Professional Documents
Culture Documents
Consolidated New
Consolidated New
ROBUST FACE
RECOGNITION
ii
ABSTRACT
Facebook to high security scenarios such as border control. Researchers find great
interest in cracking the challenges of face recognition such as pose, expression and
accurate as the technology evolves. Due to the non-contact and remote sensing
capabilities of the face biometric, it is regarded as the only viable biometric trait to
systems is significant for security applications, especially at night time when the
subject is far away from the camera. Cross Spectral face recognition is generally
required for nighttime surveillance as the registered face images in the gallery are
usually visual face images (VIS) of good quality but the probe images are in Near
iii
Infra-Red (NIR) modality. Images acquired in different spectral bands have a lot of
the differences in standoff distances termed as Cross Distance face matching creates
an unequal image quality scenario due to illumination variations and this also have
to be accounted for. Considering the increased threats at night time, the proposed
feature-based representation for both NIR images as well as VIS images. This paves
way for bridging the modality gap and also compensates the various other
The long distance and night time face recognition system comprises
of major stages like face detection, feature extraction and feature matching. The
probe image is one of the long distance night time images from the database and
the gallery data consists of visible images taken at short indoor distances. The NIR
face images are detected using the popular Viola Jones face detection algorithm
and pre-processed to improve their quality so as to match the gallery data. Then
highly discriminative features are extracted from both the gallery data as well as
the probe data. For a long distance and night time face recognition system, the
stage that helps in bridging the cross-spectral and cross-distance gap has been
The proposed method for long distance and night time face
recognition was evaluated using the LDHF database. The database consists of
several classes (100 classes) each with single sample and therefore classifiers were
not used for matching. The results show that the proposed method performs better
TABLE OF CONTENTS
LIST OF TABLES
LIST OF FIGURES
4.7 (a) DoG filtered VIS image (b) DoG filtered NIR 77
image
5.1 Stages of feature extraction 81
5.2 Wavelet decomposition upto 3 levels 84
5.3 Three level wavelet decomposition of sample face 92
images (a) 1 m and (b) 150 m
xiv
Symbols
σ - standard deviation
θ - threshold
Φ(x) - mother wavelet
i(x,y) - input image
ii(x,y) - integral image
Id(x) - Photometric Normalized Image
Gσ - Gaussian Kernel
h - decay parameter
k(x,y) - smoothing kernel
p - polarity
wi - weights
Abbreviations
2D - Two Dimensional
3D - Three Dimensional
AISS - Anisotropic Smoothing
AISSS - Anisotropic Smoothing Stable
ANLM - Adaptive Non Local Means
ASSR - Adaptive Single Scale Retinex
CCD - Charge Coupled Devices
DARPA - Defense Advanced Research Projects Agency
DCT - Discrete Cosine Transform
DoG - Difference of Gaussian
xv
SC - Structural Content
SG - Steerable Gaussians
SSR - Single scale Retinex
SSSQI - Single Scale Self Quotient Image
SWIR - Short Wave Infra-Red
TT - Tan and Triggs
VIS - Visual
WF - WeberFaces
WN - Wavelet Normalization
1
CHAPTER 1
INTRODUCTION
1
2
are interested in it, but neuroscientists and psychologists also. There is also a
general opinion that advances in computer vision research will deliver useful
perceptions to neuroscientists and psychologists into how human brain works, and
vice versa (Montag et al 2016).
2
3
Input Image
Pre-Processing
Feature
Extraction
Recognized
Classification
output
Database
3
4
4
5
5
6
identity (Y.Moses et al 1994). As it is evident in Figure 1.2, the same person, with
the similar facial expression is likely to appear strikingly different when the
direction of the light varies. These variations are exaggerated even more greater
by added factors including facial expression, hair styles, perspiration, cosmetics,
and even deviations due to aging.
6
7
The issues under pose variation has been divided in to three categories:
Modest case with small rotation angle; most regularly addressed case, when there
is a set of training image pairs, frontal and rotated images; most problematic case,
when the required training image pairs are not accessible (Z.Lei et al 2009).
7
8
8
9
9
10
10
11
environment, and to the variations of the heat patterns of the face that are affected
due to various factors, e.g. stress, changes in temperature of the surrounding
environment, physical activity etc. The prominence of MWIR in face recognition
technology has been lately proposed by T.Bourlai et al. (2012).
(a) (b)
Figure 1.7 Example showing (a) VIS image (b) NIR image
11
12
groups have studied face recognition using NIR illumination and have also offered
various NIR face detection and recognition methods.
(a)
(b)
12
13
2010). In such cases, there is a need for matching images of two different spectra
in contrast to intra-spectral matching where two or more images of the same
spectra are compared.
13
14
1. To apply and test the performance of the benchmarked Viola Jones Face
detection Algorithm on Long Distance and Night Time Face Images.
2. To develop a novel approach to carry out pre-processing such that the
differences in modality, distance and illumination are compensated.
3. To develop a novel approach for feature extraction which yields highly
discriminative features and thus enhances the recognition rate for long
distance and night time face images.
14
15
15
16
CHAPTER 2
LITERATURE SURVEY
2.1 INTRODUCTION
16
17
general statement of the issue can be resolved, formulated and observed first. Any
face recognition system comprises three main parts of that are pre-processing,
feature selection and classification (Seo et al. 2011).
17
18
18
19
19
20
are discussed in this chapter. This chapter covers several articles from different
reputed journals and conferences for literature review and has discussed the
various existing methods of face recognition. Advantages and disadvantages of
these various works are also discussed briefly in order to identify the suitability of
these techniques for in achieving a better recognition rate.
20
21
21
22
He presented that the system works in two main stages: first, face detection based
on AdaBoost with Haar cascade is used and in second, face recognition based on
fast and simple PCA and LDA is executed. The system is tested on 500 images
and after face detection the images are stored in JPEG format in 100 x 100 matrix
size for face recognition. This approach achieves accurate results for general
purpose online attendance system. Wood et al (2006) proposed a robust face
detection based on lighting-variable adaboosting, which is adaptive for fluctuating
illuminations and also depends on multiple features such as global and local
intensity variations. The system is tested on standard datasets Caltech-101, that
successfully achieve overall 95% accuracy in different lighting levels.
Robert Viola and Michael Jones (Paula Viola et al, 2001) have
defined the machine mastering approach concerning the visual object detection
that has the ability of processing photos enormously in a greater speed and
accomplishing high recognition rates. The work done by them has been
distinguished by simply three important contributions. Their first contribution was
the introduction of any new imprint representation referred to as the Integral
Image permitting the features exploited by the detector to get calculated rapidly.
The next contibution was the learning algorithm which is based on a concept
known as AdaBoost. This strategy selects a small amount of precarious visible
features at a larger extent and reassures tremendously useful classifiers. The third
was gradually utilizing more complex classifiers inside a “cascade”, thus allowing
the background regions of the image to get swiftly thrown away while bombing
out more calculation on promising object-like areas. Even though the Viola Jones
algorithm was proposed almost a decade ago, the algorithm still stands the best
accepted both commercially and scholastically.
22
23
23
24
24
25
25
26
features. This may not be needed in all cases, as many applications require only
specific or relevant data from the image.
26
27
27
28
SWIR (short wave IR) spectrum. NIR has the advantage that it can capture face
images at large standoff during night time which contains sufficient information
for face recognition. However, the literatures show that there is a great need for
improvement in the area of long distance and night time face recognition, which
is the main goal of this project.
The challenges with regards to long distance and night time facial
recognition technologies are: expensive long range IR cameras, image quality
(e.g., image resolution, compression, blur, and noise), time span (facial aging),
occlusion, and demographic information (e.g., gender, race/ethnicity, or age),
variations in pose, expression, including illumination that depends on the
operational environment, below par performance of FR algorithms and software
packages. The existing face recognition systems have addressed some of the
challenges through new image acquisition set ups, cross spectral and cross
distance type face databases and face recognition algorithms. However, the
problems related to image quality, image variations and recognition accuracy still
exist, which are the main objectives of this project work to be addressed.
2.6 CONCLUSION
This review benevolence the various literatures pertaining to
different stages of long distance and night time face recognition. A number of
authors have experimented on face recognition systems under controlled
28
Table 2.1 Summarization of Literatures related to Long Distance and Night Time Face Recognition
Author Image Acquisition Database Pre- Feature Matching Merits & Demerits
& Setup processing Extraction
Publica
tion
details
Kang et Canon 600D DSLR LDHF Geometric and combination of RS-LDA 150m, still images, cross-
al, 2014. with telephoto lens Photometric three filters (DoG, spectral and cross distance
100 subjects,
and RayMax300 Normalization, CSDN and matching, rank 1 accuracy at 60,
1m, 60m,
illuminator Image Gaussian) + 100m & 150m is 82%, 69% &
100m &
Restoration descriptors (SIFT, 28%
150m.
MLBP)
Bourlai Canon EOS 5D Mark VIS, SWIR, Geometric and LBP, LTP PCA, PCA+LDA, BIC, intra-spectral, cross spectral and
et al, II, Canon PowerShot MWIR and Photometric ML, MAP cross distance, identification
2012. SX110, Goodrich NIR Normalization rates at rank 1 - .998, .996, .968,
SU640 and the (CLAHE, .952 (IS), .988, ..985, .939, .922
30, 60, 90,
XenICS Xeva-818, SSRlog, (CS & CD)
FLIR Systems, NIR 120m; SSRatan)
Camera with PTZ
platform
Bourlai NIR Camera with PTZ unknown geometric LBP, LTP PCA, PCA+LDA, BIC, Baseline - G8 identification rate
et al, platform normalization, ML, MAP 100%, intraspectral & cross
2012. masking, ditance – LDA (CMC) with
histogram 80% training set.
equalization,
pixel
normalization
30
30
31
31
environments but with uncontrolled conditions not much work has been focused.
In this report, a list of various image acquisition setups, databases, pre-processing
techniques, feature descriptors and classifiers are provided. Among them most of
the work has been done using still images and only frontal view. Similarly, less
work has been focused on long distance face recognition.
Briefing generally the methods discussed in this chapter, each has its
own pros and cons and each one is effective in its own field of usage. Though
some schemes become intricate and their computational cost in terms of time and
space might go up high, the trade-off is made on the functionality. Fundamentally
main emphasis of conducting this survey is to group together all the work related
to Long distance and night time face images into one document.
32
CHAPTER 3
FACE DETECTION
3.1 INTRODUCTION
33
Knowledge-based methods employ human-coded rules to represent facial
features like two symmetric eyes, a nose in the middle and a mouth
underneath the nose.
Feature invariant methods attempt to discover facial features that are
invariant to pose, lighting condition or rotation. Skin colors, edges and
shapes fall under this category.
Template matching methods compute the correlation between a test image
and pre-selected facial templates.
Appearance-based methods embrace machine learning methods to extract
discriminative features from a pre-labeled training set. The Eigenface
method is the most basic method under this category. Face detection
algorithms that were proposed in the recent years such as support vector
machines, neural networks (Rowley-Balaujam, 1998), statistical classifiers
(Schneiderman-Kanade, 2000) and AdaBoost-based face detection also fit
into this category.
34
between 780- 1100 m. CASIA NIR-VIS 2.0 face database (Mohsen et al, 2017)
contains 725 subjects which has a resolution of 640˟480 pixels in both NIR and
VIS images. The number of images available in this dataset is 17500.
The very first step of the Viola-Jones face detection algorithm (Paula
Viola et al, 2004) is to convert the input image into an integral image which can
also be called as a summed area table, designed quickly and competently
computing the sum of values in a rectangle subset of a pixel grid. The integral
image at location (x, y) contains the sum of the pixels above and to the left of (x,
y), as given in Equation 3.1
where i(x, y) is the pixel value of the original image and ii(x, y) is the
corresponding image integral value.
35
1 1 1 1 2 3
1 1 1 2 4 6
1 1 1 3 6 9
With the help of the integral image to calculate the sum of any rectangular area is
extremely efficient, as shown in Figure 3.1. The sum of the pixels in a rectangle
ABCD can be estimated with only four values from integral image as shown in
Equation 3.2
36
3.2.2 AdaBoost Learning
1 𝑝𝑓(𝑥) < 𝜃
ℎ(𝑥, 𝑓, 𝑝, 𝜃) = { (3.3)
0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
where f denotes the feature value, θ is the threshold and p is the polarity indicating
the direction of the inequality. The different steps in the implementation of the
AdaBoost learning procedure is as follows:
1. Given a training sample images (x1, y1), ...,(xn, yn), where yi = 0, 1 for
negative and positive examples respectively.
1 1
2. Initialize the classifier count t = 0 and the sample weights 𝑤𝑖 = ,
2𝑚 2𝑙
for yi = 0, 1 respectively, where m and l are the number of negative and positive
samples.
3. While the number of negative samples rejected is less 50%:
(a) Increment t = t + 1.
𝑤𝑖
(b) Normalize the weights 𝑤𝑖 = ∑ .
𝑗 𝑤𝑗
(c) Select the best weak classifier with respect to the weighted
error
𝑚𝑖𝑛
∈𝑡 = 𝑓,𝑝,𝜃 ∑𝑖 𝑤𝑖 |ℎ(𝑥𝑖 , 𝑓, 𝑝, 𝜃) − 𝑦𝑖 | ( 3.4)
37
(d) Define ht(x) = h(x, ft, pt, θt) where ft, pt and θt are the
minimizers of ∈𝑡 .
1−𝑒𝑖 ∈𝑡
(e) Update the weights as 𝑤𝑖 = 𝑤𝑖 𝛽𝑡 where 𝛽𝑡 = and
1−∈𝑡
38
given sub-window is definitely not a face or may-be a face. When a sub-window
is classified to be a non-face by a given stage it is immediately discarded.
Contrariwise, a sub-window classified as a may-be-face is passed on to the next
stage in the cascade. It follows that more the stages a given sub-window passes,
the higher the chance the sub-window contains a face.
This research work considers only the NIR images taken at night and
at various standoff distances as the probe image while the 1 m VIS images are
39
considered as the gallery data. At each distance, 100 images are taken into account
for experimenting. Even though the Viola Jones face detection algorithm is a
proven algorithm for face detection, its performance for long distance images have
not been discussed in the literatures. In this work the performance of the Viola
Jones face detector on the LDHF database was analyzed.
Table 3.1 The detection rates at different distances
S.No Distance Perfect Partial No Wrong Total
& Type Detection Detection Detection Detection Detection
% % % % %
1. 1m VIS 89 11 0 0 94.5
2. 1m NIR 90 10 0 0 95
3. 60m VIS 22 78 0 0 61
4. 60m NIR 51 49 0 0 75.5
5. 100m VIS 8 92 0 0 54
6. 100m NIR 38 60 1 1 68
7. 150m VIS 19 81 0 0 59.5
8. 150m NIR 44 24 25 7 56
The detection rates at various distances are shown in Table 3.1 which
indicates that the performance of the bench mark detector decreases as the standoff
distance increases. The Figure 3.3 shows the perfect detection of NIR images at
60 meters. Even though the detection rate at 60m, the face detection using Viola
Jones algorithm show tremendous accuracy than at 100 m and 150 m still it has
false acceptances as shown in Figure 3.4. In Figure 3.5 shows the accurate
detection achieved at 100m is shown. In Figure 3.6 shows an example for false
acceptance is shown, i.e. two non faces are detected along with the correctly
detected face image. Figure 3.7 shows the false rejection in 100m. Figure 3.8
shows the accurate detection of NIR images at 150m. Figures 3.9 and 3.10 shows
the false acceptance shows the false rejection at 150m respectively.
40
Table 3.1 shows the level of accuracy and efficiency of the Viola Jones
algorithm on NIR images. It is to be noted that the perfect detection rate at 100m
is found to be less than 150m. This is because, the discrimination gets affected by
the background features merging with the facial features. For better comparison
the graphical representation of Table 3.1 is shown in Figure 3.2.
100
80
Accurate
60 Detection.
40 False
Acceptance.
20 False Rejection
0
1 60 100 150
meter meter meter meter
41
Figure 3.4 Partial detection at 60m
42
Figure 3.7 No detection at 100m
43
Figure 3.10 No detection at 150m.
3.4 CONCLUSION
44
orientation and pose. In particular, the seminal work by Viola and Jones has made
face detection practically feasible in some applications such as digital cameras and
photo organization software.
In this work the effect and the performance of Viola Jones face
detection algorithm on NIR long distance images are described. It is exhilarating
to see Viola-Jones face detection algorithm is increasingly being used in face
recognition as well as other object recognition problems, but when it is used
towards NIR images the difficulty for the face detection is on the rise. Only the
frontal upright face images are detected by Viola Jones algorithm but it shows the
ineffectiveness towards the NIR images due to dark pixels as shown by the
experimental results. Hence, the features extracted from the NIR images need to
be improve.
45
CHAPTER 4
PRE-PROCESSING
4.1 INTRODUCTION
46
between spectral modes and differences in distances in the probe and the gallery
image is accomplished using the photometric normalization and DoG filtering
techniques respectively, which is elucidated in the following sections.
NIR face images appear different from VIS face images. Further,
NIR images become blurred as the stand-off distance increases. The gallery
images are acquired indoor while the probe images are captured outdoor. To
address the problems mentioned in the previous section, preprocessing is done as
shown in figure. applied to images to reduce the appearance difference between
NIR and VIS images, and enhance the image quality.
Difference of
Median Wavelet
Gaussian
Filtering Normalization
Filtering
Photometric Normalization
47
4.2.1 Median Filtering
The median filter takes into account each pixel in the image in turn and
looks at its nearby neighbors to decide whether or not it is representative of its
surroundings. In spite of simply replacing the pixel value with the mean of
neighboring pixel values, it replaces it with the median of those values. The
median is estimated by first sorting all the pixel values from the adjoining
neighborhood ѡ into numerical order and then replacing the pixel being
considered with the middle pixel value.
48
where ѡ represents a neighborhood defined by the user, centered around
location [m,n] in the image. If the neighborhood under consideration contains an
even number of pixels, the average of the two middle pixel values is used. Figure
4.2 illustrates an example calculation. As can be seen, the central pixel value of
150 is rather unrepresentative of the surrounding pixels and is replaced with the
median value: 124. A 3×3 square neighborhoods is used here. It is to be noted that
larger neighborhoods will result in more severe smoothing and thus affects edges.
The median is a more vigorous average than the mean and so a single
very unrepresentative pixel in a neighborhood will not affect the median value
significantly. As the median value must essentially be the value of one of the
pixels in the neighborhood, the median filter does not generate new unrealistic
pixel values when the filter straddles an edge. For this reason, the median filter is
much better at conserving sharp edges than the mean filter.
49
Inverse
Wavelet Approximation Histogram
Wavelet
Transform Coefficients Equalization
Transform
Detail Scalar
Coefficeients Multiplication
50
domain. The transfer functions of DoG filter are two Gaussian differences with
different widths. In this approach the DoG filter acts as bridge between the VIS
and NIR images. It suppresses the variations and highlights the similarities
between the two different modes specifically the VIS and the NIR images. Also
it acts as a feature enhancement procedure as it can be utilized to increase the
visibility of edges.
The DoG (V.Struc, 2009) is in reality a band-pass filter that can remove
high frequency components that represents noise, and also some low frequency
components that represents the homogeneous areas in the image. The frequency
components of the passing band are assumed to be associated to the edges in the
image. The DoG normalized technique can be constructed using the equations 4.2.
51
real world applications. Predominantly when we see illumination, it is the major
problem in face recognition. Different researchers express the problem due to
illumination as variation between images of different faces can be smaller than the
variations between images of the same face under different illumination. It can
even be shown that illumination causes larger variation in face images than pose
(V. Štruc et al,2009).
2
𝐺 ||𝐼 (𝛺𝑥)−𝐼𝑛 (𝛺𝑧)|| 2
1 − 𝜎 𝑛
𝑤 (𝑧, 𝑥 ) = 𝑒 2
ℎ (4.4)
𝑍(𝑧)
52
where, Gσ represents a Gaussian kernel with the standard deviation σ,
Ωx and Ωz denotes the local neighborhoods of the pixels at the locations x and z,
respectively, h stands for the parameter that controls the decay of the exponential
function, and Z(z) represents a normalizing factor.
1
𝐼𝑖𝑐 (𝑥 ) = log[ ] (4.6)
𝐼𝑐 (𝑥)
53
(𝐼𝑖𝑐 (𝑥)−𝐼𝑖𝑐𝑚𝑖𝑛 )
ℎ(𝑥 ) = [ ] − (ℎ𝑚𝑎𝑥 + ℎ𝑚𝑖𝑛 ) (4.7)
𝐼𝑖𝑐𝑚𝑎𝑥 −𝐼𝑖𝑐𝑚𝑖𝑛
where Iicmax and Iicmin denote the maximum and minimum value of the
inverted contrast image Iic(x), respectively, and hmax and hmin stand for the target
maximum and minimum values of the decay parameter h.
54
𝑅′ (𝑥, 𝑦) = ∑𝑀
𝑖=1 𝑤𝑖 (𝑙𝑜𝑔𝑙 (𝑥, 𝑦) − log[𝑙(𝑥, 𝑦) ∗ 𝐾𝑖 (𝑥, 𝑦)] (4.9)
where Ki(x,y) denotes a Gaussian kernel at the i-th scale, and wi stands
for the weight associated with the i-th Gaussian kernel Ki(x,y).
Adaptive single scale retinex techniue was proposed (X.Xie et al, 2006)
to tackle the halo effects often encountered with the original single scale retinex
technique by incorporating an adaptive smoothing procedure with a discontinuity
preserving filter into the single scale retinex algorithm with the goal of robustly
estimating the images’ luminance.
(4.10)
and
𝐿(𝑡+1)(𝑥, 𝑦) = max{𝐿(𝑡+1)(𝑥, 𝑦), 𝐿(𝑡) (𝑥, 𝑦)} (4.11)
55
4.3.6 Isotropic Smoothing (ISS)
2
𝐽(𝐿 (𝑥, 𝑦)) = ∬(𝐿(𝑥, 𝑦) − 𝐼(𝑥, 𝑦)) 𝑑𝑥𝑑𝑦 + 𝜆 ∬(𝐿2𝑥 (𝑥, 𝑦) + 𝐿2𝑦 (𝑥, 𝑦)𝑑𝑥𝑑𝑦
(4.12)
where the first term forces the luminance L(x,y) to be close to the
original input image I(x,y), and the second term imposes a smoothing constraint
on L(x,y), and the parameter λ controls the relative importance of the smoothing
constraint
56
4.3.8 Anisotropic Smoothing Stable (AISSS)
where M and N denote the dimensions of the input image I(x,y) and μ
is chosen near the mean value of I(x,y).
57
4.3.10 Difference of Gaussian (DoG)
58
I’(x,y) in accordance to the reflectance of the image as there are no direct
subtraction of the luminance performed. Nevertheless, the result estimates the
reflectance because the effect of the luminance was condensed and that of the
reflectance was accentuated through the filtering operation in the frequency
domain.
𝐼(𝑥,𝑦)
𝑄(𝑥, 𝑦) = ≅ 𝑅(𝑥, 𝑦) (4.18)
𝐼(𝑥,𝑦)𝐹(𝑥,𝑦)
59
where, F(x,y) is a low pass filter. A low pass filter is used here as
illumination can be considered as a low frequency component. It should be noted
that the properties of Q(x,y) are dependent on the kernel size of F(x,y). If
too small, then Q(x,y) will approximate one, and the Albedo information will be
lost; if too large, there will appear halo effects near edges. SQI is usually defined
as the intrinsic property of face images of a person. This technique is mainly used
in cases where it requires removing of shadows form the images.
𝐼(𝑥,𝑦)
𝑄𝑘 (𝑥, 𝑦) = 1 ; 𝑘 = 1,2, … … , 𝑛 (4.20)
(𝑁𝑊𝑘 𝐺𝑘 )∗𝐼(𝑥,𝑦)
60
experimentally studied that the ratio of the increment threshold to the background
intensity is a constant. This can be mathematically represented as in Equation 4.21
∆𝐼
=𝑘 (4.21)
𝐼
(4.22)
61
component while large-scale Weberfaces include global information and less
detail component. Multi-scale Weberfaces which is calculated as Equation 4.23
where F1 and F2 are the filter adaptation factors at pixel p, Iin is the
intensity of the image, * is the convolution factor, G1 and G2 are 2D Gaussian
filters with standard deviation σ1 and σ2 respectively, Imax is the maximal value of
image intensity and IIa1 is the light adapted image.
62
4.3.18 Steerable Gaussians (SG)
63
level transformation that replaces each pixel with intensity I in the image with Iγ
for 0 < γ < 1 or if γ+0. A gamma correction increases the dynamic range of an
image. This cannot remove all effects of intensity gradients such as shading
effects. Hence a DoG filter which acts as a band pass filter to remove all shading
effects and other noises is used. The final step of the preprocessing chain is to
perform contrast equalization which globally rescales the image intensities to
standardize a robust measure to overall contrast or intensity variation.
𝐼
𝐼= 1 (4.26)
(𝑚𝑒𝑎𝑛(𝐼 𝛼 ))𝛼
𝐼
𝐼= 1 (4.27)
(𝑚𝑒𝑎𝑛(min(𝜏,𝐼 𝛼 )))𝛼
𝐼 = 𝜏 ∗ tanh(1⁄𝜏) (4.28)
The output of above steps is an image with pixel intensity in the ranges
(-τ, τ). In our work the value of τ is considered as ten.
64
4.4 PERFORMANCE MEASURES FOR EVALUATING PHOTOMETRIC
NORMALIZATION TECHNIQUES
4.4.1 Entropy
Entropy (S.K. Pal et al, 1983) is used to measure the content of an
image with higher value indicating an image with richer details. Higher the value
of entropy, higher will be the richness of details in the image.
65
4.4.4 Histogram Flatness Measure
2. ContrastRMS 𝑀 𝑁
3. Histogram Spread 𝐻𝑆
(3𝑟𝑑 𝑞𝑢𝑎𝑟𝑡𝑖𝑙𝑒 − 1𝑠𝑡 𝑞𝑢𝑎𝑟𝑡𝑖𝑙𝑒)𝑜𝑓 ℎ𝑖𝑠𝑡𝑜𝑔𝑟𝑎𝑚
=
(𝑚𝑎𝑥 − 𝑚𝑖𝑛)𝑜𝑓 𝑡ℎ𝑒 𝑝𝑖𝑥𝑒𝑙 𝑣𝑎𝑙𝑢𝑒 𝑟𝑎𝑛𝑔𝑒
4. Histogram 𝑔𝑒𝑜𝑚𝑒𝑡𝑟𝑖𝑐 𝑚𝑒𝑎𝑛 𝑜𝑓 𝑃
𝐻𝐹𝑀 =
Flatness measure 𝑎𝑟𝑖𝑡ℎ𝑚𝑒𝑡𝑖𝑐 𝑚𝑒𝑎𝑛 𝑜𝑓 𝑃
P denotes the histogram count of the image I(x,y) of size MxN and µ represents
the mean of the intensity values of the image.
66
4.5 EVALUATION SCHEME FOR MODALITY GAP REDUCTION
USING DIFFERENCE OF GAUSSIAN FILTERING
1
RMSE =
𝑀×𝑁
√∑𝑀 𝑁
𝑖=1 ∑𝑗=1(𝑋𝑖,𝑗 − 𝑌𝑖,𝑗 )
2
(4.29)
67
quality index of the images defined as the ratio between the maximum power of a
signal and the noise power. PSNR is usually expressed in terms of the logarithmic
decibel scale due to many signals having very wide dynamic range. The higher
the PSNR value, more the similarity with the original image, from a perceptual
point of view. PSNR is given by Equation 4.30
1
MAE = ∑M N
i=1 ∑j=1|x(i, j) − y(i, j)| (4.31)
MN
∑M N
i=1 ∑j=1(y(i, j))
2
SC = ──────────
∑M N
i=1 ∑j=1 x(i, j))
2
4.32
68
4.5.5 Normalised Cross Correlation (NK)
The closeness between two digital images can also be quantified in terms
of correlation function. The Normalised Cross Correlation is also complementary
to the difference based measures. All the correlation based measures tend to 1, as
the difference between two images tend to zero.
M N
NK = ──────────────
∑M N
i=1 ∑j=1(x(i, j))
2
(4.33)
NAE = ───────────────
M N
∑ ∑(x(𝑖, j))
I=1 j=1
(4.34)
69
4.6 EXPERIMENTAL RESULTS AND DISCUSSIONS
Tables 4.2, 4.3 and 4.4 show the values of entropy, root mean square
contrast, histogram spread and histogram flatness measure of sample images.
Higher the value of entropy, higher is the quality of the image comparatively. Also
when the root mean square contrast is high the contrast of the resulting image is
high. The values of histogram spread and histogram flatness measure lies in the
range between 0 and 1. The nl means and adaptive nl means normalization give a
moderate boost to the parameters. The single scale retinex and multiscale
techniques almost doubles the contrast level but the adaptive single scale retinex
does not show up a considerable improvement in the result. The original isotropic
smoothing techniques provides better normalization compared to the anisotropic
and the anisotropic stable smoothing algorithm.
70
a good amount of progress whereas multiscale weberfaces, Tan & Triggs,
steerable Gaussian and retina model display adequate improvement. The wavelet
normalization provides the highest values compared to all other techniques. The
analysis shows that the wavelet based normalization provides good results for all
distances. These pre-processed results are sure to provide better recognition rates
for long distance face recognition system.
Figure 4.4 Photometric Normalized outputs for a sample image at 150m (a)
anisotropic smoothing, (b) anisotropic smoothing stable, (c) adaptive nl
means, (d) adaptive single scale retinex, (e) DCT normalization, (f) DoG, (g)
homomorphic filtering, (h) isotropic smoothing, (i) lssf normalization, (j)
multiscale retinex, (k) multiscale self-quotient image, (l) multi scale
weberfaces, (m) nl means, (n) retina modeling, (o) steerable gaussians, (p)
single scale retinex, (q) single scale self-quotient image, (r) tan & trigs, (s)
weberfaces, (t) wavelet normalization .
71
Figure 4.5 Photometric Normalized outputs for a sample image at 100m (a)
anisotropic smoothing, (b) anisotropic smoothing stable, (c) adaptive nl
means, (d) adaptive single scale retinex, (e) DCT normalization, (f) DoG, (g)
homomorphic filtering, (h) isotropic smoothing, (i) lssf normalization, (j)
multiscale retinex, (k) multiscale self-quotient image, (l) multi scale
weberfaces, (m) nl means, (n) retina modeling, (o) steerable gaussians, (p)
single scale retinex, (q) single scale self-quotient image, (r) tan & trigs, (s)
weberfaces, (t) wavelet normalization .
72
Figure 4.6 Photometric Normalized outputs for a sample image at 60m (a)
anisotropic smoothing, (b) anisotropic smoothing stable, (c) adaptive nl
means, (d) adaptive single scale retinex, (e) DCT normalization, (f) DoG, (g)
homomorphic filtering, (h) isotropic smoothing, (i) lssf normalization, (j)
multiscale retinex, (k) multiscale self-quotient image, (l) multi scale
weberfaces, (m) nl means, (n) retina modeling, (o) steerable gaussians, (p)
single scale retinex, (q) single scale self-quotient image, (r) tan & trigs, (s)
weberfaces, (t) wavelet normalization
73
Table 4.2 Performance measures attained for sample images taken at
distance 150m
Image 0003_150_n Image 0007_150_n
S.No Technique
Entropy C_RMS HS HFM Entropy C_RMS HS HFM
Without
pre-
1 processing 5.89 17.52 0.15 0.06 5.80 16.12 0.29 0.38
2 NLM 6.99 34.47 0.16 0.32 6.68 32.12 0.13 0.24
3 ANLM 7.11 36.89 0.18 0.37 6.84 32.51 0.15 0.27
4 SSR 7.10 36.97 0.18 0.34 6.89 37.53 0.15 0.33
5 ASSR 5.80 33.59 0.12 0.18 5.01 25.11 0.06 0.09
6 MSR 7.03 35.27 0.18 0.31 6.80 35.20 0.14 0.29
7 ISS 7.18 37.97 0.16 0.43 6.94 35.24 0.13 0.41
8 AISS 7.03 35.92 0.14 0.41 6.78 32.58 0.11 0.36
9 AISSS 6.51 24.05 0.10 0.17 6.56 27.39 0.09 0.24
10 DCT 6.98 32.36 0.15 0.32 6.58 26.47 0.11 0.22
11 DoG 6.13 20.30 0.08 0.09 6.93 32.59 0.14 0.33
12 HMF 5.93 18.74 0.07 0.12 5.80 18.11 0.06 0.11
13 Lssf 7.10 35.02 0.18 0.29 7.00 35.66 0.15 0.31
14 SSSQI 7.42 36.97 0.27 0.67 7.13 54.69 0.20 0.54
15 MSSQI 7.57 58.75 0.32 0.71 7.29 56.43 0.24 0.61
16 MSWF 6.95 32.48 0.15 0.25 6.88 34.37 0.13 0.33
17 RM 6.56 25.05 0.11 0.17 6.69 29.31 0.11 0.26
18 SG 5.86 16.26 0.07 0.07 6.36 23.53 0.10 0.16
19 TT 6.65 25.91 0.12 0.17 6.84 31.87 0.12 0.33
20 WN 7.92 76.03 0.53 0.99 7.91 73.21 0.50 0.98
21 WF 7.22 38.29 0.19 0.37 7.16 39.35 0.15 0.44
74
Table 4.3 Performance measures attained for sample images taken at
distance 100m
75
Table 4.4 Performance measures attained for sample images taken at
distance 60m
76
4.5.2 Results of Modality Gap Reduction using DoG Filtering
Figure 4.7 (a) DoG Filtered VIS images (b) DoG Filtered NIR Images
77
Table 4.5 Similarity measures before and after DoG filtering
4.7 CONCLUSION
78
capable of recognizing faces from long distance at night time. Future work would
be used discriminative feature extractors to represent features and use them for
classification, thus yielding to a complete face recognition system.
79
CHAPTER 5
5.1 INTRODUCTION
80
chapter describes in detail about the novel approach used for feature extraction
and matching.
HOG
Wavelet
Transform
LBP
Feature
Dataset
Pre-processed
Image
Normal Fitting Parameters
81
In order to represent the images in the right way, a combination of
feature extractors is used here that are clearly shown in Figure 5.1. The proposed
methodology uses a combinative approach which uses a global feature, a local
feature, a statistical feature and a combination of global and local feature. This
combinative approach employs the advantage of all types of feature descriptor and
envisages a stronger feature vector that constructs an efficient recognition process.
Initially, the pre-processed face images are decomposed into three levels using
wavelet transform.
82
5.2.1 Image Representation using Wavelet Transform
The dilated and translated Mother function which is also called as the
analyzing wavelet Φ(x), define an orthogonal basis as in Equation 5.1
𝑠
𝛷(𝑠,𝑙) (𝑥 ) = 2−2 𝛷(2−𝑠 𝑥 − 𝑙) (5.1)
Here the variables s and l are integers that helps to scale and dilate the
mother function Φ to generate wavelets, such as a Daubechies wavelet family
(Daubechies et al,1997). The scale index s is used to specify the wavelet’s width,
and l gives the location index. In order to span the image at different resolutions,
we employ the analyzing wavelet in a scaling equation as in Equation 5.2,
𝑊 (𝑥 ) = ∑𝑁−2 𝑘
𝑘=−1(−1) 𝑐𝑘+1 𝛷(2𝑥 + 𝑘) (5.2)
83
where W(x) is the scaling function for the mother function Φ, and ck are the
wavelet coefficients. The wavelet coefficients must gratify linear and quadratic
constraints of the form as in Equation 5.3,
∑𝑁−1
𝑘=0 𝑐𝑘 = 2 ( 5.3)
∑𝑁−1
𝑘=0 𝑐𝑘 𝑐𝑘+2𝑙 = 2𝛿𝑙,0 (5.4)
L3 L3
LL HL L2
L3 L3 HL
LH HH Level 1
HL
L2 LH L2
HH
Level 1 LH Level 1
HH
84
is applied across the columns of the transformed image from the first step. As
represented in Figure 5.2, the effect of these two sets of operations gives a
transformed image with four distinct bands: LL, LH, HL and HH. Here, L denotes
low-pass filtering and H denotes high-pass filtering. The LL band approximately
refers to a down-sampled version of the original image (Cohen A et al, 1992). The
LH band is likely to preserve the localized horizontal features, whereas the HL
band preserves the localized vertical features in the original image.
85
The basic idea for developing the LBP operator was that two-
dimensional surface textures can be described by two complementary measures:
local spatial patterns and gray scale contrast. The LBP feature vector, in its
simplest form, is computed using the following steps: Divide the examined
window into cells (e.g. 16×16 pixels for each cell). For each pixel in a cell,
compare the pixel to each of its 8 neighbors (on its left-top, left-middle, left-
bottom, right-top, etc.). Track the pixels along a circle, i.e. clockwise or counter-
clockwise. Where the center pixel's value is greater than the neighbor's value,
write "0". Otherwise, write "1". This gives an 8-digit binary number (which is
usually converted to decimal for convenience). Compute the histogram, over the
cell, of the frequency of each "number" occurring (i.e., each combination of which
pixels are smaller and which are greater than the center). This histogram can be
seen as a 256 size feature vector. Concatenate histograms of all cells. This gives a
feature vector for the entire window.
86
2. Discretize each cell into angular bins according to the gradient orientation.
3. Each cell's pixel contributes weighted gradient to its corresponding angular
bin.
4. Groups of adjacent cells are considered as spatial regions called blocks.
The grouping of cells into a block is the basis for grouping and
normalization of histograms.
5. Normalized group of histograms represents the block histogram. The set of
these block histograms represents the descriptor.
1 1 𝑥−𝜇 2
𝑓(𝑥 ) = 𝑒𝑥𝑝 [− ( ) ] (5.5)
𝜎 √2𝜋 2 𝜎
−∞<𝑥<∞
where m denotes the distribution mean, and σ denotes the distribution standard
deviation. The distribution mean can be computed from the Equation 5.6,
∑𝑛
𝑖=1 𝑥𝑖
𝜇̂ = 𝑥̅ = (5.6)
𝑛
where n is the sample size. The distribution standard deviation is computed from
the Equation 5.7,
2
𝑛 ∑𝑛 2 𝑛
𝑖=1 𝑥𝑖 −(∑𝑖=1 𝑥𝑖 )
𝜎̂ = √ (5.7)
𝑛(𝑛−1)
87
5.3 MATCHING
88
5.3.2 Cosine Similarity
The Manhattan distance between two images is given by the sum of the
differences of their corresponding components (Bugatti et al, 2008). The distance
d between x= (x1, x2, …….., xn) and y = (y1, y2, ………, yn) with n variables is
given by the Equation 5.12,
𝑑 = ∑𝑛𝑖=1|𝑥𝑖 − 𝑦𝑖 | (5.12)
89
5.3.5 Chebyshev Distance
The chord distance between two vectors x and y is the measure of the
distance (Bugatti et al, 2008) between the projected vectors of x and y onto the
unit sphere, which can be calculated by the Equation 5.15,
𝑥 𝑦
𝐷 (𝑥, 𝑦) = ‖‖𝑥‖ − ‖ (5.15)
‖𝑦‖ 2
90
5.3.8 Pearson’s Correlation Coefficient
1 1
where 𝑢 = ∑𝑛𝑖=1 𝑥𝑖 and 𝑣 = ∑𝑛𝑖=1 𝑦𝑖
𝑛 𝑛
91
Figure 5.3 Three level wavelet decomposition of sample face images (a) 1m
and (b) 150 m
The significant details in a face image are usually contained in its high frequency
components. Hence, the wavelet transform decomposes the signal into low and
92
high frequency i.e. the approximations and details; details alone are used to extract
HoG and LBP features, where approximate coefficients does not find any use.
The greatest challenge faced in this work is that that the gallery
contains only a single image of the target, whereas in most of the face recognition
databases the gallery contains a minimum of three images for every subject.
However, this way of comparing a single image against 100 images with a single
version of every subject using the proposed combination of feature vectors is
93
found to be quite successful with a recognition rate of 32%, 78% and 72% at
150m, 100m and 60m, respectively as shown in Table 5.1. It is to be noted that
though the quality of images at 60m is higher than that 100m and 150m, the
recognition rate of 60m is found to be lesser than that of 100 m. This is because
all images are resized to a common size and due to this the quality of the 60m
images has come down which results in lower recognition rate comparatively.
5.5 CONCLUSION
Feature Extraction being the most significant phase in the entire face
recognition system needs to represent the image in a highly discriminative way
such that matching process works well. This combinative approach in the research
paves way for a better representation of the features and thus enhancing the
recognition rate.
94
CHAPTER 6
CONCLUSION
6.1 INTRODUCTION
This research work has dealt with improving the recognition rate of a
long distance and night time face recognition system. This research has come out
with a different approach for pre-processing the different modality image and a
combinative approach for feature extraction. This chapter summarizes the work
done and discusses the contributions. Finally, the practicable direction for future
scope of work has been discussed.
95
Distance Heterogeneous Face Database. Experimental results of each module are
tested and analyzed to highlight the importance of the proposed methodology.
96
edges increases. The work emphasizes the necessity of a robust system that is
capable of recognizing faces from long distance at night time. Future work would
be used discriminative feature extractors to represent features and use them for
classification, thus yielding to a complete face recognition system.
Feature Extraction being the most significant phase in the entire face
recognition system needs to represent the image in a highly discriminative way
such that matching process works well. This combinative approach in the research
paves way for a better representation of the features and thus enhancing the
recognition rate.
The database consists of several classes (100 classes) each with
single sample and therefore classifiers were not used for matching. The results
show that the proposed method performs better at higher distances like 150 m and
100 m. In future, this approach can be tested with larger databases and with images
that have variations in pose, expression and occlusion
The research work emphasizes the need for an improved Viola Jones
algorithm in future that can perfectly detect different modality images taken from
different distances at night time. A suitable technique to identify common features
from NIR and VIS images that yields to better recognition rates can be introduced.
97
REFERENCES
Abdullah, M., Wazzan, M. and Bo-saed, S., 2012. “Optimizing face recognition
using PCA”, International Journal of Artificial Intelligence & Applications, 3(2),
pp.23-31.
Aggarwal, C.C., Hinneburg, A., Keim, D.A. (2001), “On the surprising behavior
of distance metrics in high dimensional space”, Lecture Notes in Computer
Science 1973, 420–434
Ancong Wu, Wei-Shi Zheng, Hong-Xing Yu, Shaogang Gong, Jianhuang Lai
(2017), “RGB-Infrared Cross-Modality Person Re-identification”, IEEE
International Conference on Computer vision (ICCV), 10.1109/ICCV.2017.575.
Bakshi, U. and Singhal, R., (2014), “A survey on face detection methods and
feature extraction techniques of face recognition”, International Journal of
Emerging Trends & Technology in Computer Science (IJETTCS), 3(3), pp.233-
237.
98
Barnouti, Nawaf Hazim. (2016), “Face Recognition Using Eigen-Face
Implemented On DSP Processor”, International Journal of Engineering Research
and General Science, 4(2), pp.107-113
Bevilacqua, V., Mastronardi, G., Melonascina, F., Nitti, D., (2006), “Stereo-
Matching Techniques Optimisation Using Evolutionary Algorithms”, LNCS 4113
© Springer-Verlag Berlin Heidelberg 2006, pp. 612-621
Bugatti, P. H., Traina, A. J. M., and Traina Jr., C. (2008), “Assessing the best
integration between distance-function and image-feature to answer similarity
queries”, In Proceedings of the 23rd ACM SAC, pages 1225–1230, New York,
NY, USA. ACM.
99
Daubechies, Wim Swelden (1997),” Factoring wavelet transforms in to lifting
steps”, Princeton University, NJ and Lucent technologies, NJ.
D. Huang, Y. Wang, Y. Wang 2007, “A robust method for near infrared face
recognition based on extended local binary pattern”, in: Proceedings of ISVC, pp.
437–446.
F. Xiaolong and V. Brijeshl, (2009), “Selection and fusion of facial features for
face recognition,” Expert Systems With Applications, vol.36,pp.7157–7169.
Hatem, H., Beiji, Z., Majeed, R., Lutf, M. and Waleed, J., (2015), “Face Detection
and Pose Estimation Based on Evaluating Facial Feature Selection”, International
Journal of Hybrid Information Technology, 8(2), pp.109-120.
100
H. Frigui, R. Krishnapuram, (1999), "A Robust Competitive Clustering Algorithm
with Applications in Computer Vision", IEEE Trans. Pattern Analysis and
Machine Intelligence, vol. 21, no. 5, pp. 450-465.
H.A. Rowley, S. Baluja, and T. Kanade, (1998) “Neural Networks Based Face
Detection”, IEEE Trans. Pattern Analysis an Machine Intelligence, vol. 20, no. 1,
pp. 22-38.
H.Wang, S. Li, Y.Wang, and J. Zhang (2004) “Self quotient image for face
recognition,” in Proceedings of the International Conference on Image Processing,
pp. 1397– 1400.
Jameel S, (2015), “Face recognition system using PCA and DCT in HMM”,
International Journal of Advanced Research in Computer Communication
Engineerin, 4, 13–18.
Kang D, Han H, Jain AK, Lee S.W (2014) “Nighttime face recognition at large
standoff: cross-distance and cross-spectral matching”, Pattern Recognition,
47(12):3750–66.
101
Lin D., Tang X. (2006) “Inter-modality Face Recognition”. In: Leonardis A.,
Bischof H., Pinz A. (eds) Computer Vision – ECCV 2006. ECCV 2006. Lecture
Notes in Computer Science, vol 3954. Springer, Berlin, Heidelberg
Lienchart R, Maydt M (2002), “An extended set of Haar-like features for rapid
object detection”, In International conference on image Processing, vol 1, pp I-
900–I-903.
L. Shen, J. He, S. Wu, S. Zheng (2012), “Face recognition from visible and near-
infrared images using boosted directional binary code”, in: Proceedings of ICIC,
pp. 404–411.
M. Ao, D. Yi, Z. Lei, and S. Z. Li. (2009) “Handbook of Remote Biometrics for
Surveillance and Security”, Advances in Computer Vision and Pattern
Recognition Series, Springer.
Maeng H, Choi H.C, Park U, Lee S.W, Jain AK (2011) “NFRAD: near-infrared
face recognition at a distance”, International joint conference on biometrics
compendium (IJCB), IEEE.
Maeng H, Liao S, Kang D, Lee S.W, Jain AK (2012) “Nighttime face recognition
at long distance: cross-distance and cross-spectral matching”, ACCV, Daejeon,
Korea.
Mohsen A., Abdul Hossen R., Abd Alsaheb Ogla M. and Mahmood Ali, (2017),
“Face Detection by Using OpenCV’s Viola Jones Algorithm based on coding
eyes”, Iraqi Journal of Science.
102
N. Kalka, T. Bourlai, B. Cukic, and L. Hornak (2011) “Cross-spectral Face
recognition in Heterogeneous Environments: A Case Study on Matching Visible
to Short-wave Infrared Imagery” In International Joint Conference on Biometrics.
Nilamani Bhoi, Mihir Narayan Mohanty. (2010), “Template Matching based Eye
Detection in Facial Image”.International Journal of Computer Applications (0975-
8887)Volume 12-No.5.
Oivind Due Trier, Anil K. Jain, Torfinn Taxt, (1996), “Feature extraction Methods
for character recognition – A Survey”, Journal of Pattern Recognition, Vol. 29,
No. 4, pp. 641-642.
Paula Viola and Michael J. Jones, (2001), Rapid Object Detection using
adaBoosted Cascade of Simple Features. IEEE CVPR.
P.J. Phillips, P.J. Flynn, J.R. Beveridge, W.T. Scruggs, A.J. O'Toole, D. Bolme,
K.W.
Bowyer, B.A. Draper, G.H. Givens, Y.M. Lui, H. Sahibzada, J.A. Scallan, W.
Samual (2009), “Overview of the multiple biometrics grand challenge”, in:
Proceedings of ICB, pp. 705–714.
Rath, S.K. and Rautaray, S.S., (2014), “A Survey on Face Detection and
Recognition Techniques in Different Application Domain”, International Journal
of Modern Education and Computer Science, 6(8), pp.34-44
103
R. S. Ghiass, O. Arandjelovi´c, A. Bendada, and X. Maldague, (2014), “Infrared
face recognition: a comprehensive review of methodologies and databases.,”
Pattern Recognition, vol. 47, no. 9, pp. 2807–2824.
Singh, A., Singh, S.K. and Tiwari, S., (2012), “Comparison of face Recognition
Algorithms on Dummy Faces”, The International Journal of Multimedia & Its
Applications, 4(4), pp.121-135
S.Z. Li, R. Chu, S. Liao, L. Zhang (2007), “Illumination invariant face recognition
using near-infrared images”, IEEE Trans. Pattern Anal. Mach. Intell. 29, 627–639.
S. Z. Li, D. Yi, Z. Lei, and S. Liao (2013). “The casia nir-vis 2.0 face database”.
In Computer Vision and Pattern Recognition Workshops (CVPRW), 2013 IEEE
Conference on, pages 348–353.
S. Zhao, R.R. Grigat (2005), An automatic face recognition system in the near
infrared spectrum, in: Proceedings of MLDM, pp. 437–444.
104
T. Bourlai and B. Cukic (2012) “Multi-spectral face recognition: identification of
people in difficult environments”. In Intelligence and Security Informatics (ISI),
IEEE International Conference on, pages 196–201.
Varsha Gupta and Dipesh Sharma, (2015) “A study of various Face Detection
Methods”, International Journal of Advanced Research in computer and
communication Engineering, Vol 3, Issue 5.
X. Chai, S. Shan, X. Chen, and W. Gao, (2007), “Locally linear regression for
pose-invariant face recognition,” IEEE Trans. Image Proc., vol. 16, no. 7, pp.
1716-1725.
105
X.Tan and B.Triggs, (2007) “Enhanced local texture feature sets for face
recognition under difficult lighting conditions,” In Proceedings of the IEEE
international workshop on analysis and modeling of faces and gestures, pp. 168–
182.
X. Xie and K.-M. Lam (2006) “An efficient illumination normalization method
for face Recognition”, Pattern Recognition Letters.
Zhao, W., Chellappa, R., Rosenfeld, A., Phillips, P.J. (2003), “Face recognition:
A literature survey”. ACM Computing Surveys 35, 399–458.
Z. Lei, S. Liao, A. K. Jain, and S. Z. Li. (2012), “oupled discriminant analysis for
heterogeneous face recognition”. IEEE Transactions on Information Forensics
and Security, 7(6):1707–1716.
106