A Survey on Image Enhancement Techniques:

Classical Spatial Filter,

Neural Network, Cellular Neural Network, and Fuzzy
Filter
Dr. D. H. Rao, Senior Member, IEEE
Principal, KLS Gogte Institute of Technology, Udyambag, Belgaum, Karnataka (INDIA)
and Patavardhan Prashant Panduranga, Student Member, IEEE
Research Scholar, Research Centre, Department of Electronics and Communication Engineering,
KLS Gogte Institute of Technology, Udyambag, Belgaum, Karnataka (INDIA)
E-mail: prashantgemini(,vahoo.coin / pat itedu, d.h.rao itedu

Abstract- Present day applications require various kinds of noise is completely characterized by its variance, and it is
images and pictures as sources of information for interpretation either superimposed or added to the image.
and analysis. Whenever an image is converted from one form to
Hence, it is a great challenge to develop algorithms
another, such as, digitizing, scanning, transmitting, storing, etc.,
that can remove noise from an image without disturbing its
some form of degradation occurs at the output. Hence, the output
content. The neighborhood averaging and smoothing by

image has to undergo a process called image enhancement which

consists of a collection of techniques that seek to improve the visual
appearance of an image. In this paper, a classical spatial filter,
neural network (NN), cellular neural network (CNN) and fuzzy
filters are presented for the noise reduction of images that are
corrupted with additive noise. A three layer neural network is
trained for few test images and is used to filter the corrupted
colour images. A single layer CNN is developed to reduce the noise
in the colour image and compared with that of the classical spatial
filter. A new fuzzy filter technique is studied with respect to noisy
gray scale images. All the techniques produce convincing results
when applied to additive (Gaussian) noisy images. Experimental
results are obtained based on the mathematical models of expert
systems and compared by numerical measures and visual
inspection. It is envisaged to train CNN using gradient descent
back-propagation algorithm for better results and extend fuzzy
filter technique to reduce noise in colour images.

Keywords: Classical Spatial Filter, Neural Network, Gradient

Descent Back-propagation Algorithm, Cellular Neural
network, Fuzzy Filter
I. INTRODUCTION
Images can be contaminated with different types of
noise, for different reasons. For example, noise can occur
because of the circumstances of recording (electronic noise in a
cameras, dust in front of the lens), because of the
circumstances of transmission (damaged data) or because of
storage, copying, scanning, etc. Impulse noise (e.g. salt and
pepper noise) and additive noise (e.g. Gaussian noise) are the
most commonly found. Impulse noise is characterized by the
fact that the pixels in an image either remain unchanged or get
one of the two specific values 0 (black) and 1 (white); an
important parameter is the noise density which expresses the
fraction of the image pixel that are contaminated. Gaussian

1-4244-0726-5/06/$20.OO '2006 IEEE

2821

ima eavraging

re

The classical

iag procesing

techniques for noise removal. The classical spatial filtering

refers to the aggregate of pixels composing an image and

operating directly on these pixels.
The NN techniques have shown to be a powerful in
formation processing tools, especially for those tasks that are
difficult to be solved, such as: pattern recognition, image
analysis, etc. [4]. These can be perfectly adapted for image
enhancement.

The CNNs proposed by Chua et.al [5] [6] are

by
6 are
aea[5
hea CNNs poste
systems
a
large
of
array
of
nonlinear
dynamic
processing units (cells) with each unit connected only to its
neighboring units in a neighborhood. The 2-dimensional
discrete time (DT) CNNs have been shown to be very
efficient in image processing and pattern recognition
applications [6], such as, shadow detection, edge detection,
feature extraction, hole filling, etc.
A very important added value of fuzzy set theory is its
ability to model and to reason with uncertainty. When
processing an image for noise reduction uncertainty occurs.
Fuzzy set theory and fuzzy logic allow to model and to work
with this uncertainty, and to improve the quality of noise
reduction.
This paper is divided into seven sections. In section II,
the paper described the basis for classical spatial filters.
Section III describes hoe noise reduction can be achieved by
training conventional NN. The DTCNN architecture for
colour image enhancement is described in section IV. In
section V a new fuzzy filter method is explained. Section VI
is devoted to compare the performance of different image
enhancement techniques and discussed with numerical
measures.

consisihng

II. CLASSICAL SPATIAL FILTERS

The principle objective of enhancement techniques is
to process an image so that the result is more suitable than the
original image for specific applications. Spatial domain
methods are procedures that operate directly on these pixels.
The principal approach to defining a neighborhood about (x, y)
(mask) is to use a square or rectangular subimage area centered
at (x, y). A spatial filtered output image g(x, y) can be formed
by discrete convolution of an input image f(x, y) with LxL
impulse response array h(x, y) according to the relation,

g(x, y) =

f(x, y)h(m + x + C, n+ y + C)

the average. In order to perform median filtering in a

neighborhood of a pixel, first sort the values of the pixel and its
neighbors, determine the median, and assign this value to the
pixel.

(1)

where C =(L +1)/2. The equation utilizes the centered

convolution, whereby the input and output arrays are centered
with respect to one another, with the outer boundary of g(x, y)
of width (L - 1)/2 pixels set to zero. Spatial filters are
broadly classified as linear and nonlinear filters and both
operate on neighborhood.

Averaging or Lowpass Spatial Filtering

This is a type of linear spatial filter. For noise
reduction, h should be of lowpass form, with all positive
elements. Several common pixel impulse response arrays of
lowpass form are listed as,

hl= -{I I 1

(2)

h3=-_10 12

(1)

III. NEURAL NETWORK

The classical noise reduction spatial filters have two
main disadvantages. First, they treat all the pixels in the
same way. This is not desirable, because not all the pixels
will be contaminated with noise in the same way. Secondly,
one should try to find an adaptive way to replace a pixel
value, taking into account characteristics of the
neighborhood ofthe pixel.
The use of NN and fuzzy technique offers a solution.
Neural network is a group of elementary processes with
strong interconnections. Based on the learning algorithm of
error back-propagation, NN can be perfectly adapted for
image enhancement. A self organizing three layered feed
forward NN is employed for image enhancement. In this
architecture, the number of neurons in each layer is equal to
number of pixels with 3 x3 neighborhood as shown in Fig. 1.

h=11 2 1(3

h3=-

1I

16

2 4 2
1

Figure 1. Neural network architecture

(4)

Input Layer

These arrays, called noise reduction masks [1], are normalized

to unit weights so that the noise reduction process does not
introduce an amplitude bias in the processed image.

Hidden Layer

Weights

Weights

OutputLayer

Median Filter
Inputs
The linear processing techniques perform reasonably
well on images with continuous noise, such as additive
Gaussian distributed noise and they tend to provide too much
soothing for impulse like noise [2]. Nonlinear techniques often
provide a better trade-off between noise smoothing and
retention of fine image detail.
Lowpass spatial filtering of the smoothing method
Figure 2. Feed forward NN structure
blurs edges and other sharp details. As the objective is to
achieve noise reduction rather than blurring, an alternative Each neuron in a layer is connected to the corresponding
median filter is developed by Tukey [1] for noise suppression. neuron in the previous layer and its neighbors via weights. The
That is the gray level of each pixel is replaced by the median of algorithm involves a high computational complexity, since all
the gray levels in a neighborhood of that pixel, instead of by the weights in the network are adapted in every iteration during

2822

self-organization. For training the NN gradient descent based

supervised learning is used, in which the change in weight is
directly proportional to the error signal and the input.

IV. CELLULAR NEURAL NETWORK

CNNs are nonlinear dynamic systems [5] [6], which
are well suited for image processing and pattern recognition.
They consist of a large array of identical processing units with
each cell connected only to its nearest neighbors in a
neighborhood. These adjacent cells can interact directly with
each other. Cells not directly connected may affect each other
indirectly, because of propagation effects of the dynamics of
the network.
Cellular neural networks were introduced in 1988 [5].
The most general definition of such networks is that they are
arrays of identical dynamical systems, the cells, that are only
locally connected [6]. In the original Chua and Yang model,
each cell is a one-dimensional dynamic system. It is the basic
unit of a CNN. Any cell is connected only to its neighbor cells,
i.e. adjacent cells interact directly with each other. Cells not in
the immediate neighborhood have indirect effect because of the
propagation effects of the dynamics in the network. The cell
located in the position (i j) of a two-dimensional MxN array is
denoted by Cij, and its r-neighborhood NVj> is defined by
N>ij = {Ckl max {lk-il,IljI} < r; 1< k < M, 1< 1 < N } (5)
where the size of the neighborhood r is a positive integer
number.
Each cell has a state x, a constant external input u, and
an output y. The equivalent block diagram of a continuous time
cell is shown in Fig. 3. The first-order non-linear differential
equation defining the dynamics of a cellular neural network
can be written as follows:
a x'J (t)
1
- xij (t) + A(ij;k,b)ykl(t) + B(ij;k,l) ukl,+ /
C
at
R
CijeNrij
CkleNrlj
(6)
(7)
1l)
Yy, (t) =1/2 (|xij (t) 1 - xx1 (t) where xij is the state of cell C11, C and R conform the
integration time constant of the system, and I is an independent
bias constant. The output equation, y11(t) = f(x11(t)), where f can
be any convenient non-linear function [3].

Similarly, if IJ(.) . 1 then yij(t) .< 1.

When used as an array-processing device, the CNN
performs a mapping

xij (0)

uF

1F.
(t)

y, (t)

(9)

U
(
where F is a function of the cloning template (A, B, I).
The functionality of the CNN array can be controlled
by the cloning template A, B, I, where A and B are (2r+1) x
(2r+1) real matrices and I is a scalar number in 2-D CNN's. In
many applications, A(i j; k, 1) and B(i, j; k, 1) are space
invariant. If A(i, j; k, 1) ==A(k, I; i, j), then the CNN is called
symmetrical or reciprocal.
There are two main cases of CNN: continuous-time
(CT-CNN) and discrete-time (DT-CNN) cellular neural
networks [8]. The equations for each cell of a DT-CNN are
(10)
xij(k) = E A(ij;k, ykl(t) + E B(ij;k, 0 ukl + Ii
C eNrij
CkleNrij
Y (k) =f ( X (k 1)
x(
f (x) = sgn(x).
A special class of two-dimensional cellular neural
network is described by ordinary differential equations of the
form [6].
a Xi (t)
(1 1)
- aij xi, (t) + E Tj, kl sat( Xkl(t) ) + Ii[

at

Yqj (t) = sat( x11(t))

where 1<1<Mi 1.<jN, aij 1/RC > 0, and xij andYij are the
states and the outputs of the network, respectively, and sat(.)
represents the activation function.
Consider zero inputs (uij= 0 for all i and j) and a
constant bias vector I
= [1,12...J............
MN]T.
Under these
circumstances, we will refer to equation (11) as a zero-input
non-symmetric cellular neural network where the n neurons are
arranged in a MxN array with n =(MxN) and the
interconnection structure is restricted to local neighborhoods of
radius r.

The matrices A(.) and B(.) are known as cloning |..............

templates. The template A(.) acts on the output of neighboring ..............
cells and is referred to as the feedback operator. The template \|
Output
B(.) in turn affects the input control and is referred to as the
lll
control operator. The templates A(.) and B(.) are application
dependent. A constant bias I and the cloning templates
lI
determine the transient behavior of the cellular non-linear
.
network. In general, the cloning templates do not have to be
/
space invariant, they can be, but it is not a necessity. A= /x(O
i)
significant feature of CNN is that it has two independent input *
capabilities: the generic input and the initial state of the cells.
Input,
Normally they are bounded by,
Figure 3. Block diagram of one cell of CNN
(8)
u(t) .l<1and lx(t)
.<i 1
2823

Fuzzy Derivative Estimation

=. l.J . . . .

Estimating derivatives and filtering can be seen as a

chicken-and-egg problem; for filtering system wants a good
vuS +lI} rindication of the edges, while to find these edges we need
filtering. In this approach, first look for the edges and try to
provide a robust estimate by applying fuzzy rules.

NW

NEW

r,rY,lW

VUSA-!)

Figure 4. The basic cell architecture of CNN for image

enhancement_l

(a)

The Fig. 4 shows the basic cell architecture of

DTCNN for image enhancement operation. The single layer
implementation with template and bias adjustment provides a
promising result and reduces the computational complexity.

(b)

Figure 5. (a) Neighborhood of a central pixel (x, y), (b) Pixel

values used to compute the "fuzzy derivative" of the central
pixel (x, y) for the NW- direction

direction

V. FUZZY FILTER

The application of fuzzy techniques in image

processing is a promising research field [9]. This section
focuses on fuzzy techniques for image filtering. Already
several fuzzy filters for noise reduction have been developed,
e.g., the well known FIRE-filter, the weighted fuzzy mean
filter, and the iterative fuzzy control based filter. Most fuzzy
techniques in image noise reduction mainly deal with fat-tailed
noise like impulse noise. These fuzzy filters are able to
outperform rank-order filter schemes (such as the median
filter). Nevertheless, most fuzzy techniques are not specifically
designed for Gaussian noise or do not produce convincing
results when applied to handle this type of noise. Two
important features are presented: first, the filter estimates a
"fuzzy derivative" in order to be less sensitive to local
variations due to image structures such as edges; second, the
membership functions are adapted accordingly to the noise
level to perform "fuzzy smoothing."
The main concern of the proposed filter is to
distinguish between local variations due to noise and due to
image structure [9]. In order to accomplish this, for each pixel
we derive a value that expresses the degree in which the
derivative in a certain direction is small. Such a value is
derived for each direction corresponding to the neighboring
pixels of the processed pixel by a fuzzy rule. The further
construction of the filter is then based on the observation that a
small fuzzy derivative most likely is caused by noise, while a
large fuzzy derivative most likely is caused by an edge in the
image. Consequently, for each direction we will apply two
fuzzy rules that take this observation into account (and thus
distinguish between local variations due to noise and due to
image structure), and that determine the contribution of the
neighboring pixel values. The result of these rules (16 in total)
is defuzzified and a "correction term" is obtained for the
processed pixel value.

) EL,

NW
W

position

(-x- by -1)

1-l)

S-- )

|?

S
E

NE

set wrt (.
{(-1)(O1,080i-1)}
| {(0 1},(0QO)I,({L-1)}
(

(,O)(

( y + 1)

{(,10) (Q,O), (-1,)}

l, y + l) |1,- t (o,) (- l, l
(x+,)
{(Q-1),(OtO)>(U1)}
+

(x+ 1y-1) (0,(-f-l(O,O)(1l1)}

1
-1)
0(-10M,(O) (1)}

| 1

Table 1. Pixels involved calculating the fuzzy derivatives in

each direction

Consider the 3x3 neighborhood of a pixel (x, y) as

displayed in Fig. 5 (a). A simple derivative at the central pixel
position (x, y) in the direction D (D C dir = {NW, W, SW, S,
SE, E, NE, N}) is defined as the difference between the pixel at
(x, y) and its neighbor in the direction D. This derivative value
is denoted by CD (x, y). For example, CN (x, y) = I (x, y -1) - I
(x, y) and CNW (x, y) = I (x - 1, y - 1) - I (x, y). Next, the
principle of the fuzzy derivative is based on the following
observation. Consider an edge passing through the
neighborhood of a pixel (x, y) in the SW - NE direction. The
derivative value CNW (x, y) will be large, but also derivative
values of neighboring pixels perpendicular to the edge's
direction can expected to be large. For example, in the NW direction the values of CNW (x, y), CNW (x - 1, y + 1) and
CNW (x + 1, y - 1) as shown in Fig. 5 (b) can be calculated.
The idea is to cancel out the effect of one derivative value that
turns out to be high due to noise. Therefore, if two out of three
derivative values are small, it is safe to assume that no edge is
present in the considered direction. This observation will be
taken into account when the fuzzy rules formulated to calculate
the fuzzy derivative values. Table 1 shows pixels involved in
calculating the fuzzy derivatives in each direction.

2824

Fuzzy Smoothing
To compute the correction term delta for the
processed pixel value, a pair of fuzzy rules for each direction is
used. The idea behind the rules is the following: if no edge is
assumed to be present in a certain direction, the (crisp)

derivative value in that direction can and will be used to

compute the correction term. The first part (edge assumption)
can be realized by using the fuzzy derivative value, for the
Figure 9. Noisy Image with salt and pepper noise
second part (filtering) have to be distinguished between
positive and negative values. The final step in the computation
For the classical spatial filter h3 mask for lowpass and
of the fuzzy filter is the defuzzification. The correction term median filter provides a better results as compared to other
delta obtained, which can be added to the pixel value of masks and techniques and are shown in Fig. 10, 11 and 12
location (x, y).

Adaptive Threshold Selection

Instead of making use of larger windows to obtain

better results for heavier noise, choose to apply the filter

iteratively. The shape of the membership function small is
adapted each iteration according to an estimate of the
(remaining) amount of noise. The method assumes that a
percentage p of the image can be considered as homogeneous
and as such can be used to estimate the noise density.

Figure 10. Filtered image of Lowpass Spatial Filter for Fig. 7

VI. RESULTS
The efficacy of various image enhancement
techniques is demonstrated through simulation results. Fig. 7,
Fig. 8, and Fig. 9 shows the noisy images for the original test
image Fig. 6.

Figure 6. Original Image

Figure

11. Filtered image of Median Filter for Fig. 7

Figure 12. Filtered image using cascaded structure of Lowpass

and Median filters for Fig. 9

Figure 7. Noisy Image with zero mean 0.01-variance Gaussian

noise

The three layer NN is trained for noise reduction with

the learning rate of 0.3 and Fig. 13 top left shows input layer
output, top right shows hidden layer output and bottom left
shows output layer output which is similar to the original test
image. Fig. 14 is an output of trained NN for the input noisy

image Fig. 8. Fig. 15 shows the output of the DTCNN for the

templates

A= 0 -4 0% B=

Figure 8. Noisy Image with zero mean, 0.1 variance Gaussian

noise

with I

=-1

The fuzzy filter is applied to grayscale test images

after adding Gaussian noise and the simulated results are
shown in Fig. 16. The image quality is measured based on the

energy content in the image as shown in Table 2. The objective

measure of the image is also done through Peak Signal to
Noise Ratio.

2825

Technique

Original

Gaussian Noise

Lowpass Spatial
Filter
Median Filter
Salt and Pepper
Neural Network

Energy

Red
1.0964 * 10
1.0993 * 10
1.6264 * 10

Green
0.8067 * :
0.8180 * 10
1.1991 * 104'

Blue
0.6143*
0.6258* 10
0.9093* 104

1.0791 * 10
1.0864 * 10

0.7965 * 10
0.8182 * 10
1.6528 * 10

0.6044* 10
0.6336* 10
0.01 16* 106

2.2494* 106

Table 2. Energy content of the images for various image

enhancement techniques.

computational requirements as the weights in each layer are to

be adjusted. The NN is trained after 140 iterations where the
error is very small in each iteration. The benefit of CNNs is
used for single layer and the computational speed is very high
as compared to NN. The fuzzy filter distinguishes between
local variations due to noise and due to image structures, using
a fuzzy derivative estimation. Tests are carried out for different
noise levels and the results are demonstrated through
simulation. It can be inferred that the image enhancement
techniques provides promising results for image analysis and

Figure 13. Filtered image of Neural Network after 140

iterations for Fig. 7

pattern recognition applications.

REFERENCES

[1] Pratt W. K., Digital Image Processing,3d ed., NewYorkWiley, 2003.

[2] Gonzalez R. C and Woods R. B., Digital Image
Processing, Pearson Education, Asia, 2001.
[3] Zurada J. M., Introduction to Artificial Neural Network
Systems, Jaico Publications, Newdelhi, 2000.
[4] Satyanarayana S. and Dalal S.," Video Colour
Enhancement Using Neural Networks", IEEE Trans.
Neural Networks, vol.7, pp. 1086-1096, 1996
[5] Chua L. 0. and Yang L., "Cellular Neural Networks:
Theory", IEEE Trans. Circuits and Systems, vol. 35, pp.
1257-1272, 1988.
[6] Chua L. 0. and Yang L., "Cellular Neural Networks:

Figure 14. Filtered image after training NN for Fig. 8

Figure 15. Filtered image for DTCNN architecture

(a)

(b)

Figure 16. (a) Cameraman test image corrupted with Gaussian

noise (b) Fuzzy filtered cameraman image

pp. 1273-1290, 1988.

[7] Lee C. and Gyvez J. P., "Colour Image Processing in a
Neural Network Environment", IEEE Trans. Neural
Networks, vol. 7, pp. 1086-1096, 1996.
[8] Rao D. H. and Hosur P. I., "Fuzzy Cellular Neural
Network for Image Enhancement", SPIE Proceedings,
vol. 2664, pp. 64-74, 1996.
[9] Dimitri, Dietrich, Etienne, Wilfried and Ignace, "Noise

VII. CONCLUSION
The classical spatial filters provide the promising
results and computational complexity involved is negligible.
Whereas a self-organizing multiplayer NN for image
enhancement with back-propagation of error based on gradient
descent has been compared. This technique involves very high

2826

Reduction by Fuzzy Image Filtering", IEEE Trans. Fuzzy

Systems, vol. 11, pp. 429-436, 2003.