Exaggerate Self Quotient Image Model For Face Recognition Enlist Subspace Method

Document Sample
Exaggerate Self Quotient Image Model For Face Recognition Enlist Subspace Method Powered By Docstoc
					                                          (IJCSIS) International Journal of Computer Science and Information Security,
                                          Vol. 9, No. 6, June 2011
      Exaggerate Self Quotient Image Model for Face
          Recognition Enlist Subspace Method
            S.Muruganantham1,                                                   T.Jebarajan2
             Assistant professor                                                  Principal
             S.T.Hindu college                                           Kings College of Engineering
         Nagercoil, India -629003                                          Chennai, India - 600105
      Email: smuru_1970@yahoo.co.in

Abstract - The attribute of reliance facial                      Up to now many prosperous algorithm
recognition technique is frequently impinge              are intend for face recognition. Principle
on variation of illumination such as shadow              component analysis ( PCA ) [1] , Independent
and illumination direction changes. In this              component analysis (ICA ) [2] , Fisher’s linear
paper, to enrich the performance of self                 discriminant ( FLD ) [3] are three basic
quotient image model for the elimination of              algorithm for subspace analysis in face
lighting effect for face recognition is recited          recognition.
and dispelled in detail. A Histogram                             Besides the QI method, much other
equalization is embrace to enhance the                   technique has been proposed for face
contrast of samples. Then normalize the                  recognition under various illumination
samples as the result of enhanced SQI. To                conditions in Holocene year. SQI perpetuate QI
apply enhanced SQI to face recognition                   to perform illumination subtraction without the
subspace analysis algorithms ( PCA, KPCA                 need for alignment and no shadow assumption
and ICA ) are tout to perform subspace                   as QI does.
analysis on normalized samples. To                               Recently some researchers paying close
prosecute experiments on CAS-PEAL face                   solitious to preprocessing of face images to
database. The face samples are preprocessed              tune-up lighting in an image by initiate some
by enhance SQI to meliorate the                          models, like the quotient image QI [4,5,6]
performance        of    subspace      analysis          based methods, illumination cone based
algorithms. Experimental result canonical                methods [7,8], total variation based methods
the preprocessing of samples could salutary              (TV) [9,10]. Shashua et al. propound the
the robust to not only lighting but also facial          quotient image which is an image ratio
expressions, masking and occlusion etc. in               between a test image and a linear unification of
face recognition domain.                                 three images illuminated by non coplanar
                                                         lights, authentic only on albedo information
       Keywords:     Face     recognition,               and therefore is illumination free.
Histogram equalization, Subspace analysis,                       QI method could improve the
illumination.                                            performance, but it exigeny a large training set
                                                         and this is not practical in many predicament.
                                                         Wang et al. proposed a self quotient image
                I. Introduction
 
                                                         model (SQI) [6].
       The illumination problem has been                         SQI is proposed based on basic
believed to be one of most difficulties in face          conception of QI. SQI implements the
recognition. The image variation from lighting           normalization by dividing the low frequency
changes is more portentous than that from                part of the prime image, since the lighting
divengent personal identities.                           effects are represented as the low frequency



                                                   264                               http://sites.google.com/site/ijcsis/
                                                                                     ISSN 1947-5500
                                                      (IJCSIS) International Journal of Computer Science and Information Security,
                                                      Vol. 9, No. 6, June 2011



part. But the edge information is over                               illumination invariant properties as ostensive
smoothed by the smoother filter. To overcome                         below using Lambertian model but with
this problem by applying lighting invariant                          shadow. When shadow present the Lambertian
preprocessing methods.                                               model equation can be epitomize a
        In order to vanquish the drawback of
SQI, a new method enhanced self quotient                                   I = max ( nT · S, 0) ------------- (3)
image is proposed and dispeled in detail. To
apply ESQI to face recognition, feature is                           Analysis
essensed using subspace method and the                                      In a self quotient image, we regard
performance of PCA, KPCA and ICA are                                 three regions with divergent shapes and
ascertained and compared. A large scale CAS-                         shadow conditions.
PEAL face database is used for check the                             Region 1:
validity of algorithms. Compared with SQI                                   In the absence of shadow and presents
face sample preprocessed, the ESQI model                             small surface normal variation. In this case,
could improve the performance of face                                nT(u,v) ≈C1 , where C1 is a constant, then we
recognition algorithm.                                               have


 
                      II. SQI frame work                             Q(u,v)=          ≈              =             ------- (4)

         The Lambertian model can be
decomposition into two parts 1. The intrinsic                                In this case Q is quantisly illumination
part. 2. The extrinsic part.                                         free and be contingent on the albedo of the face
                                                                     Region 2:
I (x, y) =        (x, y) n(x, y)T · S -------------- (1)                     In the absence of shadow but present
             = F (x, y) · S                                          large surface normal variation. In this case
                                                                     nT(u,v) S is not a constant. The SQI is given
                                                                     by
Where        is albedo and n is the surface normal.
        The main problem for accurate face
recognition is to separating the two factors and                     Q(u,v) =           =                         -------- (5)
detach the extrinsic factors are a key to
attaining robust face recognition.                                   In this case Q is contingent on surface normal,
                                                                     albedo and illumination
                    III. Self quotient image                         Region 3
               




                                                                             In shadow regions. In this case the gray
     We define the self quotient image as an                         value of image is relatively low and does not
intrinsic property of face image. The self                           vary. Let us feigned that light is harmoniously
quotient image Q of image I is ipsofacto by                          scattered from all directions in shadow regions.
                                                                     The summation of the dot product between n
     Q= =                    ---------------- (2)                    and Si is constant in such region.

Where * is convolution operator, I is the                            NT(u,v)·          Si(u,v)=          nT(u,v)·Si(u,v)=C2
smoothed I and F is the smoothing kernel and
the division id point-wise as in the prime                           Where C2 is a constant and
quotient image. The self quotient image has


                                                               265                                http://sites.google.com/site/ijcsis/
                                                                                                  ISSN 1947-5500
                                                 (IJCSIS) International Journal of Computer Science and Information Security,
                                                 Vol. 9, No. 6, June 2011



        Thus I(u,v) in shadow region can be
written as I(u,v) ≈C2 (u,v), then we have


Q(u,v)          =              =      -----(6)                                       ESQI
                                                                                     Fig-1
                                                                        From Fig – 1 HE method does discard
        In such region Q is illumination free                   most of impact of lighting from face images.
and also secluds the shadow effect.                             But HE has some obstacle. The face image
        The advantage of SQI method is 1.                       preprocessed by HE method some important
Does not need alignment procedure. 2. No                        regions are blocked. Face feature like eyes nose
training images are exigency. 3. SQI is good                    and mouth are not persist in these blocked
for eradicate shadow. 4. Lightning source can                   regions. In our proposed method ESQI to
be any type.                                                    persist some important features after HE
                                                                method. Fig -1 also indicates some examples of
              IV. Proposed method                               normalized face samples got by our model
                                                                (row marked with ESQI). Analogous with HE
       In our proposed method we use a                          some consequential regions are not blocked.
weighted Gaussian filter, given by                              Analogous with ancestral HE method and SQI
                                                                method these advantages of our proposed
         F = W G --------------------- (7)                      model make normalized face samples got by
                                                                our model more apt to face recognition on large
Where W is the weight and G is the Gaussian                     scale face database.
kernel.                                                         Given the training set X={x1,x2,……… xN}
         In this application we adopt this filter               where N is the number of samples.
with a multi scale version. Histogram                                   In order to enhance contrast of every
equalization      (HE) method has          been                 sample xi wield histogram equalization, a flat
countenanced to be a cogent method to adjust                    histogram H1 with gray level K1 is procreated
lighting in images.         Some Examples of                    where i = 1,2…………. , N.
preprocessed images by HE are illustrated in
fig -1( row marked with HE)                                     H1 =                 * (n2/k1) ---------------- (8)

 Accessory Distance Expression Lighting
                                                                        A gray scale transformation T (· ) is
                                                                chosen to minimize
                                                                        │C1 (T (K) ) ─ C0 ( K ) │
                                                                Where C0 (·) is the cumulative histogram of xi.
                    Original                                            To procure the grayscale transformation
                                                                T(·) minimization of equation (8) may be
                                                                rewritten to minimize the error intervening the
                                                                optimal and actual cumulative histogram.
            Histogram equalization(HE)                                    The error is defined by

                                                                Err={CT          -           }+            (      ) --- (9)


                       SQI


                                                          266                               http://sites.google.com/site/ijcsis/
                                                                                            ISSN 1947-5500
                                              (IJCSIS) International Journal of Computer Science and Information Security,
                                              Vol. 9, No. 6, June 2011



        Where H0 is the actual histogram of xi,              (595 males and 445 females) with varying
K0 is the gray level of H0, C0 is the actual                 pose, expression, accessory and lighting. In the
cumulative histogram, C1 is desired cumulative               following experiment all the faces are rotated
histogram. Every sample xi is preprocessed to                resized and cropped to 64 X64 with 256 gray
map its gray value from the actual K0 gray                   levels according to the coordinate of two eyes.
level to the desired K1 gray level and to enrich
its contrast as                                                     Four probe sets in the frontal subsets
                                                             expressions, lighting, accessory and distance
Xi = T (xi) -------------------------- (10)                  are chosen to the algorithm. To test the validity
                                                             using our proposed ESQI model in Face
For these enhanced face samples xi appertain to              recognition.
the famous SQI method.
                                                                    The following Table-1 illustrates the
Subspace analysis                                            preciseness of recognition rate at rank -1 of
                                                             various approaches on the CAS-PEAL face
        We appertain PCA, KPCA and ICA to
                                                             database respectively. It clearly show that our
perform subspace dialytic on these normalized
                                                             proposed model improve the performance of
face samples. Bartlett et al [2] has proposed
                                                             ICA2, KPCA and PCA against accessory,
and discussed two architecture about how to
                                                             distance and facial expression.
organize samples in the training set and they
also corroborate that the second architecture
                                                                           L1      L2        Cos       Md      L1      L2          Cos    Md
about how to organize samples could                                                  Accessory                              Distance

outperform the first one. So in this paper, the                 ICA2      48.93   41.79     52.74     40.75   78.18   72.56       85.46   71.96

second architecture about how to organize                     ESQI+ICA2   45.51   49.85     58.45     50.24   80.36   82.36       92.36   82.00

samples in the training set for ICA ( ICA2) [2]                 KPCA      40.20   32.60     20.18     46.00   86.11   63.45       30.15   74.64


is chosen.                                                    KPCA+ICA2   48.97   38.30     36.30     50.32   90.12   90.45       75.32   81.66

                                                                PCA       46.15   25.50     15.06     40.35   84.18   54.78       26.44   75.64


        In order to codify features extracted by              PCA+ICA2    51.55   35.72     32.91     50.81   91.22   81.05       76.05   84.37

                                                                                    Expression                              Lighting
subspace analysis methods (PCA, KPCA and                        ICA2      55.03   54.18     60.10     56.63   8.22    7.98        10.39   10.67

ICA2) the nearest neighbor classifier is adopted              ESQI+ICA2   52.91   65.03     71.09     63.15   5.62    10.89       19.35   8.24

with four famous distances (L1, L2, cos and                     KPCA      71.55   65.45     37.33     56.61   8.11    3.00         0.5    11.98


mid) as affinity measurement for classifier in                KPCA+ICA2   75.11   70.11     65.98     64.26   12.00   6.05         7.09   10.00


face recognition.                                               PCA       68.58   62.50     30.48     53.17   6.10    2.16         0.57   10.98

                                                              PCA+ICA2    75.12   63.85     65.35     64.20   12.29   6.50         3.77   10.05


The four distances are
L1 = DL1(x, y) = ∑ i │xi – yi │.                                    Table-1. The accuracy recognition
L2 =DL2(x, y) = (x, y)T (x, y).                              rates(%) at rank 1 of various approaches on
Cos = Dcos(x, y) = −xTy ║x║║y║.                              the CAS-PEAL face database.
Mid = Dmid(x, y) = (x─y)T ∑-1(x, y).
                                                                     Fig -2 gives the accuracy recognition
                                                             rate of PCA, KPCA and ICA2 employing our
    IV. Experimental result                                  proposed ESQI model at rank 1 -50. According
             
                                                             to fig -1 ICA2 achieved the best performance
       We investigate our algorithm on the
                                                             for accessory, distance and lighting test sets.
famous CAS-PEAL face database is chosen to
                                                             KPCA actualized the best performance for
verify and validate of various algorithms. It
                                                             expression testing set.
contains 99,594 images of 1040 individuals


                                                       267                                          http://sites.google.com/site/ijcsis/
                                                                                                    ISSN 1947-5500
                                      (IJCSIS) International Journal of Computer Science and Information Security,
                                      Vol. 9, No. 6, June 2011



       If the average preciseness rate is            Fig-2 (a, b,c,d). The accuracy recognition rates
considered ICA2 could outperform the other           of PCA, KPCA and ICA employing our
two algorithms advocated ICA2 employing our          proposed ESQI model at rank 1-50 with the
                                                     similarity measurement which has achieved the
proposed ESQI model for real application.
                                                     max accuracy recognition rate in Table-1

                                                                   VI. Conclusion
                                                      
                                                             The new proposed model enhanced self
                                                     quotient image rub-on to face recognition
                                                     employing sub space analysis method. To
                                                     practice the experiments on the CAS-PEAL
                                                     face database the samples are preprocessed by
                                                     our proposed model could improve the
               Fig-2a                                performance of some notorious subspace
                                                     analysis algorithm. And these result confirm
                                                     that the samples preprocessed by our model
                                                     could make PCA, KPCA and ICA2 not only
                                                     potent to light but also potent to facial
                                                     expressions, masking and occlusion etc. Thus
                                                     the new proposed model is apt to real
                                                     applications.

                                                                      References
                                                      
               Fig-2b
                                                     [1]. M. Turk and A. Pentland, “Eigen faces for
                                                     face recognition”, Cognitive Neuroscience,
                                                     Vol.3, 1991, pp. 71-86.

                                                     [2]. M. S. Bartlett, J. R. Movellan, and T. J.
                                                     Sejnowski, “Face recognition by independent
                                                     component analysis”, IEEE Transactions on
                                                     Neural Networks, Vol 13, 2002, pp. 1450-
                                                     1464.

                Fig-2c                               [3]. P. N. Belhumeur, J. P. Hespanha, and D. J.
                                                     Kriegman, “Eigen faces vs. Fisherfaces:
                                                     Recognition using class specific linear
                                                     projection”, IEEE Transaction on Pattern
                                                     analysis and Machine Intelligence, Vol 19,
                                                     1997, pp. 711-720.

                                                     [4]. S. Shan, W. Gao, B. Cao and D. Zhao,
                                                     “Illumination normalization for robust face
                                                     recognition against varying lighting condition”,
                Fig-2d                               in proceedings of IEEE International workshop


                                               268                               http://sites.google.com/site/ijcsis/
                                                                                 ISSN 1947-5500
                                           (IJCSIS) International Journal of Computer Science and Information Security,
                                           Vol. 9, No. 6, June 2011



on analysis and Modeling of faces and
gestures, 2003, pp. 157-164.

[5]. A. Shashua and T. Riklin Raviv, “The
quotient image: class based re-rendering and
recognition with varying illumination”, IEEE
Transaction on pattern analysis and Machine
intelligence, Vol 23, 2001, pp. 129-139.

[6]. H. Wang, S.Z. Li and Y. Wang,
“Generalized quotient image “,in proceedings
of IEEE international conference on Computer
vision and pattern recognition, Vol. 2, 2004,
pp. 498-505.

[7]. P. N. Belhumeur, and D. J. Kriegman,
“What is the set of images of an object under
all possible lighting condition?”,in proceedings
of IEEE international conference on Computer
vision and pattern recognition, 1996, pp. 270-
277.
[8]. A.S.Georghades, P.N. Belhumeur and
D.J.Kriengman,”From         few     to    many:
Illumination cone model for face recognition
under variable lighting and pose”, IEEE pattern
analysis and Machine intelligence, vol 23,
2001, pp. 643-660.

[9]. T. Chen, W. Yin, X.S. Zhou, D. Comaniciu
and T.S.Huang, “Illumination normalization
for face recognition and uneven background
correction using total variation based image
models”, in proceedings of IEEE international
conference on Computer vision and pattern
recognition, Vol 2. 2005, pp. 532-539.

[10]. T. Chen, W. Yin, X.S. Zhou, D.
Comaniciu and T.S.Huang, “Total variation
model for variable lighting face recognition”,
IEEE Transaction on Pattern analysis and
Machine Intelligence, Vol 28, 2006, pp. 1519-
1524.
 




                                                    269                               http://sites.google.com/site/ijcsis/
                                                                                      ISSN 1947-5500