A NEW METHOD FOR NOISE REDUCTION BASED ON MODE FILTER ALGORITHM

Abstract

When an image is acquired by a camera or other imaging system, often the vision system for which it is intended is unable to use it directly. The image may be corrupted by random variations in intensity, variations in illumination, or poor contrast that must be dealt with in the early stages of vision processing. The variations in intensity values are called "noise". There are many well-established filtering techniques for image enhancement aimed at eliminating these undesirable characteristics. In this study, a new filter is designed for noise removal. A special computer program is developed to demonstrate the relative performance of the well-known filters and the newly designed one. The new filter is found to be quite successful.

Full text

Page 1

IJCI Proceedings of Intl. XII. Turkish Symposium on Artificial Intelligence and Neural Networks, Vol.1, No.1, July 2003


327
A NEW METHOD FOR NOISE REDUCTION BASED ON MODE FILTER ALGORITHM

Özgür Uysal1, Mehmet ?evkli2, Ali Türkyılmaz3, M.Sabih Aksoy4

1 2 3 Fatih University, Dep’t of Industrial Engineering, 34500 Buyukcekmece, Istanbul, Turkey
Tel: +90 212 889 08 10, Fax: +90 212 889 08 32
e-mail: ouysal@fatih.edu.tr

4 Sakarya University, Dep’t of Industrial Engineering, Sakarya, Adapazari, Turkey

Keywords: Image Processing, Noise Removal, Computer Vision, Filter

ABSTRACT
When an image is acquired by a camera or other imaging
system, often the vision system for which it is intended is unable
to use it directly. The image may be corrupted by random
variations in intensity, variations in illumination, or poor
contrast that must be dealt with in the early stages of vision
processing. The variations in intensity values are called "noise".
There are many well-established filtering techniques for image
enhancement aimed at eliminating
characteristics. In this study, a new filter is designed for noise
removal. A special computer program is developed to
demonstrate the relative performance of the well-known filters
and the newly designed one. The new filter is found to be quite
successful.
these undesirable
1
INTRODUCTION

Digital Image Processing 1.1
Digital image processing is a fascinating and challenging field.
It is fascinating because of the possibilities it offers, and it is
challenging because, in the final analysis, it is often trying to
"understand" and "improve" a digital image, which is just a
large array of numbers, and this has proven to be no simple task.
In addition, digital image processing is a rapidly growing field.
This is due to recent advances in desktop and personal
computers. Useful functions can now be performed in relatively
small and inexpensive computer systems in offices, labs, and
manufacturing locations. [1]
1.1.1
Digital Image
A digital image is an image in digital format and is obtained by
partitioning the area of the image into a finite two dimensional
array of small, uniformly shaped, mutually exclusive regions
called "resolution cells" and assigning a representative image
value to each such spatial region.
1.1.2
Pixel
A pixel (picture element or pel) is a pair whose first member is a
resolution cell or (row, column) spatial position and whose
second member is the intensity value or vector of image values
associated with the spatial position. The brightness or "image
intensity" at a pixel is usually represented by an integer between
0 and 255 (0 = black, 255 = white) [4].
There are, broadly speaking, three distinct classes of operations
in digital image processing: point operations, neighborhood
operations, and geometric operations.
2
NOISE REDUCTION IN IMAGE PROCESSING

Introduction 2.1
Images are often corrupted by random variations in intensity
values, called "noise". Some common types of noise are salt and
pepper noise, impulse noise and Gaussian noise. Salt and pepper
noise contains random occurrences of both black and white
intensity values. However, impulse noise contains only random
occurrences of white intensity values. Unlike these, Gaussian
noise contains variations in intensity that are drawn from a
Gaussian or normal distribution and is a very good model for
many kinds of sensor noise, such as the noise due to camera
electronics [8].
2.2
Filters
In this family of operations, the gray level value of a pixel is
transformed by a rule which takes into account the
neighborhood of the pixel. Typically, in a two-dimensional
image the gray level of the center pixel of an nxn matrix of
pixels will be replaced by reference to the gray levels of nxn-1
neighbors as well as to its own original value.
2.2.1
Smoothing Filters
One of the simplest picture-enhancement techniques is that of
noise "removal" or smoothing. Noise removal is important when
resolution and recording limitations are severe, as is the case in
astronomy, and when images are historically important [2].
Usually, in developing a picture matrix description of a scene,
some noise will be picked up; causing various elements of the
matrix to deviate from their correct value. If the noise is random,
such that noise in adjacent elements of the picture matrix is
uncorrelated, then spatial averaging or smoothing technique may
be applied to reduce it [7].
The most popular smoothing filters are: Mean Filter, Square
Mean Filter, Mode Filter, and Median Filter.
2.2.1.1
Mean Filter
The size and shape of the window over which the mean is
computed can be selected. Mean filter does not depend on

Page 2

IJCI Proceedings of Intl. XII. Turkish Symposium on Artificial Intelligence and Neural Networks, Vol.1, No.1, July 2003


328
22
)()(
1
YSXT
−+−
=
identification and removal of the noise. Instead, this filter
weakens the noise by applying some type of averaging to the
picture.
Since averaging blurs the picture, this filter must be applied with
care, to avoid degrading sharp detail such as lines or edges.
2.2.1.2
Median Filter
The median filter was introduced in 1974 by Tukey [14], who
used the moving median as a smoothing technique in time series
analysis. Rabiner [13] used the median filter to smooth speech
waveforms in a qualitative study; and reported favorable results.
Median filtering has also come into recent use for enhancing
images. Its application to image processing was first published
by Pratt [11] and Frieden [5]. Median filter is computationally
efficient; suppresses impulse noise [6]. Pratt made a rather
qualitative study of two-dimensional median filters of various
sizes and shapes; and examined application of one-dimensional
medians to a picture corrupted by impulsive noise [12]. He
concludes that the median filter is extremely useful for
suppressing impulsive and "salt-and-pepper" noise [3].
In particular, the median filter selects the medium value. Since
binary noise completely changes the gray value, it is very
unlikely that it will show the medium gray value in the
neighborhood. In this way, the medium gray value of the
neighborhood is used to restore the gray- value of the distorted
pixel [9]. It is one of the better edge preserving smoothing filters
[10].
2.2.1.3
Mode Filter
Having considered the mean and the median of the local
intensity distribution as candidate intensity values for noise
smoothing filters, it also seems relevant to consider the mode of
the distribution. Indeed, it might be imagined that this is if
anything more important than the mean or the median, since the
mode represents the most probable value of any distribution.
3
A NEW NONLINEAR FILTER FOR NOISE
REDUCTION: NOK-1 (NOISE KILLER-1)

Introduction 3.1
This chapter presents a new nonlinear filter for noise reduction
called NOK- 1 (NOise Killer-1). In fact, NOK-1 is obtained by
modifying mode filter algorithm. However, this modification is
not a simple one. Two new unique features are integrated on
mode filter and the result is NOK-1.
3.2
Filter Description
In mode filter, pixels are replaced by their most common
neighbor in the filter mask. Mode filter has important features as
follows:
i)
Averaging labels makes no sense, but mode filter
may clean up isolated points.
ii)
iii)
It enhances images by sharpening edges.
It can remove noise most effectively.
In spite of the fact that all of these features are important
compared to other filters, mode filter degrades the images. The
first feature integrated in mode filter to obtain NOK-1 decreases
degrading in image depends on “sensitivity level” given by the
user. Sensitivity level can be in the range 0-999.
In NOK-l, if this biggest frequency is greater than the be mask
sensitivity level given by the user, central pixel (x, y) in the
mask is replaced by a pixel which has the biggest frequency
among its neighbors, otherwise there will be no change in color
of the changing this sensitivity level.

Figure 3-1 Sensitivity Level Box
The second feature of NOK-1 that is different from Mode Filter
is the inverse distance coefficient matrix. It is used instead of
one count of a pixel in frequency calculation, and changes
according to distance from the center of the mask and calculated
for a (T, S) pixel as shown in Figure 3.2.
( T , S )
(X , Y )

Figure 3-2 Calculation of inverse distance coefficient matrix.
Increase in frequency of the type of pixel (T,S)

In this case, if there are two groups of pixels in the neighbors,
and each group has equal number of elements in it, the
frequency of a group which has elements near the center will be
higher than the others, because of distance dependent coefficient
matrix used in calculating frequency of one pixel type.
In short, a pixels supplement to its type of total frequency is
directly proportional with the inverse of its distance from the
center of mask. By using this second feature, the importance of
pixels near the center of mask increases. This feature also
decreases degrading in detail of the image as in the first feature
of NOK-1.

Page 3

IJCI Proceedings of Intl. XII. Turkish Symposium on Artificial Intelligence and Neural Networks, Vol.1, No.1, July 2003


329
3.3
Results and Discussion
According to three performance criteria, NOK-1 is tested
together with other filters by using NOD. In this section, NOK-1
is discussed according to each performance measure.
3.3.1
Process Time
Runtimes of different filters on a sample image are given in
Table 3.2. Compared to other filters, the score of NOK-1 is
above average.
3.3.2
Degrading in Image
Smean filter has a better performance compared to NOK-1. But,
only smean uses 3X3 mask in this test while the others use 5X5
mask. Because of this feature smean filter takes very bad score
in noise level performance measure.
3.3.3
Percentage of Reduction in Noise Level
Mode and median filters show a better performance in this field
compared to NOK-1’s. However, the performance of NOK-1 is
average and can reduce noise to a sufficient level.
3.3.4
Process Time
On a Intel Pentium Celeron 333 Mhz. processor, 64 Mb Ram
PC, the timing results performed on the image baby.bmp are as
follows:
Mode (3x3) 19 sec. Mode (5x5) 40 sec
NOK-1 (3x3) 20 sec NOK-1 (5x5) 42 sec.
Mean (3x3) 32 sec Mean (5x5) 79 sec
Median (3x3) 38 sec. Median (5x5) 104 sec.
Table 3-1 Runtime Performance of Different Filters
3.4
Summary
As a conclusion, NOK-1 is an average filter for all three criteria.
If a user wants an adequate overall performance from a noise
filter, and enjoys the good sides of mode filter, and doesn't want
its degrading effect, NOK-1 is a good choice.
4
IMPLEMENTATION

Nod 1.0 Software 4.1
Noise reduction process is applied to several images using both
the well-known filters and the newly developed one.
A special software called NOD (Noise Destroyer) is developed
for this purpose. In this program, first an image file is opened.
Then, an arbitrary amount of noise is added onto this image.
100, 500, 2000 and 5000 pixels of random noise can be added at
one time. This noisy image is used as an input for noise
reduction process.
The type and size of the filter (e.g. Mode Filter as an 3x3
operator) is defined using the menu commands. Finally, noise
reduction process is started. Thus, an output image is obtained.
Several types and sizes of filters can be used respectively for a
given input image and an output image is obtained for each of
them. These can be utilized for a comparative study of different
filters.
REFERENCES
[1] Aksoy M.S. (1995). Industrial Visual Inspection, An
Overview. Sakarya University Engineering Faculty, Department
of Industrial Engineering,
[2] Auyeung Ch. and Mersereau R.M. (1991). A Dual Approach
to Signal Restoration. Springer Series in Information Sciences,
Vo1.23, A.K.Katsaggelos (ed.): Digital Image Restoration,
Springer-Verlag Berlin
[3] Bovik A.C., Huang T.S. and Munson D.C. (1983). A
Generalization of Median Filtering Using Linear Combinations
of Order Statistics. IEEE Transactions on Acoustics, Speech,
and Signal Processing, Vol. ASSP-31, No. 6, pp 1342-1350.
[4] Frieden B.R. (1976). A New Restoring Algorithm for the
Preferential Enhancement of Edge Gradients. J. Opt. Soc.
Amer., Vol.66,
[5] Huang T.S., Yang G.J. and Tang G.Y. (1979). Fast Two-
Dimentional Median Filtering Algorithm. IEEE Trans. Acoust.,
Speech, Signal Process., Vol. ASSP-27, pp. 13-18, Feb.
[6] Jackson P.C. (1974). lntroduction to Artificial lntelligence.
Mason/Charter Publishers
[7] Jain R., Kasturi R. and Schunck B.G. (1995). Machine
Vision. McGraw-Hill
[8] Jahne B. (1993). Digital Image Processing: Concepts,
Algorithms and Scientific Aplications. Springer-Verlag Berlin
Heidelberg.
[9] Niblack W. (1986). An Introduction to Digital Image
Processing. Prentice Hall
[10] Pratt W.K. (1975). Median Filtering. Semianual Report,
Image Proc. Institute, Univ. of Southern California, pp. 116-123,
Sept.
[11] Pratt W.K. (1978). Digital Image Processing. Wiley, New
York,
[12] Rabiner L.R., Sambur M.R., and Schmidt C.E. (1975).
Applications of a Nonlinear Smoothing Algorithm for Speech
Processing. IEEE Trans. Acoust., Speech, Signal Processing,
Vol. ASSP-23,
[13] Tukey J.W. (1974). Nonlinear Methods For Smoothing
Data. Conf. Rec., Eascon, p. 673.
Technical Report.
Heidelberg.
pp. 280-283.
Inc., New York.
Inc.
International (UK) Ltd.
pp. 552-557.