IJCSNS International Journal of Computer Science and Network Security, VOL.9 No.

2, February 2009 305

Automatic Image Segmentation using Wavelets

H C Sateesh Kumar 1, K B Raja2, Venugopal K R2 and L M Patnaik3

1
Department of Telecommunication, Dayananda Sagar college of Engineering, Bangalore
2
Department of Computer Science and Engineering
University Visvesvaraya College of Engineering, Bangalore University, Bangalore
3
Vice Chancellor, Defense Institute of Advanced Technology, Puna, India

Abstract
Model-Based image segmentation plays a dominant role in The pixel intensity based image segmentation is obtained
image analysis and image retrieval. To analyze the features using Histogram-Based method, Edge-Based method,
of the image, model based segmentation algorithm will be Region-Based method and Model-Based method. Model-
more efficient compared to non-parametric methods. In Based segmentation algorithms are more efficient
this paper, we proposed Automatic Image Segmentation compared to other methods as they are dependent on
using Wavelets (AISWT) to make segmentation fast and suitable probability distribution attributed to the pixel
simpler. The approximation band of image Discrete intensities in the entire image. To achieve close
Wavelet Transform is considered for segmentation which approximation to the realistic situations, the pixel
contains significant information of the input image. The intensities in each region follow Generalized Gaussian
Histogram based algorithm is used to obtain the number of Distribution (GGD). Some of the practical applications of
regions and the initial parameters like mean, variance and image segmentation are Medical Imaging to locate tumors
mixing factor. The final parameters are obtained by using and other pathologies, locate objects in satellite images
the Expectation and Maximization algorithm. The viz., roads, forests, etc., automated-recognition system to
segmentation of the approximation coefficients is inspect the electronic assemblies, biometrics, automatic
determined by Maximum Likelihood function. It is traffic controlling systems, machine vision, separate and
observed that the proposed method is computationally track regions appearing in consequent frames of an image
efficient allowing the segmentation of large images and sequence and real time mobile robot applications
performs much superior to the earlier image segmentation employing vision systems.
methods.
Keywords: Discrete Wavelets, Image Segmentation, Motivation: Image segmentation plays an important role in
Histogram, Generalized Gaussian Distribution, EM biometrics as it is the first step in image processing and
Algorithm, ML Estimation. pattern recognition. Model based algorithms are used for
efficient segmentation of images where intensity is the
1. Introduction prime feature. The problem of random initialization is
overcome by using Histogram based estimation. The
In image processing the input is an image and the output is Wavelet transform solves the problem of resolution which
either an image or parameters related to the image is used can indicate the signal without information loss and
to solve identification problems, such as forensic medicine reduces the complexity. The segmentation is faster since
or creating weather maps from satellite pictures. Image approximation band coefficients of DWT are considered.
segmentation is a process of extracting and representing
information from an image in order to group pixels Contribution: In this paper, we introduced Wavelet
together into regions of similarity. Image segmentation is concept for image segmentation which reduces the
classified into three categories viz., i) Manual i.e., computation time by considering approximation band of
supervised or interactive in which the pixels belonging to an image which is small in dimensions and contains
the same intensity range pointed out manually and significant information of original image. The initial
segmented, the disadvantage is that it consumes more time parameters and final parameters are obtained by applying
if the image is large. ii) Automatic i.e., unsupervised which Histogram based algorithm and Expectation and
is more complex and algorithms need some priori Maximization algorithm respectively. GGD model is
information such as probability of the objects constructed and segmented by Maximum Likelihood
estimation of each approximation coefficient.
Having a special distribution to carry out the segmentation.
iii) Semi-automatic is the combination of manual and Organization: The rest of the paper is organized into
automatic segmentation. following sections. Section 2 is an overview of related

Manuscript received February 5, 2009

Manuscript revised February 20, 2009
306 IJCSNS International Journal of Computer Science and Network Security, VOL.9 No.2, February 2009

work. Section 3 describes model of AISWT and section 4 of scales and localize these edges, which is extended to
discusses the algorithm. Performance analysis of the multi-scale segmentation using anisotropic diffusion
model is presented in section 5 and conclusion is given in scheme.
section 6.
Jitendra et al., [7] proposed cue integration in image
2. Related Work segmentation by using an operational definition of textons,
the putative elementary units of texture perception and an
Sharon et al., [1] introduced fast multi-scale algorithm algorithm for partitioning the image into disjoint regions
which uses a process of recursive weighted aggregation to of coherent brightness and texture. The method finds
detect the distinctive segments at different scales. It boundaries of regions by integrating peaks in contour
determines an approximate solution to normalized cuts in orientation energy and differences in texton densities
time domain i.e., linear in the size of image with few across the contour by cue integration.
operations per pixel. The disadvantage is that the
segmented image fails to give smoother boundaries. Felzenswalb and Huttenlocher [8] described image
Prasad Reddy et al., [2] proposed a color image segmentation based on pairwise region comparison. The
segmentation method based on Finite Generalized algorithm makes simple greedy decisions and produces
Gaussian Distribution (FGGD). The observed color image segmentations that obey the global properties of being not
is considered as a mixture of multi-variant densities and too coarse and not too fine according to a particular region
the initial parameters are estimated using K-Means comparison function. The method is time linear in the
algorithm. The final parameters are estimated using EM number of graph edges and is fast in practice.
algorithm and the segmentation is obtained by clustering
according to the ML estimation of each pixel. However, Jianbo and Malik [9] proposed normalized cuts and image
computational time is more because of complex segmentation. Normalized cuts measures both the total
calculations. dissimilarity between the different groups as well as total
Donnell et al., [3] introduced a phase-based user steered similarity with in the groups, which is used for
segmentation algorithm using Livewire paradigm that segmentation.The method is optimized using generalized
works on the image features. Livewire finds optimal path eigen value problem.
between users selected image locations thus reducing
manual effort of defining the complete boundary. The Wu et al., [10] proposed a color image segmentation
phase image gives continuous contours for the livewire to method based on Finite Gaussian Mixture (FGM) model.
follow. The method is useful in medical image The observed color image is considered as a mixture of
segmentation to define tissue type or anatomical structure. multi-variant densities and the mixture parameters are
estimated using the EM algorithm. K-means algorithm is
James et al., [4] proposed color image segmentation for used to initialize the Gaussian mixture parameters. The
interactive robots based on color thresholding, nearest number of mixture components is automatically
neighbor classification, color space thresholding and determined by implementing the Minimum Message
probabilistic methods that are useful in real time mobile Length (MML) criteria into the EM algorithm. The method
robot applications. Each pixel is classified into a full is totally unsupervised because it integrates the parameter
resolution captured color image to merge regions upto 32 estimation and model selection in a single algorithm.
colors, the centroid bounding box and area at 30 hertz are
then obtained. The method is useful to accelerate low level Yamazaki [11] introduced a color image segmentation
vision in real time applications. method based on the ML estimation. The EM algorithm is
used for calculating the ML estimates and initial
Zhixin and Govindaraju [5] proposed hand written image parameters for the EM algorithm are estimated by the
segmentation using a binarization algorithm for camera multi-dimensional histogram and the minimum distance
images of old historical documents. The algorithm uses a clustering methods. Since no random selection is used for
linear approximation to determine the flatness of the parameter estimation, the method is useful for
background. The document image is normalized by unsupervised image segmentation applications. Mena and
adjusting the pixel values relative to the line plane Malpica [12] proposed supervised method for solving
approximation. Sumengen and Manjunath [6] proposed automatic extraction problem in segmentation of color
multi-scale edge detection approach to achieve good images for extracting information from terrestrial, aerial or
localization and good detection of edges. The technique is satellite images. The Dempster-Shafer theory of evidence
based on first finding an edge representation at each scale is applied in order to fuse the information using Texture
and then combining them using certain heuristics. The Progressive Analysis (TPA) so as to obtain the greatest
objective is to find favour edges that exist at a wide range amount of information from the different order statistics.
IJCSNS International Journal of Computer Science and Network Security, VOL.9 No.2, February 2009 307

Sumengen and Manjunath [13] introduced variational performance. i) Initialization of the system with two units
segmentation cost functions and associated active contour rather than one ii) Reducing the number of iterations with
methods that are based on pairwise similarities or no adverse effect and random selection among winning
dissimilarities of the pixels. A curve evolution framework vectors in case of a tie. A very high number of clusters
known as Graph Partitioning Active Contours is used lead to over segmentation which is reduced using
which is computationally efficient and a good solution for thresholding and rival penalization.
natural images.
Timothee et al., [20] presented a spectral segmentation
Marian [14] used active contours for image segmentation. algorithm based on multiple scales of the image by
An active contour is an energy minimizing spline that capturing both coarse and fine level details without
detects specified features within an image and is a flexibile iteration. The method focused on the Propagation of local
curve which can be dynamically adapted to required edges grouping cues that helps in detecting coherent regions with
or objects in the image by interactive process. The user faint boundaries. The multi-scale graph is introduced in
must suggest an initial contour, which is quite close to the parallel so that information propagates from one scale to
intended shape. The disadvantages are the fact of active another. The complexity of the algorithm is linear in the
contour dependency on the initial points of the contour, number of pixels. The long-range graph connections
type of the picture and computation difficulty. significantly improve the running time and quality of
segmentation. Ozden et al., [21] proposed a color image
Geraud et al., [15] proposed segmentation of color images segmentation method based on low-level features
using a classification in the 3-D color space. A classifier including color, texture and spatial information in a
that relies on mathematical morphology and more nonparametric framework. Discrete Wavelet Frames
precisely on the Watershed algorithm is used. Since (DWF) that provide translation invariant texture analysis.
Connected Watershed Algorithm (CWA) is applied in the The method integrates additional texture feature to the
histogram space, the method contrasts with the “classical” color and spatial space of standard mean shift
use of the Watershed algorithm in the image space as a segmentation.
segmentation tool. The main drawback is CWA consumes
more memory and time. 3. Model
Wesolkowski et al., [16] introduced color image In this section we discussed definitions and AISWT model
segmentation algorithm based on shading and highlight
invariant transformation. The transformation coupled with A. Definitions:
the mixture of Principal Components Algorithm is able to
cluster colors irrespective of highlights and shading under
i. Mean: The average intensity of a region is defined as
the condition of white balancing. The disadvantages are
the mean of the pixel intensities within that region.
the problems caused by small RGB values and illumination
The mean μz of the intensities over M pixels within a
color changes.
region K is given by Equation (1)
M
Wenchao et al., [17] proposed a hybrid segmentation 1
algorithm that combines prior shape information with μz =
M
∑x i ---------------- (1)
normalized cut to correctly segment the target whose i =1
boundary may be corrupted by noise. The shape model is
used for finding corresponding Eigen Vectors to segment Alternatively, we can use formulation based on the
the image. Leah et al., [18] showed that the segmentation normalized intensity histogram p(zi) where i=0,
and blind restoration are tightly coupled and can be 1,2,…,L-1 and L is the number of possible intensity
successfully solved together. Mutual support of the values as given by Equation (2)
segmentation and blind restoration is processed by a joint L
variational framework which integrates Mumford-Shah μ = ∑ z i p(z i ) ----------------- (2)
segmentation with parametric blur-kernel recovery and i =1
image deconvolution. The function is formulated using the
convergence approximation and is iteratively optimized ii. Variance: The variance of the intensities within a
via the alternate minimization method. region K with M pixels is given by Equation (3).

Mavrinac [19] proposed a color image segmentation using

1 M
a competitive learning clustering scheme. Two σ z2 = ∑ (xi − μ z )2 ------------- (3)
fundamental improvements are made to increase the speed M i =0
308 IJCSNS International Journal of Computer Science and Network Security, VOL.9 No.2, February 2009

Using histogram formulation the variance is given by

Equation (4) Input Image
L
σ2 = ∑(zi −μ)2 p(zi ) ---------------- (4)
i=0 DWT

iii. Probability Distribution Function (PDF) of the

intensities: The PDF P(z), is the probability that an Initial Parameters
intensity chosen from the region is less than or equal to a Estimation
given intensity value z. As z increases from -∞ to +∞, P (z)
increases from 0 to 1. P(z) is monotonic, non-decreasing
in z and thus dP/dz ≥ 0.
Shaping Parameter
iv. Shaping parameter P: Shaping parameter defines the
peak ness of the distribution which varies from 1 to ∞.
The GGD becomes Laplacian Distribution if P = 1, EM
Gaussian Distribution if P=2 and Uniform
Distribution if P→ +∞.
Generalized Gaussian
Distribution Model
v. Computational Time: Time required for the
Execution of the algorithm
Segmentation using ML
B. Block diagram of AISWT Estimation
The Figure 1 gives the block diagram of AISWT

I. Input image: The input images are of different formats, K Image Segments
sizes and types. The image pixel intensity in the entire
image is a Random Variable and follows a GGD. Fig 1: Block diagram of AISWT

II. DWT: The Wavelet Transform is created by repeatedly

filtering the image coefficients on a row by row and i) Histogram is constructed by dividing approximation
column by column basis. A two-dimensional DWT band coefficients into intervals and counting the number of
decomposition of image contains various band information elements in each subspace, which is called as bin. The K-
such as low-low frequency approximation band, high-low highest density bins are selected and the average of the
frequency vertical detail band, low-high frequency observed elements belonging to the bins is calculated to
horizontal detail band and high-high frequency diagonal derive a Centroid. If K centroids are not obtained because
detail band. We assume each coefficient of approximation of narrow intervals, the Histogram is rebuilt using wider
band is a Random Variable z and also follow GGD. The intervals and the centroids are recalculated.
approximation band is used for the segmentation purpose,
which is quarter the size and has significant information of ii) The minimum distance clustering is used to label all the
the original image. Hence the computation time required observed elements by calculating the distance D between
reduces. each centroid and all the elements of the histogram as
given in Equation (5)
III. Initial parameters Estimation: Initial parameters like
mean μ, variance σ and mixing parameter α are Dj = min||Ci – Yj|| ------------------------------- (5)
determined using Histogram based initial estimation which
is a clustering algorithm. The initial parameters are Where
calculated in two steps: Ci is ith centroid for i = 1 to K
Dj is minimum distance between Ci and jth element Yj for j
= 1 to N
IJCSNS International Journal of Computer Science and Network Security, VOL.9 No.2, February 2009 309

[ ( ( ))] (
The Histogram based initial estimation do not use random
( ) )
N K
selection for initial parameters, thus the method is stable Q θ ,θ (i ) = ∑∑ log αt f zs ,θt(i ) tt zs ,θ (i )
and useful for unsupervised image segmentation s=1 t =1
applications. The obtained mean, variance and mixing ------------ (9)
parameter for the k regions are considered as the initial
parameters for EM algorithm.
Where ( )
tt zs ,θ ( i ) is a Posterior Probability and is given
by Equation (10)
IV. Shaping parameter P: The Shaping parameter defines
the peakness of the distribution. In GGD, the three
tt (z s ,θ ) =
α t (i ) f z s , θ t (i )
(i ) ( )
h(z s ,θ (i ) )
parameters mean, variance and shaping parameter ------------- (10)
determines the PDF. The optimal shaping parameter [22,
23] is determined using initial parameters and the absolute K
mean value E [ z ] . The absolute mean is given by h ( z s , θ (i ) ) = ∑α
t =1
t f ( z s , θ t(i ) ) --- (11)
Equation (6)

ii. M-step: It finds the (i+1)th estimation θ

( i +1)

E[ z ] =
1 N k by updating
N
∑∑ z
i =1 j =1
i − μ j ------------------ (6) mixing parameter, mean and variance using Equations
(12), (13) and (14) respectively to maximize Log-

P is estimated using Equation (7)

Likelihood function Q θ , θ ( (i )
)
P = M −1 (ρ ) ---------------------------------- (7)
∑ t (z ,θ ( ) )
N
1
α t (i+1) = t s
i
---------------- (12)
Where ρ is given by N s =1

E2[z ]
∑ t (z , θ ( ) ) z
N
i
ρ= t s s
σ 2
μt (i + 1 )
= s =1
-------------- (13)
∑ t (z , θ )
N
(i )
M is the Generalized Gaussian ratio function given by t s
Equation (8) s =1

⎛2⎞
Γ 2 ⎜⎜ ⎟⎟ Choose an Initial
M (ρ ) = ⎝ρ⎠ -------------------- (8)
⎛1⎞ ⎛3⎞
Parameter θ (0)
Γ ⎜⎜ ⎟⎟ Γ ⎜⎜ ⎟⎟
⎝ρ⎠ ⎝ρ⎠
E-step
V. Expectation and Maximization: The EM algorithm is an
efficient iterative procedure to compute the ML estimate in
(
Q θ , θ (i ) )
the presence of missing or hidden data. For obtaining the
EM algorithm a sample of the coefficients z1, z,..., zn, are
drawn with PDF f (z,θ) given in Equation (15) where θ M-step
is set of initial parameters (
Q θ ,θ (i +1) )
i.e., θ = (α , μ ,σ 2 , P ).
Is
Each iteration of the EM algorithm consists of two steps as (
Q θ ,θ (i+1)
) ~ Q(θ,θ ( ) ) < 0.001 i No
shown in Figure2 Converged?

i. E-step: It computes the expected complete data Log- Yes

Likelihood function Q θ , θ ( (i )
) given by Equation (9)
Fig 2: Flow chart of EM
.
310 IJCSNS International Journal of Computer Science and Network Security, VOL.9 No.2, February 2009

1 foreground and background. The pixel intensities of

⎡ N (i ) ⎛ Γ(3 P ) ⎞ (i ) P ⎤
1

( )
p
segmented region obtained follow a corresponding GGD.
⎢∑ t s t z , θ ⎜
⎜ PΓ(1 P ) ⎟ s ⎟ z − μ t ⎥
σt (i +1)
=⎢
s =1 ⎝ ⎠ ⎥
⎢ ⎥
( )
N
4. Algorithm
⎢ ∑ tt z s , θ (i )
⎥
⎢⎣ s =1 ⎥⎦ Problem definition:
----- (14)
Consider an image, the objectives are to
The EM algorithm will converge when the difference of i. Segment the given image using DWT.
the old estimates and the new estimates is less than the ii. Initial parameters are obtained using
threshold value 0.001. The EM algorithm used for Histogram and EM
estimating the final parameters is heavily dependent on iii. Segmentation is done using ML
number of segments and the initial parameters of the Estimation
model.
Assumption: Each coefficient of approximation band is a
VI. Generalized Gaussian Distribution model: The GGD Random Variable z and also follows GGD
model is obtained using the final parameters. The
approximation band coefficients of each image region Table 1 gives the AISWT Segmentation algorithm in
follow a particular distribution such as Gaussian, which approximation band of image DWT is used to
Laplacian, Uniform etc., and characterize the GGD Model estimate parameters which are required for segmentation.
with shaping parameter P. The PDF is given by Equation
(15) Table 1 AISWT Algorithm

( zs −μi ) •
P
−
Input : Image of variable size
1
f (z,θ ) = e A( P,σ ) --- (15) • Output : Segmented regions
⎛ 1⎞
2Γ⎜1 + ⎟ A(P,σ ) 1 DWT is applied on an image and
⎝ P⎠ approximation band is considered.
Where, 2 Histogram Based method is applied to
s=1 to N and i=1 to K obtain initial parameters like mean,
1
variance and mixing parameter
⎡ 2 ⎛ 1 ⎞⎤2
⎢ σ Γ⎜ P ⎟ ⎥
for σ > 0, A(P, σ ) = ⎢
⎝ ⎠⎥ 3 Shaping parameter P is determined
⎢ Γ⎜ 3 ⎞⎟ ⎥
⎛ 4 Expectation and Maximization
⎢⎣ ⎝ P ⎠ ⎥⎦ algorithm is used to get updated final
parameters.
The function A(P, σ ) is a scaling factor and P is the
5 PDF of Generalized Gaussian
shape parameter. The GGD becomes Laplacian Distribution is determined
Distribution if P = 1, Gaussian Distribution if P = 2 and
Uniform Distribution if P→ +∞. 6 Segmentation is obtained using
Maximum-Likelihood estimation.
VII. Segmentation using Maximum Likelihood Estimation:
The segmentation is carried out by assigning each
coefficient into proper cluster according to the ML 5. Performance Analysis
estimation given by Equation (16)
Images Sea, Flower, Starfish and Boat of sizes 150*94,
L = maxt { f ( zs ,θ t )} -------------------- (16) 127*127, 800*600 and 800*603 respectively are
considered for performance analysis. If the number of
segments are selected as two i.e., K=2 foreground and
VIII. K image segments: The K segmented regions are background can be differentiated in an image. Figures 3, 4,
obtained and for K=2, the image is segmented into 5 and 6 gives the segmentation results of Sea, Flower,
Starfish and Boat in which the original image, mask,
IJCSNS International Journal of Computer Science and Network Security, VOL.9 No.2, February 2009 311

segment 1 and segment 2 are shown in Figures a, b, c and

d respectively.

Table 2 gives the comparison of computational time

between FGM, FGGD and AISWT. It is observed that as
dimension of the image increases, the computational time
increases. The existing algorithms FGM and FGGD
require more time compared to the proposed algorithm
AISWT. AISWT requires 30% less time compared to a b
FGM and 70% less time compared to FGGD.

Table 2: Comparison of Computational time for FGM,

FGGD and AISWT

Images FGM FGGD AISWT

(sec) (sec) (sec)
Sea 1.53 2.97 0.9
Flower 1.57 2.8 0.9 c d
Starfish 39.9 96.43 26.5
Boat 38.84 100.12 27.2 Fig 4 a) original Flower (127*127) b) mask
c) Segment 1 d) Segment 2

Table 3 gives the comparison of Image quality Index

between FGM, FGGD and AISWT. It is observed that
proposed model AISWT has higher Image Quality Index
compared to existing algorithms FGM and FGGD.

a b

c d a b
Fig 3 a) Original Sea (150*94) b) Mask
c) Segment 1 d) Segment 2

Table 3: Comparison of Image Quality Index for FGM,

FGGD and AISWT

Images FGM FGGD AISWT c d

Fig 5 a) original Starfish (800*600) b) Mask
Sea 93.7 93.8 97.3
c) Segment 1 d) Segment 2
Flower 95.5 96 98.6
Starfish 88 93.1 95
Boat 90.9 83 91.3
312 IJCSNS International Journal of Computer Science and Network Security, VOL.9 No.2, February 2009

Intensity Normalization,” SPIE Proceedings on Center of

Excellence for Document Analysis and Recognition,
Document Recognition and Retrieval, vol. 5676, pp. 167-
174, January 2005.
[6] B. Sumengen and B. S. Manjunath, “Multi-Scale Edge
Detection and Image Segmentation,” Proceedings of
European Signal Processing Conference, September 2005.
[7] J. Malik, S. Belongie, J. Shi and T. Leung, “Textons,
Contours and Regions: Cue Integration in Image
Segmentation,” Proceedings of Seventh International
a b Conference on Computer Vision, pp. 918-925, September
1999.
[8] P. F. Felzenswalb and D. P. Huttenlocher, “Efficient Graph-
Based Image Segmentation,” Proceedings of International
Journal of Computer Vision, vol. 59, no. 2, pp. 167-181,
2004.
[9] J. Shi and J. Malik, “Normalized Cuts and Image
Segmentation,” IEEE Transactions on Pattern Analysis and
Machine Intelligence, vol. 22, no. 8, pp. 888-905, 2000.
[10] Y. Wu, X. Yang and K. L. Chan, “Unsupervised Color
c d Image Segmentation based on Gaussian Mixture Models,”
Fig 6 a) Original Boat (800*603) b) Mask Proceedings of Fourth International Joint Conference on
c) Segment 1 d) Segment 2 Information, Communications and Signal Processing, vol.1,
no. 15, pp. 541-544, December 2003.
[11] T. Yamazaki, “Introduction of EM Algorithm into Color
6. Conclusion Image Segmentation,” Proceedings of International
Conference on Intelligent Processing Systems, pp. 368-371,
In this paper, we proposed fast segmentation algorithm 1998.
AISWT. The approximation band of an image DWT is [12] J. B. Mena and J. A. Malpica, “Color Image Segmentation
considered as a mixture of K-Component GGD. The initial using the Dempster-Shafer Theory of Evidence for the
Fusion of Texture,” International Archives of
parameters are estimated using Histogram based method. Photogrammetry Remote Sensing and Spatial Information
Through EM algorithm, the final parameters are obtained. Sciences, vol. 34, pp. 139-144, 2003.
The segmentation is done by ML estimation. The AISWT [13] B. Sumengen and B. S. Manjunath, “Graph Partitioning
algorithm is computationally efficient for segmentation of Active Contours for Image Segmentation,” IEEE
large images and performs much superior to the earlier Transactions on Pattern Analysis and Machine Intelligence,
image segmentation methods FGM and FGGD in terms of vol.28, no. 4, pp. 509-521, April 2006.
computation time and image quality index. [14] M. Bakos, “Active Contours and their Utilization at Image
Segmentation,” Proceedings of Fifth Slovakian-Hungarian
Joint Symposium on Applied Machine Intelligence and
Reference
Informatics, pp. 313-317, January 2007.
[1] E. Sharon, A. Brandt and R. Basri, “Fast Multi-Scale Image
[15] T. Geraud, P. Y. Strub and J. Darbon, “Color Image
Segmentation,” Proceedings of the IEEE Conference on
Segmentation based on Automatic Morphological
Computer Vision and Pattern Recognition, vol. 1, pp. 70-77,
Clustering,” Proceedings of IEEE International Conference
2000
on Image Processing, vol. 3, pp. 70-73, October 2001.
[2] P. V. G. D. Prasad Reddy, K. Srinivas Rao and S.
[16] S. Wesolkowski, S. Tominaga and R. D. Dony, “Shading
Yarramalle, “Unsupervised Image Segmentation Method
and Highlight Invariant Color Image Segmentation,” SPIE
based on Finite Generalized Gaussian Distribution with EM
Color Imaging: Device-Independent Color, Color Hardcopy
and K-Means Algorithm,” Proceedings of International
and Graphic Arts VI, pp. 229-240, January 2001.
Journal of Computer Science and Network Security, vol.7,
[17] W. Cai, J. Wu and A. C. S. Chung, “Shape-Based Image
no. 4, pp. 317-321, April 2007.
Segmentation using Normalized Cuts,” proceedings of
[3] L. O. Donnell, C. F. Westin, W. E. L. Grimson, J. R. Alzola,
Thirteenth International Conference on Image Processing,
M. E. Shenton and R. Kikinis, “Phase-Based user Steered
pp. 1101-1104, October 2006.
Image Segmentation,” Proceedings of the Fourth
[18] L. Bar, N. Sochen, and N. Kiryati, “Variational Pairing of
International Conference on Medical Image Computing and
Image Segmentation and Blind Restoration,” Proceedings of
Computer-Assisted Intervention, pp. 1022-1030, 2001
the European Conference on Computer Vision, vol. 3022,
[4] J. Bruce, T. Balch and M. Veloso, “Fast and Inexpensive
pp. 166-177, 2004.
Color Image Segmentation for Interactive Robots,”
[19] A. Mavrinac, “Competitive Learning Techniques for Color
Proceedings of the IEEE International Conference on
Image Segmentation,” Proceedings of the Machine
Intelligent Robots and Systems, vol. 3, pp. 2061-2066, 2000.
Learning and Computer Vision, vol. 88, no. 590, pp. 33-37,
[5] Z. Shi and V. Govindaraju, “Historical Handwritten
April 2007.
Document Image Segmentation using Background Light
IJCSNS International Journal of Computer Science and Network Security, VOL.9 No.2, February 2009 313

[20] T. Cour, F. Benezit and J. Shi, “Spectral Segmentation with Bangalore University and Ph.D. in Computer Science from
Multi-Scale Graph Decomposition,” IEEE Computer Indian Institute of Technology, Madras. He has a distinguished
Society Conference on Computer Vision and Pattern academic career and has degrees in Electronics, Economics, Law,
Recognition, vol. 2, pp. 1124-1131, February 2005. Business Finance, Public Relations, Communications, Industrial
[21] M. Ozden and E. Polat, “Image Segmentation using Color Relations, Computer Science and Journalism. He has authored 27
and Texture Features,” Proceedings of Thirteenth European books on Computer Science and Economics, which include
Signal Processing Conference, September 2005. Petrodollar and the World Economy, C Aptitude, Mastering C,
[22] J. A. D. Molina and R. M. R. Dagnino, “A Practical Microprocessor Programming, Mastering C++ etc. He has been
Procedure to Estimate the Shape Parameters in the serving as the Professor and Chairman, Department of Computer
Generalized Gaussian Distribution,” IEEE Transactions on Science and Engineering, University Visvesvaraya College of
Image Processing, pp. 1-18, 2003. Engineering, Bangalore University, Bangalore. During his three
[23] K. Sharifi and A. L. Garcia, “Estimation of Shape Parameter decades of service at UVCE he has over 200 research papers to
for Generalized Gaussian Distribution in Subband his credit. His research interests include computer networks,
Decomposition of video,” IEEE Transactions on Circuits parallel and distributed systems, digital signal processing and
and Systems for Video Technology, vol. 5, no.1, pp. 52-56, data mining.
1995.

Bibliography L M Patnaik is a Vice Chancellor, Defence

Institute of Advanced Technology (Deemed
Mr. H.C. SATEESH KUMAR is a University), Pune, India. During the past 35
Assistant Professor in the Dept of years of his service at the Indian Institute of
Telecommunication Engg, Dayananda sagar Science, Bangalore. He has over 500
college of Engg, Bangalore. He obtained his research publications in refereed
B.E. degree in Electronics Engg from International Journals and Conference
Bangalore University. His specialization in Proceedings. He is a Fellow of all the four
Master degree was “Bio-Medical leading Science and Engineering Academies in India; Fellow of
instrumentation from Mysore University and the IEEE and the Academy of Science for the Developing World.
currently he is pursuing Ph.D. in the area of He has received twenty national and international awards;
Image segmentation under the guidance of Dr.K.B.Raja, notable among them is the IEEE Technical Achievement Award
Assistant Professor, Dept of Electronics and Communication for his significant contributions to high performance computing
Engg, University Visvesvaraya college of Engg, Bangalore. His and soft computing. His areas of research interest have been
area of interest is in the field of Signal Processing and parallel and distributed computing, mobile computing, CAD for
Communication Engg. He is the life member of Institution of VLSI circuits, soft computing, and computational neuroscience.
Engineers (India), Institution of Electronics and
Telecommunication Engineers and Indian society for Technical
Education.

K B Raja is an Assistant Professor, Dept. of

Electronics and Communication Engg,
University Visvesvaraya college of Engg,
Bangalore University, Bangalore. He
obtained his Bachelor of Engineering and
Master of Engineering in Electronics and
Communication Engineering from
University Visvesvaraya College of
Engineering, Bangalore. He was awarded
Ph.D. in Computer Science and Engineering from Bangalore
University. He has over 35 research publications in refereed
International Journals and Conference Proceedings. His research
interests include Image Processing, Biometrics, VLSI Signal
Processing, computer networks.

K R Venugopal is currently the Principal and

Dean, Faculty of Engineering, University
Visvesvaraya College of Engineering,
Bangalore University, Bangalore. He obtained
his Bachelor of Engineering from University
Visvesvaraya College of Engineering. He
received his Masters degree in Computer
Science and Automation from Indian Institute
of Science Bangalore. He was awarded Ph.D. in Economics from