Person identification system for future

() IEEE Transactions on Consumer Electronics, Vol 53, No 1, FEBRUARY 2007 Manuscript received January 15, 2007 0098 3063/07/$20 00 © 2007 IEEE 218 Person Identification System for Future Digital TV wi[.]

IEEE Transactions on Consumer Electronics, Vol 53, No 1, FEBRUARY 2007

Manuscript received January 15, 2007 0098 3063/07/$20.00 © 2007 IEEE

218

Person Identification System for Future Digital

TV with Intelligence

Min-Cheol Hwang, Le Thanh Ha, Nam-Hyeong Kim, Chun-Su Park,

and Sung-Jea Ko, Senior Member, IEEE

Abstract — Intelligent digital TV (iDTV) is a future digital

TV with intelligence which can automatically provide

personalized services for each audience For the

user-personalized services, the iDTV should recognize audiences in

real-time In this paper, we define a novel structure of the

iDTV and propose a real-time person identification system in

the iDTV that analyzes captured images and recognizes

audiences The proposed system consists of three processing

units: preprocessing for reducing computational costs of the

proposed system, face detection using a statistical approach,

and face recognition using Support Vector Machines (SVMs)

Experimental results show that the proposed system achieves

efficient performance with high recognition accuracy of 90%

or higher at the speed of 15~20 fps, which is suitable for the

iDTV 1

Index Terms — Intelligent digital TV, personalized services,

audience identification, face detection, face recognition

I INTRODUCTION

Recently with the development of compression technology

and the digitization of TV programs, the digital TV (DTV)

provides not only high quality audio-visual effects as in high

definition TV (HDTV) but also datacasting and multimedia

interactive services According to the increase of the amount

of services provided by the DTV, conventional methods of

channel selection such as browsing become impractical [1]

The electronic program guide (EPG) can help viewers check

future programs in advance However, the multi-channel DTV

service delivers more the number of programs than that of

programs which viewers can handle That results in

information overload for users [2] The ability of the DTV to

provide user-personalized services for the individual person

automatically is required In order to provide the personalized

service, a future DTV with intelligence (iDTV) needs to

identify the audiences

Considering the above requirements and problems, we

define a novel structure of the iDTV as illustrated in Fig 1

The proposed iDTV consists of two main components:

intelligent environment sensor manager (iESM) and intelligent

EPG manager (iEPGM) The iESM equipped with various

sensors investigates the indoor environment including

audiences, light, illumination, and noise The iESM performs

the person identification by using the gathered information

1 M.-C Hwang, L T Ha, N.-H Kim, C.-S Park, and S.-J Ko are with the

Department of Electronic Engineering, Korea University, Seoul, 136-701,

Korea (e-mail: sjko@dali.korea.ac.kr)

through the sensors The iEPGM manages the list of programs which is appropriate for audiences and provides the user-personalized service to each person by using the information received from the iESM For example, for those who like to watch the sports, the iEPGM turns on the sports channel automatically

Person Identifier

iESM

Personalized Service Provider

iEPGM

Sensor (Camera) Input images

DB

iDTV

1 Moving object tracking

3 Personalized services

2 Person ID

Audience

Fig 1 The novel structure of the iDTV

Various identification technologies have been widely used for commercial purpose The most common personal verification and identification methods are the password/PIN (personal identification number) and token systems [3] Because those systems are vulnerable to forgery, theft, and lapses in users’ memory, biometric identification systems using pattern recognition techniques are attracting considerable interest For example, fingerprint, retina, and iris are used for identification technologies The above biometric technologies have high person identification accuracy for a large group of authorized members However, these techniques are not appropriate for the iDTV due to their inevitable drawbacks

Compared with the above identification methods such as the fingerprint and iris recognition, face recognition has several appropriate characteristics, such as non-intrusive and user-friendly interfaces, low-cost sensors, easy setup, and active identification, for consumer applications [4] Generally, the system using human faces for person identification is separated into two parts: face detection and face recognition There exists much effort to improve the performance of face detection [5]-[7] and face recognition [3][4][8]-[14] Although considerable successes have been reported, it is still a difficult task to design an automatic face recognition system in real-time In this research, we focus only on the real-time person identification module in the iESM, and the others are left for

M.-C Hwang et al.: Person Identification System for Future Digital TV with Intelligence 219

future research The proposed system consists of three phase:

preprocessing for reducing computational costs of the

proposed system, face detection using a statistical approach,

and face recognition using Support Vector Machines (SVMs)

The paper is organized as follows In Section II, the proposed

person identification system is explained Section III presents

the experimental results and the performance analysis of the

overall system Finally, the paper concludes in Section IV

II P ROPOSED P ERSON I DENTIFICATION S YSTEM FOR T HE

I DTV

Fig 2 portrays the processing flow of the proposed person

identification system for the iESM The proposed system

consists of three processing units The first unit is the

preprocessing for reducing computational costs of next unit In

the preprocessing, moving objects are detected and tracked The

second is the face detection unit using a statistical approach with

Haar-like features The region of moving objects are checked

whether they contain faces or not independently in the face

detection unit Haar-like features are efficient to detect faces

with low costs, especially frontal faces [7] These Haar-like

features are suitable for the proposed system because most

captured faces of audiences are frontal The last is the face

recognition using SVMs [11] Compared with other learning

methods such as eigenfaces [12], hidden markov models

(HMM) [13], and neural network (NN) [14], the training time of

SVMs is short enough to be applied to the real-time iDTV Our

proposed system can add a set of face images with new identity

in the database and retrain it in real-time The details of each

phase are described in the following subsections

A Preprocessing

The detector is applied to every location in the input image in

order to find faces regardless of the position of the faces To

detect faces larger or smaller than the window size, the input

image is scaled and the detector is applied to all the scaled

images As a result, the computational complexity increases

drastically In the preprocessing, we propose the methods which

can reduce the search range and the number of image scaling

iterations for the face detection

Generally, people are considered as moving objects We are

able to reduce the search range of face detection by considering

only the area related to the moving objects Background

subtraction [15] which has been used to extract moving objects

in many applications is applied In our system, background is

easily estimated because the camera is attached to the iDTV in

the house Background subtraction and thresholding are performed to produce difference images The moving object is extracted from the difference image using a morphological opening, erosion followed by dilation, to remove small clusters This computational cost is much lower than the face detection method described in next subsection Background subtraction is not suitable where many objects are moving at the same time However, our iDTV considers a small family composed of only

a few members That performance is efficient enough to be adopted into our proposed system

Using the stereo image recognition system, we can obtain the distance information of each object from the cameras Face size

is estimated by using the relationship between the distance from cameras to an object and the size of the object Therefore the face detector deals with only one scaled input image The detail algorithm used for the distance detection is presented in [16] and [17]

Fig 3 Extended set of Haar-like features: (a) edge features (b) line features (c) center-surround features

B Face Detection

We use a statistical approach for faces detection, the approach originally developed by Viola and Jones [7], and then analyzed and extended by Lienhart and Maydt [18] A cascade of boosted tree classifiers as a statistical model and Haar-like features which are computed similar to the coefficients in Haar wavelet transforms are used in this method A highly accurate or strong classifier can be produced through the linear combination of many inaccurate or weak classifiers [19] Every weak classifier is trained by using a single feature and checks whether an object region at a certain location looks like a face or not

Each feature is described by the template, its coordinate relative to the search window origin and the size of the feature

In [18], the 14 different templates are used as shown in Fig 3

Fig 2 The processing flow of the proposed person identification system

IEEE Transactions on Consumer Electronics, Vol 53, No 1, FEBRUARY 2007

220

Each feature consists of two or three joined “black” and

“white” rectangles either up-right or rotated by 45o The value

of each feature is calculated as a weighted sum of two

components: the pixel sum over the black rectangle and the

sum over the whole feature area which are all black and white

rectangles The weights of these two components are of

opposite signs and for normalization, their absolute values are

inversely proportional to the areas: for example, the black

features in Fig 3(c) have weight black = -9 × weight whole

There exist hundreds of features in real classifiers

Computing pixel sums over multiple small rectangles is too

slow to detect faces in real-time Viola [7] introduced an

elegant method to compute the sums very fast First, an

integral image, summed area table (SAT), is computed over

the whole image I, where

,

x X y Y

< <

= ∑ (2)

The pixel sum over a rectangle r = {(x,y), x 0 x<x 0 +w,

y 0 y<y 0 +h} can then be computed using SAT by using just

the corners of the rectangle regardless of size:

This is for up-right rectangles For rotated rectangles, a

separate rotated integral image whose axes are diagonal lines

is used

The computed feature value, x i = w i,0 RecSum(r i,0 ) +

w i,1 RecSum(r i,1), is used as input to a weak classifier that

usually has just two terminal nodes, i.e.,

i i i

i i

f

⎧

= ⎨− <⎩ (4)

or three terminal nodes:

i

f

else

⎧

= ⎨−

⎩ (5) where the response +1 means the face and -1 means the

nonface

In the next step, a complex and robust classifier is built out

of multiple weak classifiers using a procedure called boosting

introduced by Freund and Schapire [19] Formally, given an

image X, a boosted classifier is defined as a signed linear

combination of the outputs of a number of weak classifiers

iteratively:

1

n

i i i

F X sign c f X

=

⎝∑ ⎠ (6)

On each iteration, a new weak classifier f i is trained and

added to the sum The smaller the error f gives on the training

set, the larger the coefficient c i that is assigned to it The weight of each training sample is updated, so that on the next iteration the role of those samples that are misclassified by the

already built F is emphasized In general, the performance of

an individual weak classifier may be only slightly better than

random It is proven in [19] that F can achieve an arbitrarily

high (<1) hit rate and an arbitrarily small (>0) false alarm rate,

if the number of weak classifiers is large enough However, in practice, this boosted classifier requires not only a number of weak classifiers but also a very large training set That results

in a slow processing speed

Instead, Viola [7] suggests building several boosted

classifiers F k with constantly increasing complexity and chaining them into a cascade with the simpler classifiers going first During the detection stage, the current search window is

analyzed subsequently by each of the F k classifiers that may

reject it or let it go through In other words, F k ’s are

subsequently applied to the face candidate until it gets rejected

by one of them or until it passes them all as shown in Fig 4

Fig 4 Face detection cascade of classifiers where rejection can happen at any stage

C Face Recognition

In this subsection, we explain our face recognition method using SVMs which are recently shown to be effective learning mechanisms for object recognition We introduce the some fundamentals of SVMs and some of our work on exploring the hypothesis space of SVMs for face recognition The very details can be referenced in [20]

1) Support Vector Classification

By defining hyperplanes in a high-dimensional feature space, SVM builds complex decision boundaries to learn the distribution of a given data set In linearly separable binary

case, given a data set {(x i ,y i)}, i=1, , ,l x∈R n, y∈ − +{ 1, 1},a hyperplane such as w x⋅ + =b 0, w∈R ,n b∈R can be oriented across the input space to perform a binary classification task minimizing the empirical risk of a hyperplane decision function f x( )=sign( w x⋅ +b) This is

achieved by changing the normal vector w, also known as the

weight vector to maximize the functional margin

2

1 ,

w

γ = (7)

M.-C Hwang et al.: Person Identification System for Future Digital TV with Intelligence 221

where exists usually a margin on either side of the hyperplane

between the two classes The corresponding problem can be

represented as follows:

( )

,

w b

⋅

The above always produces hypothesis with the perfectly

linear separable data set {(x i ,y i)} It cannot deal with the noisy

data, there exists no linear separation To do that, we need to

introduce slack variables ξ to allow the margin constraints to

be violated, and a coefficient C to penalize these variables

And we obtain the C-SVM problem as follows:

,

1

l i

w b

i

C

ξ

ξ ξ

=

Its Lagrangian is

( )

1

1 ( , , , , )

2

l i i

L w b r w w C

y w x b r

=

∑

(10)

with αi( 0) and r i( 0) are the Lagrangian multipliers The

corresponding dual is obtained by differentiating with respect

to w and b and resubstituting the relations into (10), and then

the dual problem is

1

1

2

i i j i j i j

l

i

=

∑ ∑

∑

(11)

where C is chosen by user, a larger C corresponds to assigning

a higher penalty to errors

The optimal hyperplane is mainly defined by the weight

vector w which consists of all the data elements with non-zero

Lagrange multipliers (αi) in (11), those elements lie on the

margins of the hyperplane They define both the hyperplane

and the boundaries of the two classes The decision function

of the optimal hyperplane is

1

l

i i i i

=

Suppose parameters α solve the above problem, then the *

weight vector and bias can be obtained by

1

*

,

2

l

i i i i

w y x

w x w x b

α

=

=

= −

∑

(13)

Only inputs, x i ’s lying closest to the hyperplane, are

corresponding non-zero α*i 's They are called support vectors

A hyperplane decision function attempts to fit an optimal hyperplane between two classes in a training data set, which will inevitably fail in cases where the two classes are not linearly separable in the input space Therefore, a high dimensional mapping:

Φ → (14)

is used to cater for nonlinear cases As decision function is

expressed in terms of dot products of data vectors x, the

potentially computational intensive mapping ( )Φ ⋅ does not

need to be explicitly evaluated A kernel function, K(x;z),

satisfying Mercer’s condition can be used as a substitute for ( )x ( )z

objective function becomes

1

1

2

l

i

=

∑ ∑

∑

(15)

with the corresponding decision function given by

1

l

i i i i

=

There exist a number of kernel functions such as polynomials and radial basis function (RBF) which have been found to provide good generalization capabilities Here we explore the use of a RBF kernel function as follows:

2 2

2

K x y

σ

(17)

Training SVM is necessary to find the Lagrangian multipliers αi to maximize the functional margin γ.There exist some learning algorithms for training an SVM Sequential minimal optimization (SMO) is considered as the most powerful algorithm among the conventional methods Because

it is ease of use and better scaling with training set size, SMO

is strong candidate for becoming the standard SVM training algorithm [21]

The primal version of SVM is for binary classifying Some methods have been proposed for SVMs in multi-classification,

IEEE Transactions on Consumer Electronics, Vol 53, No 1, FEBRUARY 2007

222

some of them are one-against-all, one-against-one, and

DAGSVM which combine several binary SVM

One-against-all strategy classifies between each class and One-against-all the remaining

and one-against-one strategy classifies between each pair The

strategy of DASGSVM is the same as that of one-against-one,

but its decision works with a directed acyclic graph (DAG)

Hsu and Lin [22] suggested that one-against-one and DAG

approaches are more suitable for practical use

Frontal view

face image

Illumination

& Histogram equalization

Resampled to 100x100

Fig 5 Normalization of face image for recognition

Fig 6 Accuracy distribution with C and gamma

2) Face Recognition using SVMs

The face image of each person is normalized by correcting

illumination and equalizing histogram, and resampled to size

of 100×100 pixels as shown in Fig 5 The face in the image

must be in frontal view

Multiclass C-SVM with kernel RBF is used with 10000

inputs corresponding to the number of pixels of the input

image The number of outputs is the number of people that

need to be classified A sample is a pair (x,y), where

x∈ [0:1]10000

and y is an integer number representing person

ID Vector x is constructed from the image by x(100×i+j) =

I(i,j), I(i,j) is the gray value at position (i,j) of the image After

trained, C-SVM automatically learns features of each subject

and can predict the person ID of the test image

We construct our own face database which has 5 subjects;

each has 24 face images for learning To explore the

hypothesis of SVMs for face recognition, we find the two

parameters C for C-SVMs and gamma for RBF kernel by

using 6-fold cross-validation method, which partitions the

input samples into subsamples such that analysis is initially

performed on a single subsample, while further subsamples

used for validating the initial analysis The distribution of correct class rate in percent is depicted in Fig 6 We choose

C=0.4 and gamma=0.5 as the optimum parameters for SVMs

in face recognition

III E XPERIMENTAL R ESULTS

The proposed person identification system has been implemented to operate at about 15~20 frames per second (fps), with a frame resolution of 320×240 pixels, using a Pentium IV PC with 2.6GHz CPU In order to evaluate the accuracy of the proposed system, we have separately examined face detection phase including preprocessing and face recognition phase Each phase has its own accuracy and the accuracy of whole system is obtained by multiplying above two kinds of accuracy

A Accuracy of Face Detection

Three testing data sets are used to examine the face detection phase SET12 which is from the IITK-CSE test sets consists of 188 images Each image in SET1 contains only one frontal face as shown Fig 7 SET23 which is from the Humanscan AG consists of 1521 images All images in SET2 contain one frontal face with background as shown Fig 8 SET3 which is created by us consists of 135 images containing a total of 283 faces as well as nature scenery as shown in Figs 9 and 10

Most faces in SET1 and SET2 can be detected with a few false detections because faces in SET1 and SET2 are only frontal faces Not only frontal faces but also slightly rotated (about ±20 degrees) and turned (about ±40 degrees) faces are well detected by the face detection module SET3 contains real world images with various sizes and appearances, and complex background Figs 9 and 10 show the examples of face detection results from SET3 In Fig 9, all the faces are detected correctly and accurately Especially the low quality image such as Fig 9(h) can also be detected Fig 10 shows some other examples of false detections and missed faces False detections occur in Figs 10(a) and (c) Faces with a hat

in Fig 10(b) and a face with complex background in Fig 10(d) are not detected Table I summarizes the detection performance for each testing data set The overall face detection performance of the face detection phase using the 1,840 images containing a total of 1,992 faces is 96.2% with

329 false detections We assume that the proposed system is attached to the iDTV False detections can be removed easily because the background may be fixed and predictable The performance of the face detection phase is good enough to operate in the proposed iDTV

2 The 188 images (are from Website: http://vis-www.cs.umass.edu/~vidit/ AI/dbase.html) have been created by Vidit Jain and Dr Amitabha Mukherjee

in cooperation with Neeraj Kayal, Pooja Nath and Utkarsh Hriday Shrivastav

3 The 1521 images are from Website: http://www.BioID.com

M.-C Hwang et al.: Person Identification System for Future Digital TV with Intelligence 223

TABLE I FACE DETECTION PERFORMANCE OF THE PROPOSED SYSTEM

Data set # faces # detected

faces

# false detections

Detection rate

B Accuracy of Face Recognition

As mentioned in the previous section, by using

cross-validation method, our system appears to have the accuracy of

recognition about 98% with suitable parameters C and gamma

An experimental program is implemented in the porposed

system It allows a person to sit in front of the screen and

automatically counts the number of frames and computes the

accuracy of recognition, while he is looking at the screen as if

he is watching the iDTV The program gives out the accuracy

about 95%

Considering the application, we also do experiment with 4

people in 3 different lighting conditions such as full, half and

low light The accuracy of our face recognition phase is

summarized in Tables II and III We use two kinds of the

training sample sets One set is obtained in full light

environment and the other is taken in low light condition The

recognition accuracy depends on the environment where we

obtain the training samples When the light condition of taking

training samples is the same as that of testing, the accuracy of

recognition is the highest The recognition accuracy of the low

light training data set is better than that of the full light one It

suggests that the training data set should be obtained in the

condition where light effects are minimized

The results of the face recognition are shown in Figs 11

and 12 The proposed face recognition module can well recognize a face in different light condition as shown in Fig

11 Two or more faces can be recognized simultaneously and independently as shown in Fig 12 Fig 12 also shows the GUI of our proposed system

TABLE II FACE RECOGNITION RESULTS USING FULL LIGHT TRAINING DATA SET Light condition P1 P2 P3 P4 AVG

TABLE III FACE RECOGNITION RESULTS USING LOW LIGHT TRAINING DATA SET Light condition P1 P2 P3 P4 AVG

IV C ONCLUSIONS

In this paper, we have defined the novel structure of the iDTV and have proposed the real-time person identification system for the iDTV To identify a person, the proposed system has used a face recognition algorithm which consists

of preprocessing, face detection, and face recognition In the preprocessing, objects are tracked and the distance for each object is calculated to reduce computational costs of detecting faces A statistical approach with Haar-like features has been used for the face detection and SVMs have been chosen for the face recognition Those are efficient to detect and recognize faces in real-time Experimental results have shown that the proposed system achieves the performance with a total

Fig 7 Face detection results from SET1

Fig 8 Face detection results from SET2

IEEE Transactions on Consumer Electronics, Vol 53, No 1, FEBRUARY 2007

224

(a) (b) (c)

(d) (e) (f) (g) (h)

Fig 9 Face detection results from SET2

(a) (b)

(c) (d)

Fig 10 Examples of false detection and missed faces from SET2

recognition rate of 90% or higher and 15~20 fps processing

speed It is efficient enough to operate in the iDTV

The major contributions of the proposed system are the

introduction of the novel structure of the iDTV as the future

intelligent electronic home devices, and researches for

merging face detection and face recognition into one system

and applying it to consumer electronic device, the iDTV

Using the proposed system, each audience can enjoy the

personalized services without any inconvenience

The current face recognition system is able to recognize

frontal faces and slightly rotated (about ±20 degrees) and

turned (about ±40 degrees) faces Large rotated and turned

faces are not detected and recognized in the proposed system The accuracy of the face recognition is about 90% Future work still remains to treat large head rotations to facilitate more natural user interfaces as well as to increase the accuracy and performance of the person identification

(a) (b) (c) Fig 11 Results of face recognition in different light condition: (a) full light (b) half light (c) dark

M.-C Hwang et al.: Person Identification System for Future Digital TV with Intelligence 225

Fig 12 The proposed system GUI and result of three people recognition

A CKNOWLEDGEMENT

This research was supported by Seoul Future Contents

Convergence (SFCC) Cluster established by Seoul

Industry-Academy-Research Cooperation Project

R EFERENCES

[1] T Isobe, M Fujiwara, and H kaneta, “Development and features of a

TV navigation system,” IEEE Trans Consumer Electronics, vol 50, no

1, pp 393-399, Nov 2003

[2] H Zhang, S Zheng, and J Yuan, “A personalized TV guide system

compliant with MHP,” IEEE Trans Consumer Electronics, vol 51, no

2, pp 731-737, May 2005

[3] A Pentland and T Choudhury, “Face recognition for smart

environments,” IEEE Computer, vol 33, no 2, pp 50-55, Feb 2000

[4] F Zuo and P H N de With, “Real-time embedded face recognition for

smart home,” IEEE Trans Consumer Electronics, vol 51, no 1, pp

183-190, Feb 2005

[5] M H Yang, D Kriegman, and N Ahuja, “Detecting faces in images: a

survey,” IEEE Trans Pattern Anal Mach Intell, vol 24, no 1, pp

34-58, Jan 2002

[6] C A Waring and X Liu, “Face detection using spectral histograms and

SVMs,” IEEE Trans Syst., Man Cybern B, vol 35, no 3, pp 467-476,

Jun 2005

[7] P Viola and M Jones, “Rapid object detection using a boosted cascade

of simple features,” IEEE CVPR, vol 1, pp 511-518, 2001

[8] R Chellappa, C L Wilson, and S Sirohey, “Human and machine

recognition of faces: a survey,” Proc IEEE, vol 83, no 5, pp 705-740,

May 1995

[9] R Brunelli and T Poggio, “Face recognition: features versus templates,”

IEEE Trans Pattern Anal Mach Intell., vol 15, no 10, pp 1042-1053,

Oct 1993

[10] M J Er, W Chen, and S Wu, “High-speed face recognition based on

discrete conise transform and RBF neural networks,” IEEE Trans

Neural Netw., vol 16, no 3, pp 679-691, May 2005

[11] G Guo, S Z Li, and K Chan, “Face recognition by support vector

machines,” Proc the IEEE Int Conf Automatic Face and Gesture

Recognition, pp 196-201, Mar 2000

[12] M Turk and A Pentland, “Eigenfaces for recognition,” J Cogn

Neurosci., vol 3, pp 71-86, 1991

[13] A V Nefian and M H Hayes III, “Hidden markov models for face

recognition,” Proc the IEEE Int Conf Acoustics, Speech, and Signal

Processing, pp 2721-2724, May 1998

[14] S Lawrence, C L Giles, A C Tsoi, and A D Back, “Face recognition:

a convolutional neural network approach,” IEEE Trans Neural Netw.,

vol 8, no 1, pp 98-113, Jan 1997

[15] J W Kim, K S Choi, B D Choi, J Y Lee, and S J Ko, “Real-time system for counting the number of passing people using a single

camera,” Lecture Notes in Computer Science, vol 2781, pp 466-473,

Sep 2003

[16] K Saneyoshi, “Driving assist system using stereo image recognition,”

IEEE Proc Intelligent Vehicles Symposium, pp 230-235, Sep 1996

[17] S Nedevschi, R Danescu, D Frentiu, T Marita, F Oniga, C Pocol, R Schmidt, and T Graf, “High accuracy stereo vision system for far

distance obstacle detection,” IEEE Proc Intelligent Vehicles

Symposium, pp 292-297, June 2004

[18] R Lienhart and J Maydt, “An extended set of Haar-like features for

rapid object detection,” IEEE Proc Int Conf Image Processing, vol 1,

pp 900-903, Sep 2002

[19] Y Freund and R E Schapire, “Experiments with a new boosting

algorithm,” Machine Learning: Proc the 13th Int Conf pp 148-156,

1996

[20] Cristiana and S Taylor, An introduction to Support Vector Machines,

Cambridge University Press, UK, 2000 [21] J Platt, “Sequentional minimal optimization: a fast algorithm for

training support vector machines” Technical Report, pp 98-14,

Microsoft Research, 1998

[22] C W Hsu and C J Lin, “A comparison of methods for

multiclasssupport vector machines,” IEEE Trans Neural Netw., vol 13,

no 2, pp 415-425, Mar 2002

Min-Cheol Hwang received the B.S degree in

Electronics Engineering from Korea University in 2003

He is now a Ph.D candidate in Electronics Engineering with the Department of Electronics Engineering at Korea University His research interests are in the areas of image processing, multimedia communications, and image compression coding such as H.264 and JPEG

2000

Le Thanh Ha reveived the B.S degree from Falcuty of

Technology, Vietnam National University (VNU) in

2002 He completed the M.S degree in Information Technology Department, College of Technology, VNU

He is now a Ph.D candidate in Electronic Engineering with the Department of Electronic Engineering at Korea University His current research interests are in the areas

of Image/Video Processing and Machine learning

Nam-Hyeong Kim received the B.S and M.S degrees

in electronics engineering with the Department of Electronics Engineering from Korea University, Seoul,

in 2000 and 2002, respectively He is currently pursuing the Ph.D degree in electronics engineering at Korea University His research interests are in the areas of video signal processing, and multimedia communications

Chun–Su Park received the B.S degree from Korea

University, in Electronics Engineering, in 2003 He is now a Ph.D candidate in electronic engineering with the Department of Electronic Engineering at Korea University His research interests are in the areas of Mobile QoS, IP QoS, Handoff, video signal processing

and multimedia communications

IEEE Transactions on Consumer Electronics, Vol 53, No 1, FEBRUARY 2007

226

Sung-Jea Ko (M’88-SM’97) received the Ph.D degree

in 1988 and the M.S degree in 1986, both in Electrical and Computer Engineering, from State University of New York at Buffalo, and the B.S degree in Electronic Engineering at Korea University in 1980 In 1992, he joined the Department of Electronic Engineering at Korea University where he is currently a Professor From

1988 to 1992, he was an Assistant Professor of the Department of Electrical and Computer Engineering at the University of Michigan-Dearborn He has published more than 300 papers

in journals and conference proceedings He also holds over 20 patents on

video signal processing and multimedia communications

He is currently a Senior Member in the IEEE, a Fellow in the IEE and a

chairman of the Consumer Electronics chapter of IEEE Seoul Section He has

been the Special Sessions chair for the IEEE Asia Pacific Conference on

Circuits and Systems (1996) He has served as an Associate Editor for Journal

of the Institute of Electronics Engineers of Korea (IEEK) (1996), Journal of

Broadcast Engineering (1996 - 1999), the Journal of the Korean Institute of

Communication Sciences (KICS) (1997 - 2000) He has been an editor of

Journal of Communications and Networks (JCN) (1998 - 2000) He is the

1999 Recipient of the LG Research Award given to the Outstanding

Information and Communication Researcher He received the Hae-Dong best

paper award from the IEEK (1997) and the best paper award from the IEEE

Asia Pacific Conference on Circuits and Systems (1996) He received the

research excellence award from Korea University (2004)