MARKOV NETWORK-BASED

UNIFIED CLASSIFIER
FOR FACE RECOGNITION
DOMAIN: IMAGE PROCESSING
GUIDED BY,
Tmt. M. BLESSY QUEEN MARY ME.,

PRESENTED BY,
AARTHI K (1218101)
PARKAVI B (1218131)
RUBY BENITA K (1218139)
SANJANA L (1218140)
1
 OBJECTIVE
 To achieve an optimal face
recognition rate by
implementing and analysing
various face recognition
algorithms.

2
 PROBLEM
STATEMENT
One – To – Many Recognition
Problem
 For a face image as a query,
“Is it possible to retrieve a
single desired face image from a
large image database ?”
Positively not easy!
 Because the accuracy in face
recognition is not satisfactory
when the input image has many
variations different from training
database. 3
LITERATURE
SURVEY-I
Title: Bayesian Face Revisited: A
Joint Formulation
Technique: A naive and Core joint
formulation along with a model
learning algorithm.
Advantages:
 92.4% test accuracy on the
challenging Labelled Face in Wild
(LFW) dataset.
Reduction on error rate by 10% in
comparison with current best
commercial system.
Disadvantages: when the training
data size is small, the naive
formulation is the worst and it
LITERATURE SURVEY-II
Title: A Scalable Formulation of
Probabilistic Linear Discriminant
Analysis(LDA): Applied to Face
Recognition.
Technique: Probabilistic linear
discriminant analysis (PLDA).
Advantages: On large Multi-PIE
database, we illustrate the gain in
performance when using more
training samples per identity
(class).
Disadvantages: It involves
inversion of
a matrix whose size grows
LITERATURE SURVEY-II
Title: A Scalable Formulation of
Probabilistic Linear Discriminant
Analysis(LDA): Applied to Face
Recognition.
Technique: Probabilistic linear
discriminant analysis (PLDA).
Advantages: On large Multi-PIE
database, we illustrate the gain in
performance when using more
training samples per identity
(class).
Disadvantages: It involves
inversion of
a matrix whose size grows
LIERATURE SURVEY-
III
Title: Learning face representation
from Scratch
Technique: A semi- automatical way
to collect face images from Internet
and builds a large scale dataset-
CASIAWebFace.
Advantages: Anyone can easily train
a high performance face recognition
engine with CASIAWebFace.
Disadvantages: Ineffective
annotation tools and algorithms.
MODULE -I
 Linear Discriminant Analysis (LDA )
is a method used in pattern
recognition to find a linear
combination of features that
characterizes two or more classes
of objects.
 LDA has lower error rates
 Works well even if different facial
express.

8
FISHERFACES ALGORITHM-TRAINING

 a1   b1   c1   d1 
       
a2  b c d
      
2 2 2
      
 
       
a b c d
 N 
 N 
2
 N 2
 N 2 2

 e1   f1   g1   h1 
       
e f2  g h

2 
 
2 
 
2

        
       
e  f N 2  g h
 N2   N 2
 N 2

9
 CONTD..
We compute the average of all faces.

 a1  b1    h1 
 
 1  a2  b2    h2 
m , where M 8
M   
 
 aN 2  bN 2    hN 2 

1
0
 CONTD…
Compute the average face of each person

 a1  b1   c1  d1 
   
 1  a2  b2  1  c2  d2 
x , y ,
2    2   
   
 aN 2  bN 2   cN 2  dN2 
 e1  f1   g1  h1 
   
 1  e2  f2   1  g2  h2 
z , w
2    2   
   
 eN 2  f N 2   gN2  hN 2 
1
1
CONTD…
And subtract them from the training faces

 a1  x1   b1  x1   c1  y1   d1  y1 
    b2      
 a2  x2  x2  c2  y2  d2  y2 
am  , bm  , cm  , d m  ,
           
       
 aN 2  xN 2   bN 2  xN 2   cN2  y N2   d N2  y N2 

 e1  z1   f1  z1   g1  w1   h1  w1 
         h2  
 e2  z2  f2  z2   g2  w2  w2 
em  , fm  , g m  , hm 
            
       
 eN 2  z N 2   fN2  z N 2   g N2  wN 2   hN 2  wN 2 
12
MODULE II-RECOGNITION
 We build scatter matrices S1, S2, S3, S4

 
   
S1  am am  bmbm , S 2  cm cm  d m d m ,


 
   
S3  em em  f m f m , S 4  g m g m  hm hm


 And the within-class scatter matrix SW

SW S1  S 2  S3  S 4
13
 CONTD…
 The between-class scatter matrix

 
S B 2  x  m  x  m   2  y  m  y  m   2  z  m  z  m   2 w  m w  m 
   

 We are seeking for a matrix which has minimized

with-in class differences and maximized between
class distances at the same time.
 From this matrix , we could recognize the highly
equivalent images.

14
 CONTD….
 Columns of W are eigenvectors of
 We have to compute the inverse of SW
 We have to multiply the matrices
 We have to compute the eigenvectors


Columns of W are eigenvectors satisfying , =

 The eigenvalues are roots of ,

 Get eigenvectors by solving , = 0

15
CONTD…
 Project faces onto the LDA-
space

xLDA W  x , y LDA W  y ,
 
z LDA W  z , wLDA W  w

 Calculate the minimum deviation

using Euclidean Distance
16
 PROBLEMS IN EXISTING
SYSTEM
 The existing system does not work well if
there are different illuminations

 It also has pose-Illumination problem

 It does not give high recognition rate in such

cases
 MODULE-III
Patch based local generic representation
+ + , i = 1,2...,S

LGR based classification:

 LGR based classification scheme can be
proposed to determine the class label of query
face z .
 The sum of the weighted residual over all

patches should be calculated.

 The query sample z is classified to the class
22
which has the minimal weighted representation
residual over all patches.
 CONTD…
We make use of multiple classifiers and
assign observation nodes to those of a
query image and hidden nodes to those of
gallery images.

For each observation-hidden node pair,the

relationship between the hidden nodes in
terms of a similarity matrix among the
retrieved gallery images.
 CONTD…
We analysemultiple classifiers by
modelling the relationship between
neighbouring classifiers, where
there is no loop for the sake of
simplicity.
We first define the nodes and the
compatibility functions, derive the
relevant equations, and then
discuss belief propagation for the
recognition task.
Belief propagation calculates the
marginal probability of the result
24
node i using
( ) = () () (,)
AR DATASET
EXPRESSION
DISGUISE
ILLUMINATION
DISGUISE + ILLUMINATION
GRAPH
 REFERENCES
1.Wonjun Hwang, Member, IEEE, and Junmo Kim,
Member, IEEE , “Markov Network-Based Unified Classifier
for Face Recognition”, IEEE Transactions on image
processing,Vol.24, no.11, November 2015
2.Pengfei Zhu1, Meng Yang2, Lei Zhang1⋆, and Il Yong
Lee, Local Generic Representation for Face Recognition
with Single Sample per person.
3.D. Chen, X. Cao, L. Wang, F. Wen, and J. Sun,
“Bayesian face revisited:A joint formulation,” in Proc.
12th Eur. Conf. Comput. Vis., vol. 7574.Oct. 2012, pp.
566–579
4.H. T. Ho and R. Chellappa, “Pose-invariant face
recognition using Markov random fields,” IEEE Trans.
Image Process., vol. 22, no. 4,pp. 1573–1584, Apr. 2013
5.S. Liao, D. Shen, and A. C. S. Chung, “A Markov
random field groupwise registration framework for face 31
recognition,” IEEE Trans. Pattern Anal. Mach. Intell., vol.
36, no. 4, pp. 657–669, Apr. 2014.
THANK YOU
32