Upload
others
View
4
Download
0
Embed Size (px)
Citation preview
125
5. Face Recognition Using Eigen Features of Multi-Scale Face Components and Artificial Neural Networks
Automated face recognition aims at exploiting face images to
identify human subjects. A number of approaches such as appearance
/holistic based, feature / component based and hybrid face recognition
approaches have been proposed in literature for automated face
recognition and these three approaches are discussed in detail in the
second chapter. The following section deals with an overview of hybrid
face recognition techniques existing in literature and the proposed
method.
5.1) Hybrid Approach for Face Recognition
In hybrid face recognition approach, either both holistic and
component information are used for recognition or it is defined as
combination of different face recognition methods. In paper [7] Mehrtash
et.al proposed a hybrid recognition system which tries to find the human
recognition behavior by efficiently combining the facial components in
the recognition task. In this method the decision system uses the
weighted majority strategy to fuse the results of whole image and facial
component classifier. Simulation studies justify the superior performance
of the proposed method as compared to that of Eigenface method.
In paper [6] Mazloom et.al presented a hybrid approach for face
recognition by handling three issues put together. For pre-processing
and feature extraction stages, a combination of wavelet transform and
PCA are applied, and for classification, the Neural Network (MLP) is
126
explored for robust decision in the presence of wide facial variations. The
experiments that have been conducted on the ORL database indicated
that the combination of Wavelet, PCA and MLP exhibited an excellent
performance. This is on account of the fact that it has the lowest overall
training time, the lowest redundant data, and the highest recognition
rates, compared to similar methods introduced so far. The proposed
method, in comparison with the present hybrid methods, enjoys a low
computation load in both training and recognizing stages.
In paper [8], Suvendu Mandal et.al has proposed a hybrid face
recognition system. The proposed method combines the structural
features with holistic features. In that experiment, the facial image is
partitioned into a number of petals considering nose tip as the origin.
From each petal, intensity profile is computed by radial projection and
then they are more compactly represented by applying Discrete Cosine
Transform (DCT) on them. To classify the images, correlation between the
vectors is used as the „similarity‟ measure. The proposed method is
simple to implement and is efficient. The performance of the proposed
method is encouraging and gives scope for further study.
A couple of hybrid approaches which use both holistic and local
features are also proposed in literature: Local Feature Analysis (LFA) is
an interesting biologically motivated method [106]. It demonstrates that
using the same number of kernels, the perceptual reconstruction quality
of LFA based on the optimal set of grids is better than that of PCA. But
127
no results on recognition performance based on LFA are reported. An
interesting shape model called Active Shape Model (ASM) [107]. In this
method both shape and grey level information are modeled and used. It
uses statistical models of the shapes of various face components which
iteratively deform to fit to an example shape in a new image.
Performance in terms of percentage of recognition for different
hybrid face recognition methods exists in literature are given by table
number 5.1, using eigen features of multi-scale face components and
artificial neural networks has shown better results compared to the
existing similar methods.
Table 5.1 – Performance Comparison of different hybrid face recognition techniques on ORL database.
S.No Method Percentage of
Recognition
1 Scale-robust feature extraction for face
recognition. (high resolution100X100) (low resolution12X14). [11]
90% 80%
2 Weighted modular principal component
analysis. [7] 87%
3 Hybrid method based on structural and
holistic. [21] 90.30%
4 Hybrid model using facial components.
[22] 88.57%
5 PCA, LDA and Neural Network for face
identification. [23] 95.80%
In the present work, a hybrid method for face recognition is
proposed in which facial components such as eyes, nose and mouth are
extracted from pre-processed face image using cropping technique. These
128
features and the whole face are down sampled with different resolutions
and a feature vector is formed, which is projected on eigen feature space
or fisher feature space to get weight vector and this vector is used as
input to the Artificial Neural Network for face classification.
5.2) Proposed Algorithm for Face Recognition
In this study, hybrid face recognition technique is proposed by
combining both componential cues (such as eyes, mouth, and nose) and
holistic information to improve the face recognition rate at different
resolutions.
In the proposed method, initially face images are pre-processed
(average filtering, histogram equalization and image resizing using Bi-cubic
Interpolation) then face features (eyes, nose and mouth) are extracted
manually by cropping techniques, then image patches of eyes, nose, mouth
and the whole face are down sampled using different sampling rates based
on component significance in recognition process (Ex eye 1:1, nose 1:2,
mouth 1:4 and whole face 1:8). These 2D face image patches of different
components are combined and represented by a matrix 𝐴 = [𝑎1 , 𝑎2 , … … 𝑎𝑖 , ]
where 𝑎1 represents a one-dimensional image column vector obtained by
scanning the two-dimensional image patches in lexicographic order and
writing them into a column vector (it is called feature vector).
Subspace analysis PCA or LDA are employed to the matrix A. PCA is
used for further dimensionality reduction and good representation, and
LDA is used to get better discrimination information.
129
The weight vector obtained by projecting training face on PCA or LDA
feature space is used as input to the Artificial Neural Network which is
trained by Back-propagation (BP) or Radial basis function (RBF) network
for recognizing faces. The flow chart and implementation steps for the
proposed face recognition technique are given as follows.
5.3) Flow chart for Proposed face recognition method using Eigen features of multi-scale face components and ANN (Figure 5.1)
5.4) Implementation of Algorithm for the Proposed Face Recognition Technique
The algorithm is described in the following steps -
ANN Training for face
classification
Face Data Base
Average filtering, histogram equalization and image
resizing
Feature Extraction by cropping face image
Multi-scaling of facial features
with different resolutions
Forming face feature vector for
each face in data base
Projecting feature vectors on
PCA/LDA space to find weight vector
130
5.4.1) Training Phase
1. Given face image is preprocessed and normalized .
2. Face image is re-sized by using Bi-cubic interpolation method.
3. The face ellipse is extracted by applying a standard mask.
4. Eyes, nose and mouth patches are cropped using image cropping
techniques and the locations of the eyes, nose and mouth patches are
kept same for all images in training and test database.
5. Dimensions of face components are reduced by down sampling face
comoponents with different resolution ratios based on component
significance in recognition process.
6. The remaining portion of the face is extracted by removing the face
component patches from the whole face and remaining portion of face or
the whole face are down sampled with higher sampling ratios as this
component is not much significant compared to the other components
such as eyes, nose and mouth.
7. Feature vector is obtained by scanning image patches obtained in (4),
(5) and (6) in lexicographic order and writing them into a single column
vector of dimension N x 1.
8. Feature weight vector is found by projecting each face feature vector
on PCA or LDA feature space.
9. Weight vector found in step 8 is used as input to artificial neural
network classifier and it is trained to classify the faces using back
propagation or RBF network algorithm.
131
5.4.2) Testing Phase
In testing phase, face image is selected from test data base and
steps 1 to 8 are carried out as it is done in training phase. But in the
last step, training of Artificial Neural Network is not required, in which
weight vector is directly fed to the already trained Artificial Neural
Network to identify the test face with one of the training face images.
5.5) Results and Discussion 5.5.1) Simulation Result
In the proposed face recognition algorithms, experiments are
conducted on both ORL and FERET databases. In ORL database 400
face images of 40 subjects are considered, performance of face
recognition techniques is measured in terms of percentage of recognition
and it is calculated by averaging results for 10 independent runs of the
program, where different combination of training and testing face images
are used. In FERET database only frontal face images are considered and
results are calculated only for one run of program. For evaluating
performance of partial face recognition, percentage recognition is
calculated with respect to resolution of face image, down sampling
resolution ratios of face components, variation in training set, number of
hidden layer nodes, and number of principal components.
Specifications of ANN architecture and training algorithms are selected as follows:
Number of Layers : 3 (input layer, hidden layer, output layer)
132
Number of input nodes : It is equal to number of elements in
feature vector
Number of output nodes : The number of face images to be
trained
Number of hidden : In RBFN, it is selected not exceeding to
layer nodes the maximum number of inputs.
In BPNN, it is limited to the maximum
number equal to the average of output
node and input nodes
Learning algorithms (BPNN) : Batch Gradient descent with
used in training process momentum (using „traingdm‟ in matlab)
Learning algorithms (RBFN) : Gradient descent
used in training process (using „newrb‟ in matlab)
Learning rate range : 0.01 to 1
Mean Square Error range : 0.01 to 0.001
Transfer functions : Logsig in BPNN
between input and Radial Basis Function in RBFN
hidden layer
Transfer functions : Logsig in BPNN
between hidden layer Logsig in RBFN
and output layer
133
Table 5.2 – Performance of face recognition w.r.t number of training faces for four different methods
a) Specifications of face recognition system for ORL database
Data Base ORL
Image Size 56X46
Multi-scaling Eye-1:1, Nose-1:2,
Mouth-1:4, Whole Face-1:8
ANN Structure 40: 50:40 (for
LDA39:50:40)
Number of Training
Faces for each class
Variable
Number of Testing Faces
Variable
Number of Principle Components
40 (for LDA 39)
b) Performance evaluation on ORL database
Number of Percentage of recognition
Train faces
Testing faces PCA
+BPNN PCA +RBFN
LDA +BPNN
LDA +RBFN
2 3 4
5 6 7
8
8 7 6
5 4 3
2
84.50 86.10 93.20
96.35 96.95 97.10
98.00
84.70 87.35 94.85
96.55 97.00 97.35
98.10
83.75 85.65 93.00
96.85 97.10 97.55
98.20
84.00 87.00 93.10
97.10 97.50 97.95
98.40
134
Figure 5.2a – Performance of face recognition w.r.t number of training faces for four different methods (ORL
database) c) Specifications of face recognition system for FERET database
d) Performance evaluation on FERET database
83
85
87
89
91
93
95
97
99
2 3 4 5 6 7 8
Pe
rce
nta
ge o
f R
eco
gnit
ion
Number of Training Faces
PCA+BPNN
PCA+RBFN
LDA+BPNN
LDA+RBFN
Data Base FERET
Image Size 56X46
Multi-scaling Eye-1:1, Nose-1:2, Mouth-1:4, Whole Face-1:8
ANN Structure 40: 50:40 (for LDA39:50:40)
Number of Training Faces for each class
Variable
Number of Testing Faces
Variable
Number of Principle
Components
40 (for LDA 39)
Number of faces Percentage of recognition
Train faces
Testing faces PCA
+BPNN
PCA
+RBFN
LDA
+BPNN
LDA
+RBFN
2
3 4
3
2 1
95.50
97.15 97.65
96.00
97.80 98
95.15
98.00 98.25
94.30
98.20 98.50
135
Figure 5.2b – Performance of face recognition w.r.t number of training faces for four different methods (FERET
database)
From the results cited in table number 5.2, the following inferences can
be made -
Face recognition percentage increases with number of training faces
in all four face recognition methods, because the Artificial neural
network adaptability capacity to recognize faces increases with
increasing number of training faces.
When number of training faces in each class is small compared to
testing faces, PCA face recognition method outperforms LDA. When
number of training faces is larger in each class, LDA outperforms
PCA.
LDA+RBFN method of face recognition outperforms the other methods
(98.5% for FERET 98.4% for ORL) as face discrimination ability of
92
93
94
95
96
97
98
99
2 3 4
Per
cen
tage
of
Rec
ogn
itio
n
Number of Training Faces
PCA+BPNN
PCA+RBFN
LDA+BPNN
LDA+RBFN
136
LDA is more compared to PCA whereas RBF network classification
efficiency is better compared to BPNN.
When numbers of training and testing faces are equal to 5 (on ORL
database), the face recognition percentage is found to be better
compared to the other similar face recognition methods in literature
(Table 5.1).
Table 5.3 – Performance of face recognition w.r.t Resolution of face components for four different methods
a) Specifications of face recognition system for ORL database
b) Performance comparison of PCA+BPNN, PCA+RBFN, LDA+BPNN and LDA+RBFN methods
Data Base ORL
Image Size 56X46
Multi-scaling Variable
ANN Structure 40: 50:40 (for
LDA39:50:40)
Number of Training Faces for each class
5
Number of Testing Faces 5
Number of Principle Components
40 (for LDA 39)
Row No.
Eye (L)
Eye (R)
Nose Mouth Whole face
Percentage of recognition
PCA +BPNN
PCA +RBFN
LDA +BPNN
LDA +RBFN
1 2
3 4 5
6
1:1 1:1
1:2 1:2 1:2
1:1
1:1 1:1
1:1 1:2 1:2
1:1
1:2 1:4
1:4 1:8 1:2
1:2
1:4 1:8
1:8 1:8 1:4
1:8
1:8 1:16
1:8 1:16 1:8
1:16
96.35 93.75
88.60 84.45 85.00
89.95
96.55 94.00
87.00 85.10 85.20
91.00
96.85 94.10
88.10 85.70 88.80
93.60
97.10 94.25
88.40 87.15 87.50
94.10
137
Figure 5.3a – Performance of face recognition w.r.t Resolution of face components for four different methods (ORL
database)
c) Specifications of face recognition system for FERET database
d) Performance comparison of PCA+BPNN, PCA+RBFN, LDA+BPNN
and LDA+RBFN methods Row No
Eye(L) Eye(R) Nose Mouth Whole face
Percentage of recognition
PCA+BPNN
PCA+RBFN
LDA+BPNN
LDA+RBFN
1 2 3 4 5 6
1:1 1:1 1:2 1:2 1:2 1:1
1:1 1:1 1:1 1:2 1:2 1:1
1:2 1:4 1:4 1:8 1:2 1:2
1:4 1:8 1:8 1:8 1:4 1:8
1:8 1:16 1:8 1:16 1:8 1:16
97.15 94.10 90 85.10 85.90 90.25
97.80 94.70 88.10 86.25 86.30 92.35
98 95.25 89 86.30 89.25 94.20
98.20 95.40 89.15 87.70 88.10 94.40
75
80
85
90
95
100
Set1 Set2 Set3 Set4 Set5 Set6
Per
cen
tage
of
Rec
ogn
itio
n
Resolution of face components
PCA+BPNN
PCA+RBFN
LDA+BPNN
LDA+RBFN
Data Base FERET
Image Size 72X48
Multi-scaling Variable
ANN Structure 40: 50:40 (for
LDA39:50:40)
Number of Training
Faces for each class
3
Number of Testing Faces 2
Number of Principle Components
40 (for LDA 39)
138
Figure 5.3b – Performance of face recognition w.r.t Resolution of face components for four different methods (FERET
database)
From the results cited in table number 5.3, the following inferences can
be made -
Percentage of face recognition decreases with decrease in resolution
for eyes, nose, mouth and the whole face for all the four methods.
LDA+RBF method of face recognition outperforms other methods with
percentage of face recognition (97.10% for ORL, 98.20% for FERET)
when resolution of eye is 1:1, nose is 1:2, mouth is 1:4 and the whole
face is 1:8
It can be seen from the 1st and 5th row of results that the change in
resolution of eye results in drastic change in the percentage of
recognition of face.
75
80
85
90
95
100
Set1 Set2 Set3 Set4 Set5 Set6
Per
cen
tage
of
Rec
ogn
itio
n
Resolution of face components
PCA+BPNN
PCA+RBFN
LDA+BPNN
LDA+RBFN
139
It can be observed from 1st and 6th row of results that the change in
resolution of mouth and whole face changes face recognition rate,
only slightly.
Table 5.4 – Performance of face recognition w.r.t Resolution of face
image a) Specifications of face recognition system for ORL database
b) Performance evaluation on ORL database:
Data Base ORL
Image Size Variable
Multi-scaling Eye-1:1, Nose-1:2,
Mouth-1:4, Whole Face-1:8
ANN Structure 40: 50:40 (for LDA39:50:40)
Number of Training Faces for each class
5
Number of Testing Faces
5
Number of Principle Components
40 (for LDA 39)
Resolution of Face
Percentage of recognition
PCA +BPNN
PCA +RBFN
LDA +BPNN
LDA +RBFN
28X23 42X35 56X46
84X69 112X92
85.10 90.05 96.35
96.60 97.00
87.30 91.25 96.55
97.00 97.15
88.00 91.40 96.85
97.50 98.10
88.30 92.00 97.10
98.00 98.25
140
Figure 5.4a – Performance of face recognition w.r.t Resolution of face image (ORL database)
c) Specifications of face recognition system for FERET database
d) Performance evaluation on FERET database:
75
80
85
90
95
100
28X23 42X35 56X46 84X69 112X92
Per
cen
tage
of
Rec
ogn
itio
n
Resolution of Face
PCA+BPNN
PCA+RBFN
LDA+BPNN
LDA+RBFN
Data Base FERET
Image Size Variable
Multi-scaling Eye-1:1, Nose-1:2, Mouth-
1:4, Whole Face-1:8
ANN Structure 40: 50:40 (for LDA
39:50:40)
Number of Training Faces
for each class
3
Number of Testing Faces 2
Number of Principle Components
40 (for LDA 39)
Resolution of Face
Percentage of recognition
PCA+
BPNN
PCA+
RBFN
LDA+
BPNN
LDA+
RBFN
24X16 48X32
72X43 144X96
192X128
85.70 90.50
97.15 97.20
97.25
88 92
97.80 97.95
98
88.60 92.20
98 98.05
98.10
89 92.50
98.20 98.30
98.40
141
Figure 5.4b – Performance of face recognition w.r.t Resolution of face image (FERET database)
From the results cited in table number 5.4, the following inferences can
be made -
Percentage of face recognition increases with increase in resolution of
face image for all the four methods.
LDA+RBF method outperforms other methods with percentage of face
recognition 98.25% when face image resolution is 112X92 for ORL
database, 98.40% for FERET database when face image resolution is
192X128.
The face recognition rate increases with the face image resolution and
levels off when arriving at one certain resolution (56X40 for ORL and
72X43 for FERET) it is concurring with conjecture that the face
discrimination information increases with the face image resolution
and remains stable when reaching one specific resolution.
75
80
85
90
95
100
24X16 48X32 72X43 144X96 192X128
Per
cen
tage
of
Rec
ogn
itio
n
Resolution of Face
PCA+BPNN
PCA+RBFN
LDA+BPNN
LDA+RBFN
142
Table 5.5 – Performance of face recognition w.r.t Variation in training set of face images for ORL database
a) Specifications of face recognition system for ORL database
b) Performance evaluation on ORL database:
Average recognition rate= 97.25 97.46 97.55 97.67
Refer section 3.5 for different face images in ORL and FERET database.
Data Base ORL
Image Size 56X46
Multi-scaling Eye-1:1, Nose-1:2, Mouth-1:4, Whole Face-1:8
ANN Structure 40: 50:40
Number of Training Faces for each class
5: by changing different faces in set
Number of Testing Faces 5: by changing
different faces in set
Number of Principle
Components
40
Set Training Set
Testing Set Percentage of recognition
PCA
+BPNN
PCA
+RBFN
LDA
+BPNN
LDA
+RBFN
1
2 3 4
5 6
1,2,3,4,5
2,3,4,5,6 3,4,5,6,7 4,5,6,7,8
5,6,7,8,9 6,7,8,9,10
6,7,8,9,10
1,7,8,9,10 1,2,8,9,10 1,2,3,9,10
1,2,3,4,10 1,2,3,4,5
96.10
97.00 97.25 98.50
98.30 96.35
96.20
97.25 97.40 98.85
98.55 96.55
96.25
97.95 98.30 98.00
98.00 96.85
96.85
98.00 98.40 98.20
97.50 97.10
143
Figure 5.5 – Performance of face recognition w.r.t Variation in training set of face images for ORL database
From the results cited in table number 5.5, the following inferences can
be made -
The face recognition rate changes based on training set and testing set
used in recognition process.
The maximum face recognition rate (98.85%) is achieved when training
set and testing set are balanced with respect to facial expression and
spectacles that is 4,5,6,7,8 images are used for training and 1,2,3,9,10
images used for testing, for the above mentioned sets PCA methods
outperforms the LDA methods as discrimination information is found to
be less in training set.
When discrimination information is high in set number 1 and 6, the
percentage of face recognition is found to be the maximum and LDA+RBF
method outperforms other methods.
94.5
95
95.5
96
96.5
97
97.5
98
98.5
99
99.5
Set1 Set2 Set3 Set4 Set5 Set6
Per
cen
tage
of
Rec
ogn
itio
n
variation in training set
PCA+BPNN
PCA+RBFN
LDA+BPNN
LDA+RBFN
144
Table 5.6 – Performance of face recognition w.r.t ANN structure for ORL database
a) Specifications of face recognition system ORL database
b) Performance evaluation on ORL database
Data Base ORL
Image Size 56X46
Multi-scaling Eye-1:1, Nose-1:2, Mouth-1:4, Whole
Face-1:8
ANN Structure Variable
Number of Training Faces for each class
5
Number of Testing
Faces
5
Number of Principle
Components
40
Number of Hidden layer
nodes
ANN Structure
Percentage of recognition
PCA +BPNN
PCA +RBFN
LDA +BPNN
LDA +RBFN
10 20
30 40 50
70
40:10:40 40:20:40
40:30:40 40:40:40 40:50:40
40:70:40
90.65 91.05
95.25 95.30 96.10
94.50
91.15 93.10
95.75 95.80 96.25
95.25
91.45 93.25
95.95 95.95 96.60
95.50
91.65 93.55
96.85 96.85 97.10
96.00
145
Figure 5.6 – Performance of face recognition w.r.t ANN structure for
ORL database
*For LDA instead of 40 inputs 39 inputs are used in ANN structure
From the results cited in table number 5.6, the following inferences can
be made -
The face recognition rate increases with the number of hidden layer
nodes, because dimension of weight matrix increases with hidden layer
nodes, but after certain dimension the percentage of recognition
decreases and maximum recognition rate is achieved for 40:50:40 ANN
structure.
LDA+RBF outperforms other methods as the discrimination information
is high in training set (1, 2, 3, 4 & 5).
90
91
92
93
94
95
96
97
98
10 20 30 40 50 70
Per
cen
tage
of
Rec
ogn
itio
n
Number of hidden layer nodes
PCA+BPNN
PCA+RBFN
LDA+BPNN
LDA+RBFN
146
Table 5.7 – Performance of face recognition w.r.t Number of principle components for ORL database
a) Specifications of face recognition system for ORL database
b) Performance evaluation on ORL database
Data Base ORL
Image Size 56X46
Multi-scaling Eye-1:1, Nose-1:2,
Mouth-1:4, Whole Face-1:8
ANN Structure Variable
Number of Training Faces for each class
5
Number of Testing
Faces
5
Number of Principle
Components
Variable
Number of principal components
ANN Structure
Percentage of recognition
PCA
+BPNN
PCA
+RBFN
LDA
+BPNN
LDA
+RBFN
15
25 30
35 40 50
100
15:40:40
25:40:40 30:35:40
35:50:40 40:50:40 50: 40:40
100:60:40
60.00
94.60 95.25
95.50 95.50 96.35
92.00
62.05
94.60 93.50
95.00 95.90 96.55
93.00
63.25
94.15 94.20
96.20 96.10 96.85
94.00
67.85
94.90 94.90
96.95 96.50 97.10
94.25
147
Figure 5.7 – Performance of face recognition w.r.t Number of principle components for ORL database
*For LDA 39:50:40 ANN structure is used.
From the results cited in table number 5.7, the following inferences can
be made -
The face recognition rate increases with principle components up to
30, which are 15% of 200 faces and levels off thereafter. But once
principle components cross 50, the percentage of recognition
decreases due to the complexity of ANN structure.
LDA+RBF outperform other methods and achieve maximum
recognition rate of 97.10% for 40 principle components, when
structure of ANN is 40:50:40.
60
65
70
75
80
85
90
95
100
15 25 30 35 40 50 100
Per
cen
tage
of
Rec
ogn
itio
n
Number of Principle Components
PCA+BPNN
PCA+RBFN
LDA+BPNN
LDA+RBFN
148
5.6) Conclusions
In the proposed work, face recognition method using feature weight
vector and Artificial Neural Network (BPNN or RBFN) classifier, resulted in
four face recognition methods: 1) PCA+BP 2) PCA+RBF 3) LDA+BP 4)
LDA+RBF and performance is evaluated for the above mentioned four
techniques on FERET & ORL databases in terms of percentage recognition
with respect to resolution of face components, resolution of face, number of
training faces, variation in training set, Artificial Neural Network structure
and number of principal components. Based on the results and analysis,
the following conclusions are made -
LDA+RBF method of face recognition outperforms the other methods
(with 98.40% recognition rate), as face discriminating ability of LDA is
higher compared to PCA, and RBF network classification accuracy is
more compared to back propagation neural network.
The face recognition percentage is at its highest with 97.10%
recognition rate for LDA+RBFN method, for low resolution face image
(56X46) with multi-scaling of eye 1:1, nose 1:2, mouth 1:4 and whole
face 1:8, when number of training and testing faces are equal to 5 on
ORL database. This percentage of recognition is the highest compared
to the other similar hybrid methods of face recognition in literature.
Percentage of face recognition increases with the number of training
faces in all four face recognition methods (PCA+BPNN, PCA+RBFN,
LDA+BPNN, LDA+ RBFN).
149
Face recognition rate increases in all four methods of face recognition
with resolution of face image and levels off, when arriving at one
certain resolution (56X40). It is concurring with conjecture that the
face discriminate information increases with the face image resolution
and remains stable when reaching one specific resolution.
Change in resolution of eyes changes the face recognition rate
drastically, whereas change in resolution of nose, mouth and whole
face changes the recognition rate slightly.
In training set of face images, if discriminant information is high, LDA
method outperforms PCA, and when discriminant information is low,
it is vice-versa.
The face recognition rate increases with number of hidden layer
nodes, because dimension of weight matrix increases with hidden
layer nodes, however, after reaching certain dimension, the
percentage of recognition decreases and maximum recognition rate is
achieved for 40:50:40 ANN structure.
The face recognition rate increases with principle components upto 30,
which are 15% of 200 faces and levels off. Once a principle component
crosses 50, the percentage of recognition decreases due to the
structure of ANN.