See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/49603948

HSV Brightness Factor Matching for Gesture Recognition System

Article · December 2010

Source: DOAJ

CITATIONS READS
60 1,342

2 authors:

Mokhtar Mohammed Hasan Pramod Kumar Mishra

University of Baghdad Banaras Hindu University
24 PUBLICATIONS 879 CITATIONS 110 PUBLICATIONS 2,133 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Mokhtar Mohammed Hasan on 02 June 2014.

The user has requested enhancement of the downloaded file.

Mokhtar M. Hasan & Pramod K. Misra

HSV Brightness Factor Matching for Gesture Recognition

System

Mokhtar M.Hasan [email protected]

Department of Computer Science
Faculty of Science
BHU-Varanasi
Uttar Pradesh-221005, India

Pramod K. Mishra [email protected]

Department of Computer Science
Faculty of Science
BHU-Varanasi
Uttar Pradesh-221005, India

Abstract

The main and primary objective of the gesture recognition research is to

establish a system which can identify specific human gestures and utilize these
identified gestures to be carried out by the human made machines, In this paper,
we are going to introduce a new method for gesture recognition that based on
the local brightness of each block of the gesture image, the input gesture image
is divided into 25x25 blocks each of 5x5 block size, and we have calculated the
local brightness of each divided block after applying colored segmentation
operation using HSV (Hue, Saturation and Value) color model, so, each gesture
produces 25x25 feature values called feature vector, our experimental shows
that more that 65 % of these features are zero value which leads to minimum
storage space, the recognition rate achieved is 91 % using 36 training gestures
and 24 different testing gestures. We build a gesture recognition system that can
communicate with the machine in natural way without any mechanical devices
and without using the normal input devices which are the keyboard and mouse
and the mathematical equations is the only translator that existed between the
human and human-made machines, we have focused in this study on the hand
gesture since hand can represents more meaning than other human-bodily
organs.

Keyword: Brightness Calculation, HSV color model, Gesture Recognition, Template Matching, Image
Segmentation, Laplacian Edge Detection.

1. INTRODUCTION
In all of the simulation processes, we tried to simulate the human abilities, in gesture recognition
system, the remarkable ability of the human vision is the gesture recognition, it is noticeable
mainly in deaf people when they communicating with each other via sign language and with
hearing people as well. In this paper we tried to simulate this ability but this time will be between
the human and human-made machines.

International Journal of Image Processing (IJIP), Volume (4): Issue (5) 456
Mokhtar M. Hasan & Pramod K. Misra

A gesture pose is a mode of communication between people that depends on the bodily
movement especially the hand motion and pose; this form of communication is established along
with spoken words in order to create a comprehensive statement to be carry out be the hearer.
Most people use gestures language represented by bodily movement in addition to spoken words
when they communicate between each other [1]; Figure (1) shows a gesture example for
helicopter signaler.

A, take off B, landing

FIGURE 1: Helicopter Signaller for Marshaling Operations [2].

The normal communication between the people is the speaking which needs the sound to convey
the meaning, while the later kind needs the space to convey the meaning [3]. The coarse
classification of gestures is two; static and dynamic, the static gesture is a specific hand pose
formed by a single image. The dynamic gesture is a moving gesture formed by a sequence of
images [3] as in Figure (2),

A, dynamic B, static
FIGURE 2: A and B Represent the Dynamic and Static Gesture
Respectively.

The application of gesture system on interactive applications produces many challenges. The first
and important challenge is the response time which should be fast [4]. There should be no
noticeable time between user gesture movement and computer replies [4]. The designed
computer vision algorithms should be reliable and work for different ethnic people [4] especially
when the color of human is changed comparing with white and black people. One more challenge
which is the cost challenge, the gesture system needs special hardware such as the camera and
sensors as necessarly, those special hardware will be the replacement of the existing hardware
devises which may considered as low cost [4] such as the keyboard and mouse, but the gesture
system with these new devices will be more worthwhile for wire-less communication.

This paper applied a new gesture recognition method for identifying the gestures for the computer
or for the telerobotic in order to understand and carry on the human teleoperations, we have
applied this novel method by windowing the image in order to recognize the input gesture and
discover the meaning for that gesture. We have applied the proposed method using six gestures
database, each of ten samples, so the total is sixty gestures database used for gesture
recognition; we used the hand gesture rather than the face because the hand is the most flexible
part of the body and can shows different meaning.

International Journal of Image Processing (IJIP), Volume (4): Issue (5) 457
Mokhtar M. Hasan & Pramod K. Misra

2. RELATED WORK
Roberto and Tomaso [5] applied a face recognition template matching; a very simple and direct
recognition technique based on the use of whole image as grey-level templates. The most direct
of the matching procedures is correlation. First, the image is normalized to obtain aunified
locations for the mouth, eye and nose. The authors in [5] applied her technique by creating a
database entry for each person contains the frontal view of that person, along with four masks
which are the eyes, nose, mouth, and face (the region from eyebrows downwards as decided by
the authors in [5]). All these four masks are relatively to the position of normalized eye position in
whole of their database. The recognition measure applied by them is the Euclidian Distance; by
matching the new presented face gesture with all the database gestures, and the database
gesture corresponds to the maximum matching score is the recognized gesture, they had used
samples taken from 47 persons with 4 gestures each.

Freeman and Roth [6] applied hand gesture recognition using orientation histogram, They had
applied some transformation T to the image data in order to create the feature vector that will be
used for recognition purpose and represents that specific gesture. To classify the gesture, they
compare the feature vector with the feature vectors from a previously generated training set. The
transformation T can be described as a polar representation for the histogram of the local
orientations of the input gesture, they use the gradient direction to calculate this orientation, the
histogram of the directions is then sketched using polar plot which represents the final features of
the input gesture and the other features will treated the same, they had used samples taken from
one person with 5-15 gestures.

K. Symeonidis in [4] applied gesture recognition method using neural network, he has used 8
different hand gestures each of 3 samples, those were for training purpose, he did not use an
exact number of gestures for testing purpose since some gestures tolerate more samples than
others, the features than he has used were the 19 elements degree numbers that been converted
from the polar representation of the orientation histogram of the input gesture, then he presents
these features for training the neural network after some preprocessing that casted the original
features into later number of features.

Xingyan [7] applied gesture recognition using fuzzy C-means algorithm, He has used image
processing methods to transform the raw image into the feature vector. The feature vector is
created by applying segmentation using HSV color model and then he has reduced the noise, the
feature vector of the image is thirteen parameters long. The first feature is the aspect ratio of the
hand’s bounding box as decided by the authors in [7]. The last 12 features are values
representing coarse parameters of the image, where each grid cell is the mean gray level value in
the 3 by 4 block division of image. Each of the 12 values calculated by the mean value of 3 by 4
partitions, which represent the mean of the brightness value. The classification phase is applied
using a recognition algorithms based on the Fuzzy C-Means (FCM) algorithm, he had used
samples taken from 6 persons with 6 gestures each and achives a time of 2-4 seconds with 86 %
recognition rate.

3. OVERALL APPROACH
Our system is composed of four main segments in order to recognize the input gesture, these
stages are summarized in Figure (3), and the details of these stages are as follows:

3.1 Collect the Input

In this phase, the input gestures are predisposed and the gesture poses are decided along with
their meaning, and the database is creating which contains the different gestures poses with
many samples per gesture, the number of the samples are limited by the speed and accuracy of
the system, as the samples increased, the system speed decreased and the accuracy increased,
and vice versa. We have chosen our database as in Figure (4) along with their meaning.

International Journal of Image Processing (IJIP), Volume (4): Issue (5) 458
Mokhtar M. Hasan & Pramod K. Misra

Collect the Image Feature Recognition

input Preprocessing Extraction Algorithm

FIGURE 3: Overview of Gesture Recognition System.

Debug Maximize Stop

Step In Minimize Step Out

FIGURE 4: System’s Database Vocabulary.

As seen by Figure (4), the database contains six gesture poses, which represent the target
recognition decision which may belong to any of these six gestures if the new presented gesture
is recognized, else, the system will announce that the ne presented gesture is not-known, these
six gestures that implied in the database have many samples for each gesture which are six
samples for each, as the number of samples increases the system accuracy increases and the
testing time increases which effects badly on the overall speed and performance of the system.

3.2 Image Preprocessing

In this phase, we have predisposed the input gestures for both training and testing purposes; this
phase is divided down into the following sub phases.

3.2.1 Segmentation
In this phase, we have applied the segmentation operation to segment the hand area in the input
gesture and isolate it from the background, all of the gesture systems depend on the perfect
segmentation of the hand gesture region, there are two main methods for segmentation: first
method is by using HSV model; which deals with the color pigment of the human skin, the
different ethnic groups have a significant property which is the different in the skin color which is
represented by the pigment concentration difference which affect the saturation of the skin [7].
The color of the skin, on the other hand, is roughly invariant across ethnic groups [7]. By deciding
a range for each of H, S and V parameters, a good segmentation method can be achieved;
Xingyan Li [8] have decided certain values for H and S parameters only. The other method used
for segmentation operation is by using clustering algorithms or thresholding technique, these
algorithms are suitable mainly for homogeneous or uniform background that has no cluttered

International Journal of Image Processing (IJIP), Volume (4): Issue (5) 459
Mokhtar M. Hasan & Pramod K. Misra

objects and the hand is the prominent object, we have applied HSV mode of segmentation for
splitting the hand area from the image gesture with some threshold value, after image
segmentation; we have to normalize the segmented image gesture in order to obtain gestures
database that are invariant against position, scale, and rotation, this will speed up the recognition
operation and reduce the number of gesture samples stored in the database.

3.2.2 Edge Detection

The areas in images with high contrast that has an intensity jump between the connected pixels
are called the edges, the non-edge areas do not has any intensity difference, noise also has
some intensity difference, so, edge detection algorithms preserve the important information in the
images that represents the edges and reduces the useless information which is classifies as non-
edges pixels, there are two types of edge methods, the gradient (first derivative) and the laplacian
(second derivative) as Figure (5), the gradient seeking for the maximum and minimum values and
by using of thresholding techniques it classifies the edges and this methods is used when the
intensity is high and is changed rapidly which provide a significant contrast, but when the gray
level changes slightly from dark to light or vice versa the second derivative is used; the second
derivative for maximum value is zero value which is the principle of the laplacian edge detector,
this methods uses zero crossing to find the edges and it is suffers from false edges produced
because of noise so it need special treatment like blurring the image before applying laplacian
edge detector, we used this laplacian to find the edges.

FIGURE 5: Edge Methods Categories.

As seen by Figure (50), the maximum value hereinabove is the edge in case of first derivative
methods; this edge can be located by using of thresholding technique which produces different
edge location depending on the threshold value and the edge thinness depends on the correct
selection of the threshold as well, but in second derivative; the intersection with x-axis is the edge
produces a unique and non-duplicated edge.

3.2.3 Normalization
In this phase of image preprocessing, the gesture is trimmed to get rid of the unnecessarily area
that surrounding the gesture area, this is done by removing this useless area from four directions.

3.3 Feature Extraction

After the preparing of the image and segmenting of the hand gesture, a black-white image is
created and represented the hand pose inset, the feature extraction phase will start, the overall
feature vector size is 625 elements which are the brightness values of each block in the gesture,
these features are stored in the database and the same algorithm is applied on the 36 gesture
database, the actual feature vector size that stored in the database is more smaller than the
calculated feature vector since the existence of zero values, this fact will be discussed hereafter.

International Journal of Image Processing (IJIP), Volume (4): Issue (5) 460
Mokhtar M. Hasan & Pramod K. Misra

3.4 Recognition Algorithm

When a new input presents to the system for classification, the input gesture is treated as
prescribed sequence and extracting the features of this input gesture, the feature vector of the
new presented input is compared against the stored database feature vectors using our new
algorithm for matching purpose, each two brightness values are considered equal in case of both
have no brightness value which means the block which is 5x5 (25 pixels) is black and this lead us
to matching between two black areas, and in case of the existence of the brightness value, i.e.
>0, we apply some threshold for considering the two blocks are equal, we use threshold of one to
obtain maximum flexibility of the system in case of changing the inset of the same gesture, this
matching algorithm yields good recognition rate comparing with other prescribed methods
because the recognition rate depends on two decisions, the black matching and the brightness
matching.

4. EXPERIMENTAL RESULTS
As we explained before, we trained the system with six different gestures each of six samples,
these gesture images have undergone in a serial operations in order to extract the features.

First step is the segmentation operation, this operation is required for splitting the hand region,
the segmentation is applied using HSV color space, the input gesture is RGB color space and
converted to HSV color space using the following equations:
Let p(x, y) represents the input pixel with R, G, and B components, and let p’(x, y) with H, S, and
B components, the following steps for converting p(x, y) to p’(x, y)[12]:

Step1: calculate M from, M = max(R, G, B)

Step2: calculate m from, m = min(R, G, B)
Step3: calculate r from, r = (M-R)/(M-m)
Step4: calculate r from, g = (M-G)/(M-m)
Step5: calculate r from, b = (M-B)/(M-m)
Step6: calculate V output from, V = max(R, G, B)
Step7: calculate S output from, if M = 0 then S = 0 and H = 180 degrees
if M <> 0 then S = (M - m) / M
Step8: calculate H output from, if R = M then H = 60(b-g)
if G = M then H = 60(2+r-b)
if B = M then H = 60(4+g-r)
if H >= 360 then H = H - 360
if H < 0 then H = H + 360
Step9: output H, S, and V color space, where H in the range [0,360], S and H in the range
[0,100]

We decided a range of values for each of H, S and V so it accepts the pigment of human skin, let
is the input gesture colored image location (x, y), and let H(x, y), S(x, y) and V(x ,y) are the H, S
and V bands for the HSV color space for the input gesture image at location (x, y); let M be the
output binary image, we set M(x, y) to 1 when Hmin<, H(x ,y)<Hmax and Smin<, S(x, y)<Smax
and Vmin<, V(x ,y)<Vmax ; otherwise, we set M(x, y) to 0; Figure (6) is the output of applying this
technique on Figure (4).

After this phase, hand boundary required to be last step in the image preprocessing, laplacian
edge detector is applied as we explained before, the mask shown in Figure (7) is used and
produces the output shown in Figure (8).

International Journal of Image Processing (IJIP), Volume (4): Issue (5) 461
Mokhtar M. Hasan & Pramod K. Misra

FIGURE 6: Binary Images.

0 1 0
1 -4 1
0 1 0

FIGURE 7: Laplacian Mask.

FIGURE 8: Edges of the Gestures.

Now, normalization operation is applied for removing the unwanted area of the gesture image,
Figure (9) shows the application of normalization operation for one gesture of the samples
gestures, the rest gestures are the same.

After this point, the gesture image is ready for feature extraction, as we said, the image output is
128x128 pixels, and block size is 5x5 pixels, so, each gesture image can produce 625 feature
vector represents the features of this gesture, Figure (10) shown the values computed from the
output gesture in Figure (9).

International Journal of Image Processing (IJIP), Volume (4): Issue (5) 462
Mokhtar M. Hasan & Pramod K. Misra

A, Edge Image B, Before Norm. C, After Norm.

FIGURE 9: Normalization Operation.

FIGURE 10: Features Calculation via Dividing the Gesture.

As seen in Figure(11), out of 625 features just 98 are white and stored and the other are
neclected which represents 15.68 % stored features out of 625, and 84.32 % is neglected,
Equation (1) shows the mathematical implementation for the gesture division.

Where i,j takes the values 2..height-3, 2.. width-3 respectively.

B: is the output brightness value.

After getting these features, and at the time of recognition during testing stage, the features of the
testing gesture are calculated and compared with database features using our suggested
algorithm, and the highest matching score is passed along with it is gesture and meaning, the
algorithm is described below:

Consider D(i,k) is the database feature k of gesture i, and T(k) is the feature k of the testing
image, M(k) is the matching status of feature k between both database gesture i and input testing
gesture, for each gesture i, if D(i, k) is black area and T(k) is black area, the set M(k) as matched
status, if the brightness value of D(i, k) > Threshold and T(k) > Threshold, then set M(k) as
matched status, otherwise, set M(k) as non-matched status, after that the number of matched
state is calculated and the recognition percentage with this database gesture is calculated via
Equation (2).

And when the matching percentage exceed 85 %, the algorithm stops immediately for saving the
time and matching is found, if not, the algorithm stays running until all gestures database are
matched and the maximum matching percentage is passed.

International Journal of Image Processing (IJIP), Volume (4): Issue (5) 463
Mokhtar M. Hasan & Pramod K. Misra

5. TESTING AND VALIDATION

We test our system with 24 different samples, 4 gestures for each pose( total of 6 poses), and the
recognition rate was 91 %, our system recognized the gestures with different transformations and
achieve this percentage of recognition, the translation and scaling solved by normalization
operation which produces a unified gesture image boundaries, rotation is managed here by
brightness value and blocking method which allows some flexibility in gesture slope, the testing
gesture is undergone in same serious of operation to get the features, after that these extracted
features will be compared with the stored database features, the maximum matching score
gesture is the equivalent gesture and it is meaning represents the meaning of the tested gesture,
Figure (11) shows the testing gestures with their maximum matching probability.

77 % 75 % 95 % 71 % 68 % 66 % 65 % 66 %

73 % 61 % 61 % 62 % 71 % 79 % 79 % 77 %

90 % 87 % 76 % 74 % 89 % 76 % 81 % 93 %
FIGURE 11: Recognition Percentage for each Tested Gesture.

In Figure (11), the underlined and bolded recognition percentage represents the non-recognized
gestures by the system, they have highest probability but referring to wrong gesture. You can
notice the rotation in these testing gestures and the system still recognizes these gestures. Figure
(12) represents the matching chart for two selected testing gestures from Figure (11) which have
matching percentage %93, %65 respectively, and first one is recognized gesture and the second
one is non-recognized gesture.

FIGURE 12: Matching Chart for Two Selected Testing Gestures.

International Journal of Image Processing (IJIP), Volume (4): Issue (5) 464
Mokhtar M. Hasan & Pramod K. Misra

In Figure (13), we have applied our recognition algorithm against all database feature vectors and
the recognition rates are shown herein below, the aim is to reveal the prominent of the recognized
gesture over all other gestures.

FIGURE 13: Matching Single Gesture against all Database Gestures.

6. CONSLUSION & FUTURE WORK

Hand gesture segmentation using HSV color model provide a promising results, this
segmentation rely mainly on the pigment color of the human skin regardless the illumination
which can be overcame by setting correct parameters for this model, also the seeking for skin
color considered to be an efficient technique specially when the input image is cluttered with
many objects, many researches commenced to enhance and develop the interaction between
the human and human-made machines, template matching uses raw data and affected by
rotation and transformation changes in addition to the illumination changes, in order to overcome
this problem we need to increase the number of samples per gesture which indicates the
increasing of the database size and in turn the processing time as well, the orientation histogram
provides a promising solution for illumination changes due to the edge direction which is not
affected with illumination variance, nevertheless, the hand object must dominate [6] the input
gesture image and the database vocabulary must be chosen to avoid some confusion [6] in
testing phase which may provide same testing result for different input gestures, by using of
neural network the number of samples must be limited due to training time and the training data
must be separable, in our method we have built a recognition system can handle a good degree
of rotation via partitioning technique which partitions the hand gesture into blocks, and then we
have applied the suggested testing algorithm.

Overall, feature selection is an important issue for gesture recognition and it is considered to be a
crucial to the recognition algorithm, image preprocessing steps also important because the
perfect steps can promise a unique and small feature vector which will reduce the number of
samples in the database and speed up the recognition time.

In this study we have achieved 91 % recognition rate using different gestures at different rotation
angles but using same conditions of illumination against uniform background, we have trained our
system with 60% of gestures and tested our system with 40%, in the future work one can use the
non-uniform background instead of uniform background.

International Journal of Image Processing (IJIP), Volume (4): Issue (5) 465
Mokhtar M. Hasan & Pramod K. Misra

7. REFERENCES
1. Wikipedia Internet Web Site

2. Internet Web Site. Available at:http://commons.wikimedia.org

3. S. Naidoo, C.W. Omlin, M. Glaser, “Vision-Based Static Hand Gesture Recognition using
Support Vector Machines”. Department of Computer Science, University of the Western
Cape, South Africa, 1999

4. K. Symeonidis. “Hand Gesture Recognition using Neural Networks”, Master Thesis,

School Of Electronic And Electrical Engineering, 2000

5. R. Brunelli, T. Poggio. “Face Recognition: Features versus Templates”. IEEE

Transactions on Pattern Analysis And Machine Intelligence, 15(10):1042-1052, 1993

6. W. T. Freeman, M. Roth. “Orientation Histograms for Hand Gesture Recognition”.

Mitsubishi Electric Research Laboratories, Cambridge, Ma 02139 USA, 1994

7. X. Li. “Gesture Recognition based on Fuzzy C-Means Clustering Algorithm”. Department

Of Computer Science The University Of Tennessee Knoxville, 2003

8. X. Li. “Vision Based Gesture Recognition System with High Accuracy”. Department of
Computer Science, The University of Tennessee, Knoxville, TN 37996-3450, 2005

9. J. Wachs, U. Kartoun, H. Stern, Y. Edan. “Real-Time Hand Gesture Telerobotic System

using Fuzzy C-Means Clustering”. Department of Industrial Engineering and
Management, Ben-Gurion University of the Negov, 1999

10. J. Triesch, C. Malsburg. “Robust Classification of Hand Postures Against Complex

Backgrounds”. IEEE Computer Society, In Proceedings of Second International
Conference On Automatic Face And Gesture Recognition, 1996

11. T. Yang, Y. Xu. “Hidden Markov Model for Gesture Recognition”. The Robotics Institute
Carnegie Mellon University Pittsburgh, Pennsylvania 15213, 1994

12. J. J. Phu, Y. H. Tay. “Computer Vision Based Hand Gesture Recognition using Artificial
Neural Network”. Faculty of Information and Communication Technology, University
Tunku Abdul Rahman (Utar), Malaysia, 2006

13. H. B. Amor, S. Ikemoto, T. Minato, H. Ishiguro. “Learning Android Control using Growing
Neural Networks”. Department Of Adaptive Machine Systems Osaka University, Osaka,
Japan, 2003

14. M. Swain and D. Ballard. “Indexing via Color Histograms”. In Proceedings of Third
International Conference on Computer Vision, 390-393, 1990

15. S. Venkataraman, V. Gunaseelan. ”Hidden Markov Models in Computational

Biology”.lectures in HMM

16. The AF Research Laboratory. “Language and Cognition”. Elsevier, Neural Networks, 22:
247-257, 2009

17. H. Gunes, M. Piccardi, T. Jan.”Face and Body Gesture Recognition for a Vision-Based
Multimodal Analyzer”. Computer Vision Research Group, University of Technology,
Sydney (UTS), 2007

International Journal of Image Processing (IJIP), Volume (4): Issue (5) 466
 Mokhtar M. Hasan & Pramod K. Misra

18. Y. Lu, S. Lu, F. Fotouhi, Y. Deng, Susan J. Brown. “A Fast Genetic K-Means Clustering
Algorithm”. Wayne State University, Kansas State University Manhattan, USA, 2000

19. B. Heisele, P. Ho, T. Poggio. “Face Recognition with Support Vector Machines: Global
versus Component-based Approach”. Massachusetts Institute of Technology Center for
Biological and Computational Learning Cambridge, 2001

20. K. Jain, R. P.W. Duin J. Mao. “Statistical Pattern Recognition: A Review”. IEEE
Transactions on Patterns Analysis and Machine Intelligence, 22(1):4-35, 2000

21. S. S. Keerthi, O. Chapelle, D. DeCoste. “Building Support Vector Machines with Reduced
Classifier”. Complexity, Journal of Machine Learning Research, 8:1-22, 2006

22. Abin – Roozgard. “Convolutional Neural Networks”. lectures in Neural Networks

23. Y. P. Lew, A. R. Ramli, S. Y. Koay, A. ali, V. Prakash. “A Hand Segmentation Scheme

using Clustering Technique in Homogeneous Background”. Student Conference on
Research and Development Proceedings, Shad Alam, Malaysia, 2002

24. C.C. Lo, S. J. Wang. “Video Segmentation using a Histogram-Based Fuzzy C-Means
Clustering Algorithm”. Institute of Information Management, National Chiao-Tung
University, Computer Standards & Interfaces, 23:429–438, 2001

25. S. Marcel, O. Bernier, J. Viallet, D. Collobert. “Hand Gesture Recognition using Input–
Output Hidden Markov Models”. France Telecom Cnet 2 Avenue Pierre Marzin 22307
Lannion, France, 1999

26. C. Karlof, D. Wagner. “Hidden Markov Model Cryptanalysis”. Computer Science Division
(EECS) Univertsity of California Berkeley, California 94720, 2004

International Journal of Image Processing (IJIP), Volume (4): Issue (5) 467

View publication stats