SOM and PCA Approach for Face Recognition - A Survey

Document Sample
SOM and PCA Approach for Face Recognition - A Survey Powered By Docstoc
					International Journal of Computer Trends and Technology- March to April Issue-2011

                               SOM and PCA Approach for
                               Face Recognition - A Survey
                                            Jayshree Ghorpade1, Siddhant Agarwal2
                     Assistant Professor, Department of Computer Engineering, MITCOE, Pune University, India
                            Student, Department of Computer Engineering, MITCOE, Pune University, India

 Abstract — wouldn’t you love to replace password based             PIN’s and passwords: birthdays, phone numbers and social
 access control to avoid having to reset forgotten password and     security numbers. Face recognition technology may solve this
 worry about the integrity of your system? Wouldn’t you like        problem since a face is undeniably connected to its owner
 to rest secure in comfort that your healthcare system does not     except in the case of identical twins. This unique feature of an
 merely on your social security number as proof of your             individual is non-transferable. The system can then compare
 identity for granting access to your medical records? Because      scans to records stored in a central or local database or even
 each of these questions is becoming more and more important,       on a smart card.
 access to a reliable personal identification is becoming
 increasingly essential .Conventional method of identification
 based on possession of ID cards or exclusive knowledge like a      Biometrics:
 social security number or a password are not all together          Biometrics refers study of methods for uniquely recognizing
 reliable. ID cards can be lost, forgotten or misplaced. But a      human based upon one or more intrinsic physical or
 face is undeniably connected to its owner. It cannot be            behavioural characteristics. Biometrics is used to identify the
 borrowed stolen or easily forged. Face recognition has always      input sample when compared to a template used in cases to
 been a fascinating research area. Face recognition is the          identify specific people by certain characteristics.
 preferred mode of identification by humans: it is natural,
 robust and non-intrusive. Face recognition being the most
 effective and natural technique to identify a person since it is
 the same as the way human does and there is no need to use
 special equipments. A lot of algorithms have been proposed
 for solving face recognition problem. Principal component
 analysis (PCA) is a classical and successful method for face
 recognition. Self organizing map (SOM) has also been used for
 face space representation. This paper makes an attempt to
 enhance these two techniques and describe the use of PCA and
 SOM in facial recognition. It also explains the performance
 result when the two algorithms are combined together. The
 advantage in combining the two techniques is that the
 reduction in data is higher but at the cost of recognition rate.

 Keyword — Face Recognition, SOM (Self Organizing Map),
 PCA (Principal Component Analysis), LDA (Linear
 Discriminant Analysis), ORL (Olivetti Research Laboratory).

                          I. I NTRODUCTION

The information age is quickly revolutionizing the way
transactions are completed. Everyday actions are increasingly
being handled electronically, instead of with pencil and paper
or face to face. This growth in electronic transactions has                              Fig. 1 Types of Biometrics.
resulted in a greater demand for fast and accurate user
identification and authentication. Access codes for buildings,      Face recognition is one of the few biometric methods, which
banks accounts and computer systems often use PIN's for             is very complicated system since the human faces change
identification and security clearances. Using the proper PIN        depending on their age, expressions etc. Face recognition is
gains access, but the user of the PIN is not verified. When         the preferred mode of identification by humans: it is natural,
credit and ATM cards are lost or stolen, an unauthorized user       robust and non-intrusive [3, 4]. Security systems, human –
can often come up with the correct personal codes. Despite          computer interaction, entertainment, criminal identification,
warning, many people continue to choose easily guessed              videos surveillance etc. are among the most common

ISSN: 2231-2803                                            -1-                                                   IJCTT
International Journal of Computer Trends and Technology- March to April Issue-2011
application of it [1, 2].Number of supervised and unsupervised      verification of the user and determine whether the match
learning techniques or methods have been reported for the           declared is right or wrong.
face recognition. These techniques have been broadly divided
into two categories namely Structured-based and Statistics-         D.   Verification:
                                                                    The verification module also consists of a pre-processing
                                                                    system. Verification means the system checks as to who the
         Structured-based Face Recognition:
                                                                    person says he or she is and gives a yes or no decision. In this
          In the structured-based approaches (Brunelli &
                                                                    module the newly obtained sample is pre-processed and
          Poggio, 1993; Wiskoot et al, 1997) recognition is
                                                                    compared with the sample stored in the database. The decision
          based on the relationship between human facial
          features such as eye, mouth, nose, profile silhouettes    is taken depending on the match obtained from the database.
                                                                    Correspondingly the sample is accepted or rejected. Instead of
          and face boundary [8].
                                                                    verification module we can make use of identification module.
                                                                    For each comparison made a match score is given. The
         Statistics-based Face Recognition:                        decision to accept or reject the sample depends on this match
          Statistics-based approaches (Turk & Pentlands 1991;       score falling above or below a predetermined threshold.
          Zhao, 2000) attempt to capture and define the face as
          a whole. Under this approach, the face is matched
          through finding its underlying statistical regularities


The Face Recognition system includes:

         An automated mechanism that scans and captures a
          digital or an analog image /of living personal
          characteristics (Enrollment module).
         Another entity which handles compression,
          processing, storage and compression of the captured
          data with stored data (Database).
         The third interfaces with the application system
          (Identification module).

The various Modules of Face Recognition System as shown in
the Figure. 2 are:

A.   User Interface:
User interface captures the analog or digital image of the
person's face. In this we take number of photos of a single
person either in analog or in a digital form.

B.   Enrollment Module:
In the enrollment module the obtained sample is pre-processed
and analyzed. This analyzed data is stored in the database for
the purpose of future comparison.                                                        Fig. 2 Components of Face Recognition system.

C.   Database:
                                                                    This paper makes an attempt to compare the two different
The database compresses the obtained sample and stores it           techniques that are SOM (an unsupervised learning algorithm)
properly. It should have retrieval property also so that it         and the popular, successful classical method PCA to see the
compares all the stored sample with the newly obtained              performance of face recognition system when the two
sample and retrieves the matched sample for the purpose of          techniques are combined together. After a brief discussion of

ISSN: 2231-2803                                           -2-                                                     IJCTT
International Journal of Computer Trends and Technology- March to April Issue-2011
SOM and PCA in section III, section IV describes the                most common ones are either one-dimensional or two-
combined approach of SOM and PCA for face recognition and           dimensional maps. The number of input connections depends
experiments analysis, V contains the application and finally        on the number of attributes to be used in the classification.
VI contain the conclusion.                                          The neuron with weights closest to the input data vector is
                                                                    declared the winner during the training. Then the weights of
                                                                    all of the neurons in the neighbourhood of the winning neuron
                    III. SOM AND PCA                                are adjusted by an amount inversely proportional to the
                                                                    distance. It clusters and classifies the data set based on the set
A. lf Organizing Map                                                of attributes used.
In this section, we will give a brief introduction on Self-         The algorithm is summarized as follows [4, 5]:
Organizing Map (SOM) [3, 4] (Kohonen, 1985). SOM has
been proposed by Kohonen in the early eighties (Kohonen,
1985) [5]. Since that time, it has been used most widely for        Step 1. Initialization:
data analysis in some areas such as economics physics,              Choose random values for the initial weight vectors wj(0) ,
chemistry or medical applications. The SOM provides an              the weight vectors being different for j = 1, 2,...,l where 1 is
orderly mapping of an input high-dimensional space in much          the total number of neurons.
lower dimensional spaces, usually one or two dimensions. As
it compresses information while preserving the most                  Step 2. Sampling:
important topological and metric relationships of the primary       Draw a sample x from the input space with a certain
data items, it can be thought to produce some kind of               probability.
abstractions of information. So it can be utilized in a number
of ways in complex tasks such as pattern classification,            Step 3. Similarity Matching:
process analysis, machine perception, control, and                  Find the best matching (winning) neuron i (x) at time steps n
communication [3, 4].                                               by using the minimum Distance Euclidean criterion

                                                                                  i(x) = arg min||x(n) – wj||,       j = 1,2,...,l

                                                                     Step 4. Updating:
                                                                    Adjust the synaptic weight vector of all neurons by using the
                                                                    update formula

                                                                           Wj (n + 1) = wj (n) + (n)hj, i(x) (n)(x(n) - wj (n))

                                                                              (n) is the learning rate parameter, and
                                                                              hj, i(x)(n) is the neighbourhood function centred
                                                                              around the winning neuron i(x).

                                                                             Both (n) and hj, i(x) (n) are varied dynamically
                                                                             during learning for best results.

                                                                    Step 5.
           Fig. 3 An example of a Two-Dimensional SOM.              Continue with step 2 until no noticeable changes in the feature
                                                                    map are observed.

T. Kohonen introduced the Self-Organizing Map (SOM) [3, 4,
6]. It is an unsupervised learning process, which learns the
distribution of a set of patterns without any class information.
It has the property of topology preservation. There is a
competition among the neurons to be activated or fired .The
result is that only one neuron that wins the competition is fired   B. Principal Component Analysis
and is called winner-takes all neuron. SOMs may be one-             The Principal Component Analysis (PCA) [10] is one of the
dimensional, two-dimensional or multidimensional, but the           most successful techniques that have been used in image

ISSN: 2231-2803                                           -3-                                                    IJCTT
International Journal of Computer Trends and Technology- March to April Issue-2011
recognition and compression. PCA is a statistical method
under the broad title of factor analysis. The purpose of PCA is
to reduce the large dimensionality of the data space (observed
variables) to the smaller intrinsic dimensionality of feature
space (independent variables), which are needed to describe
the data economically. This is the case when there is a strong
correlation between observed variables.

The jobs which PCA can do are prediction, redundancy
removal, feature extraction, data compression, etc. Because
PCA is a classical technique which can do something in the
linear domain, applications having linear models are suitable,    Normal (original) images              Eginefaces images
such as signal processing, image processing, system and
control theory, communications, etc.                                            Fig. 4 Example of Eginefaces [1, 10].

Kirby and Sirovich [7] developed a technique based upon
which the idea of eigenfaces was used. The main idea of using                  IV. COMBINING SOM AND PCA
PCA for face recognition is to express the large 1-D vector of
pixels constructed from 2-D facial image into the compact         As discussed earlier that SOM [3, 4] is an unsupervised
principal components of the feature space. This can be called     learning process and this technique has the property of
eigenspace projection. Eigenspace is calculated by identifying    topology preservation. It defines a mapping from an input
the eigenvectors of the covariance matrix derived from a set of   space onto a set of nodes in a space that has dimension much
facial images (vectors). These eigenfaces represented the faces   lower than that of the input space. The set of nodes is
efficiently using Principal Component Analysis. An image          topologically ordered. An image, divided into sub blocks, is
was treated as a vector in a very high dimensional space. Only    mapped to a lower dimensional space with topologically
the best eigenfaces (eigenvectors of the covariance matrix of a   ordered set of nodes thereby providing dimensionality
set of images) those that had the largest eigenvalues were used   reduction. Further feature extraction is provided with the
to approximate the face [5].                                      method known as Karhunen - Loeve (KL) [7] transform via
                                                                  Principal Component Analysis (PCA). A set of orthogonal
Consider a set of N sample images {1,  2 ...  N} taking        axes of projections called as principal components or the
values in an n-dimensional image space. Let us also consider a    eigenvectors are generated by PCA. PCA is applied to the
linear transformation mapping the original n-dimensional          weight matrix generated by mapping the image onto lower
image space to m-dimensional feature space where m < n.           dimensional space using SOM. For the further reduction of
The new feature vectors Yk are defined by the following           dimensionality, the eigenvectors with smaller eigenvalues are
linear transformation [5]:                                        ignored and the eigenvectors corresponding to the largest
                                                                  eigenvalues are retained for image reconstruction.
                         Yk =  T k                              In PCA method [1, 9, 10] the pixels of each row of the
                                                                  training images, taken one at a time, are concatenated
where,                                                            vertically to form a single vector containing all the pixel
          is a matrix with orthonormal columns                   values of an image thereby producing a matrix, each column
        .                                                         of which represents an image and there are as many numbers
The covariance matrix is defined as                               of columns as the number of training images [7]. Nearly half
                                                                  of the total number of eigenvectors, corresponding to the
                      N                                           largest eigenvalues, obtained from the covariance matrix of
                  C =  (k - Ψ)(k - Ψ) T                        the training images, is retained. The test images are
                     K=1                                          reconstructed after finding out its KL coefficients using the
where,                                                            retained eigenvectors and matching is done using Euclidean
         Ψ is the mean image of all the samples. Only m-          norm as the similarity measure.
number of n-dimensional eigenvectors [V1, V2, ... Vm] of C is
                                                                  Whereas in SOM method [5], the training images are mapped
chosen that correspond to the m largest eigenvalues [5].
                                                                  to lower dimension using SOM and the weight matrix of each
                                                                  training image is stored. At the time of recognition, the
Example of eginefaces:                                            training images are reconstructed using the weight matrices

                                                                  and matching is done with the test image using Euclidean
                                                                  norm as the similarity measure. The same procedure is
                                                                  adopted for the third technique when the two SOM & PCA are

ISSN: 2231-2803                                         -4-                                                    IJCTT
International Journal of Computer Trends and Technology- March to April Issue-2011
combined. The eigenvectors of the weight matrix are found              varied from 10 to 20 to 40 and then the recognition rate
which are obtained through SOM training and sixty four                 was found. Table 1 clearly show the recognition rate (%)
percent of the total number of eigenvectors corresponding to           as the number of classes is varied. Table 1observations
largest eigenvalues and hence the KL coefficients is retained          indicates that as the number of classes is increased then
and at the time of recognition, the images are reconstructed           the recognition rate is decreases and hence automatically
and matching is done with the test images. In the second case,         performance of the system is decreased. The results
the weight matrix was obtained by using the eigenvectors and           shows that the recognition rate of SOM is slightly better
the KL coefficients and the weight vectors corresponding to            that that of PCA technique. While for the combined SOM
first twenty neurons were considered for reconstruction of the         & PCA (1) method, sixty four percent of eigenvectors
image thereby reducing the memory space required to store              corresponding to the largest eigenvalues were retained for
the image.                                                             reconstruction of the image.

                                                                  B. In the second experiment the performance of the face
                    TABLE 1                                          recognition system is checked when the sub block size is
RECOGNITION RATE OF THE FACE RECOGNITION SYSTEM                      changed. Table 2 shows the recognition rates as the size
                                                                     of the sub block is changed. It is clear from Table 2 that
                                                                     there is a little change in the recognition rate for SOM
                   Recognition Rate (%)
                                                                     and SOM & PCA combined (1) techniques whereas for
                                  Number of Classes                  SOM & PCA combined (2), it is less as compared with
         Method                                                      the first two techniques and it reduces more rapidly as the
                              10               20          40        size of sub block is increased.
   SOM (5 × 5)               94.06            90.72       89.92
   SOM (10 × 10)             94.06            91.86       90.82
   PCA                       93.39            90.25       89.51                       V. APPLICATIONS
   SOM+PCA (1)               77.75            72.08       62.64
                                                                  The natural use of face recognition technology is the
                                                                  replacement of PIN, physical tokens or both needed in
                                                                  automatic authorization or identification schemes. Additional
                     TABLE 2                                      uses are automation of human identification or role
                                                                  authentication in such cases where assistance of another
                SUB BLOCK SIZE [5]
                                                                  human needed in verifying the ID cards and its beholder.
                   Recognition Rate (%)                           There are numerous applications for face recognition
                                 Size of Sub Block
      Method                                                      A.   Government Use:
                      ( 4 × 4)     ( 8 × 8)           (16×16)             Law Enforcement: Minimizing victim trauma by
   SOM (5 × 5)         94.06        94.06              95.95               narrowing mugs hot searches, verifying identify for
   SOM +PCA (1)        77.75        77.17              72.83               court records, and comparing school surveillance
   SOM +PCA (2)        68.29        62.17              54.89               camera images to know child molesters.
                                                                          Security/Counterterrorism.     Access      control,
                                                                           comparing surveillance images to Know terrorist.
                                                                          Immigration: Rapid progression through Customs.
Experimentation Analysis
For the experimentation [5, 9] on combined SOM and PCA            B.   Commercial Use:
the ORL [4, 11] (Olivetti Research Laboratory) face database
has been used. This ORL database basically contains 10                    Day Care: Verify identity of individuals picking up
different images of 40 distinct subjects. Now Two –                        the children.
dimensional SOM was trained and finally one weight matrix                 Residential Security: Alert homeowners of
of 25×16 was generated and after that PCA was applied to the               approaching personnel.
transpose of the weight matrix and eigenvalues were retained              Voter verification: Where eligible politicians are
for reconstruction of image. The experiments are as follows:               required to verify their identity during a voting
                                                                           process this is intended to stop ‘proxy’ voting where
                                                                           the vote may not go as expected.
A. In the first experiment to see the performance of face                 Banking using ATM: The software is able to quickly
   recognition system the number of classes was varied from                verify a customer’s face.
   one value to another value. The ORL databases were

ISSN: 2231-2803                                             -5-                                              IJCTT
International Journal of Computer Trends and Technology- March to April Issue-2011
         Physical access control of buildings areas, doors, cars             [8]    Face Recognition Using Self-Organizing Maps, Qiu Chen, Koji Kotani,
                                                                                     Feifei Lee and Tadahiro Ohmi, Tohoku University, Japan.
          or net access.
                                                                              [9]    “Face Recognition using PCA versus ICA versus LDA cascaded with
                                                                                     the Neural Classifier of concurrent Self-Organizing Maps”, Victor-
                                                                                     Emil    Neagoe, Senior Member IEEE, Ioan Anton Stanculescu,
                          VI. CONCLUSION                                             University of Bucharest.

Thus Face recognition is a fascinating research area. In this                 [10]   “Face Recognition Using Improved Fast PCA Algorithm”, Neerja,
                                                                                     Ekta Walia, University Patiala (Punjab), INDIA.
paper information about two very innovative techniques i.e.
SOM and PCA is given in detail. And also explained how                        [11]   AT&T Laboratories Cambridge, The database of faces at ttp: //
these techniques can be used for face recognition. Both                     research/ dtg/attarchive/facesataglance.html.
techniques are very useful in face recognition. In this paper
we have explained SOM and PCA approach for face
recognition. With the help of this paper a new idea was
proposed. New idea is nothing but combining both SOM and
PCA together. PCA was combined with SOM for
dimensionality reduction and feature extraction. This is true
for all the three methods i.e. SOM, PCA and SOM & PCA
combined. If we compare combined SOM & PCA method
with other two methods then we will come to know that the
decrease is more in case of combined method. The
experimental results also shows that if the size of the sub
block is varied or changed, then there is only small significant
change in the performance of the recognition system for SOM
and for the technique where the two SOM & PCA are
combined. For the combined technique where lesser number
of neurons of the weight matrix is used for image
reconstruction, the recognition rate is less as compared to the
earlier two techniques and decreases more rapidly. The
reduced number of neurons and hence reduced number of
features result in decrease in the recognition rate, but at the
same time, it results in saving in the memory space required
for storing the images.


[1]    W.Zhao, R. Chellapa, A. Rosenfeld, P.J Phillips,, "Face Recognition:
      A Literature Survey ," ACM Computing Surveys, Vol.35, No.4, pp.
      399-458, 2003.

[2]   R. Chellapa, C.L.Wilson, S.Sirobey, "Human and Machine
      Recognition of Faces: A Survey," Proc. IEEE, Vol.83, No.5, pp. 705-
      740, 1995.

[3]   Santaji Ghorpade, Jayshree Ghorpade, Shamla Mantri, ‘Pattern
      Recognition using Neural Networks’, IJCSIT, Vol 2, No 6, Dec- 2010,
      pg.92-98, India.

[4]   Santaji Ghorpade, Jayshree Ghorpade, Shamla Mantri, Dhanaji
      Ghorpade ”Neural Network For Face Recognition using SOM”, IJSCT,
      Vol. 1, Iss ue 2, Dec-2010, pg.65-67, India.

[5]   “Face Recognition using Self-Organizing Map and Principal
      Component Analysis”, Dinesh Kumar, C.S. Rai and Shakti Kumar, 0-
      7803-9422-4/05, 2005 IEEE.

[6]   Kohonen, T.(1985). Self - Organizing Maps, Springer –Verlag, Berlin.

[7]   M.Kirby and L.Sirovich, "Application of the Karhunen-Loeve
      Procedure for the Characterization of Human Faces," IEEE Trans.

ISSN: 2231-2803                                                   -6-                                                             IJCTT

Shared By:
Description: International Journal of Computer Trends and Technology(IJCTT)