Welcome to Scribd, the world's digital library. Read, publish, and share books and documents. See more
Download
Standard view
Full view
of .
Look up keyword
Like this
5Activity
0 of .
Results for:
No results containing your search query
P. 1
face recognition using local dct

face recognition using local dct

Ratings: (0)|Views: 289|Likes:
Published by Akrm Sied
face recognition using local dct
face recognition using local dct

More info:

Published by: Akrm Sied on Dec 27, 2009
Copyright:Attribution Non-commercial

Availability:

Read on Scribd mobile: iPhone, iPad and Android.
download as PDF, TXT or read online from Scribd
See more
See less

03/04/2011

pdf

text

original

 
AbstractThis paper presents the results of a project investigating theuse of Discrete-Cosine transform (DCT) to represent facialimages of an identity recognition system. The objective was toaddress the problem of sensitivity to variation in illumination,faced by commercially available systems. The proposedmethod uses local, block-wise DCT to extract the features of the image. Results show that this appearance based methodgives high identification rates, improved tolerance to variationin illumination, and simple implementation.
I.
 
I
NTRODUCTION
 Biometric signatures, which characterize a personwith respect to their body features, cannot be lost, stolen, orshared with others [1], unlike commonly used access cardsand passwords. Consequently, there has been an increasingmotivation to development identification systems based onthese signatures. Applications for such systems includeperson identification for use by law enforcement agencies[1], facilitation of human-machine interactions in robotics,and identification for access to secure areas or sensitiveinformation. While a range of commercial systems arecurrently available, they are documented to demonstratepoor performance under extreme variations in environment[2], and therefore there is a need for further development of these systems. Additionally, for applications such as smartcards for ATMs, training needs to be able to be done overthe life of the system. Therefore independence from thelarger database is preferable.Principle Component Analysis (PCA) and LinearDiscriminant Analysis (LDA) are baseline methods oftenused for comparison. They are based on the Karhunen-Loeve Transform (KLT), which was first used forrepresenting faces by Kirby and Sirovich (1990). InPCA[3], Turk and Pentland reduced the calculationcomplexity of their approach by only selecting the mostsignificant coefficients to represent the face. In LDA[4],the Fisher’s linear discriminant was applied to the featurevector space found via KLT to find the most discriminatingfaces (rather than the most expressive ones). The KLTminimizes the mean squared error for data compression,and has the most information in the fewest number of transform coefficients, however it is sensitive to variationsin face pose, position, and illumination. It also has theadded disadvantage that its basis set is data-dependent, withdetermination of this set requiring knowledge of thecomplete reference database. As a result, training is alsocomputationally expensive.As an alternative to KLT for representing the featurespace, the Discrete-Cosine Transform (DCT) has foundpopularity due to its comparative concentration of information in a small number of coefficients, andincreased tolerance to variation of illumination. The DCT-based approach offers other advantages over KLT-basedmethods including improved processing time, potentialdimensionality reduction and compatibility with dataalready encoded via DCT coefficients such as jpegcompressed digital video and images. Unlike KLT-basedmethods, the DCT basis vectors are data independent,making it much more suitable for systems where thereference database is dynamic over the life of the system.A number of different implementations using DCTsto represent facial images have been developed in recentyears. In Hafed and Levine [5] the DCT of the entire,normalized image is extracted. A subset of the mostsignificant coefficients were retained as the feature vector,with nearest-neighbour classification then used to identifythe face. Using the first 64 coefficients, which have low tomid range frequencies and the most variance, a very highrecognition rate was reported.The 2-D DCT method described in Sanderson [6]uses a block-by-block approach. The image is broken into8x8 pixel blocks, overlapping by 50%. The first 15 2D DCTcoefficients (ordered in a zigzag pattern –see Figure 1) werefound and combined at the decision level. Variations alsoproposed in Sanderson [6] included discarding of the firstthree DCT coefficients used to represent each block; andthe use of delta functions to replace the first three,illumination sensitive coefficients in each block.In an alternative method, Ekenel and Stiefelhagen[7] used DCTs for localized appearance-based featurerepresentations, where the detection of face features such aseyes, etc. is not required. In their method, the image wasbroken into 8x8 pixel non-overlapping blocks, which werethen represented by their DCT coefficients. The firstcoefficient (DC coefficient) was discarded and of theremaining coefficients, only the most significant were kept.In this paper, results from a project implementing afacial recognition system using local appearance DCTfeature extraction are presented. It is shown that theapproach used provides improved tolerance to variation inillumination, with a good identification rate using nearest-neighbour classification. In section II, the proposed local-DCT feature extraction method is described. Section IIIshows results of testing, and finally in section IV,conclusions and recommendations for future work aregiven.
Local-DCT features for Facial Recognition
Signal Processing Laboratory, Griffith University, Brisbane, AUSTRALIA.Belinda Schwerin, Kuldip Paliwal.
 
    
+
    
+=
= =
qn M  pm A B
 M m nmnq pq p
2)12( cos2)12( cos
101-N0.
π π α α 
II.
 
M
ETHODS
 
Discrete-Cosine Transform
The main features of the DCT which make itattractive for face recognition are:1.
 
data compression (energy compaction) property,commonly used in image/video compression, and2.
 
efficient computation, due to its relationship to theFourier transform.The DCT was first introduced by Ahmed, Natarajan, andRao in 1974, with categorization into 4 variations by Wang(1984). This invertible transform represents a finite signalas a sum of sinusoids of varying magnitudes andfrequencies. For images, the 2-dimensional DCT is used,with the most visually significant information beingconcentrated in the first few DCT coefficients (ordered in azigzag pattern from the top left corner - see Figure 1).
Figure 1: Labelling of DCT coefficientsof 5x5 image block.
Information concentration in the top left corner isdue to the correlation between image and DCT properties.For images, the most visually significant frequencies are of low-mid range, with higher frequencies giving finer detailsof the image. For the 2-D DCT, coefficients correspond tosinusoids of frequencies increasing from left to right andtop to bottom. Therefore those in the upper corner are thecoefficients associated with the lowest frequencies. Thefirst coefficient, known as the DC coefficient, represents theaverage intensity of the block, and is the most affected byvariation in illumination.For the DCT, any image block 
 A
m,n
 
, of size MxN,can be written as the sum of the MN functions of the form:, 2)12( cos2)12( cos
    
+
    
+
 N qn M  pm
q p
π π α α 
 for
1-Mp0
,
10
q
.These are the basis functions of the DCT.The DCT coefficients
 B
 p,q
are the weights applied toeach basis function for reconstruction of the block. Thecoefficients are calculated as:for
10
M- p
,
10
N-q
,where
==
11 ,  0 , 
21
 M- p p
 M  M  p
α 
,
==
11 ,  0 , 
21
 N-qq
 N  N q
α 
,and
 A
mn
 
is the MxN input matrix.The result is MN 2-D DCT coefficients.
Proposed Local-DCT method:
The proposed feature extraction method uses a localappearance based approach. This provides a number of advantages. Firstly, the use of local features allows spatialinformation about the image to be retained. Secondly, foran appropriately selected block size, illumination affectingthe block can be assumed quasi-stationary. Consequently,the resulting recognition system can be designed forimproved performance in environments where variation inillumination occurs across images.The local-DCT method operates on normalizedimages where faces are aligned (e.g. by eyes) and of thesame size. To extract the features of the image, blocks of size
b
x
b
are used. Neighbouring blocks overlap by(
b
-1)×(
b
-1) pixels, scanning the image from left to right andtop to bottom. The number of blocks processed is then:
))1(())1((
b N b M 
.For each block, the 2-dimensional DCT is then found and
c
 f 
 
coefficients are retained. The coefficients found for eachblock are then concatenated to construct the feature vectorrepresenting the image. For
c
 f 
retained coefficients andsquare image, the dimensionality of this feature vector isthen:
=
 f 
cb N 
×
2
))1((
 where
= feature vector dimensionality
 N 
x
 N 
= original image resolution(
 D
= original image dimensionality =
 N 
2
),
b
x
b
= block size, and
c
 f 
= number of coefficients retainedfrom each block.The retained coefficients from each block are those whichcontain the most significant information about the image.Since the first coefficient is the most affected by variationin illumination, it is excluded. Coefficient selection, choiceof block size, and number of coefficients retained isdiscussed in part III.A brief description of other methods used for comparison isgiven below.
Whole Image method
The image intensities themselves can be used to construct afeature vector by concatenating the columns of the image.0 1 5 6 142 4 7 13 153 8 12 16 219 11
1
720 22
1
018 19 23 24
 
SS
 B
The resulting vector has dimensionality MN for imageresolution MxN.
PCA Eigenfaces
For a reference database of R images, each of size
 
NxN, thecolumns of each image are concatenated to form columnvectors
i
 f 
, for
i
=1,2,…,R. The average face
m
 f 
is thencalculated, and subtracted from each of the training faces inthe database to give zero-mean vectors
i
g
=
i
 f 
-
m
 f 
, for
i
= 1,2,…,R. The mean training set is then represented asthe matrix G = [
 ...
21
], and the covariance matrix isthen
=
G G
.
For reduced computation, the eigenvectorsof C can be equivalently found by multiplying
G
by theeigenvectors of 
G
G
. For a feature space with
 dimensions, the
most significant eigenvectors (with thelargest eigenvalues) are selected and describes the featurespace
]...[
21
uuu
=
, where
i
u
are the eigenfaces.Finally, the zero-mean vectors of each face are projectedonto this feature space to find their feature vectors
ii
g
=
 β 
,
i
= 1,2,…,R.
LDA Fisherfaces
For a database of R images, comprising P different people,and image dimensionality of N
2
, the LDA method requiresthat R
N
2
for the within-class scatter matrix to besingular. For smaller databases, PCA must first be appliedto reduce the dimensionality of each face to R, then theLDA method can be applied as follows.For a database of faces represented in columnvectors
i
 f 
, for
i
=1,2,…,R, the average of all faces
m
 f 
,and the average face for each person
 j
 p
,
 j
= 1,2,…,P,training faces are represented as the difference from theiraverage face
 jii
p f h
=
. The scatter matrices for eachperson (or class) can then be constructed as
ii j
hhS
=
, for each image vector
i
h
of person
 j
.The within-class scatter matrix S
W
and the between-classscatter matrix S
B
can then be found:
=
 j j
SS
, for
 j
= 1,2,…,P.
m j jm j B
 p f  pS
)()(2
=
, for
 j
= 1,2,..,P.The matrix W which maximizes hascolumns which are the eigenvectors of 
.
1
 B
SS
Since thecolumns of 
are eigenvectors satisfying :
iwii B
wSwS
λ 
=
, the eigenvalues
i
λ 
are found by solving
0
=
wi B
SS
λ 
and the eigenvectors are found by solving
( )
.0
=
iwi B
wSS
λ 
Matrix
then defines the LDA-space. The feature vector for each person in this space isthen found by projecting faces onto it.III.
 
E
XPERIMENTS
&
 
D
ISCUSSION
 The Purdue AR Face database [8], contains 100 individuals,each with 14 images (excluding those with occlusions suchas sunglasses and scarves) of various illuminations andfacial expressions. For each person, 7 of these are used fortraining and 7 for testing. The original images have aresolution of 165x120 pixels. Before testing, each has itsillumination normalized via histogram equalization, and isthen resized to an asymmetrical resolution of 80x80 pixels(or other required testing size). Figure 2 shows a sampleimage set from the database.Example training setExample testing set
Figure 2: Example image set of one person in the ARdatabase.
Experiments performed on this database implement theidentified feature extraction method, then nearest-neighbourclassification with the Euclidean distance measure toidentify the class of each test image. The identificationrate, which gives the percentage of test images correctlyidentified, is then calculated.
Coefficients of Significance
The local-DCT method uses a selection of the 2-dimensional DCT coefficients from each block to constructthe feature vector for an image. For the best identificationrates, the coefficients selected should be the mostsignificant in representing important and distinguishingimage details. Using one coefficient to represent eachblock, all images of the database were represented as

Activity (5)

You've already reviewed this. Edit your review.
1 thousand reads
1 hundred reads
murthy4318014 liked this
Naimababou liked this

You're Reading a Free Preview

Download
scribd
/*********** DO NOT ALTER ANYTHING BELOW THIS LINE ! ************/ var s_code=s.t();if(s_code)document.write(s_code)//-->