A NEW APPROACH TO IMPULSE NOISE REMOVAL FOR COLOR IMAGE
Qing Xu, Rui Zhang
Tianjin University
School of Computer Science and Technology
China
Mateu Sbert
Universitat de Girona
Institut d’Informatica i Aplicacions
Spain
ABSTRACT
This paper presents a new two-stage approach to impulse noise
removal for color image. The first step is the impulse detec-
tion, in which the so-called Rank-Ordered Absolute Differ-
ences (ROAD) is used to determine whether a pixel is im-
pulsive to be removed. The second step does noise suppres-
sion based on maximizing the proposed pixel similarity mea-
sure that considers both magnitude and angular information
of pixel vectors’ difference. Experimental results show that
our method can remove impulse noise effectively, and in the
meantime can preserve chromaticity and image details very
well.
1. INTRODUCTION
Color images are often corrupted by impulse noise during ac-
quisition and transmission. Vector Median Filter (VMF) [1]
provides efficient noise attenuation against impulses. How-
ever, it treats all pixels of a color image in the same way
and tends to modify pixels that are not corrupted by noise,
which leads to blurring of details and change of signal struc-
ture. To avoid the damage of uncorrupted pixels, two-step
filtering methods (e.g.[2]) can be used. The idea is based on
impulse noise detection and noise removal.
We present a new two-step method for impulse noise re-
moval. The first stage uses the Rank-Ordered Absolute Dif-
ferences (ROAD), which is proposed by Roman Garnett et
al. [3] to do impulse detection. In the second stage, a simi-
larity function using the magnitude and angular information
of pixel vectors’ difference is advanced and the pixel which
maximizes this similarity function is the output in a filter win-
dow. Experimental results show that our new two-step algo-
rithm achieves good performance.
The remainder of this paper is organized as follows: Sec-
tion 2 introduces ROAD. Section 3 describes our new ap-
proach. Section 4 shows experimental results. Section 5 gives
conclusions and our future work.
2. RANK-ORDERED ABSOLUTE DIFFERENCES
Let
i =(i
1
,i
2
) be the location of a pixel under consideration,
and let
Ω
i
(n)= {
i +(g,h): -n ≤ g,h ≤ n} (1)
be the set of points in a (2n + 1) × (2n + 1) filter window
centered at
i for some positive integer n.
Ω
0
i
=Ω
i
(n) -{
i} (2)
represents the set of points in a (2n +1) × (2n +1) deleted the
center point at position
i. For each point
j ∈ Ω
0
i
, define d
(
i,
j)
as the absolute difference in vector value of pixels between
position
i and
j ,
d
(
i,
j)
= ‖x
i
- x
j
‖
2
(3)
where x
i
is the vector value of pixel at position
i. Sort the
d
(
i,
j)
values in an ascending order and define
ROAD
m
(
i)=
m
k=1
r
k
(
i) (4)
where 2 ≤ m ≤ (2n + 1) × (2n + 1) - 2, r
k
(
i) is kth
smallest d
(
i,
j)
for
j ∈ Ω
0
i
. Experiments show that n=1 and
m=4 will gain good performance, n=2 and m=12 will have a
good result.
The ROAD statistic provides a measure of how close a
pixel value is to its most similar neighbors. The logic under-
lying the statistic is that unwanted impulses will vary greatly
from most or all of their neighboring pixels, so an impul-
sive pixel will result in a large ROAD value. However, pixels
composing the actual image should have at least half of their
neighboring pixels of similar value. Even if a pixel is on the
edge, it will have neighbors with similar value and result in a
significantly low ROAD value.
3. OUR NEW APPROACH
3.1. Impulse Noise Detection
In this stage, we calculate ROAD value for each pixel and
compare it with a threshold T. If the ROAD value of a pixel is
1667 1-4244-1017-7/07/$25.00 ©2007 IEEE ICME 2007