1742

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 18, NO. 8, AUGUST 2009

Geometric Features-Based Filtering for Suppression

of Impulse Noise in Color Images
Zhengya Xu, Member, IEEE, Hong Ren Wu, Bin Qiu, Senior Member, IEEE, and Xinghuo Yu, Fellow, IEEE

AbstractA geometric features-based filtering technique,

named as the adaptive geometric features based filtering technique (AGFF), is presented for removal of impulse noise in
corrupted color images. In contrast with the traditional noise
detection techniques where only 1-D statistical information is used
for noise detection and estimation, a novel noise detection method
is proposed based on geometric characteristics and features (i.e.,
the 2-D information) of the corrupted pixel or the pixel region,
leading to effective and efficient noise detection and estimation
outcomes. A progressive restoration mechanism is devised using
multipass nonlinear operations which adapt to the intensity and
the types of the noise. Extensive experiments conducted using a
wide range of test color images have shown that the AGFF is superior to a number of existing well-known benchmark techniques,
in terms of standard image restoration performance criteria,
including objective measurements, the visual image quality, and
the computational complexity.
Index TermsColor image restoration, impulse noise detection,
progressive filtering.

I. INTRODUCTION
A. State-of-the-Art
ONLINEAR filtering techniques [1][8] have been extensively researched in the last decade due to their effectiveness in restoration of impulse noise corrupted color images. The
median filter is usually used to remove impulse noise [1], [9].
Compared with linear filters, the median filter is more robust in
that a single corrupt or noisy pixel in the filtering window will
not affect the median value significantly. For removal of noise in
color images, various vector median filters have proven relevant
and effective. Amongst the early publications, the most wellknown vector filters for color image denoising include the vector
median filter (VMF) [10], the vector directional filter (VDF)
[11], and the directional distance filter (DDF) [12]. While these
vector filters perform well in suppressing the impulse noise, they
introduce image distortions such as blurring around edges and
in detail areas which feature high spatial frequency contents and

Manuscript received August 14, 2008; revised April 05, 2009. First published
May 05, 2009; current version published July 10, 2009. This work was supported
by an Australian Research Council Discovery Project grant (DP0988654). The
associate editor coordinating the review of this manuscript and approving it for
publication was Dr. Ercan Kuruoglu.
Z. Xu, H. R. Wu, and X. Yu are with the School of Electrical and Computer
Engineering, Royal Melbourne Institute of Technology, Melbourne, Victoria
3001, Australia (e-mail: [email protected]; [email protected];
[email protected]).
B. Qiu is with the Faculty of Information Technology, Monash University,
Victoria 3800, Australia (e-mail: [email protected]).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TIP.2009.2022207

variations. Different types of weighted nonlinear filtering techniques have been investigated over the years to achieve better
performance in both noise suppression and detail preservation
[13][16].
Recently, fuzzy filtering techniques have been developed,
achieving robust image denoising performance [17][38]. A
class of chromatic filters for image restoration in the
color space was also proposed to achieve better chromatic
smoothness [39]. Adaptive filters have demonstrated their
effectiveness in image restoration considering various types
of noise with different distributions and image structures
[40][53].
In a sharp contrast with the additive noise that contaminates
all image pixels, the impulse noise destroys only some portion of an image and leaves other pixels noise-free. Detectionbased vector filtering techniques such as the adaptive vector median filter (AVMF) [42], the adaptive vector LUM (lower-uppermiddle) smoother (AVLUM) [54], modified weighted vector
median filter (MWVM) [55], and the adaptive selection center
weighted vector direction filter (ACWVDF) [56] were specially
designed to remove the impulse noise from color images. They
utilize a series of weighted median vector filters to perform binary noise detection and switch between the output of an identity filter and that of a weighted median vector filter, according to
the detection results. A survey of nonlinear vector filtering was
presented in [57] for impulse noise removal from color images.
B. Statistical Models of Impulse Noise
Color images may be contaminated by various types of noise
and impulse noise is the noise model frequently used and reported in digital restoration literature [1][8]. Impulse noise corruption often occurs in digital image acquisition or transmission
process as a result of photo-electronic sensor faults or channel
bit errors [14], [22], [58].
Image transmission noise may be caused by various sources,
such as car ignition systems, industrial machines in the vicinity
of the receiver, switching transients in power lines, lightning in
the atmosphere and various unprotected switches. This type of
transmission noise is often modeled as the impulse noise [22],
[33], [37]. The impulse noise can also be introduced into images during acquisition of the images. For example, the impulse noise may be introduced during fingerprint acquisition in
real-life border security check. For more background information about the physical model of the impulse noise, we refer
readers to [58]. The two most common impulse noise types are
fixed-value impulse noise (also known as the salt-and-pepper
noise) and random-value impulse noise [14], [22], [33], [58].
denote
Let
the set of pixel coordinates of a color image, where and are

1057-7149/$25.00 2009 IEEE

XU et al.: GEOMETRIC FEATURES-BASED FILTERING FOR SUPPRESSION OF IMPULSE NOISE

the height and the width of the image, respectively. At each pixel
, a multivariate value vector in the RGB color
coordinate
, is used to represent the
space,
RGB (Red, Green, Blue) pixel values.
Two approaches as reported in the literature are used in this
paper to model the impulse noise for color image restoration. In
the first approach, the impulse noise corruption of color images
in the RGB space is expressed by a multivariate model [11],
[41], [45], [59]
with probability
with probability

(1)

and in the second approach the impulse noise corruption of color

images in the RGB space is expressed by a multivariate model
[22], [25], [27], [32], [34], [53], [58]
with probability
with probability

(2)

where
and
represent the original and the observed
pixel (vector) values at coordinate , respectively, and the value
or
is generated by substituting at least one color
of
component of the pixel
by a distinct value in both (1)
is the impulse noise ratio; a factor
and (2). In (1),
is used to simulate the channel correlation for each corrupted
pixel, namely, if at least one of the three components of the pixel
is corrupted by the impulse noise, its remaining noise-free components will have a 50% probability to be corrupted [41], [45],
[59]. The second approach (2) is a more generalized impulse
is the
noise model of color images, where
impulse noise ratio for each channel of a corrupted color image,
assuming that the image is corrupted by the impulse noise in a
channel independent manner.
or
,
In (1) and (2), if , the component value of
equals the maximum or the minimum value of the digital image
(e.g., 255 or 0 for an 8-bit channel of the 24-bit color image in
the RGB space), the impulse noise is referred to as the salt-andpepper impulse [45]. Each pixel of the image may be corrupted
by either the pepper or the salt impulse with unequal probabilities. However, if the amplitudes of the impulse are distributed
randomly with, e.g., the uniform or the Gaussian distribution, in
the range of [0, 255], a more general type of the impulse noise
is generated and named as the random impulse noise [13], [58].
The impulse noise can be represented by a joint probability
distribution describing the spatial distribution of the impulses
as well as their amplitudes. As is typically the case, these two
quantities are considered to be independent [58].
In this paper, an adaptive geometric features-based filtering
technique (AGFF) with a low computational complexity is
proposed for restoration of digital color images corrupted by
the impulse noise. This technique uses a set of novel noise
detection criteria for detection of the corrupted pixels, which
are based on 2-D geometric and dimension features of the noisy
pixel or the noisy region of images. This is in contrast with the
traditional noise detection techniques where only 1-D statistical
information is used for estimation of the noise ratio and the
noise statistical distribution model [2][4], [10][12], [45],
[60][63]. Based on the result of the estimation, an adaptive

1743

progressive filtering operation is employed in combination

with optimized dimension and shape of processing windows.
Computational efficiency of the AGFF is also investigated.
Denoising performance of the AGFF is evaluated to demonstrate noticeable gains against that of a number of well-known
benchmark techniques mentioned above, in terms of standard
objective measurements, perceptual image quality and computational complexity, especially for suppression of the impulse
noise in medium- and large-size color images.
The organization of this paper is as follows. Section II
presents the dimensional and geometric features of the impulse
noise. Section III addresses the principles of the adaptive
geometric features-based filtering technique (i.e., AGFF) and
issues regarding its implementation and computational complexity. Experimental results are presented in Section IV. Final
conclusions are drawn in Section V.
II. DIMENSIONAL AND GEOMETRIC FEATURES
OF IMPULSE NOISE
A major problem in restoration of color images to date is
the destruction of detailed image structures due to inability of
denoising filters to distinguish a cluster of corrupted pixels from
a cluster of pixels presenting fine (detailed) image structures and
the incorrect removal or modification of pixel segments [21],
[22], [26], [42], [44], [56], [59], [60]. This section describes
a novel technique which detects and removes, effectively and
efficiently, impulses in color images.
and
are
As defined in [64], any two pixels at
called 4-neighbors, if they have a city block distance
from each other. Similarly, 8-neighbors are two pixels with a
chessboard distance
. The city block distance is defined
and the chessas
board distance is defined as
[64][66]. For example, each color image pixel
in Fig. 1(d) is represented by the coordinates of its position, i.e.,
a pair of integers (column number, row number). Given a pixel
(3,3), for instance, its 4-neighbours are (2,3), (3,4), (3,2) and
(4,3), and its 8-neighbours are its 4-neighbours plus (2,2), (4,4),
(2,4) and (4,2) [66].
Careful examination of a variety of color images corrupted
by the aforementioned impulse noise models reveals that most
of uncorrupted pixels or pixel regions in a natural color image
always demonstrate a certain degree of smoothness. This means
that the color intensities of a pixel always change gradually in
all its 8-neighbors directions (e.g., in a smooth area), or change
gradually at least in one (edge) direction (e.g., in a boundary
area) [see Fig. 1(a) and (b)]. In contrast with normal or uncorrupted pixels of images, impulse noise corrupted pixels always
stand out as an isolate spot or a cluster by its very un-harmonious colors, shapes and sizes compared with those of its neighborhood [see Figs. 1(c) and (d) and 2]. Even in the boundary
(or edge) area, uncorrupted objects in natural color images have
different types of edges from those corrupted by the impulses
[see Figs. 1(c) and (d) and 2(a)].
It is observed that almost all impulses only have sharp step
edges and, in contrast, almost none of the uncorrupted objects
has this type of edges in its vicinity. The borders of the uncorrupted objects still have a narrow transitional region of a

1744

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 18, NO. 8, AUGUST 2009

Fig. 1. Geometric features of impulse noise for image corrupted by impulse with low ratio noise. (a) Original color image Boats. (b) Zoomed portion of the test
color image Boats. (c) The zoomed test color image Boats in (b) corrupted by 10% of the salt-and-pepper impulse noise using model (2). (d) The further zoomed
test color image within the black rectangle of Fig. 1(c).

Fig. 2. Geometric features of impulse noise for images corrupted by impulse with high ratio noise. (a) Zoomed portion of the test color image Parrots corrupted
by 40% of the salt-and-pepper impulse noise. (b) The further zoomed test color image within the black rectangle of (a).

few pixels, even in the gradient direction of a sharp changing

boundary area.
In cases where images corrupted by the impulse with a low
noise ratio, the sizes of the corrupted pixels (i.e., corrupted
areas) are most likely represented by isolated individual pixels
or a short line with one-pixel width [see Fig. 1(c) and (d)]. The
pixels of the line may be adjacent in the diagonal direction. With
the increase of the noise ratio, corrupted pixel regions/clumps
with two-pixel width in two perpendicular directions may occur
along with the individual impulses and smaller impulse regions
as shown in images corrupted with the low impulse ratio. The
shapes of the noise regions may be an isolated point, a short thin
line, a cross of two short thin lines or other small round-shaped
blocks (see Fig. 2). In other words, with the increase of the
noise ratio, the noise may appear isolated or clustered with
more different sizes and shapes.
According to the above observations and analysis of color,
shape and sizes of impulse noise corrupted pixels/regions, and

the type of edges which form the borders of the noise regions,
a novel impulse noise detection method is devised here based
on 2-D geometric features of the impulses, instead of the 1-D
rank ordered statistical information used by other well-known
filtering techniques, to determine if each pixel in a color image
is corrupted or clean. One of the geometric properties of the impulse noise is the edge feature of its boundary. An edge can
be defined as a local discontinuity in color component or illumination intensity function and the edge orientation is defined as edges of an octagonally shaped object whose amplitude is higher or lower than its background [65]. Therefore,
the criteria for identifying the edge feature around the pixel are
based on two types of derivatives, which are approximated by
pixel differences in digital color images. Given that
is the vector containing color component functions of a color image [66], the two special types of
and
,
partial derivatives are denoted as
at
is approximated by
,
respectively.

XU et al.: GEOMETRIC FEATURES-BASED FILTERING FOR SUPPRESSION OF IMPULSE NOISE

the differences between the pixel and its 4-neighbors for each
component of the color pixel, and defined as follows:

(3)

where
,
and the default value
is 1, for
, and superscript represents the
of
transpose operation.
When a derivative is only considered in the diagonal direcis approximated by
, the differences betion,
tween the pixel and its other 8-neighbors, for each component
of the color pixel, and defined as follows:

(4)

where
,
and the default value of
is 1, for
.
The two special derivatives,
and
, will be used to measure the edge feature (sharpness) and other geometric properties
is corrupted
to determine whether the center pixel at
or not in the AGFF technique.

1745

corruption detectors are used in [19], [21], [42], [45], [56], [59],
[70] based on statistical information about the neighborhood of
each local pixel to locate impulse noise and to preserve clean
pixels. Time-consuming multiple reference filtering and complex parameter training process highly limit the usage of these
filters in real-time application. The new criteria presented in this
paper also depart from recently developed fuzzy impulse noise
filtering techniques [22], [32], [33], [36], [37]. For example, the
fuzzy noise detection method [22], [32], [37] is mainly based
on calculation of fuzzy gradient values and fuzzy reasoning,
and the fuzzy membership function representing the impulse
noise is a simplification of the obtained noise histogram [22],
[32].
A. Two-Dimensional Impulse Noise Detection
A key component of the AGFF technique is a novel impulse
detection scheme based on the 2-D geometric information of
the corrupted pixels. First, we define the edge feature-identification threshold, , which represents the value of a derivative to
distinguish the sharp step edges from other types of edges [3].
Since very short thin lines usually form impulse noise pixels,
the length of a line is also used as a feature to distinguish a short
noise line from a fine line in color images. The length threshold,
, may be defined according to the noise ratio.
Second, in terms of the pixel coordinates of a color image, ,
a set of corrupted pixels is defined as

III. PRINCIPLE OF AGFF TECHNIQUE

In detecting and removing impulse noise, a filter may make
three main types of mistakes. Type I error (miss) occurs when
there is a corrupted pixel which the filter does not detect. Type
II error (false alarm) happens when the filter detects an impulse
noise pixel which is actually clean. When the filter removes
an impulse noise and replaces it with a value determined by a
certain restoration strategy, Type III error (over- or under-correcting error) is defined as the difference between the resultant
value after the restoration process and the true pixel value as the
noise-free pixel was.
Different types of the so-called switching filters and fuzzy
filtering techniques [3], [21], [22], [25][28], [30][35], [38],
[41][45], [53], [56], [57], [59], [60], [63], [67][70] have
been developed over the years, achieving good performance in
both noise suppression and detail preservation. Similar to other
well-known benchmark techniques including the so-called
switching filters [41][45], [59], [60], [63], [67][70] and
fuzzy-based filtering techniques [21], [22], [25][27], [33],
the AGFF technique described in this section consists of two
components, i.e., impulse detection and impulse removal. The
novel criteria used by the AGFF for noisy pixel detection are
based on a combination of the 2-D edge, geometric and size
features of the noisy pixel/region in the images. They depart
from traditional noise detection techniques used by other existing filters, which only use some properties of the edge of a
noisy pixel or 1-D rank ordered statistical information around
the noisy pixel [2], [3], [10], [11], [22], [25], [33], [41], [42],
[67][70]. For example, multiple threshold framework and

for
(5)
is used to define corrupted pixel-sizes in
and its
where
default value is 2. According to (5), IF the two partial derivaand
of a pixel have the same sign while their magtives
is 1 and
nitudes are greater than a preset threshold , when
is 1 or 2 or 3, or
(its components may have different
values), THEN the pixel belongs to . Set
includes individual impulse pixels, slant noise lines with one-pixel width and
the pixels of the lines only adjacent to each other in diagonal direction within the defined length of .
Third, a set of corrupted pixels, which include individual impulse pixels, straight noise lines with one-pixel width and the
pixels of the lines being only 4-neighbors to each other within
the defined length of , is defined as

for
(6)
is used to define corrupted pixel-sizes in this set and
where
its default value is 2. According to (6), IF the two partial derivaand
of a pixel have the same sign while their magtives
is 1 and
nitudes are greater than a preset threshold , when
is 1 or 2 or 3, or
(its components may have different values), THEN the pixel belongs to . Next, a set of

1746

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 18, NO. 8, AUGUST 2009

TABLE I
PERFORMANCE OF THE CORRUPTED PIXEL SETS AS DEFINED IN (5)(9) FOR DETECTING RANDOM IMPULSE NOISE IN A
VARIETY OF STANDARD TEST COLOR IMAGES CORRUPTED WITH NOISE RATIO IN NOISE MODEL DEFINED BY (1)

corrupted pixels is defined as , which include noisy pixels/regions within 3-pixel width in any direction except the noisy
and/or , i.e.,
pixels already in

(7)
where
is 2. Thus,

, is 2 or 3, and the default value for

can be represented as

and
(8)

for
and
in (8). According to (8), IF the
where
two partial derivatives
and
of a pixel have the same sign
while their magnitudes are greater than a preset threshold ,
and
are 2 or 3 and the pixel is not in
or ,
when
THEN the pixel belongs to .
Finally, according to observations and analysis of a variety
of natural images corrupted with the impulse noise, a protrusive
point in a border area with high possibility of being a corrupted
pixel is defined as

(9)
and
of a
According to (9), IF the two partial derivatives
pixel have the same sign while their magnitudes are greater than
a preset threshold , with the partial derivatives indexed by
containing only three out of the four distance settings, and the
partial derivatives indexed by being either {2, 3} or {3, 4} or
{4, 1} or {1, 2} and equal to , when , , , and
are
1, THEN the pixel belongs to . Since an impulse noise ratio
,

, where

, in the current design.

in

(5), (6), (7), and (9) may be set at different values.

Figs. 1 and 2 are used to illustrate some geometric shapes of
impulse pixels in each of the sets defined by (5), (6), (7) and (9).
Each color image dot in Figs. 1(d) and 2(b) is represented by its
position, i.e., a pair of integers (column number, row number).
is 2 and
is 40, in Fig. 2(b), the
According to (5)(9), if
red component of (4,6) is in ; the red component of (10,4)

and the blue component of (6,2) are in ; the red component

of (12,5) and the green components of (6,3) and (7,2) are in
; the red component of (6,6) is in ; the blue components
of (11,5), (8,7) and (13,7), the red component of (12,9) and the
and . For
green components of (4,5) and (7,5) are in both
another example, the green components of (6,3) and (7,2) in
of two pixel long (its
Fig. 2(b) form a slant noise line in
). Using the same setup for
and ,
length within
in Fig. 1(d), the green component of (10,7) in an edge region
or . Because if it were in , the values of all
is not in
components of
for the green component would, at least, have
to be either greater than 40 or less than 40. Actually, although
for the green component are
all the component values of
greater than 0, their minimal value is only 6 in this example.
In yet another example, in a sharp edge in Fig. 1(d), the RGB
components of pixels (8,2)(8,6), which represent a wall, are
nor
nor
nor , because the values of some
neither in
or
, for the given
, are either less
RGB components of
than 40 or greater than 40.
Image signals are statistical in nature due to random changes
and noises, and it is advantageous to treat the image function
as realization of a stochastic process [71]. In such an approach,
mathematical analysis of the AGFF can be performed using statistical descriptors. Without loss of generality, the proposed definitions of the corrupted pixel sets in color images were tested
by experiments using typical test images, which included different types of real-life images [72]. Table I presents the experimental results for detection of random impulse noise using ,
, and on a variety of original clean (without corruption)
test images, which include well-known standard images such
as, Airplane, Boats, Flower, Girl, Goldhill, Moon, Pen, Soccer,
Zelda, and Yacht with image resolutions of 500 362, 512
512, 787 576, 720 576, 768 512, and 1986 1986 pixels,
respectively (see Fig. 3).
In Table I, FA stands for false alarm rep,
,
resenting Type II error, and
and
denote false alarm ratios
(
,
). The false alarm ratios are very low for the
where
test images, especially for the high resolution color images.
Using the above definitions of the corrupted pixel sets, therefore, the noisy pixels in color images can be detected effectively
and efficiently using the characteristics of shape, size and edge
type of the corrupted pixels, by adapting appropriate parameters/thresholds to the noise ratio and the types.

XU et al.: GEOMETRIC FEATURES-BASED FILTERING FOR SUPPRESSION OF IMPULSE NOISE

1747

Fig. 3. Test images. (a) Airplane; (b) boats; (c) Flower; (d) Girl; (e) Goldhill; (f) Moon; (g) Pen; (h) Soccer; (i) Yacht; (j) Zelda.

Since the shapes and the sizes of corrupted pixels depend on

the noise ratio, and the restoration of the corrupted pixels requires the statistical information about the noise density, the estimation of the noise ratio and the noise type is important in
order to minimize both Type I and Type II errors. The criteria
used by the AGFF technique for classifying the degree of the
impulse noise corruption are the ratio and the size of the largest
noise corrupted region as well as the values of corrupted pixels.
The strategy of the progressive restoration for the AGFF is,
first, to restore corrupted individual pixels or noise regions of
small size. If it made either Type II or Type III errors, it should
not introduce any new impulse noise region bigger than the existing ones. Then, further operations are carried out around large
noise corrupted regions to restore areas of the images associated
with noise regions of considerable size reliably.
In order to take advantages of the median filter and to avoid
its drawbacks (i.e., causing a number of artifacts for uncorrupted
pixels) [9], [73], a detection scheme is described in this section
for use before the median filtering for restoration. As a result, the
proposed restoration method based on the restricted median can
keep the image unchanged when the filter processing window
moves across the uncorrupted image details.
Clearly, it may become expensive to perform a sort on pixels
within a large rectangular window. If the width of processing
window is larger than three, a modified median filter can be applied alternatively in the AGFF technique. The modified median
will be based on the part of the pixels which are from the outline of the window or the noise-free pixels within the processing
window [3], since part of pixels inside the window may have
been corrupted. For example, the noisy central pixel and its corrupted eight neighborhood pixels, if detected, will be excluded
from the set for median filtering. The modified median filter increases the reliability of the restoration and reduces the computational cost, especially for removing impulses of high noise
ratio. If the channel correlation factor, , for each corrupted pixel
as defined in (1), approaches 1, the modified vector filtering is
recommended.

B. AGFF Implementation
Implementation of the principles of the AGFF is described as
, a square filter processing
follows. At each coordinate
window is defined, which centers at the coordinate and conpixels, where
is an odd number. The width of the
tains
and must be a positive integer,
window is represented as
i.e., 3, 5, 7, and so on, in the recommended implementation.
are imThe operations for removing impulses in , , and
plemented by a 3 3 processing window, and the operation for
removal of impulses in is implemented by a 5 5 processing
window. Each processed pixel in a color image is at the center
of the symmetric window.
The restoration technique used in this work is based on
the modified median as described previously, with which the
destruction of noise-free pixels (i.e., Type II error) becomes
increasingly severe with the increase of the processing window
size while less likely with the increase of the edge feature
threshold . In order to achieve the best performance of the
proposed filter, in terms of both visual quality and objective
measurements, the design of the processing windows has to depend on the shapes and sizes of corrupted pixels/pixel regions.
The threshold to discriminate the corrupted pixel and its
neighborhood is usually set to less than 20 in magnitude for
the filter, in order to improve perceptual image quality (i.e.,
reducing Type I error; see Figs. 7 and 8). However, with the
decrease of the threshold , the filter removes more and more
uncorrupted details of the image during the operation. It is important that Type II error is also under good control in order to
preserve the uncorrupted details, structures and features of the
image as much as possible. A balance of removing noise and
preservation of the details must be made based on optimization
for detecting the
of objective measurements. The threshold
sharp step edge type for color images can be determined by
a greedy algorithm searching for the least MSE (mean square
error) [13] using a set of test color images [72] with different
impulse noise ratios and types in an off-line experiment. In the
following implementation, the threshold
is chosen from the
range of [15], [40] adapted to the size of processing window.

1748

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 18, NO. 8, AUGUST 2009

Fig. 4. Set of fuzzy membership functions as given by (10)(14).

The estimation of the impulse noise ratio and the type of a

corrupted image is obtained by analyzing the ratio and values
or
detected from the image, through a fuzzy ruleof
or
are similar
based approach. The estimation results by
to each other and, therefore as an example, the estimation by
is described as follows. A set of membership functions are
, lowgiven by expressions (10)(14) for the noise-free
noise-ratio
, medium-noise-ratio
, high-noiseand very-high-noise-ratio
of the fuzzy
ratio
set

Fig. 5. Relationship between the impulse noise ratio and R for different test
images which were corrupted with different random impulse noise ratios using
the noise model defined by (1).

(10)

(11)

(12)

(13)
Fig. 6. Algorithm 1impulse suppression by AGFF.

(14)
is the ratio of
(except
) detected
where
from RGB channels of the image to the image size, i.e.,
. Fig. 4
illustrates the shape of the membership functions in (10)(14).
Without loss of generality, the recommended parameters of
fuzzy membership functions were obtained by experiments
using widely used test images, including both real-life images
and images containing artificial objects such as Airplane, Girl,
and Pens (see Figs. 3 and 5) [72].
The number of fuzzy membership functions associated with
each variable depends on the denoising operations and the sum
of the fuzzy membership values where the functions overlap is
recommended to be one or less than one. Because the AGFF can
tolerate the estimation deviation of the noise ratios, the simple
trapezoidal shape is chosen as the functions in the fuzzification process. The maximum method is used in defuzzification.

Fig. 7. Relationship between the chosen values of the parameter T and the
average MSE of the output images, where the test images were corrupted with
different random impulse noise ratios in noise model defined by (2).

The noise type of the salt-and-pepper can be determined by the

or .
values of
Table II demonstrates, the ratio of the number of salt-and(except
) to the
pepper noise corrupted pixels in

XU et al.: GEOMETRIC FEATURES-BASED FILTERING FOR SUPPRESSION OF IMPULSE NOISE

1749

TABLE II
RATIO (%) OF THE NUMBER OF PIXELS IN S AS THE SALT-AND-PEPPER NOISE TO THE NUMBER OF PIXELS IN S FOR
DIFFERENT COLOR IMAGES, WHERE THE TEST COLOR IMAGES WERE CORRUPTED BY 1% OR 70% OF RANDOM IMPULSE NOISE,
OR 1% OR 30% OF SALT-AND-PEPPER IMPULSE NOISE USING THE NOISE MODEL (1)

TABLE III
COMPUTATIONAL COMPLEXITY ANALYSIS OF THE AGFF COMPARED WITH OTHER TECHNIQUES

TABLE IV
EXECUTION TIME OF THE PROPOSED FILTER AGFF COMPARED WITH OTHER TECHNIQUES, WHERE THE 256 256 COLOR IMAGE LENA
CORRUPTED BY 10% IMPULSE NOISE WAS USED, AND ALL FILTERS RUN WITH A 3 3 FILTER WINDOW

TABLE V
STATISTICAL ANALYSIS OF PERFORMANCE OF THE AGFF USING VARIETY OF TEST IMAGES CORRUPTED BY
DIFFERENT LEVELS OF RANDOM IMPULSE RATIOS USING THE NOISE MODEL DEFINED IN (2)

number of all corrupted pixels in (except

), using various real-life images with different impulse noise densities. In
, as a discrimination feature, can be used to divide
Table II,
clearly a variety of color images corrupted by different impulse
noise types with different impulse noise densities into disjoint
, the salt-and-pepper noise
subsets. Therefore, according to
can be reliably discriminated from the random impulse noise for
any image corrupted by the impulse noise. The recommended
is 10 in the AGFF.
threshold ratio
A design principle for the following operations, which are
adapted to different noise ratios and types, is to use as small a
size of the window and as less a number of the passes as possible, as long as the impulse noise can be removed (to ensure
preserving image details as much as possible). The number of
passes was determined for removal of a noise region based on

the worst case scenario within the estimated maximal size of the
noise region. The operations designed for removing impulses
from different corrupted pixel sets in natural digital color images, are defined as follows.
Operation I

Operation II

Consists of a two-pass filtering to restore

color images with a low noise ratio. In the
first pass, it restores impulse corrupted pixels
in . In the second pass, it restores impulse
corrupted pixels in .
Consists of Operation I and Operator D. The
Operator D is designed to remove corrupted
pixels in . Operation II is designed to
restore color images with a medium or high
noise ratio.

1750

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 18, NO. 8, AUGUST 2009

Fig. 8. Zoomed difference between portion of the original image Lena and the output images, which are reconstructed by the AGFF with different values of its
parameter T . d) Difference between original image Lena and a), e) Difference between original image Lena and b), f) Difference between original image Lena
; b) output of AGFF with T
c) output of AGFF with T
.
and c). a) Output of AGFF with T

= 15

=5

Operation III Consists of two passes of Operation I,

Operator D and one-pass filtering to restore
pixels in . First, one-pass filtering of
pixels in
is applied, which is followed
by Operation I. The second pass includes
Operator D which is followed by another
Operation I. Operation III is designed to
restore corrupted images with a high noise
ratio.
Operation IV Consists of two passes of Operation III and
to restore
one-pass filtering of pixels in
color images with impulse of very high
noise ratio. It applies Operation III and then
restores impulse corrupted pixels in , where
for in (7). Finally, it repeats
Operation III.
The filter operation defined in the above Operations is the
modified median filter described in Section III. An adaptive progressive multipass filtering algorithm/process is introduced in
Fig. 6 in order to implement the principles described above by
the operations. Denote the salt-and-pepper impulse as SP and
the random distribution impulse as R.
Unlike other partition-based filters [35] which use the information of the local sample spread tracked by the fuzzy ranking
through its adaptive fuzzy sample similarity function, the proposed Operations of the AGFF do not use any ranking to detect

= 40

Fig. 9. Average MSE of the output images from the operations used in the
.
AGFF on different noise intensities represented by the parameter

Rs

noise corruption in color images, leading to a comparatively low

computational cost.
In the case of SP noise contamination, each Operation in Algorithm 1 is applied twice.
The recommended parameters of the algorithm may be further optimized or simplified (see Fig. 9). For example, setting
all to 3 will show little difference in the results. Actually, the
scales (levels) of the noise ratio and the operations may be simplified. For instance, if all Operation I is replaced by Operation
II (Merging Operation I and Operation II into Operation II), it
will only cause a little more artifacts for the images or even no
significant difference for many of the test images. As a result,
the adaptive filter is very robust and tolerant to the estimation

XU et al.: GEOMETRIC FEATURES-BASED FILTERING FOR SUPPRESSION OF IMPULSE NOISE

1751

TABLE VI
IMPULSE SUPPRESSION PERFORMANCE OF THE AGFF COMPARED WITH OTHER TECHNIQUES, WHERE THE TEST
COLOR IMAGE LENA WAS CORRUPTED BY DIFFERENT LEVELS OF RANDOM IMPULSE IN NOISE MODEL (2)

deviations for impulse noise ratio of corrupted images. Besides,

is actually nonsensitive to different types of images and the
noise density (see Figs. 7 and 8) and, therefore, it may be fixed
corresponding to the Operations as the algorithm defines to reduce parameters.
Excellent convergence was observed by using the AGFF.
Three or two passes of the operations may lead to a zero
response for further repetition. That means, as an alternative,
the operations in the above algorithm can be applied iteratively
until reaching the stop criterion of zero response, without
causing too many artifacts to the images (see Table VI). This
property of the operations is very useful to ensure perceptual
quality and at the same time to minimize the risk of causing
Type II error.
Although the filter implemented by the above algorithm deals
with noise ratio within 80% using the noise model defined by
(1) to ensure good perceptual image quality, a larger processing
window size can be considered with the increase of noise ratio in
color images, in order to remove all impulse noise which can be
detected by the human eyes. However, this may result in destruction of fine image details, in terms of objective image quality
measurements.
C. Computational Complexity Analysis
The AGFF only performs restoration operation on corrupted
parts of the images, except for simple comparisons to detect
corrupted impulse pixels. As a matter of fact, for most pixels
in a color image, only few subtractions and comparisons are
required to determine if a pixel is corrupted or not (see the
detection criteria defined by (5)to (9), which likely leads to
early stage reject in most cases). A framework can be used to
analyze the computational requirements of image filters [12],
[23][25], [35], [59]. The framework evaluates the computational requirements of a special filter by counting the total elwindow which
ementary operation required to process an
contains
(here, is the width of the processing window
) vector samples. Basic operations which are
and
frequently taken into account in the analysis include additions
(ADDs), multiplications (MULTs), divisions (DIVs), square root

Fig. 10. Performance of the filters in terms of NCD compared with other techniques, where the test image Parrots was corrupted by different random impulse noise ratios using noise model defined by (1).

(SQRTs), comparisons (COMPs), exponents (EXPs), and arc

cosines (ARCCOSs).
For the AGFF, in contrast with other filters, only comparison
and addition/subtraction operations are involved, and the computational complexity of the AGFF is mainly dependent on the
restoration operations. Since the restoration method is chosen
as the modified median, the computational complexity for the
modified median in the processing window is
where
. In the worst case scenario, the maximal number
and
is 16 as analyzed
of subtractions/additions, in
above, and is independent of processing window size. However, if the vector filtering is used, the maximal number of
. Table III [59] presents the
subtractions/additions, , is
elementary operations required for a number of vector filters
in FGPG [53] and FMGPF [28],
and the AGFF. In Table III,
is defined as follows:
(15)
where is the impulse noise ratio of a test image;
is 2 or
;
, are defined respectively in [28]
3;
and [53] as the probability of distances needed to declare the
central pixel as corrupted or noncorrupted if the central pixel is
are defined respectively in [28] and [53] as
noncorrupted;
the probability of distances needed to declare the central pixel
as corrupted or noncorrupted if the central pixel is corrupted.

1752

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 18, NO. 8, AUGUST 2009

TABLE VII
IMPULSE SUPPRESSION PERFORMANCE OF THE AGFF COMPARED WITH OTHER TECHNIQUES, WHERE THE TEST IMAGE PEPPERS AND
TIFFANY WERE CORRUPTED WITH DIFFERENT RANDOM IMPULSE NOISE RATIOS USING NOISE MODEL DEFINED BY (1), AND THE TEST IMAGE
PARROTS WAS CORRUPTED WITH THE SALT-AND-PEPPER NOISE OF DIFFERENT NOISE RATIOS USING NOISE MODEL DEFINED BY (1).
(A) COLOR IMAGE PEPPERS CORRUPTED BY DIFFERENT LEVELS OF RANDOM IMPULSE. (B) COLOR IMAGE PARROTS CORRUPTED BY DIFFERENT
LEVELS OF SP (SALT-AND-PEPPER) IMPULSE. (C) COLOR IMAGE TIFFANY CORRUPTED BY DIFFERENT LEVELS OF RANDOM IMPULSE

Although the AGFF uses multipass operation and the time

consumed by the filter is dependent on the noise ratio, the
computational complexity of its algorithm is actually very low.
Table IV [59] lists the execution times of a number of filters
256 RGB image Lena corrupted by
in processing the 256
10% random impulse noise using the noise model (1). The
program was written in C++ and ran on the same platform (P4
3.0 GHz/2096 MB DDR/Window XP Pro) without any code
optimization. Time consumed in the experiment for the AGFF
is a fraction of one second and much less than those by others
(see Table IV). The AGFF is potentially valuable, especially

in restoration of large size color images for various online

applications.
IV. EXPERIMENTS
The AGFF technique has been evaluated by an extensive
range of tests and its performance is compared with a number
of prior-art filtering techniques in the area of removing impulse
noise from color images. Several objective criteria are used
in the tests to measure the distortion in restored images. The
objective criteria include the Mean Square Error (MSE) and the
Mean Absolute Error (MAE) defined in the RGB color space

XU et al.: GEOMETRIC FEATURES-BASED FILTERING FOR SUPPRESSION OF IMPULSE NOISE

1753

TABLE VIII
IMPULSE SUPPRESSION PERFORMANCE OF THE AGFF COMPARED WITH BDND [70], WHERE THE TEST IMAGE LENA WAS CORRUPTED BY
THE SALT-AND-PEPPER IMPULSE WITH DIFFERENT NOISE DENSITIES USING NOISE MODEL DEFINED BY (1)

TABLE IX
IMPULSE SUPPRESSION PERFORMANCE OF AGFF-1 (VECTOR FILTERING) AND AGFF-2 (CHANNEL INDEPENDENT FILTERING) COMPARED
WITH OTHER TECHNIQUES, WHERE THE TEST COLOR IMAGE LENA WAS CORRUPTED BY DIFFERENT LEVELS OF RANDOM IMPULSE

Fig. 11. Performance of AGFF-1 (vector filtering) and AGFF-2 (channel independent filtering) using a variety of test images corrupted by 20% noise ratio with
different channel correlation factor (r ) values for each corrupted pixel.

[13], [56], [59], the Peak Signal-to-Noise Ratio (PSNR) [70]

and the Normalized Color Difference (NCD) [13], [56] which
measures the color distortion in perceptual uniform CIELUV
color space.
A. Mathematical Analysis of AGFF and its Parameters
Since it is advantageous to treat the image function as realization of a stochastic process [71], the theoretical analysis of the
AGFF is conducted using a statistical approach. In Table V, the
sample images include well-known standard test images such
as, Airplane, Boats, Flower, Girl, Goldhill, Moon, Pen, Soccer,
Tiffany, Zelda, Parrots, Peppers, Lena, and Yacht with an image
362 or 512
512 or 787 576 or 720
resolution of 500
576 or 768 512 or 1986 1986 pixels (see Fig. 3); is
the expectation value of the AGFF performance, and is the
standard deviation. The results in Table V show that the AGFF
works effectively and reliably (robustly) for a variety of real-life
images over a large range of noise ratios.
Analysis of parameters for the AGFF is presented as follows. Without loss of generality, the recommended empirical/
predetermined parameters (thresholds) of the algorithm were
obtained by experiments using a variety of test color images
based on a greedy searching method. The test images included

human faces, animals, natural scenes and artificial objects, a majority of which were selected from the still test images set (see
Fig. 3) [72].
In Fig. 7, the test images, Airplane, Girl, Boats, Soccer,
Moon, and Pens, corrupted with different random impulse noise
ratios, were restored by the AGFF with different values of its
parameter . The averages of the MSEs for these test images
for
show that AGFF performs better over a narrow range of
different noise intensity.
Fig. 8 using the image Lena, which was corrupted with 15%
random impulse noise using noise model defined by (1), restored by the AGFF using different values of the parameter ,
demonstrates the different visual quality, though the quality is
not very sensitive to the change of the parameter . In Fig. 9,
the test images, which are Airplane, Girl, Boats, Soccer, Moon,
and Pens corrupted with different random impulse noise intensity, are restored by the AGFF Operations. The averages of the
MSEs for these test images show that each operation has an optimal performance over a range of noise intensity, although in
certain ranges the differences between them are not large. The
correspond to the
ranges of noise intensity represented by
parameters of the fuzzy member function in (10)(14).
In addition, Figs. 79 demonstrate that the AGFF is fairly robust and tolerant to inaccurate estimation of the type and the

1754

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 18, NO. 8, AUGUST 2009

= 20%

Fig. 12. Reconstruction of proposed filter compared with other techniques, where the test image Lena is corrupted by random impulse with p
in noise
model defined by (1). (i), (j), (k) and (I) show the zoomed portion of the test image restored by different techniques. (a) Original image Lena; (b) 20% random
corruption; (c) DDF output; (d) HBTM output; (e) ACWMF output; (f) AGFF output; (g) AGFF_i output; (h) difference between (a) and (b); (i) difference between
(a) and (c); (j) difference between (a) and (d); (k) difference between (a) and (e); (l) difference between (a) and (f).

XU et al.: GEOMETRIC FEATURES-BASED FILTERING FOR SUPPRESSION OF IMPULSE NOISE

1755

Fig. 13. Reconstruction of proposed filter compared with other techniques, where the test image Parrots is corrupted by salt-and-pepper impulse with p = 15%
noise in noise model defined by (2). (a) Original image Parrots; (b) 15% salt-and-pepper corruption; (c) MWC output; (d) SAHVF output; (e) VMF output;
(f) AGFF output.

ratio of the impulse noise, and its performance is not very sensitive to the change of its parameters.
B. Comparison With Other Techniques
All impulse corruptions were generated according to the
noise model defined by (1) or (2), using the random impulse
noise or the salt-and-pepper noise, and a noise ratio or
varied from 0% to 80%. The evaluation of impulse suppression
was conducted using five 24-bit RGB images [72], Lena,
Peppers, Tiffany, Boats and Parrots, with image resolutions
256, 512
512, 787
576, and 1536
1024
of 256
pixels, respectively, which have been widely used by prior-art
impulse filtering techniques due to their representative color
characteristics and image structure.
Two classes of filters were used in the impulse suppression
tests. The first includes classic filters such as VMF, DDF
and AHDF [77]. The second is the state-of-the-art techniques
recently developed for the impulse and the mixed noise suppression, including AVMF [44], SAA [43], AVLUM, ACWVDF,
SCWVDF [56], ACWMF [47], SWVDF [74], MWC [78],
HBTM [45], SAHVF [41], PBTVM [59], VWMKNNF [60],
[63], VMF_FAS [44], and BDND [70]. It includes,
especially, a number of recently developed techniques such as
FMPGF, FRVP [35], SVMF [50], FTSCF [38], FIVF [25], and
SDDF [27].
Table VI presents the experimental results for suppression of
impulse noise using the noise model in (2) on a test 512 512

RGB image Lena. The impulse noise ratios in Table VI are for
the noise model in (2). Other techniques used a 3 3-filtering
window except for the ACWMF at 20% using a 5 5 window
and the MWC. AGFF_i means that AGFF is applied iteratively
until reaching the stop criterion of zero response.
Table VII presents the experimental results for suppression of
impulse noise on a 256 256 RGB test image Peppers, a 1536
1024 RGB test image Parrots and a 512
512 RGB test
image Tiffany. The performance of the most relevant state-ofthe-art techniques, which include the PBTVM, the SAHVF, the
SAA filter (with the actual noise ratio) and the SWVDF (with
) and other filters with their recommended
parameters
parameter settings, is listed in the table for a comparison. The
filtering weights of the PBTVM were trained using different
test images and noise corruption [59], while the noise ratio is
50% in Tables VII(b) and VII(c), other techniques used a 5
5-filtering window except for the MWC. While the noise ratios
are under 50% in Tables VII(b) and VII(c), other techniques
used a 3 3-filtering window except for the ACWMF at 30%
or 40% which used a 5 5-window and the MWC.
Table VIII presents the experimental results for restoration of
an extremely corrupted 512 512 RGB test image Lena, which
was corrupted by either pepper or salt impulse with unequal
probabilities. In Table VIII, the image is corrupted by the impulse noise in a channel independent manner [70] and the total
noise density of the salt-and-pepper noise for the test image is
either 70% or 80%.

1756

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 18, NO. 8, AUGUST 2009

Fig. 14. Reconstruction of proposed filter compared with other techniques, where the zoomed portion of the test image Parrots is corrupted by salt-and-pepper
impulse with p = 40% noise in noise model defined by (1). (a) Original image Parrots; (b) 40% salt-and-pepper corruption; (c) MWC output; (d) VMF output;
(e) HBTM output; (f) AGFF output.

In Fig. 10, the AGFF performs better than other recently

developed filtering techniques such as FRVP, FTCST [38],
FBTVM, SVMF [50] and ACWVDF in terms of majority of
the performance measures amongst which the FRVP filter [35]
was trained on Lena with 10% impulse noise.
Table IX presents the experimental results for a seriously corrupted 512 512 RGB test image Lena restored by the AGFF-1
using vector filtering and AGFF-2 using color channel indepen(Mandent filtering. For the calculation of vector distance,
hattan distance) was used to improve computational efficiency.
In Table IX, the noise model defined by (1) was used with the
different noise ratios, while its channel correlation factor
for each corrupted pixel, i.e., if a pixel was corrupted its RGB
components were corrupted together; when the noise ratios were
under 40%, ACWMF and HBTM used a 3 3-filtering window;
when the noise ratios was 40%, ACWMF [47] and HBTM [45]
used a 5 5-filtering window; VMF used a 3 3-filter window
and was applied twice. The AGFF-1 and AGFF-2 have shown
much better performance than other techniques under above
conditions. It can also been seen in Table IX that the AGFF-1
has demonstrated better performance than the AGFF-2 under
the selected experimental conditions. As shown by Fig. 11, the
AGFF-1 generally performs better than the AGFF-2 in terms of
the performance criteria when approaches 1.
In Fig. 11, the noise model defined by (1) was used and the
channel correlation factor
for each corrupted pixel took on
values of 0.5, 0.75, and 1 respectively; the test images, Air-

plane, Tiffany, Peppers, and Flower which were corrupted by

random impulse with a noise ratio of 20% and different , were
restored by the AGFF-1 using vector filtering and the AGFF-2
using color channel independent filtering. Fig. 11 shows that
in terms of the averaged MSE, MAE and NCD criteria using
the aforementioned test images, the AGFF-1 outperforms the
AGFF-2 when approaches 1. Therefore, vector filtering is recommended for color image restorations under such conditions.
Besides the excellent objective performance measurements,
the AGFF also achieved a consistently better performance in
perceptual image quality than other impulse filtering techniques.
Figs. 1214 demonstrate the performance of the AGFF compared with other typical and state-of-the-art techniques. A test
image Lena with 20% random impulse corruption generated by
the noise model defined in (1) was selected to reveal the detail preservation capability of the AGFF (see Fig. 12) and a
test image Parrots with 15% salt-and-pepper impulse corruption
generated by the noise model defined in (2) was selected to reveal the detail preservation capability of the AGFF (see Fig. 13).
In Fig. 12, other methods used a 3 3-filtering window, except for the ACWMF which used a 5 5 filtering window. In
Fig. 13, other methods used a 3 3-filtering window, except for
the VMF which used a 5 5 filtering window. It is observed that
the AGFF provides an almost noise free reconstruction while
the image detail structures are much better preserved than those
by other techniques. The AGFF performs well in both smooth
background areas and detail areas with sharp changes [see the

XU et al.: GEOMETRIC FEATURES-BASED FILTERING FOR SUPPRESSION OF IMPULSE NOISE

1757

Fig. 15. Reconstruction by the AGFF compared with other techniques, where the test image Fingerprint is corrupted by real, not-approximated, impulsive noise.
(a) Original image fingerprint; (b) AGFF output; (c) MWC output; (d) VMF output; (e) HBTM output; (f) AHDF output.

differences in Fig. 12(c)(f), especially in Lenas hair area].

In Fig. 14, zoomed portions of the original, filtered and corrupted images of the test image Parrot are presented for better
visual comparison, highlighting the image restored by the AGFF
and other prior-art techniques, where a salt-and-pepper impulse
noise contamination is considered. Fig. 14 shows the results of
the compared techniques which use a 5 5-filtering window
except for the MWC [78]. It can be seen that some pepper-like
impulses are still left or sharp image detail structures are blurred
in the reconstructions by a number of other prior-art techniques,
while a much better result is achieved by the AGFF which provides an almost noise-free restoration while the fine stripes and
the eye of the bird are clearly preserved [see Fig. 14(f)].
In addition to the simulated noisy images, the AGFF was also
tested with real-life images with real, not-approximated (i.e.,
not-simulated by software) noise. In terms of noise suppression
and image detail preservation, the AGFF shows much better performance than that of others to suppress real impulse noises in
practical application (see Fig. 15).
V. CONCLUSION
A geometric features-based filtering technique has been
proposed for removing impulse noise from corrupted digital
color images. The special contribution of the new filtering
technique is its novel impulse detection method, which uses
2-D geometric features (shape and edge type) and the size of the
impulse corrupted pixel/pixel region, instead of 1-D statistical
information, to identify the impulse in an effective and efficient
manner. The other novelty is its progressive adaptive restoration
mechanism, where a carefully selected set of sizes and shapes
of processing windows are employed, adapting to noise ratio
and type to recover the corrupted pixels step by step through
a reliable multipass process of low computational complexity.
This technique also provides a very reliable impulse noise type

and ratio discrimination method. Through extensive experiments conducted using a wide range of natural color images, the
proposed filtering technique has demonstrated superior performance to that of well-known benchmark techniques, in terms of
standard objective measurements, visual image quality and the
computational complexity, in removing the salt-and-pepper and
the random impulse noise which are commonly considered in
color image restoration. The technique is very useful for online
applications to suppress impulse noise especially for medium
and large sized color images. It can be further integrated with
other benchmark techniques to suppress a mixed Gaussian and
impulse noise contamination for color images to improve their
performance.
REFERENCES
[1] I. Pitas and A. N. Venetsanopoulos, Nonlinear Digital Filter: Principles and Applications. Norwell, MA: Kluwer, 1990.
[2] E. Abreu, M. Lighstone, S. K. Mitra, and K. Arakawa, A new efficient
approach for the removal of impulse noise from highly corrupted images, IEEE Trans. Image Process., vol. 5, no. 6, pp. 10121025, Jun.
1996.
[3] C. Kenney, Y. Deng, B. S. Manjunath, and G. Hewer, Peer group
image enhancement, IEEE Trans. Image Process., vol. 1, no. 2, pp.
326334, Feb. 2001.
[4] V. I. Ponomaryov, Real-time 2D-3D filtering using order statistics
based algorithms, J. Real-Time Image Process., vol. 1, no. 3, pp.
173194, 2007.
[5] X. Li, On modeling interchannel dependency for color image denoising, Int. J. Imag. Syst. Technol., vol. 1, no. 3, pp. 163173, Oct.
2007.
[6] D. Keren and A. Gotlib, Denoising Color Images using regularization
and correlation terms, J. Vis. Commun. Image Represent., vol. 9, no.
4, pp. 352365, Dec. 1998.
[7] O. Lezoray, A. Elmoataz, and S. Bougleux, Graph regularization for
color image processing, Comput. Vis. Image Understand., vol. 10, no.
12, pp. 3855, Jul.Aug. 2007.
[8] A. Elmoataz, O. Lezoray, and S. Bougleux, Nonlocal discrete regularization on weighted graphs: A framework for image and manifold
processing, IEEE Trans. Image Process., vol. 1, no. 7, pp. 10471060,
Jul. 2008.

1758

[9] R. H. Chan, C. Ho, and M. Nikolova, Salt-and-pepper noise removal

by median-type noise detectors and detail-preserving regularization,
IEEE Trans. Image Process., vol. 14, no. 10, pp. 14791485, Oct. 2005.
[10] J. Astola, P. Haavisto, and Y. Neuvo, Vector median filter, Proc.
IEEE, vol. 78, no. 4, pp. 678689, Apr. 1990.
[11] P. E. Trahanias and A. N. Venetsanopoulos, Vector direction filter: A
new class of multichannel image processing filter, IEEE Trans. Image
Process., vol. 2, no. 10, pp. 528534, Oct. 1993.
[12] D. G. Karakos and P. E. Trahanias, Combining vector median and
vector direction filters: The directional-distance filter, in Proc. IEEE
Int. Conf. Image Process., Washington, DC, Oct. 1995, vol. 1, pp.
171174.
[13] R. Lukac and K. N. Plataniotis, A taxonomy of color image filtering
and enhancement solutions, in Advances in Imaging and Electron
Physics, P. W. Hawkes, Ed. New York: Elsevier, 2006, vol. 140, pp.
187264.
[14] R. Lukac, B. Smolka, K. Martin, K. N. Plataniotis, and A. N. Venetsanopoulos, Vector filtering for color imaging, IEEE Signal Process.
Mag., vol. 22, no. 1, pp. 7486, Jan. 2005.
[15] R. Lukac, B. Smolka, K. N. Plataniotis, and A. N. Venetsanopoulos,
Selection weighted vector directional filter, Comput. Vis. Image Understand., vol. 94, no. 1-3, pp. 140167, Apr. 2004.
[16] Y. Dong and S. Xu, A new directional weighted median filter for removal of random-valued impulse noise, IEEE Signal Process. Lett.,
vol. 1, no. 3, pp. 193196, Mar. 2007.
[17] V. Chatzis and I. Pitas, Fuzzy scalar and vector median filters based
on fuzzy distances, IEEE Trans. Image Process., vol. 8, no. 5, pp.
731734, May 1999.
[18] H. H. Tsai and P. T. Yu, Genetic-based fuzzy hybrid multichannel
filters for color image restoration, Fuzzy Sets Syst., vol. 114, pp.
203224, 2000.
[19] R. Lukac, K. N. Plataniotis, B. Smolka, and A. N. Venetsanopoulos,
cDNA microarray image processing using fuzzy vector filtering
framework, J. Fuzzy Sets Syst., vol. 152, no. 1, pp. 1735, May 2005.
[20] L. Khriji and M. Gabbouj, Adaptive fuzzy order statistic-rational hybrid filters in color image processing, Fuzzy Sets Syst., vol. 128, no. 1,
pp. 3546, Mar. 2000.
[21] E. S. Hore, B. Qiu, and H. R. Wu, Improved color image vector
filtering using fuzzy noise detection, Opt. Eng., vol. 42, no. 6, pp.
16561664, Jun. 2003.
[22] S. Schulte, V. D. Witte, M. Nachtegeal, D. V. Weken, and E. E. Kerre,
Fuzzy two-step filter for impulse noise reduction from colour images,
IEEE Trans. Image Process., vol. 15, no. 11, pp. 35683579, Nov.
2006.
[23] K. Arakawa, Median filter based on fuzzy rules and its application to
image restoration, Fuzzy Sets Syst., vol. 7, no. 1, pp. 313, Jan 1996.
[24] V. F. Kravchenko, V. I. Ponomaryov, and V. I. Pustovoit, Algorithms
of the 3D filtration with using fuzzy-set theory for color image
sequences degraded by noise, Doklady Phys., vol. 53, no. 7, pp.
363367, 2008.
[25] S. Morillas, V. Gregori, G. Peris-Fajarnes, and P. Latorre, A fast impulsive noise color image filter using fuzzy metrics, Real-Time Imag.,
vol. 11, no. 56, pp. 417428, Nov.Dec. 2005.
[26] S. Morillas, V. Gregori, G. Peris-Fajarnes, and A. Sapena, New adaptive vector filter using fuzzy metrics, J. Electron. Imag., vol. 16, no.
3, Jul.Sep. 2007, paper number 033007.
[27] S. Morillas, V. Gregori, G. Peris-Fajarnes, and A. Sapena, Local
self-adaptive fuzzy filter for impulsive noise removal in color images,
Signal Process., vol. 88, no. 2, pp. 390398, Feb. 2008.
[28] J. G. Camarena, V. Gregori, S. Morillas, and A. Sapena, Fast detection
and removal of impulsive noise using Peer groups and fuzzy metrics,
J. Vis. Commun. Image Represent., vol. 19, no. 1, pp. 2029, Jan. 2008.
[29] S. Morillas, V. Gregori, and G. Peris-Fajarnes, Isolating impulsive
noise pixels in color images by peer group techniques, Comput. Vis.
Image Understand., vol. 110, no. 1, pp. 102116, Apr. 2008.
[30] Y. Nie and K. E. Barner, The fuzzy transformation and its application
in image processing, IEEE Trans. Image Process., vol. 15, no. 4, pp.
910927, Apr. 2006.
[31] S. Schulte, V. De Witte, M. Nachtegael, D. Van der Weken, and E. E.
Kerre, Fuzzy random impulse noise reduction method, Fuzzy Sets
Syst., vol. 158, no. 3, pp. 270283, Feb. 2007.
[32] S. Schulte, V. De Witte, M. Nachtegael, D. Van der Weken, and E.
E. Kerre, Histogram-based fuzzy colour filter for image restoration,
Image Vis. Comput., vol. 25, no. 9, pp. 13771390, Sep. 2007.
[33] S. Schulte, S. Morillas, V. Gregori, and E. E. Kerre, A new fuzzy
color correlated impulsive noise reduction method, IEEE Trans. Image
Process., vol. 16, no. 10, pp. 25652575, Oct. 2007.

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 18, NO. 8, AUGUST 2009

[34] S. Schulte, B. Huysmans, A. Pizurica, E. E. Kerre, and W. Philips,

A new fuzzy-based wavelet shrinkage image denoising technique,
in Proc. Advanced Concepts for Intelligent Vision Systems, 2006, vol.
1, pp. 1223.
[35] Z. Ma, H. R. Wu, and D. Feng, Fuzzy vector partition filtering technique for color image restoration, Comput. Vis. Image Understand.,
vol. 107, no. 12, pp. 2637, Jul.Aug. 2007.
[36] S. Schulte, V. De Witte, and E. E. Kerre, A fuzzy noise reduction
method for color images, IEEE Trans. Image Process., vol. 16, no. 5,
pp. 14251436, May 2007.
[37] S. Schulte, M. Nachtegael, V. De Witte, D. Van der Weken, and E. E.
Kerre, A fuzzy impulse noise detection and reduction method, IEEE
Trans. Image Process., vol. 15, no. 5, pp. 1153162, May 2006.
[38] A. Rosales-Silva, V. Ponomaryov, and F. Gallegos-Funes, Impulse
denoising using fuzzy and directional processing on a DSP, J. Lecture
Notes Adv. Artif. Intell., vol. LNAI 5317, pp. 554561, 2008.
[39] L. Lucchese and S. K. Mitra, A new class of chromatic filters for color
image processing: Theory and applications, IEEE Signal Process.
Mag., vol. 22, no. 1, pp. 7486, Jan. 2005.
[40] K. N. Plataniotis, S. Vinayagamoorthy, D. Androutsos, and A. N.
Venetsanopoulos, An adaptive nearest neighbor multichannel filter,
IEEE Trans. Circuits Syst. Video Technol., vol. 6, no. 6, pp. 699703,
Dec. 1996.
[41] Z. Ma, H. R. Wu, and B. Qiu, An structure adaptive hybrid vector
filter for the restoration of digital color images, IEEE Trans. Image
Process., vol. 14, no. 12, pp. 19902001, Dec. 2005.
[42] R. Lukac, Adaptive vector median filtering, Pattern Recognit. Lett.,
vol. 24, no. 12, pp. 18891899, Aug. 2003.
[43] B. Smolka, A. Chydzinski, K. Wojciechowski, K. N. Plataniotis, and
A. N. Venetsanopoulos, Self-adaptive algorithm for impulsive noise
reduction in color images, Pattern Recognit., vol. 35, pp. 17711784,
Aug 2002.
[44] B. Smolka, R. Lukac, A. Chydzinski, K. N. Platanioyis, and K. Wojciechowski, Fast adaptive similarity based impulse noise reduction
filter, Real-Time Imag., vol. 9, no. 4, pp. 261276, Aug. 2003.
[45] Z. Ma and H. R. Wu, A histogram based adaptive vector filter for color
image restoration, in Proc. IEEE ICICS-PCM, Singapore, Dec. 1518,
2003, pp. 1A3.4.11A3.4.5.
[46] R. S. Lin and Y. C. Hsueh, Multichannel filtring by gradient information, Signal Process., vol. 80, no. 2, pp. 279293, Feb. 2000.
[47] T. Chen and H. Wu, Adaptive impulse detection using center-weighted
median filter, Signal Process. Lett., vol. 8, no. 1, pp. 13, Jan. 2001.
[48] 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.
[49] K. N. Plataniotis, D. Androutsos, and A. N. Venetsanopoulos,
Multichannel filters for image processing, Signal Process.: Image
Commun., vol. 9, no. 2, pp. 143158, Jan. 1997.
[50] L. Jin and D. Li, A switching vector median filter based on the
CIELAB color space for color image restoration, Signal Process.,
vol. 87, no. 6, pp. 13451354, Jun. 2007.
[51] S. Q. Yuan and Y. H. Tan, The solutions of equation-based noise detector for an adaptive median filter, Pattern Recognit., vol. 39, no. 11,
pp. 22522257, Nov. 2006.
[52] S. Q. Yuan and Y. H. Tan, Impulse noise removal by a global-local
noise detector and adaptive median filter, Signal Process., vol. 86, no.
8, pp. 21232128, Aug. 2006.
[53] B. Smolka and A. Chydzinski, Fast detection and impulsive noise removal in color images, Real-Time Imag., vol. 11, no. 56, pp. 389402,
Nov.Dec. 2005.
[54] R. Lukac and S. Marchevsky, Adaptive vector LUM smoother,
in Proc. IEEE int. Conf. Image Processing, Oct. 2001, vol. 1, pp.
878881.
[55] B. Smolka, M. Szczepanski, K. N. Plataniotis, and A. N. Venetsanopoulos, On the modified weighted vector median filter, in Proc.
IEEE Int. Conf. Digital Signal Processing, Santorini, Greece, Jul.
2002, vol. 1, pp. 939942.
[56] R. Lukac, Adaptive colour image filtering based on center weighted
vector direction filters, Multidimens. Syst. Signal Process, vol. 15, no.
2, pp. 169196, Apr. 2004.
[57] M. E. Celebi, H. A. Kingravi, and Y. A. Aslandogan, Nonlinear vector
filtering for impulsive noise removal from color images, J. Electron.
Imag., vol. 16, no. 3, Jul.Sep. 2007.
[58] , K. N. Plataniotis and A. N. Venetsanopoulos, Eds., Color Image Processing and Applications. Berlin, Germany: Springer, 2000.
[59] Z. Ma, H. R. Wu, and D. Feng, Partition based vector filtering technique for suppression of noise in digital colour images, IEEE Trans.
Image Process., vol. 15, no. 8, pp. 23242342, Aug. 200.

XU et al.: GEOMETRIC FEATURES-BASED FILTERING FOR SUPPRESSION OF IMPULSE NOISE

[60] V. I. Ponomaryov, F. J. Gallegos-Funes, and A. Rosales-Silva, Realtime color imaging based on RM-filters for the impulsive noise reduction, J. Imag. Sci. Technol., vol. 49, no. 3, pp. 205219, May/Jun.
2005.
[61] R. Lukac, K. N. Plataniotis, and A. N. Venetsanopoulos, Color image
denoising using evolutionary computation, Int. J. Imag. Syst. Technol.,
vol. 15, no. 5, pp. 236251, Oct. 2005.
[62] R. Lukac, K. N. Plataniotis, B. Smolka, and A. N. Venetsanopoulos,
Generalized selection weighted vector filters, EURASIP J. Appl.
Signal Process., vol. 2004, no. 12, pp. 18701885, Sep. 2004.
[63] J. Camacho, S. Morillas, and P. Latorre, Efficient impulsive noise suppression based on statistical confidence limits, J. Imag. Sci. Technol.,
vol. 50, no. 5, pp. 427436, Sep./Oct. 2006.
[64] M. Sonka, V. Hlavac, and R. Boyle, Image processing, Analysis, and
Machine Vision. New York: Thomson, 2008.
[65] W. K. Pratt, Digital Image Processing. New York: Wiley, 2001.
[66] R. C. Gonzalez and R. E. Woods, Digital Image Processing, 2nd ed.
Englewood Cliffs, NJ: Prentice-Hall, 2002.
[67] R. Lukac, K. N. Plataniotis, A. N. Venetsanopoulos, and B. Smolka,
A statistically-switched adaptive vector median filter, J. Intell. Robot.
Syst., vol. 42, no. 4, pp. 361391, Apr. 2005.
[68] R. Lukac, B. Smolka, K. N. Plataniotis, and A. N. Venetsanopoulos,
Vector sigma filters for noise detection and removal in color images,
J. Vis. Commun. Image Represent., vol. 17, no. 1, pp. 126, Feb. 2006.
[69] D. Dang and W. Luo, Impulse noise removal utilizing second-order
difference analysis, Signal Process., vol. 87, 2007.
[70] P. Ng and K. K. Ma, A switching median filter with boundary discriminative noise detection for extremely corrupted images, IEEE Trans.
Image Process., vol. 15, no. 6, pp. 15061516, Jun. 2006.
[71] H. H. Barrett and K. J. Myers, Foundation of Image Science, ser. Pure
and Applied Optics. Hoboken, NJ: Wiley, 2004.
[72] [Online]. Available: http://www.hlevkin.com
[73] A. C. Bovic, Streaking in median filtered images, IEEE Trans.
Acoustic., Speech, Signal Process., vol. 35, pp. 493503, Oct. 1985.
[74] R. Lukac, B. Smolka, K. N. Plataniotis, and A. N. Venetsanopoulos,
Selection weighted vector directional filter, Comput. Vis. Image Understand., vol. 94, no. 13, pp. 140167, Apr. 2004.
[75] B. Smolka, K. N. Plataniotis, and A. N. Venetsanopoulos, Nonlinear
techniques for color image processing, in Nonlinear Signal and
Image Provessing: Theory, Methods, and Applications. Boca Raton,
FL: CRC, Dec. 2003, pp. 445505.
[76] M. Barni and V. Cappellini, On the computational complexity of multivariate median filters, Signal Process., vol. 71, no. 1, pp. 4554, Nov.
1998.
[77] M. Gabbouj and F. A. Cheikh, Vector median-vector directional hybrid filter for color image restoration, in Proc. EUSIPCO, Trieste,
Italy, Sep. 1013, 1996, vol. 2, pp. 879881.
[78] E. S. Hore, B. Qiu, and H. R. Wu, Prediction based image restoration
using a multiple window configuration, Opt. Eng., vol. 41, no. , pp.
18551865, Aug. 2002.
Zhengya Xu (M06) received the M.S. degree in
information technology from the Vrije Universiteit
Brussel, Brussels, Belgium, in 1994, and the Ph.D.
degree from IRIS (Control and Automation), Swinburne University of Technology, Victoria, Australia,
in 2000.
He was a Research Engineer with the Aeronautics Research Institute, Ministry of Aviation and
Aerospace Industry, Beijing, China, an Associate
Researcher with Swinburn Technology University, a
Senior Software Engineer with Digital Image, Inc.,
and a Senior Research Fellow with Monash University, Australia. He works
currently as a Senior Research Fellow with Computer and Network Engineering, School of Electrical and Computer Engineering, RMIT University,
Australia. His research interests include image and video processing, computer
vision, biometric pattern recognition, object tracking, software and embedded
systems development.

1759

Hong Ren Wu received the B.Eng. and M.Eng. from

the University of Science and Technology, Beijing
(formerly Beijing University of Iron and Steel
Technology), China, in 1982 and 1985, respectively,
and the Ph.D. degree in electrical and computer
engineering from the University of Wollongong,
N.S.W. Australia, in 1990.
He worked on the academic staff of the Chisholm
Institute of Technology and then Monash University,
Melbourne, Australia, from April 1990 to January
2005, last as an Associate Professor in digital systems. He has been with the Royal Melbourne Institute of Technology, Australia,
since February 2005, as a Professor of Visual Communications Engineering
and Discipline Head, Computer and Network Engineering, School of Electrical
and Computer Engineering. His research interests include fast DSP algorithms,
digital picture compression and quality assessment, video processing and
enhancement, embedded DSP systems, and their industrial applications. He is
a co-editor of the book Digital Video Image Quality and Perceptual Coding
(CRC, 2006).

Bin Qiu (M93SM00) received the B.Eng degree

from Beijing Jiaotong University in 1982 and the
M.Sc. and Ph.D. degrees from University of Manchester, Institute of Science and Technology, U.K.,
in 1987 and 1989, respectively.
He was a Lecturer at Victoria University, Australia,
from 1990 to 1995. He joined Monash University,
Australia, in 1995 as a Senior Lecturer, where he is
currently an Associate Professor. Since 1987, he has
participated and headed research projects in the fields
of digital and computer communication systems and
networks, signal processing, image processing, and the applications of neurofuzzy techniques in the above areas. He is also interested in the design and implementation of digital systems.

Xinghuo Yu (M91SM98F08) received the

B.Eng. and M.Eng degrees from the University of
Science and Technology of China, Hefei, China, in
1982 and 1984, respectively, and the Ph.D. degree
from South-East University, Nanjing, China, in
1988, respectively.
He is now with RMIT University, Australia,
where he is the Director of the RMIT Platform
Technologies Research Institute and Professor of
information systems engineering. His research interests include variable structure and nonlinear control,
signal processing, complex, and intelligent systems. He has published over 300
refereed papers in technical journals, books, and conference proceedings, as
well as co-edited ten research books.
Prof. Yu served as an Associate Editor of the IEEE TRANSACTIONS ON
CIRCUITS AND SYSTEMS PART I (20012004) and the IEEE TRANSACTIONS
ON INDUSTRIAL INFORMATICS (20052008), and is serving as an Associate
Editor of IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS and several
other scholarly journals. He has been on the program committees of many
international conferences and co-chaired several international conferences.
He was the sole recipient of the 1995 Central Queensland University Vice
Chancellors Award for Research. He is a Fellow of the Institution of Engineers
Australia and was made Emeritus Professor of Central Queensland University
Australia in 2002 for his long-term significant contributions.