Fundamentals of Face Recognition Techniques
Fundamentals of Face Recognition Techniques
(eyes, nose and mouth) into column vector and this vector is used as
or fisher feature space (LDA) and this weight vector is used as input to
In the proposed partial face recognition system, the above cited two
considering only partial face components from eyes, nose and mouth. In
all the face recognition techniques proposed in this work require pre-
respectively.
72
face image pre-processing and feature extraction stages, with the help of
flow charts. Second part of this chapter explains about PCA, LDA and
related to the face recognition process are explained as fallows with the
FACE DATABASE
SELECT A FACE-SUBJECT
HISTOGRAM EQUALIZATION
+
NORMALIZATION
IMAGE RESIZING
face, the region of the image where the face is located is cut out from the
image and only this area is used in the process of face recognition. By
using cropping technique only main face can be extracted and the
that it includes main features of face such as eyes, nose, and mouth. All
using same corner points („imcrop‟ function in MATLAB is being used for
the dynamic range (contrast range) of the image and as a result, some
k nj
Sk = T rk = j=0 n (3.1)
'nj' is the number of pixels with gray level 'rk' ,and 'L' is the total number
Two histogram plots are given in Figure 3.3. The histogram on the
left is of the original face image (between 6-250) and the one on the right
recognition. The entire high and low illumination levels are adjusted so
face image is performed to get zero mean value and unity standard
deviation value.
76
The process of image resizing changes the size of the image, in this
work, the size of the image is scaled down to reduce the resolution of face
that pixel, which has got the flavor of all the 16 replaced pixels.
3.1.5) Masking
By using a mask, which simply has a face shaped region, the effect
Figure 3.4 – The shape of the face mask used in the pre-processing
77
maintaining the quality of face image. The 5*5 filter is used for this
process.
1 25
R= 25 i=1 𝑍𝑖 (3.2)
'z' is the mask, 'i' are mask elements. The mask is then convolved with
Mouth Remaining
Eyes 1:1 Nose 1:2 portion 1:8
1:4
emphasis on different parts of face e.g. eyes, nose, cheeks, forehead and
facial components - two eyes, nose and mouth of the face are extracted
from each face image). Dimensionality of these face components are then
combining them into a column vector. The size of the final image column
the four image patches, N depends on the size of face image, resolution
and down sampling ratios and which is very much less than the original
full image data size. In the proposed work, this feature vector is directly
79
3.2.2) Implementation
patches are kept same for all images in training and test
database(section 3.2).
extracted in step-4 and down sample this portion of face or whole face
(4), (5) and (6) in lexicographic order and combining them into a single
feature extraction for sample face image with dimention of 192 x 128 as
follows:
L
Converted into a single image vector: R 350 x 1
N
M
Re
3.3) Feature Dimensional Reduction Techniques using PCA / LDA
and Classification Techniques using Neural Networks
and the testing phases of BPNN and RBF network in detail and
classification accuracy, which also useful to make the system work faster
According to the PCA or LDA, the relative places between the data
and only the axes are being changed which handles the data from a
“better” point of view i.e., the generalization for PCA and discrimination
for LDA.
the pattern.
image data is not being distributed randomly and exist similarity in the
face image distribution involving the entire image space length N2,
space.”
Considering the training set of face images to be Γ1, Γ2,.., ΓM , the average
Φi = Γi – Ψ ----- (3.4)
83
distribution of the data. The kth vector, Uk, is chosen such that
1
𝜆𝑘 = 𝑀
𝑛=1 𝑈𝑘𝑇 Φn 2
------- (3.5)
𝑀
is a maximum, subjected to
1, 𝑖𝑓 l = 𝑘
𝑈𝐼𝑇 𝑈𝑘 = 𝛿lk = ------- (3.6)
0, 𝑜𝑡𝑒𝑟𝑤𝑖𝑠𝑒
was investigated.
Where, it can be found that Avi are the eigenvectors and 𝜇𝑖 are the
eigenvalues of C= A AT.
from the order of image dimension (N2) to order of the number of images
in training set (M). Most of the face recognition cases, the training set of
<< N2) and calculations also get reduced to manageable level. In order to
distinguish the variation among the images, the eigenvalues are useful to
rank the eigenvectors. Sirovich & Kirby [24] evaluated a limited version of
are digitized in a controlled manner and found that the 40 eigenfaces (M'
= 40) are sufficient for describing the face image. In practice, as accurate
input face image. Henceforth, the major idea is to trace the face
𝜀𝑘 = Ω − Ω𝑘 ------- (3.13)
Testing Set
Classifier
(Euclidean Distance)
Decision Making
separating the images of different classes take place due to the fisher
the number of classes of images). For example, two sets of points are
depending on the direction the points are either mixed (Figure 3.9a) or
separated (Figure 3.9b). The fisher discriminate to find the line which
best separates the points i.e., in order to identify the input test image,
the comparison of the projected test image with each training image
takes place after which the test image as the closest training image can
be identified.
being projected into a subspace. The test images being projected at the
same subspace can be identified using a similarity measure and the only
image and the LDA method discriminates different face classes in order
to find the subspace as shown in Figure 3.9. The within class scatter
minimizes the distance between the face images of same class i.e.,
𝐶 𝑁𝑗
𝑗 𝑗 𝑇
𝑆𝑤 = Γ𝑖 − 𝜇𝑗 Γ𝑖 − 𝜇𝑗 − − − − − − − (3.14)
𝑗 =1 𝑖=1
𝐶
𝑇
𝑆𝑏 = 𝜇𝑗 − 𝜇 𝜇𝑗 − 𝜇 − − − − − − (3.15)
𝑗 =1
90
𝑊 𝑇 𝑆𝑏 𝑊
𝑤 = 𝑎𝑟𝑔𝑚𝑎𝑥 = − − − − − − (3.16)
𝑊 𝑇 𝑆𝑤 𝑊
Testing Set
Classifier
(Euclidean Distance)
Decision Making
the face image takes place closely within classes, whereas the
classes take place from each other. During the projection of face
the subspace characterizing the face space. The face which has the
minimum distance with the test face image is labeled with the
Figure 3.11 shows the testing phase of the LDA approach. Figure 3.12
networks are used as the most powerful tools to perform complex and
pattern recognition.
93
The main characteristics of neural networks are that they have the
used for data clustering and feature mapping. The learning process
systems, such as the brain, process information. The key element of this
connected to form a layer for nonlinear separable input data. Each layer
in network gets input from the previous layer and feed its output to the
next layer but the connections to the neurons in the same or previous
layers are not permitted. Figure 3.13 represents the architecture of the
artificial neural network classifier. These methods are called PCA-NN and
the target vector is calculated, and the weights of the network are
the output vector and the target vector. Input for hidden layer is
calculated as
𝑛
1
𝑚 = − − − −(3.18)
1 + exp
(−𝑛𝑒𝑡𝑚 )
96
1
𝑜𝑘 = − − − −(3.20)
1 + exp
(−𝑛𝑒𝑡𝑘 )
𝑘
1
𝐸= (𝑜𝑖 − 𝑡𝑖 )2 − − − − − (3.21)
2
𝑖=1
compared to the predefined limit else the weights are updated till
range [0.01, 1.0], hj is the output jth neuron in the hidden layer
and δi is given by
𝛿𝑖 = 𝑡𝑖 − 𝑜𝑖 𝑜𝑖 l − 𝑜𝑖 − − − − − (3.23)
value
training.
Initial Weights
This initial weight will influence whether the net reaches a global
or local minima of the error and if so how rapidly it converges. To get the
best result the initial weights are set to random numbers between -1
and 1.
98
Training a Net
the error begins to increase, the net is starting to memorize the training
If the activation function can vary with the function, then it can be
units. If more number of hidden layers are present, then the calculation
for the ‟s are repeated for each additional hidden layer present,
summing all the ‟s for units present in the previous layer that is fed into
Learning rate
current gradient and the previous gradient. A small learning rate is used
identify the given image as a face image or non face image and then
recognizes the given input image .Thus the back propagation neural
Though the Radial Basis Function (RBF) network has its own
global network. The global and local networks are differentiated based on
Figure 3.14 as
The input layer of this network consists of n units, and accepts the
input vector xn goes as input to the l hidden function, i.e., output of the
100
hidden function, and is later multiplied by the weighting factor wij , i.e.,
𝑁𝑘
1
𝜇𝑘 = 𝑥𝑘𝑖 , 𝑘 = 1, 2, 3, … . , 𝑚 − − − −(3.27)
𝑁𝑘
𝑖=1
𝑓
pattern 𝑥𝑘 belong to itself:
𝑓 𝑓
𝑑𝑘 = 𝑥𝑘 − 𝜇𝑘 , 𝑘 = 1,2, … . . , 𝑚 − − − −(3.28)
found by the activation function of the hidden layer units, since RBF
chosen as the activation function of the RBF units (hidden layer unit) as
dimensional input feature vector, μi is the center of the RBF unit, which
101
given as:
𝑦𝑗 𝑥 = 𝑖 𝑥 𝑤𝑖𝑗 − − − − − (3.30)
𝑖=1
where wij is the connection weight of the i -th RBF unit to the j -th output
node.
the input data elements xi should be equal to the number of the feature
the number of the face classes in the face image database considered for
training.
function network is done in two stages. In the first stage of learning, the
number of radial basis functions and their center values are calculated.
follows [100,101]:
1. Initial weights wij are assigned with small random values and the
radial basis function hj are defined, using the mean μj, and standard
deviation σj.
to train the weights. The biases may/may not be present in the output
layer.
the input vectors exists and cluster the similar input vector. The hidden
node will be activated when the input vector (x) lies near the center of a
receptive field (μ). No activation of both the hidden nodes takes place in
case the input vector lies between two receptive fields where as the input
vector lie far from all the receptive fields. Hence, here the RBF output is
equal to that of the output layer bias values [102] which is a local
global mapping i.e., all inputs cause on output whereas an RBF performs
103
a local mapping which causes the activation of all inputs near the
The experiments are conducted on the FERET & ORL face image
recognition techniques.
photo sessions, and in this study upto 500 frontal face images of 100
3.5) Conclusions
PCA and LDA algorithms are explained. Under the databases used in
proposed work, ORL and FERET databases have been explained. In the