively outperforms
other techniques. Furthermore, there are still many other filters
proposed for removing impulse noise. For instance, Fung et
al. [8] propose an improved algorithm for removing impulse
noise based on long-range correlation in an image. Harja et al.
[9] propose a recursive weighted median filter with negative
weights. Khriji et al. [10] propose a Median-Rational Hybrid
Filter (MRHF) that is based on the CWM filter for removing
additive impulse noise. Abreu et al. [11] propose an efficient
nonlinear algorithm to suppress impulse noise from highly
corrupted images while preserving details and features. The algorithm is based on the detection-estimation strategy called the
Signal-Dependent Rank Ordered Mean (SD-ROM) filter. The
SD-ROM filter can achieve an excellent tradeoff between noise
suppression and detail preservation and outperform a number
of well-known techniques for highly corrupted images. The
Weighted Fuzzy Mean (WFM) filter [12] has a better ability to
remove high impulse noise, especially when the noise rate is
above 0.5. The Adaptive Weighted Fuzzy Mean (AWFM) filter
[13] not only improves the the WFM filters incapability in a
low noisy environment but also retains its capability of processing in the heavily noisy environment. An intelligent Fuzzy
Image Filter (FIF) [14] proposed by Lee et al. can automatically

Fig. 5. Fuzzy interval F

Fig. 6.

mean for fuzzy mean process.

Fuzzy sets for luminance difference fuzzy variable.

decide the number of fuzzy numbers based on image features to

overcome the drawbacks of the AWFM filter. Russo [15], [16]
presents the hybrid neuro-fuzzy filters for images, which are
highly corrupted by impulse noise. The network structure of

696

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART B: CYBERNETICS, VOL. 35, NO. 4, AUGUST 2005

Fig. 7. (a) Encoding of the parameters of fuzzy sets. (b) Encoding of fuzzy rules. (c) Encoding of the linguistic modifiers of the linguistic terms.

Fig. 8.

Position of each linguistic term of luminance.

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 get a significant performance that is better than the
other approaches. Wang et al. [17] present a Histogram-based
Fuzzy Filter (HFF) to the restoration of noise-corrupted images,
which is particularly effective at removing highly impulsive
noise while preserving image details. Lukac [18] proposes an
adaptive vector median filter for impulse noise suppression and
outlier rejection in multichannel images. Pok et al. [19] propose
a decision-based, signal-adaptive median filtering algorithm
for the removal of impulse noise. Chang et al. [20] propose a
classifier-augmented median filter for impulse noise removal
from images. Liu [21] proposes a fuzzy neural network filter
for impulse noise removal from images. Lin et al. [22] propose
a multichannel filtering by gradient information for impulse
noise removal from images.
In this paper, we propose a genetic-based fuzzy image filter
(GFIF) to remove additive identical independent distribution
(i.i.d.) impulse noise from highly corrupted images. The proposed filter consists of a fuzzy number construction process, a
fuzzy filtering process, a genetic learning process, and an image
knowledge base. First, the fuzzy number construction process
receives sample images or the noise-free image and then constructs the image knowledge base for the fuzzy filtering process.
Second, the fuzzy filtering process contains a parallel fuzzy inference mechanism, a fuzzy mean process, and a fuzzy decision
process to perform the task of noise removal. Finally, the genetic learning process adjusts the parameters of an image knowledge base based on the genetic algorithm. We will apply the genetic learning approach proposed by Cordon et al. [23] to tune
the parameters of the membership functions for removing ad-

ditive impulse noise from highly corrupted images. The rest of

this paper is organized as follows. In Section II, we propose the
structure of genetic-based fuzzy image filter. The algorithm for
the fuzzy number construction process is presented in this section. Section III describes the parallel fuzzy inference mechanism for the fuzzy filtering process. Section IV focuses on the
genetic learning process for GFIF. The experimental results for
GFIF are described in Section V. Finally, Section VI contains
conclusions.
II. STRUCTURE OF GENETIC-BASED FUZZY IMAGE FILTER
In this paper, we propose a GFIF to remove impulse noise
from highly corrupted images. Fig. 1 shows the structure of
GFIF.
There are three main processes, including a fuzzy number
construction process, a fuzzy filtering process, and a genetic
learning process in Fig. 1. In addition, the image knowledge
base should be constructed before performing the fuzzy filtering
process. In this paper, we use the trapezoidal function to be
membership functions of fuzzy sets. Equation (1) denotes the
of fuzzy set
membership function

(1)
.
The trapezoidal membership function of fuzzy set is de. Fig. 2 illusnoted by the parameter set
trates an example for luminance fuzzy variable with five linguistic terms. The membership degree is usually a value in the

LEE et al.: GENETIC-BASED FUZZY IMAGE FILTER AND ITS APPLICATION TO IMAGE PROCESSING

Fig. 9.

Fig. 10.

Graphic representation for linguistic terms normal

697

DK , very DK , and more-or-less DK .

Structure of genetic learning process.

range
, 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
, dark
,
sets for an image, including very dark
, bright
, and very bright
, shown
medium
,
,
in Fig. 2. The membership functions of fuzzy sets
,
and
are denoted as
,
,
,
, and
, respectively. The fuzzy
sets describing the intensity feature of a noise-free image can
be derived from the histogram of the source image. Then,
the algorithm for the fuzzy number construction process is as
follows:
Algorithm for Fuzzy Number Construction
Process:
Input:
The histogram of the sample image or
noise-free image.
Output:
The parameter set of membership functions.
Method:
Step 1: Decide the overlap ranges of the
be the
fuzzy sets, respectively. Let
th gray level of image
and
be the

number
in
Step
that
Step
that
Step

of pixels with the

1.1: Set
,
1.2: Set
,
1.3: Set

th gray level

be the first
.
be the last
.

such
such

, where
denotes the value of the Interval_Range
and
,
is the number
between
.
of fuzzy sets and
,
.
Step 1.4: Set
Step 1.5: Set
,
.
Step 2: Decide the parameter values of
of fuzzy set
the membership function
:
.
Step 3: Decide the parameter values of
of fuzzy set
the membership function
by the following sub-steps:
Step 3.1: Set
.
.
Step 3.2: Set
.
Step 3.3: Set
Step 4: Decide the parameter values of
of fuzzy set
the membership function
by the following sub-steps:
.
Step 4.1: Set
.
Step 4.2: Set

698

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART B: CYBERNETICS, VOL. 35, NO. 4, AUGUST 2005

Fig. 11. (a). Original Lena image. (b) Noisy image corrupted by salt-and-pepper impulse noise (corruption rate 0.4).(c) Result yielded by genetic learning after
50 generations.
TABLE I
PARAMETERS OF FUZZY SETS CONSTRUCTED BY GFIF FOR SALT-AND-PEPPER NOISY LENA IMAGE WITH A CORRUPTION RATE 0.4

Step 4.3: Set

.
Step 4.4: Set
.
Step 5: Decide the parameter values of
of fuzzy set
the membership function
by the following sub-steps:
Step 5.1: Set
.
.
Step 5.2: Set
.
Step 5.3: Set
Step 6: Decide the parameter values of
of fuzzy set
the membership function
:
.
Step 7: Stop.
III. PARALLEL FUZZY INFERENCE MECHANISM FOR FUZZY
FILTERING PROCESS
In this section, we will describe the parallel fuzzy inference
mechanism for the fuzzy filtering process. We adopt the 8-neighborhood of an image for the filter. Fig. 3 illustrates the neigh-

borhood in an image. Fig. 4 shows the architecture of the fuzzy

filtering process for impulse noise removal.
The fuzzy filtering process consists of a parallel fuzzy inference mechanism, a fuzzy mean process, and a fuzzy decision
process. Now, we describe them as follows.
A. Parallel Fuzzy Inference Mechanism
The structure of parallel fuzzy inference mechanism consists
of three layers. There are three kinds of nodes in this model, including the fuzzy linguistic node for the Fuzzy Linguistic Layer,
the fuzzy term node for the Fuzzy Term Layer, and the rule node
for the Fuzzy Rule Layer. A fuzzy linguistic node represents a
fuzzy variable. A fuzzy term node represents the mapping degree of the fuzzy variable. A rule node represents a rule and decides the final firing strength of that rule during inferring. Now,
we briefly describe each layer as follows.
Layer 1 (Fuzzy Linguistic Layer): The fuzzy linguistic nodes
in this layer directly transmit input values to the next layer. If
, then
is denoted as
the input vector is

LEE et al.: GENETIC-BASED FUZZY IMAGE FILTER AND ITS APPLICATION TO IMAGE PROCESSING

699

TABLE II
PARAMETERS OF FUZZY SETS CONSTRUCTED BY GFIF FOR ADDITIVE MIDDLE-TAILED IMPULSE NOISY LENA IMAGE WITH A CORRUPTION RATE 0.4

the input value of the th pixel in the neighborhood. Then, the

output for this layer will be

(2)
where
is the output value of the th fuzzy linguistic term for
the th pixel.
Layer 2 (Fuzzy Term Layer): Each fuzzy variable of the
second layer appearing in the premise part is represented
with a condition node. Each of the outputs of the condition node is connected to rule nodes in the third layer to
constitute a condition specified in some rules. This layer
performs the first inference step to compute matching degrees. If the input vector of this layer is
, then the
output vector will be

(a)

(b)

(3)
where
is the membership degree of the th fuzzy term
for the th pixel.
Layer 3 (Fuzzy 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 matching
of fuzzy logical rules. Each linguistic term of every fuzzy rule
, where
, and
contains a binary weights
. The
will be introduced in the next section. The
fuzzy inference rules are denoted as follows:

Fig. 12. (a) Fuzzy sets of the Lena image constructed by the fuzzy number
construction process. (b). Tuned fuzzy sets by the fuzzy filtering process and the
genetic learning process.

Rule VDK: IF ( is
with weight
is
with weight
) AND AND (
) THEN
is
weight
..
.
with weight
Rule VBR: IF ( is
with weight
) AND AND ( is
is
THEN

is

) AND (
with

) AND ( is
with weight

700

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART B: CYBERNETICS, VOL. 35, NO. 4, AUGUST 2005

TABLE III
PARAMETERS OF FUZZY SETS CONSTRUCTED BY GFIF FOR ADDITIVE LONG-TAILED IMPULSE NOISY LENA IMAGE WITH A CORRUPTION RATE 0.4

Fig. 13.

Type 1 images. (a) Albert. (b) Baboon. (c) Cameraman.

Equation (4) shows the computing mechanism for the output

of Fuzzy Rule Layer, where
, and we have
if
otherwise.
(4)

C. Fuzzy Decision Process

,
There are five computation functions including
,
,
, and
and two membership functions
and
utilized in the fuzzy decision process.
including
Now, we briefly describe them as (6)(10), shown at the bottom
of the next page.
and large
Fig. 6 shows fuzzy sets small
for the fuzzy decision process. The parameters
and for fuzzy sets small and large are defined as follows:

B. Fuzzy Mean Process

The fuzzy mean process performs the fuzzy mean of input
variables. Equation (5) denotes the computing process with the
fuzzy interval [24] F_mean for the fuzzy mean process
if
otherwise
where

is shown in Fig. 5.

(5)

(11)
The final output of the fuzzy decision process is the com. The membership functions
and
puting result of
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 function
can be interpreted as a measure function of the correction
process. If this measure is large, a full correction is allowed. If

LEE et al.: GENETIC-BASED FUZZY IMAGE FILTER AND ITS APPLICATION TO IMAGE PROCESSING

Fig. 14.

701

Type 2 images. (a) Sailboat. (b) Bridge. (c) Boats. (d) House. (e) Pentagon. (f) Airplane.

this measure is small, on the contrary, the correction is further

reduced in order to better preserve the quality of fine details and
textures.

IV. GENETIC LEARNING PROCESS FOR GFIF

the linguistic terms. Fig. 7 shows the chromosome that is com,

, and
.
posed of the following subparts:
1)
is a 23-gene that encodes the parameters of fuzzy sets.
is a group of nine-genes that directly encode
2)
, where
, and
the binary weights
. For example, the set of binary weight

This section will introduce the genetic learning for the fuzzy
image filter. We adopt a supervised learning method based on
the genetic learning for the GFIF system. The important questions when using genetic learning are how to encode each solution, how to evaluate theses solutions, and how to create new
solutions from existing ones. We apply the learning approach
proposed by Cordon et al. [23] to learn the image knowledge
base containing image DB stored in the parameters of fuzzy sets
and the image RB stored in the parameters of fuzzy rules.
Now, we describe the main components of the image knowledge base as follows. The three components of the image knowledge base to be encoded are the membership functions of the
fuzzy variables, the fuzzy rules, and the linguistic modifiers of

identifies the pattern defined by the set of indexes

, and the corresponding
fuzzy rules are denoted as follows.
AND
is
AND
Rule VDK: IF ( is
is
) THEN
is
..
.
AND
is
AND
Rule VBR: IF ( is
is
AND
is
) THEN
is
.
is a five-gene that encodes the linguistic modifiers
3)
of the linguistic terms.
The factors of the luminance considered here are the fuzzy
,
,
,
, and
. In this
linguistic terms

where the index makes

value for

be the minimum
(6)
(7)
(8)
(9)
(10)

702

Fig. 15.

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART B: CYBERNETICS, VOL. 35, NO. 4, AUGUST 2005

Image distance between the Lena image and the testing images.

(a)

Fig. 17. Values of fitness obtained during the learning process and effects of
different choices of genetic parameters for the Lena image.

(12)

(b)
Fig. 16. (a) MSE curves of Type 1 images by GFIF. (b). MSE curves of Type
2 images by GFIF.

paper, we define the encoding restrictions for the fuzzy variable

as follows:

The restrictions are used to preserve meaningful fuzzy sets.

In contrast to the variation intervals proposed by Cordon et al.
[23], the restrictions make the tuning range large to improve the
training efficiency for the image knowledge base. In fact, the
above restrictions are not suitable to train the membership functions of the luminance. The luminance is variable for the images;
,
,
,
, or
. Thus, the linguistic
it may be
of the luminance may be located at different gray
term
levels for a different chromosome. In this paper, we propose
an adaptive restriction algorithm that can solve this problem.
The Adaptive Restriction Algorithm for the fuzzy variable luminance is as follows (see Fig. 8).
Adaptive Restriction Algorithm
Set the adaptive restrictions of fuzzy
variable luminance for genetic tuning.
Input:
The chromosome from the gene population.
Output:

LEE et al.: GENETIC-BASED FUZZY IMAGE FILTER AND ITS APPLICATION TO IMAGE PROCESSING

703

(a)

(a)

(b)

(b)

Fig. 18. (a) Values of fitness obtained during the learning process and effects
of different choices of genetic parameters for Albert. (b) Values of fitness
obtained during the learning process and effects of different choices of genetic
parameters for Cameraman.

Fig. 19. (a) Values of fitness obtained during the learning process and effects
of different choices of genetic parameters for Pentagon. (b) Values of fitness
obtained during the learning process and effects of different choices of genetic
parameters for Airplane.

The restriction set of the linguistic

terms of the luminance.
Method:
Step 1: Retrieve the parameters of the
luminance from the chromosome.
Step 1.1:

,
and
.
the linguistic term with
Step 2.1:
minimum mountain value.
the linguistic term
Step 2.2:
with med-minimum mountain value.
the linguistic term with
Step 2.3:
medium mountain value.
the linguistic term
Step 2.4:
with med-maximum mountain value.
the linguistic term
Step 2.5:
with maximum mountain value.
Step 3: Generate the adaptive dynamic restrictions as follows:
Step 3.1: Generate the restriction
.
Step 3.2: Generate the restriction

.
Step 1.2:
.
Step 1.3:
.
Step 1.4:
.
Step 1.5:
.
Step 2: Sort the positions of the linguistic terms mountain values
,
,

704

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART B: CYBERNETICS, VOL. 35, NO. 4, AUGUST 2005

(a)

(a)

(b)

(b)
Fig. 20. (a) MAE curves of GFIF before tuning and after tuning on the images
corrupted by salt-and-pepper impulse noises with the noise corruption rate p,
where p = 0 to 0.9. (MAE curves of Cameraman image). (b) MAE curves of
GFIF before tuning and after tuning on the images corrupted by salt-and-pepper
impulse noises with the noise corruption rate p, where p = 0 to 0.9. (MAE
curves of Airplane image).

Step 3.3: Generate the restriction

.
Step 3.4: Generate the restriction
.
Step 3.5: Generate the restriction
.
Step 4: End.
Next, a linguistic modifier used in the image knowledge base
is a function with the parameter that lets us alter the membership functions. Two of the most well-known modifiers are the
and the dilation linerosion linguistic modifier very
guistic modifier more-or-less
[23], [25]. Equations
(13) and (14) denote the functions of these two modifiers used
in this paper:
(13)
(14)
Fig. 9(a)(c) show the graphic representation of the specific
linguistic terms normal
, very DK
, and
more-or-less DK
, respectively. The initial population for the gene pool is composed of four groups with the same

Fig. 21. (a) MSE curves of GFIF before tuning and after tuning on the images
corrupted by salt-and-pepper impulse noises with the noise corruption rate p,
where p = 0 to 0.9. (MSE curves of Cameraman image) (b) MSE curves of
GFIF before tuning and after tuning on the images corrupted by salt-and-pepper
impulse noises with the noise corruption rate p, where p = 0 to 0.9. (MSE
curves of Airplane image).

part and
part. The first group is comnumber
part and the original
part
posed of the original
with the binary weight (
,
,
)
and the unit modifier
. The second group is composed
part and the randomized
part.
of the original
The third group is composed of the randomized
part and
the original
part. The fourth group is composed
of the randomized
part and the randomized
part. Next, we describe the genetic operators as follows.
Fig. 10 shows the flowchart of genetic learning process [25].
The genetic learning process initiates the population by the
encoding schema and restrictions and then records the initiation
population as the current population. The chromosome in
the current population is evaluated by the fitness function. If
the evaluation does not suit the fitness function, then genetic
learning processes the generation. The elitism is used in the
learning process. In the beginning of the selection, the best two
chromosomes in the current population are selected to a new
population without crossover and mutation. After the elitism,
we produce the other population by means of reproduction,
crossover, and mutation operators. The individuals having
the best fitness have more chances to be reproduced. The
one-point crossover method is adopted, and the crossover point

LEE et al.: GENETIC-BASED FUZZY IMAGE FILTER AND ITS APPLICATION TO IMAGE PROCESSING

705

Fig. 22. Results of salt-and-pepper noisy image Albert with noisy corruption rate p, where p = 0:2 and 0.8. (a) Noise corruption rate 0.2. (b) Result of GFIF
before tuning fuzzy inference alone. (c) Result of GFIF after tuning fuzzy inference with genetic learning. (d) Noise probability 0.8. (e) Result of GFIF before
tuning fuzzy inference alone. (f) Result of GFIF after tuning fuzzy inference with genetic learning.

is randomly placed. The mutation process is applied to each

offspring after the crossover. In this paper, the
part and the
part are mutated in the restriction of (12), binary
weight, and the set
, respectively. The object function
, which measures the fitness of each individual, is based on
the mean-absolute error (MAE) between the processed image
and the original noise-free image:

(15)
is the number of processed pixels. The learning
where
process stops when an assigned number of generations has been
evolved or when a satisfactory value of fitness has been obtained.
V. EXPERIMENTAL RESULTS
In this section, we compare the capability of GFIF with the
famous filters. The compared filters are the NASM filter [7],
the FIF filter [14], the SD-ROM filter [11], Russos filter [15],
[16], the FNN filter [21], and the HFF filter [17]. The noise-free
image is corrupted by additive identical independent distribution (i.i.d.) impulse noise with the corruption rate , and the impulses take on positive and negative values with an equal
,
i.e., the is a Bernoulli random variable [13], as follows:
with corruption rate
with corruption rate
with probability
(16)

is the gray level of the noise-free pixel on locawhere

tion
,
is the noise amplitude corrupted on location
, and
is the gray level of the noisy pixel for
.
The amplitudes
of the middle-tailed impulse noise and
long-tailed impulse noise adopted in this paper are 100 and
200, respectively. In the beginning, we analyze the properties of
GFIF and then verify the noise removal capability of GFIF by
comparing it with the other filters. To decide the parameter set of
GFIF for the experiment, we adopt the well-known 256 256
Lena image that has 256 gray levels to be the sample image
for the fuzzy number construction process to construct the image
knowledge base. In addition, we also produce a salt-and-pepper
noisy Lena image with a corruption rate 0.4 for the fuzzy filtering process and the genetic learning process. We have chosen
a small population of 20 individuals and set the parameters of
genetic learning as follows: Crossover probability 0.6, mutation
rate 0.1, and 50 generations. Fig. 11(a)(c) show the noise-free
Lena image, noisy Lena image with a corruption rate 0.4,
and result image by GFIF, respectively.
Fig. 12(a) illustrates the fuzzy sets of the Lena image constructed by the fuzzy number construction process. The tuned
fuzzy sets by the fuzzy filtering process and the genetic learning
process are shown in Fig. 12(b).
Tables IIII show the parameters of fuzzy sets constructed by
GFIF for the noisy Lena image with various adaptive i.i.d. impulse noise, including salt-and-pepper noise, middle-tailed impulse noise, and long-tailed impulse noise, respectively. In addition, the corruption rate of these three noisy images is 0.4.
To analyze the properties of GFIF, we adopt two kinds of
images to be the testing data. The first kind of images, called
Type 1 images, include Albert, Baboon, and Cameraman,
which are shown in Fig. 13(a)(c).

706

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART B: CYBERNETICS, VOL. 35, NO. 4, AUGUST 2005

TABLE IV
PSNR VALUES OF THE COMPARED APPROACHES FOR ADDITIVE LONG-TAILED
IMPULSE NOISY LENA IMAGE WITH THE CORRUPTION RATE p,
WHERE p = 0:1 TO 0.7

(a)

TABLE V
PSNR VALUES OF THE COMPARED APPROACHES FOR ADDITIVE LONG-TAILED
IMPULSE NOISY BRIDGE IMAGE WITH THE CORRUPTION RATE p,
WHERE p = 0:1 TO 0.7

(b)
Fig. 23. (a) PSNR curves of the compared approaches corrupted by long-tailed
impulse noises with the noise corruption rate p, where p = 0:1 to 0.7. (PSNR
curves of Lena image). (b) PSNR curves of the compared approaches
corrupted by long-tailed impulse noises with the noise corruption rate p, where
p = 0:1 to 0.7. (PSNR curves of Bridge image).

The second kind of images are called Type 2 images, including Sailboat, Bridge, Boats, House, Pentagon,
and Airplane are shown in Fig. 14(a)(f). They are all
256 256 images having 256 gray levels.
We use the Euclidean distance [26] to estimate the image
distance between the Lena image and the testing images including Type 1 and Type 2 images. The image distance estimation function is denoted as follows:
if
(17)
where
and denote the histograms of sample image and
the other testing image , respectively. Fig. 15 shows the image
distance between the Lena image and the testing images.
To analyze the behavior of GFIF on image distance for similar
images, we produce two kinds of salt-and-pepper impulse noise
with a corruption rate of 0.2 and 0.8 on the testing images and
then filter them by GFIF with tuned parameters of Lena shown
in Table I. Fig. 16(a) and (b) show the MSE curves of Type 1
and Type 2 images by GFIF, respectively. By this experimental

result, we observe that the small image distance will generate

the better performance for GFIF.
In order to analyze the behavior of the genetic learning
process, we choose smaller image distance Albert and Pentagon from the Type 1 and Type 2 images and larger image
distance Cameraman and Airplane from the Type 1 and
Type 2 images, respectively, to test the convergence for GFIF.
In addition, we also choose a small population of 20 individuals
and 50 generations and set the crossover probability 0.9 or 0.6
and the mutation rate 0.05 or 0.1. Fig. 17 shows the fitness
curves of GFIF with various parameters for the Lena image.
Fig. 18(a) and (b) show the fitness curves of Type 1 images
for Albert and Cameraman, respectively. Fig. 19(a) and (b)
shows the fitness curves of Type 2 images for Pentagon and
Airplane, respectively.
By this experimental result, we can see that the genetic
learning process of GFIF is robust for various parameters and
images. Next, we analyze the filtering capability of the fuzzy
filtering process for GFIF. We utilize two kinds of fuzzy sets
including original parameters obtained by the fuzzy number
construction process and tuned parameters obtained by the genetic learning process for GFIF. The detailed parameter sets for
salt-and-pepper impulse noise, additive middle-tailed impulse

LEE et al.: GENETIC-BASED FUZZY IMAGE FILTER AND ITS APPLICATION TO IMAGE PROCESSING

707

TABLE VI
RUNTIME (IN SECONDS) CONSUMED AT VARIOUS NOISE DENSITIES p USING THE PROPOSED GFIF AND OTHER FILTERS BASED ON THE LENA IMAGE

(a)

(b)

(c)

(d)

Fig. 24. (a) MAE curves of the compared approaches corrupted by additive impulse noises with the noise corruption rate p, where p = 0:4 to 0.9. (a) MAE
curves of salt-and-pepper noisy Boats image with various corruption rate p. (b) MAE curves of additive long-tailed impulsive noisy Baboon image with various
corruption rate p. (c) MAE curves of additive long-tailed impulse noisy Pentagon image with various corruption rate p. (d) MAE curves of additive middle-tailed
impulse noisy Albert image with various corruption rate p.

noise, and long-tailed impulse noise are shown in Tables IIII,

respectively.

This experiment is performed on the images corrupted by the

salt-and-pepper impulse noise with the corruption rate from 0 to

708

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART B: CYBERNETICS, VOL. 35, NO. 4, AUGUST 2005

(a)

(b)

(c)

(d)

Fig. 25. PSNR curves of the compared approaches corrupted by additive impulse noises with the noise corruption rate p, where p = 0:4 to 0.9. (a) PSNR curves
of salt-and-pepper noisy Boats image with various corruption rate p. (b) PSNR curves additive long-tailed impulse noisy Baboon image with various corruption
rate p. (c) PSNR curves of additive long-tailed impulse noisy Pentagon image with various corruption rate p. (d) PSNR curves of additive middle-tailed impulse
noisy Albert image with various corruption rate p.

0.9. We also adopt the Cameraman and Airplane images to

be the testing data. Fig. 20(a) and (b) show the MAE curves of
GFIF.
Fig. 21(a) and (b) shows the MSE curves of GFIF with two
kinds of parameters. By the experiments, we declare that the
filtering capability of GFIF with tuned parameters is more effective than GFIF with original parameters.
Since it is difficult to judge the performance of an image
noise-removal processing algorithm based solely on quantitative analysis, we show the filtered images for subjective evaluation in Fig. 22. Fig. 22(a)(c) shows the salt-and-pepper noisy
Albert image with a corruption rate 0.2, the result of GFIF
before tuning, and the result of GFIF after tuning, respectively.
Fig. 22(d)(f) show the salt-and-pepper noisy Albert image
with a corruption rate 0.8, the result of GFIF before tuning, and
the result of GFIF after tuning, respectively.
After analyzing the properties of GFIF, we compare the noiseremoval capabilities of CWM, SD-ROM, Russo, NASM, FIF,
FNN, and GFIF in the following experiments. The parameters
,
of Russos method are setting as follows:
, crossover probability 1.0, mutation rate 0.005, population
size with 40 individuals and 50 generations. For SD-ROM, the
threshold values are 12, 25, 40, and 50. The extrapolated peak

signal-to-noise ratio (PSNR) [7] curves resulted from using various filters at different noise densities, ranging from 0.1 to 0.7,
are shown in Fig. 23. The proposed GFIF significantly outperforms other filtering schemes when the noise rate is above 0.5.
The extrapolated PSNR value of the Lena image and the
Bridge image resulting from using various filters at different
noise densities, ranging from 0.1 to 0.7, are shown in Tables IV
and V, respectively.
The runtime analysis of the proposed GFIF and other concerned filters were conducted for the Lena image using a
Pentium IV 2.4-GHz Personal Computer and documented in
Table VI.
Fig. 24(a) shows the MAE curves of all compared approaches
for the salt-and-pepper impulse noise on the Boats images.
Fig. 24(b) and (c) shows the MAE curves of all compared
approaches for the additive long-tailed impulse noise on the
Baboon images and the Pentagon images, respectively.
Fig. 24(d) shows the MAE curves of all compared approaches
for the additive middle-tailed impulse noise on Albert images.
Fig. 25 shows the PSNR curves of all compared approaches.
Figs. 26(a) and (b) show the MAE and MSE curves of all
compared approaches for the uniformly distributed impulsive
noise on the Lena images, respectively. We observe that GFIF

LEE et al.: GENETIC-BASED FUZZY IMAGE FILTER AND ITS APPLICATION TO IMAGE PROCESSING

709

(a)

(b)

= 0%

Fig. 26. MAE and MSE curves of the compared approaches corrupted by uniformly distributed impulse noise with the noise corruption rate p, where p
to
10%. (a) MAE curves of uniformly distributed impulsive noisy Lena image with various corruption rate p. (b) MSE curves of uniformly distributed impulsive
noisy Lena image with various corruption rate p.

=06

Fig. 27. Results of salt-and-pepper noisy Boats image by salt-and-pepper noise with corruption rate p, where p
: . (a) Original image 256
256 pixels.
(b) Original image 100 100 pixels. (c) Corruption rate 0.6. (d) Russo result. (e) SD-ROM result. (f) FIF result. (g) FNN result. (h) HFF result. (i) GFIF result.

710

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART B: CYBERNETICS, VOL. 35, NO. 4, AUGUST 2005

Fig. 28. Results of additive long-tailed impulse noisy Baboon image by additive long-tailed impulse noise with corruption rate p, where p = 0:6. (a) Original
image (256 256 pixels). (b) Original image (100 100 pixels). (c) Corruption rate 0.6. (d) Russo result. (e) SD-ROM result. (f) FIF result. (g) FNN result. (h)
HFF result. (i) GFIF result.

work effectively when the noise corruption rate is below 5% for

the mixed impulsive noise.
To verify the capability of GFIF for fine detail preservation,
we adopt the 100 100 pixel Boats and Baboon images
with which to experiment. Figs. 27 and 28 show the results of
the salt-and-pepper noisy Boats image and the additive longtailed impulse noisy Baboon image with a corruption rate of
0.6, respectively. By the results, we can see that GFIF preserves
the fine details and textures better than the other approaches for
the additive i.i.d. impulse noise.

VI. CONCLUSION
In this paper, a genetic-based fuzzy image filter has been presented. GFIF contains four parts, including a fuzzy number construction process, a fuzzy filtering process, a genetic learning
process, and an image knowledge base to perform efficient and

effective impulse noise removal. The fuzzy number construction process receives sample images or the noise-free image and
then constructs the image knowledge base for the filter. Then,
the fuzzy filtering process refers to the image knowledge base to
execute the fuzzy inference. Since it is very relevant to remove
noise without degrading the image structure, a fuzzy set-based
mechanism is also embedded in the filter in order to preserve
the quality of fine details and textures. The genetic learning
process adjusts the parameters of fuzzy sets for getting the optimal image knowledge base. From the experimental results, we
observe that PSNR, MSE, and MAE curves of GFIF achieve the
most effective results than other approaches including CWM,
SD-ROM, Russo, NASM, FIF, FNN, and HFF for removing
heavily corrupted additive i.i.d. impulse noise. Subjective evaluation of GFIF also shows a high-quality restoration of filtered
images for the noise model. In the future, we will extend GFIF
to process color images. Moreover, the uniform distribution impulsive noise model (mixed impulsive noise model) will be further studied.

LEE et al.: GENETIC-BASED FUZZY IMAGE FILTER AND ITS APPLICATION TO IMAGE PROCESSING

ACKNOWLEDGMENT
The authors would like to express their gratitude to the anonymous reviewers for their comments that improved the quality of
this paper.

REFERENCES
[1] J. W. Tukey, Exploratory Data Analysis. Reading, MA: Addison-Wesley, 1971.
[2] J. Astola and P. Kuosmanen, Fundamentals of Nonlinear Digital Filtering. Boca Raton, FL: CRC, 1997.
[3] I. Pitas and A. Venetsanopoulos, Nonlinear Digital Filters: Principles
and Application. Norwell, MA: Kluwer, 1990.
[4] K. Arakawa, Median filter based on fuzzy rules and its application to
image restoration, Fuzzy Sets Syst., vol. 77, pp. 313, 1996.
[5] D. Brownrigg, The weighted median filter, Commun. Assoc. Comput.,
pp. 807818, Mar. 1984.
[6] S. J. Ko and S. J. Lee, Center weighted median filters and their applications to image enhancement, IEEE Trans. Circuits Syst., vol. 15, no.
9, pp. 984993, Sep. 1991.
[7] H. L. Eng and K. K. Ma, Noise adaptive soft-switching median filter,
IEEE Trans. Image Process., vol. 10, no. 2, pp. 242251, Feb. 2001.
[8] Y. H. Fung and Y. H. Chan, An improved algorithm for removing impulse noise based on long-range correlation in an image, in Proc. IEEE
Int. Conf. Multimedia Expo, vol. 1, Aug. 2002, pp. 157160.
[9] O. Y. Harja, H. Huttunen, A. Niemist, and K. Egiazarian, Design
of recursive weighted median filters with negative weights, in Proc.
IEEEEURASIP Workshop Nonlinear Signal Image Process., Baltimore, MD, Jun. 2001.
[10] L. Khriji and M. Gabbouj, Median-rational hybrid filters, in Proc. Int.
Conf. Image Process., Chicago, IL, 1998, pp. 853857.
[11] E. Abreu and S. K. Mitra, A signal-dependent rank ordered mean (SDROM) filter, in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process.,
Detroit, MI, 1995, pp. 23712374.
[12] C. S. Lee, Y.-H. Kuo, and P.-T. Yu, Weighted fuzzy mean filters for
image processing, Fuzzy Sets Syst., vol. 89, pp. 157180, 1997.
[13] Y. H. Kuo, C.-S. Lee, and C.-L. Chen, High-stability AWFM filter for
signal restoration and its hardware design, Fuzzy Sets Syst., vol. 114,
no. 2, pp. 185202, 2000.
[14] C. S. Lee, C. Y. Hsu, and Y. H. Kuo, Intelligent fuzzy image filter for
impulse noise removal, in Proc. IEEE Int. Conf. Fuzzy Syst., vol. 1, May
2002, pp. 431436.
[15] F. Russo, Hybrid neuro-fuzzy filter for impulse noise removal, Pattern
Recognit., vol. 32, pp. 18431855, 1999.
, Noise removal from image data using recursive neurofuzzy fil[16]
ters, IEEE Trans. Instrum. Meas., vol. 49, no. 2, pp. 307314, Apr.
2000.
[17] J. H. Wang, W. J. Liu, and L. D. Lin, Histogram-based fuzzy filter for
image restoration, IEEE Trans. Syst., Man, Cybern. B, vol. 32, no. 2,
pp. 230238, Apr. 2002.
[18] R. Lukac, Adaptive vector median filtering, Pattern Recogn. Lett., vol.
24, pp. 18891899, 2003.
[19] G. Pok, J. C. Liu, and A. S. Nair, Selective removal of impulse noise
based on homogeneity level information, IEEE Trans. Image Process.,
vol. 12, no. 1, pp. 8592, Jan. 2003.
[20] J. Y. Chang and J. L. Chen, Classified-augmented median filters for
image restoration, IEEE Trans. Instrum. Meas., vol. 53, no. 2, pp.
351356, Apr. 2004.
[21] P. Liu, Representation of digital image by fuzzy neural network, Fuzzy
Sets Syst., vol. 130, pp. 109123, 2002.

711

[22] R. S. Lin and Y. C. Hsueh, Multichannel filtering by gradient information, Signal Process., vol. 80, pp. 279293, 2000.
[23] O. Cordon, F. Herrera, and P. Villar, Generating the knowledge base
of a fuzzy rule-based system by the genetic learning of the data base,
IEEE Trans. Fuzzy Syst., vol. 9, no. 4, pp. 667674, Aug. 2001.
[24] H. J. Zimmermann, Fuzzy Set Theory and Its Applications. Boston,
MA: Kluwer, 1991.
[25] C. S. Lee and C. Y. Pan, An intelligent fuzzy agent for meeting scheduling decision support system, Fuzzy Sets Syst., vol. 142, pp. 467488,
2004.
[26] V. Castelli and L. D. Bergman, Image Database Theory and Application. New York: Wiley, 2002.

Chang-Shing Lee received the B.S. degree in information and computer engineering from the Chung
Yuan Christian University, Chung-Li, Taiwan,
R.O.C., in 1992, the M.S. degree in computer science and information engineering from the National
Chung Cheng University, Chia-Yi, Taiwan, in 1994,
and the Ph.D. degree in computer science and
information engineering from the National Cheng
Kung University, Tainan, Taiwan, in 1998.
He is currently an Associate Professor with the Department of Information Management, Chang Jung
Christian University (CJCU), Tainan. His research interests include intelligent
agent, ontology engineering, knowledge management, Web services, semantic
Web, and soft computing systems. He holds several patents on ontology engineering, document classification, and image filtering. He guest edited a special
issue of the Journal of Internet Technology.
Dr. Lee received the MOEs Campus Software Award in 2002, CJCUs Outstanding Research Achievement Award in 2003, and the Outstanding Teacher
Award from CJCU in 2004. He is a Member of TAAI.

Shu-Mei Guo received the M.S. degree from the

Department of Computer and Information Science,
New Jersey Institute of Technology, Newark, in
1987 and the Ph.D. degree in Computer and Systems
Engineering from University of Houston, Houston,
TX, in May 2000. Since June 2000, she has been an
assistant professor with the Department of Computer
System and Information Engineering, National
Cheng-Kung University, Tainan, Taiwan, R.O.C.
Her research interests include various applications on evolutionary programming, chaos systems,
Kalman filtering, fuzzy methodology, sampled-data systems, and computer and
systems engineering.

Chin-Yuan Hsu was born in Chiyai, Taiwan, R.O.C.,

in 1978. He received the B.S. degree in information
management from the Kun Shan University of Technology, Tainan, Taiwan, in 2002 and the M.S. degree
in computer science and information engineering
from National Cheng-Kung University, Tainan, in
2004. His research interests include digital image
processing, content-based image retrieval, and web
service.