A MINI PROJECT REPORT

ON
NON LINEAR TOPOLOGICAL COMPONENT ANALYSIS:
APPLICATION TO AGE INVARIANT FACE RECOGNITION
IN
I M.TECH II-SEMESTER
BY
B.KRANTHI SANDHYA
(15761D4501)

DEPARTMENT OF ELECTRONICS AND COMMUNICATION ENGINEERING

LAKIREDDY BALIREDDY COLLEGE OF ENGINEERING
(AUTONOMOUS)

(Approved by AICTE, New Delhi, Affiliated to JNTU Kakinada)

Mylavaram, Krishna District-521230, Andhra Pradesh

2015-2016

LAKIREDDY BALIREDDY COLLEGE OF ENGINEERING

(AUTONOMOUS)

(Approved by AICTE, New Delhi, Affiliated to JNTU Kakinada)

Mylavaram, Krishna District-521230, Andhra Pradesh

2015-2016

CERTIFICATE

This is to certify that the mini project entitled Non Linear Topological

Component Analysis: Application to Age Invariant Face

Recognition has been submitted by B.KRANTHI SANDHYA (15761D4501) in
M.Tech II-SEM of Electronics and communication engineering of LBR college of engineering
,Mylavaram during the academic year 2014-2015.
Date:
Place: Mylavaram

Mini project in charge

Head of the department

Mr.P.Rakesh Kumar

prof:

B.Ramesh Reddy

ACKNOWLEDGEMENT

First and foremost I sincerely thank our institution LakiReddy BaliReddy College of
Engineering for giving this opportunity for fulfilling my dreams of becoming engineers .I
express my special gratitude to our director Dr.E.V.Prasad, Ph.D., who made this endeavor
possible .
I own my profound gratitude to our principal DR.N.R.M.Reddy, Ph.D., who took keen interest
towards our project and provided us with unlisted encouragement and more over his timely
support by providing necessary resource
I have the immense pleasure in expressing my thanks and deep sense of gratitude of
Prof.B.Ramesh Reddy,(Ph.D)., Head of the Department of Electronics

and

Communication engineering for extending necessary facilities for the completion of the
project.
I am highly thankful to our mini project coordinator Mr. P. Rakesh Kumar,(Ph.D),B.Santhosh
kumar MTech and other faculty members Dr E.V.Krishna Rao,Ph.D,Mr.T.Anil Raju ME ,
Smt.B.Rajeswari M.tech,Mr P.Ravi kumar,(Ph.D), for providing the knowledge to deal
with problem at every phase of our mini project in a systematic manner.
I would like to extend my deepest gratitude to support rendering by the Teaching and NonTeaching staff of electronics and communication department during the course of the
work .Finally I thank each and every one who directly or indirectly contributed his or her
help to complete this mini project
B.KRANTHI SANDHYA
15761D4501

LIST OF CONTENTS
LIST OF FIGURES

LIST OF TABLES

II

TITLE NAME
PAGENO
1. ABSTRACT

2. INTRODUCTION

3.SOME BASICS
3.1 KERNEL BASICS

3.2 -SHAPES FORMALISM

11

4. KRBFS DIMENSIONALITY REDUCTION

4.1 MAPPING DATA SPACE TO LATENT SPACE
4.2.LATENT SPACE PROBABILITY DISTRIBUTION

12
14

5.KRBF PARAMETERS
5.1 EXPECTATION-MAXIMIZATION PARAMETERS
5.2. NUMBER OF PARAMETERS IN KRBF

15
17

6. TOPOLOGICL FETURE EXTRACTION

18

7. CLASSIFICATION/TESTING

19

8. FACIAL IDENTIFICATION ACROSS AGE

8.1 PROBLEM STATEMENT AND BACKGROUND
8.2 FACIAL FEATURE AND DATA SET GENERATION

21
23

9 . EXPERIMENT AND RESULTS

10. SOFTWARE REQUIRED
10.1.INTRODUTION TO MATLAB

27

10.2 APPENDIX TO MATLAB CODE

28

11.CONCLUSION

31

12.REFERENCES

32

I.LIST OF FIGURES
TITLE NAME

PAGE NO

1. ALPHA SHAPES

11

2. VOIDS AND GAPS INFEATURES

13

3. KRBF LATENT VARIABLE

4. NON LINEAR MAPPING
5.STEPS IN KRBF
6.HISTOGRAM OF TOPOLOGICAL FEATURES
7.FACE IMAGE OF SAME INDIVIDUAL
8.BASIC LBP OPERATOR
9.EXTENDED LBP
10.SAMPLE OF FACE IMAGES FROM GEORGIA TECH
11.SAMPLE OF FACE IMAGES IN MORPH DATA BASE
12.ALPHA SHAPES IN FG NET DATA BASE

II.LIST OF TABLES
TITLE NAME

PAGE NO

1. NUMBER OF POINTS USED DURING TRAINING, PROBE

2. ACCURACY (%) OF TDBN AND NTCA USING DIFFERENT
TRAINING/GALLERY/PROBE FOLDS (RANK 1)

16
21

3. IDENTIFICATION ACCURACY (%) OF TDBN AND NTCA USING FIVE

FOLDS ON TWO DIFFERENT AGING DATABASES (RANK 1)

26

1. ABSTRACT
We introduce a novel formalism that performs dimensionality reduction and
captures topological features (such as the shape of the observed data) to conduct pattern
classification. This mission is achieved by:
1) Reducing the dimension of the observed variables through a kernelized radial basis function
technique and expressing the latent variables probability distribution in terms of the observed
variables;
2) Disclosing the data manifold as a 3-D polyhedron via the -shape constructor and extracting
topological features; and
3) Classifying a data set using a mixture of multinomial distributions. We have applied our
methodology to the problem of age-invariant face recognition. Experimental results obtained
demonstrate the efficiency of the proposed methodology named nonlinear topological component
analysis when compared with some state-of-the-art approaches.

2. INTRODUCTION
PRINCIPAL component analysis (PCA) remains an invaluable tool when one acquires
measures on a number of observed variables and wish to generate a smaller number of artificial
variables (called principal components) that will account for most of the variance. This variable
reduction is motivated by the belief that some observed variables measure the same construct and
therefore exhibit significant correlations. The principal components may then be utilized as a
predictor set in subsequent analyses. PCA has been exploited in a nonlinear way by means of the
kernel trick to derive kernel-based PCA (KPCA) . Other prominent nonlinear techniques include
manifold learning such as Isomap, locally linear embedding (LLE) , Hessian LLE , Laplacian
Eigenmaps, and local tangent space alignment , which have been proposed in the machine
learning literature.
These techniques build a low-dimensional data representation using a cost function that
maintains local properties of the data. These approaches can be perceived as defining a graphbased kernel for KPCA. Likewise, graph and network embedding techniques have also been
introduced as a general framework for dimensionality reduction for the purpose of visualization.
Other formalisms such as generative Shown in Fig. 1(a).

Topographic mapping have been put forward as an attempt to reveal the inverse
nonlinear relationship between the set of latent variables and the set of observed variables.
Unlike PCA, these latter mappings assume an underlying causal structure within the observed
variables and thus represent a building block of an exploratory factor analysis. However, most of
these techniques are oriented toward the search of a dimension reduction criterion; they have
difficulty providing a means that permits the disclosure of the low-dimensional shape formed by
the data set. Unveiling shapes would allow the exploitation of mathematical concepts such as
homeomorphism, homotopy equivalence, and topological invariance , which are precious in
many pattern recognition problems such as classification of 3-D protein folds or brain repair after
7

Injury.
The main motivation of this research consists of extending dimensionality reduction
techniques to derive topological features (e.g., as the shape of the observed data) and to merge
them with statistical information. The approach that we propose is viewed as: 1) an integration of
a kernelized radial basis function (KRBF) dimensionality reduction method; 2) a construction of
-shapes and extraction of topological features (signatures); and 3) an object classification based
on a mixture of multinomial distributions.

One would be able to gain an insight into the morphology of shapes formed by the
data set in a latent space for a sharper classification task. We show how statistical information
can be merged seamlessly with topological knowledge using a multinomial mixture probability
distribution. This allows comparing several shapes in the latent variable space. The entire
methodology that we have named nonlinear topological component analysis (NTCA) is
undertaken by:
1) Mapping the original data set to a Hilbert space using a quadratic function: this allows to
obtain a more accurate separation of the observations;
2) Mapping the obtained set to a set spanned by KRBF;
3) Projecting linearly this latter set on a latent space;
4) Building a hierarchy of -shapes associated to the points of the latent space; and
5) Extracting topological signatures from the -shapes for the purpose of classification.
In other words, 1)3) represent a dimensionality reduction technique (that we have named:
KRBF), whereby each point of the original data set is mapped nonlinearly to a point in a lowdimensional latent vector space. The cloud of points obtained in the latent space is assigned a
hierarchy of -shapes that represents the original data set. Finally, a group of topological traits
are extracted by averaging over the -shapes to form a template for object identification.
Applications of the NTCA formalism are diverse:
Biometrics and Forensics can benefit from NTCA since the original data set
depicting topological manifold can be built from different images of the same biometric
Sample (e.g., fingerprint, face) of an individual. The proposed approach maps
This multidimensional manifold to a 3-D -shape, which is a 3-D-polyhedron
(N-polytope). Another application would consist of reconstructing 3-D-MRI images of
Damaged organs (such as the brain) to identify the causes of the injury, and find its location
using the geometrical concept of -shapes

3.SOME BASICS
3.1 Kernel Basics:
We consider the data set D as a set composed of N points t1, t2, . . . , tn, in which each ti RD.
Definition 1: A kernel k(., .) on a domain D is a real positive function k : D D R+
that is continuous, symmetric, and positive semidefinite in a certain sense. Let us denote the data
matrix as: D = (D (1)|D(2)|, . . . , |D(D)) in which each column of D corresponds to a variable
D (i) = (t1i , t2i, . . . , tni ) Rn, and each row is a data point ti = (ti1, ti2, . . . , tiD) RD.
Given a set of data points t1, t2, . . . , tn D, and the (n n) matrix Ki, j = k(ti , t j ) (called the
Gram matrix with respect to t1, t2, . . . , tn), we have the following definition.
Definition 2: The kernel function k is positive definite, if n N, its Gram matrix with
respect to all t1, t2, . . . , tn is a positive definite matrix (1, 2, . . . ,n) Rn,_ i, j ij Ki, j 0,
and if i, j ij Ki, j = 0 1 = 2 . . . = n = 0.
Theorem 1 (Kernel Trick Theorem): Given a positive definite kernel k(., .) on a domain
D, one can build a mapping from D into a Hilbert space Hk , in which the function k represents
a dot product <.> such that <(t1), (t2)> = (t1)T (t2) = k(t1, t2) (t1, t2) D2.
The idea behind the kernel trick is to replace a nonlinear model defined in the original space by a
linear model defined in the target space.
The original space is mapped to the target space using a nonlinear transformation based
on suitably chosen basis functions. Therefore, the inner product of basis functions (t1)T (t2)
is replaced by a kernel function k(t1, t2), between instances in the original input space.
Therefore, instead of mapping two instances t1 and t2 to the Hilbert space and performing a dot
product there, one directly applies the kernel function in the original space. Among the most used
kernels, one can cite the Gaussian kernel k(x, y) = k(//xy//) = exp{(xy)T1(xy)}, which is
an example of radial basis function kernel and the polynomial kernel k(x, y) = (axT y+c)d,
where a is the slope parameter, c is a constant term, and d is the polynomial degree. There are
several other kernels proposed ; however, it is important to underscore that the choice of a best
kernel depends on the type and structure of the data at hand.
9

3.2. -Shapes Formalism:

The concept of -shapes represents an efficient means for creating shapes at different
levels of resolution out of point sets. The -shapes formalism can be viewed as a soft
generalization of the standard notion of convex hull assigned to a set of points. In the sequel, we
provide some basic definitions that explain briefly shapes and how they are constructed.
Definition 3: Let E be a finite set of points in R3, the Voronoi cell of u E (or Voronoi
Complex) is defined as Vu = {x R3: v E {u},
xu
x v }. The set of
Vu represents convex polyhedral or an empty set since the set of points that is equidistant from
two points in E forms a hyperplane.
Definition 4: Let E be a finite set of points in R3, the union of balls centered at the points
in E with radius for 0 is denoted by B(E) B(E) = {x R3 : u E such that //u x//
}.
Definition 5: Let E be a set of points in R3, and 0, the -complex of E is the dual
complex of the Voronoi diagram of E restricted to the union of balls B(E). The restricted
Voronoi cell of u E is: V u = Vu B(E). The -shape is the underlying space corresponding
to the -complex. The -shape concept represents a formalization of the intuitive notion of shape
for spatial point set data.
An -shape is a concrete geometric object that is uniquely defined for a particular set of
points. The parameter controls the desired level of details of the shape. There are several
algorithms that construct a family of -shapes for a given set of size n in a worst case time
complexity equal to O(n2). The -shapes define a hierarchy of shapes from a set of points that
allows features multiscale modeling that is very useful in macromolecule structure exploration
as well as in facial aging (identifying changes of human facial compartments: a human face is
made of compartments of fat). The -shapes insert a ball of radius around each point and
builda simplicial complex that respects the intersections among.
The -shapes define a hierarchy of shapes from a set of points that allows features
multiscale modeling that is very useful in macromolecule structure exploration as well as in
facial aging (identifying changes of human facial compartments: a human face is made of
compartments of fat). The -shapes insert a ball of radius around each point and build a
simplicial complex that respects the intersections these balls.

10

Fig. 1. Discrete set of data points that depicts the -shape associated to the
letter a from finer ( = 0) to cruder ( =) levels of details.
The simplicial space formed is defined as the -shapes.
1) -Shapes Construction: The construction of -shapes is based on the notion of ball known
also as generalized disk. Given a set of points and a specific value, the -shape is constructed
using the following scheme.
1) Each point Xu in the embedded set is assigned a vertex u.
2) An edge is created between two vertices u and v whenever there exists a generalized disk of
radius containing the entire set of points and which has the property that Xu and Xv lie on its
boundary. However, the notion of generalized disk is defined as follows.
1) If > 0 but very large, then the generalized disk is a half-plane (very large radius!).
2) If > 0 but not large, then the generalized disk is a closed ball of radius.
3) If < 0, the generalized disk is the complement of a closed ball of radius.

Fig. 2. Voids and gaps formed by the alphabetical letters. An -shape in R3

can have voids and gaps, but it can also have tunnels.
11

4. KRBFS DIMENSIONALITY REDUCTION

4.1.Mapping Data Space to Latent Space:

The mission in KRBF is to find a representation for the distribution p(x) of lowdimensional latent variables in terms of a number D of data variables t = (t1, t2, . . . , tD). The
computation of this distribution allows the creation of the set of points X that represents the data
set in the latent space. Fig. 2.
Voids and gaps formed by the alphabetical letters. An -shape in R3 can have voids and gaps,
but it can also have tunnels. The assumption of the existence of a low-dimensional data manifold
governed by some latent variables is explained by
The fact that:
1) Data exhibit substantial correlations between the variables and
2) Some data are noisy due to defective sensors.
A nonlinear mapping between the original data set and the latent variable space
X is performed by first mapping the data set D of RD to a subset De (e stands for extended) of a
high dimensional Hilbert space Hk (endowed with a positive definite kernel function k on D)
using a feature space mapping .
The subset De is in turn mapped to a subset B spanned by radial basis functions,
and finally, B is projected on the latent space RL via a nonlinear transformation to derive the set
X. The image x of an element t D is computed as x = ( )(t), whereby z = (t) De, (t
D), and ((z)) = W(z), where = (i ) (i = 1, . . . , Q) designates a set of fixed nonlinear basis
functions, and W is an (L Q) matrix of parameters.
Finally, the -shape geometric descriptor is invoked to compute a family of Ldimensional -shapes (defined in A()) assigned to the subset of data points X (Fig. 3). The
dimensionality L of the latent space is less than the dimensionality D of the data space and
therefore less than the dimensionality H of th Hilbert space. Topological signatures extracted
from the -shapes (polyhedra) characterize the data manifold. These signatures can be used for a
classification or a regression task (Fig. 4). It should be borne in mind that the motivation for this
paper is twofold:
1) It provides a 3-D visualization in the form of -shapes (in the latent space) of the data
topological manifold and
2) It creates a methodology that merges seamlessly statistical data and topological features
within a single probabilistic framework.

12

Fig. 3. Nonlinear dimensionality reduction mapping (KRBF) the latent

variable data of which are used to build the -shapes of A().

Fig. 4. Nonlinear mapping between the data topological manifold and an

-shape (3-D polyhedron) via a latent space.

13

4.2. Latent Space Probability Distribution:

We start by defining a probability distribution p(t) in which t D, this will infer a
corresponding probability distribution p(x) that generates the set X in the latent space RL

(i = 1, . . . , Q), where zi = (ti ) and i = 1/N2Max2j (zi j ).Using the kernel trick theorem, the
nonlinear basis functions can be written as

14

5. KRBF PARAMETERS
5.1Expectation-Maximization Parameters Estimation:
For a given latent variable set = {x1, . . . , xn}, one can determine the parameter
matrix W, and the inverse variance using maximum-likelihood estimation. It is often
convenient to maximize the log-likelihood.

which is equivalent to
Finally, the likelihood function that needs to be maximized with respect to W and is, L(W,)

Since the density model consists of a sum of M mixtures, the EM algorithm is more suitable for
parameter estimation. The maximization of the likelihood L(W, ) can be viewed as a missing
data problem in which the mixture j that generated a latent variable is unknown. In the E-step of
the EM algorithm, we use Wc and c (the superscript c stands for current) to evaluate the
responsibilities Resp. (posterior probabilities) of each Gaussian mixture j for every latent point
xn associated to a data point tn using Bayes rule

We now consider the expectation of the complete data loglikelihood in the form

15

Maximizing (1) with respect to the variable W, one obtains

(18)
This latter expression can be written as Similarly, the maximization of the complete loglikelihood with respect to yields the following:

The initial values for the parameters of the matrix W are computed so that initially the NTCA
models performance gets as close as possible to the PCAs performance.

16

5.2. Number of Parameters in KRBF:

We define, in this section, all parameters involved in the KRBF dimensionality
reduction method. We also show how they are computed.
1)D is the dimension of the input data vectors ti s; this dimension value depends on the
application at hand.
2) H = D (D + 1)/2 represents the dimension of the vector zi : it is the sum of an arithmetic series.
3) Q is the dimension of the vectors (zi ), which consists of the numbers of groups obtained via
a clustering procedure of the vectors zi s. It corresponds to the cardinality of the radial basis set.
4) The initial elements of the (L Q) matrix W are the L eigenvectors of the covariance matrix
of the vectors i (z) computed using PCA (minimize information loss during the mapping from
the i (z) in the Q-dimensional space to the new L (L < Q) -dimensional space RL ).
5) L is the dimension of the space generated by the xi vectors; its value is set to three for
visualization purposes
.
6) i = 1/N2Max2j (zi j ), for scaling the exponential.
7) M = N is the number of vectors ti = |D|.
8) is initially set to one.

17

6. TOPOLOGICAL FEATURE EXTRACTION

As outlined in Section 2, the mission in this step consists of exploiting

dimensionality reduction techniques to reveal the -shape low-dimensional layout (3-D
polyhedron) of the data topological manifold living in RD. The thrust consists of unwrapping
this data manifolds shape in the latent space RL (L is often equal to 3 for visualization) and
represents these data in terms of topological features extracted from this -shape.
Any modification in the data manifold is viewed as a deformation of this L-dimensional shape polyhedron. One can extract signatures of -shapes such as metric properties: (volume,
area, and length), combinatorial properties: (number of tetrahedral, number of triangles, number
of edges, number of vertices), and topological properties: (number of components, number of
independent tunnels, number of voids, and number of gaps).
These signatures (or features) characterize an -shape. A void might be encountered in a
connected -shape in R2, whereas a gap is encountered only if we allow for nonconnected
shapes. However, the notion of tunnel (or open doors) exists only in the 3-D -shape context.All
these signatures are computed as average over a set of predefined values.

18

7. CLASSIFICATION/TESTING
In addition to the fact that NTCA is endowed with the power of representing the input
data set in a low-dimensional latent space; it also allows classification of this data set when this
set defines a physical object. The -shape formalism provided a way to represent any set by its
shape and thereby renders classification feasible. The NTCA classification consists of assigning
one and only one class * (from a set of c predefined classes in _) to this -shape signature
pattern. Given a set of data points D, the NTCA methodology starts by invoking the KRBF
dimensionality reduction procedure. This latter requires the estimation of the probability density
function p(x) defined in the latent space X. Once the set X RL assigned to the set D is
generated in the latent space, the -shapes geometric constructor is invoked to draw the 3-Dpolyhedron associated with the nonlinear topological manifold representing the data set D.
Finally, topological signatures, inherent to the manifolds -shape characterizing the input object
to classify, are extracted and stored into a set S. The classification problem can be recast into:
determine the class * among c target classes (from a gallery) assigned to an input (or probe)
data point set D = {t1, . . . , tN
such that = argmaxi P[i |t1, t2, . . . , tN ]. (23)
Because the data set D is mapped nonlinearly to the latent set X = {x1, . . . , xn}RL , this
classification problem is equivalent to the determination of the class *such that
= argmaxi P[i |{x1, x2, . . . , xn}] (24)
Where the set {x1x2, . . . , xn} is generated using the KRBF algorithm. However, because our
goal is to disclose topological information conveyed by the -shapes, a set of signatures is
extracted for the purpose of classification. It is worth highlighting the fact that the entire subset X
RL is mapped to only one -shape the topological signature set Sj of which contains average
values of signatures of the -shape at some different resolutions . Finally, one can restate the
classification problem with respect to the signatures set as follows: determine the class *
among c target classes assigned to a set of points D = {t1, . . . , tN} RD such that
W* = argmaxi Pi |Sj _, (i, j = 1, . . . , c), 0. (25)
Equation (25) requires the estimation of the posterior probability model
P[i |Sj]. Through Bayes rule, one can transform the problem into an estimation of the class
conditional probability

To obtain an expression of the class conditional probability, we make an analogy between the
models, we are investigating and the word/document model. In this analogy, an -shape
described by a set of signatures corresponds to a document represented by a set of words. As
outlined in Section II, signatures are features such as independent tunnels, voids, triangles, edges,
and vertices; they are the words of an -shape. These words are assumed to be generated
independently (bag of words document model). This latter assumption favors the use of a
multinomial distribution. Each signature is a word; therefore, signature histograms are term
frequency distributions. Likewise, because statistical similarities among the signature histograms
of -shapes designating similar objects are observed, we can assume that signature histograms of
similar -shapes are generated by similar multinomial distributions. Consequently, a multinomial
mixture model should be adequate to model the topological signature distribution in each class
19

i _. In other words, we make the assumption that the distribution of an -shape signature set
in each class i is constituted of p multinomial distributions.
1) Let A() = {a1, . . . , a|A()|} be the set of -shapes, each of which is described by topological
signatures.
2) Let Sj= {s1 , . . . , sT} be the set of topological signatures assigned to the -shape with a
specific value.
Fig. 6 shows the topological feature frequency distribution in which n i j represents the
frequency of the topological signature sj in the -shape ai at a given value.
1) Let C = {c1, . . . , cP} be the set of P mixtures (clusters) within the class i . The
probability P(a I i ) that the -shape ai (at resolution ) of a class i is generated by
a multinomial mixture is

Fig. 6. Unimodal topological signatures histogram [number of: components(com), independent

tunnels (ide tun), voids (voi), triangles (tri), edges (edg),and vertices (ver)] of an -shape.

ck,i
T
P (ck,i )

is the mixture coefficient for the kth mixturein class i ;

is the number of topological signatures;
is the probability that the -shape a i assigned to the class i belongs to the
clusterck,i ;
P(sj|ck,i )
is the probability that the topological signature sj appears in the cluster ck,i ;
P
are probabilities P (ck,i ) and P T probabilities P(sj |ck,i ) are the
parameters of the multinomial mixtures. For each class i , they are subject to

The maximum-likelihood estimation consists of maximizing subject to the constraints in the

system of (28). The EM algorithm is used for parameters estimation. It is worth noting that this
classification problem is enriched by the topological knowledge Sj revealed by the 3-D -shape
Polyhedron defined in the latent space. This synergy between probability and topology is
exhibited by the NTCA formalism to perform a classification or a regression task.
20

8. FACIAL IDENTIFICATION ACROSS AGES

To validate the proposed methodology, we have selected one of the most difficult
problems in robust biometrics known as face identification across ages.

8.1. Problem Statement and Background:

Given a face sample of an individual at age interval A0, one investigates if this
input face is associated with any of a large number of enrolled faces of individuals. However,
some face images of the same individual at different age intervals Ai are among the enrollees. In
other words, given two face images, can one infer that they represent the same individual but at
different age intervals (Fig. 7)?
Several features that are responsible for a facial appearance are:
1) The 3-D structure of human faces;
2) The reflective peculiarity of facial skin; and
3) The bilateral symmetry in the structure of facial features in combination with scene-centric
attributes such as viewpoint and illumination. One of the main challenges in face identification
stems from the alteration of facial appearances.
This modification is explained by the combination of several factors such as
illumination, facial expression, occlusions, and age. However, the power of face identification
across ages is mostly dependent on the ability of modeling some singularities caused by face
wrinkles as well as bones structural deformation (e.g., jaw bones) of the enrolee face. The
process of aging remains the major cause of significant deformation in the appearance (face
texture) and anatomy of human faces. Aging is, therefore, one of the most compelling challenges
for automatic human identification systems and forensics investigation.

Fig. 7. Face images of the same individual, at different age intervals [Afghan Girl: National
Geographic June 1985].

21

A manifold assigned to an individual is described by images (set of points) of

the same individual at different ages of his/her life. -shapes computed at different resolutions
capture this manifold through an averaged signature set. Each individual has his/her own shape.
Hence, the mission consists of:
1)Extracting facial features by applying a traditional facial features extraction method. These
facial feature vectors represent the data points of the set D RD;
2) Building for each enrolee the set X using the probability distribution p(x) of the data manifold;
3)Computing the -shape (3-D polyhedron) assigned to each individuals nonlinear manifold
forVisualization; and
4) Extracting -shape topological signatures for a face identification task. In the next section, we
show how these low-level facial features are built to fill out the data set D RD.

Fig. 8. Basic LBP operator: the neighbors are thresholded, and their binary sequence is converted
into a decimal LBP label.

22

8.2. Facial Feature and Data Set Generation:

To illustrate the entire NTCA approach, we extract a set of facial features using the
traditional LBP method. A face image is first divided into nonoverlapping k small rectangular
regions. For each pixel in a region, one compares the pixel with each of its eight neighbours. If
the center pixels value is greater than the neighbours value, this pixel value is replaced by 1,
otherwise, it is replaced by 0. The thresholded neighbours provide an eight-digit binary
sequence, which is converted into a decimal number (called LBP label), (Fig. 8).
A total of 256 possible patterns ranging from 0 to 255 can be extracted in the case of
8 neighbors. An LBP bit sequence corresponding to thresholded pixel intensities matrix is said to
be uniform if the bits change less than two times (e.g., 11100001 has two changes, between the
third and the forth bit and between the seventh and the eighth bit). This operator has been
extended whereby different neighborhood sizes are considered. In general, LBPP,R refers to the
neighbors size of P equally spaced pixels on a circle of radius R which form a circularly
symmetric neighbor set.
As these points do not necessarily fall in the center of a pixel of the image, their values
are obtained using bilinear interpolation (on-grid versus off-grid) (Fig. 9). Every region has an
associated histogram depicting the frequency of each LBP label. Each histogram has 59 bins:
there are 56 bins that correspond to 56 uniform bit sequence that have exactly 2-b changes, 2
bins that correspond to the 2-b sequences that undergo no change. These two sequences are
assigned to bin 0 and bin 58. The nonuniform sequences are assigned to bin 59. The feature
vector is a concatenation (putting side by side) of all histograms corresponding to the k regions
of an image and therefore has 59 components.

Fig. 9. Extended LBP8,1 corresponding to P = 8 and R = 1.

23

9. Experiment and results:

Fig. 10. Sample of face images from the GeorgiaTech database.

Fig. 11. Sample of face images from (a) MORPH-album 1, (b) MORPHalbumm2, and (c)
FGNET. The light blue row in the bottom of each sequence represents the ages.

24

Fig. 12. shapes of two different individuals computed and stored as prototypes in the FGNET
gallery.

25

26

TABLE IV:
COMPARISON OF KRBF DIMENSIONALITY REDUCTION TECHNIQUEWITH SOME
STATE-OF-THE-ART APPROAC

9.SOFTWARE REQUIRED
9.1.Introduction to MATLAB:

MATLAB is a high-performance language for technical computing. It integrates

computation, visualization, and programming in an easy-to-use environment where problems
and solutions are expressed in familiar mathematical notation. Typical uses include

Math and computation

Algorithm development

Data acquisition

Modeling, simulation, and prototyping

Data analysis, exploration, and visualization

Scientific and engineering graphics

Application development, including graphical user interface building

MATLAB is an interactive system whose basic data element is an array that does not
require dimensioning. This allows you to solve many technical computing problems,
especially those with matrix and vector formulations, in a fraction of the time it would take to
write a program in a scalar non interactive language such as C or FORTRAN.

27

9.2.APPENDIX: MATLAB CODE

clc;
clear all;
close all;
%% Face recognition
% This algorithm uses the eigenface system (based on pricipal
component
% analysis - PCA) to recognize faces. For more information on this
method
% refer to http://cnx.org/content/m12531/latest/
%% Download the face database
% You can find the database at the follwoing link,
% http://www.cl.cam.ac.uk/research/dtg/attarchive/facedatabase.html
The
% database contains 400 pictures of 40 subjects. Download the zipped
% database and unzip it in the same directory as this file.
%% Loading the database into matrix v
w=load_database();
%% Initializations
% We randomly pick an image from our database and use the rest of the
% images for training. Training is done on 399 pictues. We later
% use the randomly selectted picture to test the algorithm.
ri=round(400*rand(1,1));
r=w(:,ri);
will use to test the algorithm
v=w(:,[1:ri-1 ri+1:end]);
images.

% Randomly pick an index.

% r contains the image we later on

N=20;
each image.
%% Subtracting the mean from v
O=uint8(ones(1,size(v,2)));
m=uint8(mean(v,2));
images.V
vzm=v-uint8(single(m)*single(O));

% Number of signatures used for

% v contains the rest of the 399

% m is the maen of all

% vzm is v with the mean removed.

28

%% Calculating eignevectors of the correlation matrix

% We are picking N of the 400 eigenfaces.
L=single(vzm)'*single(vzm);
[V,D]=eig(L);
V=single(vzm)*V;
V=V(:,end:-1:end-(N-1));
% Pick the eignevectors
corresponding to the 10 largest eigenvalues.
%% Calculating the signature for each image
cv=zeros(size(v,2),N);
for i=1:size(v,2);
cv(i,:)=single(vzm(:,i))'*V;
% Each row in cv is the signature
for one image.
end
%% Recognition
% Now, we run the algorithm and see if we can correctly recognize the
face.
subplot(121);
imshow(reshape(r,112,92));title('Looking
for ...','FontWeight','bold','Fontsize',16,'color','red');
subplot(122);
p=r-m;
% Subtract the mean
s=single(p)'*V;
z=[];
for i=1:size(v,2)
z=[z,norm(cv(i,:)-s,2)];
if(rem(i,20)==0),imshow(reshape(v(:,i),112,92)),end;
drawnow;
end
[a,i]=min(z);
subplot(122);
imshow(reshape(v(:,i),112,92));title('Found!','FontWeight','bold','Fon
tsize',16,'color','red');
function

X1 = LBP(Resimage)

% Returns a rotation invariant LBP (uniform patterns) histogram (10

bins)
% of picture X.
% the size of picture X must be at least 3x3 pixels
w1 = (1/sqrt(2))^2; %formula
w2 = (1-1/sqrt(2))*(1/sqrt(2));
[sy sx] = size(Resimage); %size
29

Xi = zeros(sy+2,sx+2); %zeros for size of image

Xi(2:sy+1,2:sx+1) = Resimage; %incrementing for full image
Xi2 = zeros(sy+2,sx+2); %different derivatives
Xi4= zeros(sy+2,sx+2);
Xi6 = zeros(sy+2,sx+2);
Xi8 = zeros(sy+2,sx+2);
p1 = zeros(sy+2,sx+2);
p2 = zeros(sy+2,sx+2);
p3 = zeros(sy+2,sx+2);
p4 = zeros(sy+2,sx+2);
p5 = zeros(sy+2,sx+2);
p6 = zeros(sy+2,sx+2);
p7 = zeros(sy+2,sx+2);
p8 = zeros(sy+2,sx+2);
p9 = zeros(sy+2,sx+2);
p1(3:sy+2,3:sx+2) = Resimage;
p2(3:sy+2,2:sx+1) = w2*double(Resimage); %combining with formula to
obtain LBP
p3(3:sy+2,1:sx) = Resimage;
p4(2:sy+1,3:sx+2) = w2*double(Resimage);
p5(2:sy+1,2:sx+1) = (1-1/sqrt(2))^2*double(Resimage);
p6(2:sy+1,1:sx) = w2*double(Resimage);
p7(1:sy,3:sx+2) = Resimage;
p8(1:sy,2:sx+1) = w2*double(Resimage);
p9(1:sy,1:sx) = Resimage;
Xi1 = w1*p1+ p2+p4
Xi2(3:sy+2,2:sx+1)
Xi3 = w1*p3 + p2 +
Xi4(2:sy+1,1:sx) =
Xi5 = w1*p9 + p8 +
Xi6(1:sy,2:sx+1) =
Xi7 = w1*p7 + p8 +
Xi8(2:sy+1,3:sx+2)

+ p5 + 0.000001; %Xi1 to the right and down from X

= Resimage;
p6 + p5 + 0.000001;
Resimage;
p6 + p5 + 0.000001;
Resimage;
p4 + p5 + 0.000001;
= Resimage;

Xi = (Xi4>=Xi)+2*(Xi5>=Xi)+4*(Xi6>=Xi)+8*(Xi7>=Xi)+16*(Xi8>=Xi)
+32*(Xi1>=Xi)+64*(Xi2>=Xi)+128*(Xi3>=Xi);
X1 = Xi(3:sy,3:sx); %making into whole form
end

30

10. CONCLUSION

We have developed a nonlinear age-invariant face recognizer that we named

NTCA. This mission has been achieved by: 1) nonlinearly mapping the data topological
manifold to a low-dimensional latent variable space; 2) extracting topological features through
the computation of the -shape polyhedron formed by these latent variables; and 3) performing
shape classification using a mixture of multinomial distributions. The results obtained have
demonstrated the potential of the NTCA approach as a whole. The benefit of expressing p(x) is
invaluable since it allows the estimation of some parameters and the generation of other points in
case where face images of the same individual are lacking. It also renders the more general
weighted -shapes concept exploitable. However, our methodology can still be improved

31

11. References

[1] H. Hoffmann, Kernel PCA for novelty detection, Pattern Recognit., vol. 40, no. 3, pp. 863
874, Mar. 2007.
[2] J. B. Tenenbaum, V. de Silva, and J. C. Langford, A global geometric framework for
nonlinear dimensionality reduction, Science, vol. 290, no. 5500, pp. 23192323, Dec. 2000.
[3] S. T. Roweis and L. K. Saul, Nonlinear dimensionality reduction bylocally linear
embedding, Science, vol. 290, no. 5500, pp. 23232326,Dec. 2000.
[4] D. Donoho and C. Grimes, Hessian eigenmaps: New tools fornonlinear dimensionality
reduction, in Proc. Nat. Acad. Sci., 2003,pp. 55915596.
[5] M. Belkin and P. Niyogi, Laplacian eigenmaps and spectral techniques for embedding and
clustering, in Advances in Neural Information Processing Systems, vol. 14, D. G. Dietterich,
S. Becker, and Z. Ghahramani, Eds. Cambridge, MA, USA: MIT Press, 2002, pp. 585591.
[6] Z. Zhang and Z. Hongyuan, Principal manifolds and nonlineardimension reduction via local
tangent space alignment, SIAM J. Sci.Comput., vol. 26, no. 1, pp. 313338, 2002
[7] S. Yan, D. Xu, B. Zhang, H.-J. Zhang, Q. Yang, and S. Lin, Graphembedding and
extensions: A general framework for dimensionalityreduction, IEEE Trans. Pattern Anal. Mach.
Intell., vol. 29, no. 1,pp. 4051, Jan. 2007.
[8] G. D. Battista, P. Eades, R. Tamassia, and I. G. Tollis, Graph Drawing:Algorithms for the
Visualization of Graphs. Englewood Cliffs, NJ, USA:Prentice-Hall, 1998.
[9] D. Harel and Y. Koren, Graph drawing by high-dimensionalembedding, in Proc. 10th Int.
Symp. Graph Drawing, London, U.K.,Feb. 2002, pp. 207219.
[10] C.-C. Lin and H.-C. Yen, A new force-directed graph drawing methodbased on edge-edge
repulsion, in Proc. 9th Int. Conf. ICCV, Beijing,China, Jul. 2005, pp. 329334

32