CHAPTER 5 FACE RECOGNITION IN COMPRESSED DOMAIN...
Transcript of CHAPTER 5 FACE RECOGNITION IN COMPRESSED DOMAIN...
91
CHAPTER 5
FACE RECOGNITION IN COMPRESSED DOMAIN USING
CANONICAL CORRELATION ANALYSIS BASED FEATURE
VECTOR OPTIMIZATION WITH MODE BASED CLASSIFICATION
5.1 Introduction
Until recently, Canonical Correlation Analysis (CCA) was a relatively
unknown statistical technique. As with almost all of the multivariate techniques, the
availability of computer programs has facilitated its increased application to research
problems. It is particularly useful in situations in which multiple output measures.
The Canonical correlation places the fewest restrictions on the types of data on
which it operates. Because the other techniques impose more rigid restrictions, it is
generally believed that the information obtained from them is of higher quality and
may be presented in a more interpretable manner. For this reason, many researchers
view canonical correlation as a final effort, to be used when all other higher-level
techniques have been exhausted. But in situations with multiple dependent and
independent variables, canonical correlation is the most appropriate and powerful
multivariate technique. It has gained acceptance in many fields and representing as a
useful tool for multivariate analysis, particularly as interest has spread to considering
multiple dependent variables.
In the recent years, the Canonical Correlation Analysis (CCA) arouse the
growing interest of experts in biometrical technologies of people recognition, as a
method which helps to relate sets of observations describing different aspects of
appearance (Jelsovka 2011). The CCA represents a high-dimensional relationship
between two sets of variables with a few pairs of canonical variables. It was intended
to describe relations between two sets of one dimensional data sequences. The CCA
method has been widely used in several fields such as signal processing (Hotelling
92
1936), medical studies, pattern recognition (W. Zhao. 2000) (M. Borga. 2001) etc.
The CCA as a novel method is applied to image processing and biometrics too. It is
compared to other projection approaches like Principal Component Analysis (PCA)
and Independent Component Analysis (ICA), CCA can concurrently deal with two
sets of data.
A face image can be considered as a vector in a high dimensional space.
This high dimensional vector has large variations between two images of the same
person, and therefore it is not directly suited for the face recognition. Principal
Component Analysis (PCA) and Independent Component Analysis (ICA) are
commonly used dimensionality reduction techniques, which give good recognition
rates. Although PCA is a good low dimensional representation for face images, it is
not able to discriminate between variations due to illumination and expression
changes. The ICA solves the illumination change problem to some extent by finding
the transformation such that it maximizes the inter-class separation and minimizes
the intra-class variations.
In (O. Friman. 2001), CCA was used for the segmentation of functional
Magnetic Resonance Images and this motivated to apply CCA for the purpose of
face recognition in this research. The Correlation analysis is useful to find a linear
relationship between two sets of variables, and CCA creates new variables for each
set such that the correlation between these variables is maximized and independent
of affine transformation (Fu-Chang Liu. 2008).
The possibility of using JPEG2000 compression scheme and Canonical
Correlation Analysis for performing face recognition in fully compressed domain is
explored in this chapter. A novel approach for efficient face recognition in
compressed domain has been proposed in this chapter using 2-dimensional Canonical
Correlation Analysis. The Matching of image data has been done by Mode based
Matching method. The experimental results proved that the proposed method
considerably improves the recognition rates and also reduces the computational time
and storage requirements.
93
5.2 Theoretical Foundations for Canonical Correlation Analysis
The Canonical Correlation Analysis (CCA) is a multivariate analysis
method used to identify and measure the association between two sets of variables
(Hotelling 1936). It first searches for a pair of linear combinations which has the
largest variation. Then CCA establishes a pair of linear combinations having the
largest correlation among all pairs uncorrelated with the initially settled pair and so
on. In fact CCA represents a high dimensional relationship between two sets of
variables with a few pairs of canonical variables (Sun Ho Lee 2007). The CCA has
various applications in pose estimation (Z. Lei 2008) and face matching (M. Slamka
2007). For two sets of variables, CCA is to create the CCA subspace to mutually
maximize the correlation between these two sets of variables.
For better understanding of the concepts and terminology used in the chapter, a
review of the important terms is given below:
Canonical correlation: It measures the strength of the overall relationships
between the linear composites (canonical variates) for the independent and
dependent variables. In effect, it represents the bivariate correlation between
the two canonical variates.
Canonical function: It is the relationship (correlational) between two linear
composites (canonical variates). Each canonical function has two canonical
variates, one for the set of dependent variables and one for the set of
independent variables. The strength of the relationship is given by the
canonical correlation.
Canonical loadings: This is a measure of the simple linear correlation
between the independent variables and their respective canonical variates.
These can be interpreted like factor loadings, and are also known as canonical
structure correlations.
Canonical roots / Eigenvalues: The Squared canonical correlations, which
provide an estimate of the amount of shared variance between the respective
94
optimally, weighted canonical variates of dependent and independent
variables.
Canonical variates/Linear composites: The Linear combinations that
represent the weighted sum of two or more variables and can be defined for
either dependent or independent variables.
5. 3 Feature Vector optimization using Canonical Correlation Analysis
The Canonical correlation analysis is a suitable and dominant technique which can
be used for exploring the relationships among multiple dependent and independent
variables. Therefore a powerful feature projection approach for facial images is
proposed based on CCA. The Correlation analysis is useful to find a linear
relationship between two sets of variables, and CCA creates new variables for each
set such that the correlation between these variables is maximized and independent
of transformation. The CCA is useful to find two sets of basic vectors, one for x and
another for y, such that the correlation between the projections of the variables onto
these basic vectors is maximized. The CCA is useful to find wx and wy which are two
pairs of vectors such that the correlation between the projections of the variables is
maximized onto the basic vector: xwx
T
xand
ywyT
y. The Projections x and y are
known as canonical variables. The projections x and y are also referred to as
canonical variates in the perspective of CCA (W. Yang 2008) (X. Wang 2004).
Officially CCA maximizes the function:
(5.1)
Whereby, E represents empirical expectation. For two random variables X and Y, the
total covariance matrix is a block matrix where Cxx∈Rp×p and Cyy∈Rq×q are the
within-sets covariance matrices of x and y respectively and Cxy∈Rp×q is the
between- sets covariance matrix (G. Kukharev 2010). Consider two random variables
yyyyT
xxxxT
xxyxT
wCwwCw
wCEwp
95
X and Y. The total covariance matrix is a block matrix where Cxx and Cyy are the
within-sets covariance matrices of x and y respectively and Cxy=Cyx T is the
between-sets covariance matrix.
(5.2)
The canonical correlations between x and y can be found by solving the
eigenvalue equations:
xyxyyxyxxx wCCCCwp112
(5.3)
yxyxxyxyyy wCCCCwp112
(5.4)
Where, eigenvalues p2 are the squared canonical correlations and eigenvectors wx
and wy are the normalized canonical correlation vectors (R. Hardon 2004). Instead
of the two eigenvalue equations (5.3) and (5.4), the problem can be formulated in
one single eigenvalue equation:
(5.5)
Where
(5.6)
Input data and the structure of CCA are schematically presented in Figure 5.1.
96
Figure 5.1 CCA calculation
5.4 Experimental Setup
The proposed methodology to perform face recognition in compressed
domain uses wavelet transform based compression scheme, Canonical correlation
analysis based feature projection technique and Mode based Matching method for
matching of images have been described here. The clear approach to carry out face
recognition in compressed domain is to utilize the coefficients extracted before
inverse transformation as input to face recognition systems. Both the stages of
inverse quantization and inverse transformations are avoided. The original image
has been transformed using the DWT and quantization and entropy coding was
done on the coefficients. The implementation has been done using MATLAB
2008B (Toolbox used are: Image Processing, JPEG 2000 toolbox and Wavelet
Toolbox), a computer system with Intel i3 Processor 2.20 GHz, memory of 3GB
and hard disk memory of 500GB. All the images were preprocessed as given in
section 5.4.2 before conducting experiments.
5.4.1 Dataset
The AR Face database of face images collected at the Computer Vision
Center in Barcelona, Spain in 1998 (Martinez. A. M. 1998) has been used for the
experiments. The AR Face Database consists of over 4000 images corresponding
97
to 126 subjects (76 males, 50 females). All images are of frontal view with
different expression, occlusions and illumination variations. Image datasets
include image feature frontal view faces with varying facial expressions (fb),
varying illumination conditions (fc) and partial occlusions (fd). The system was
trained with 4 images per subject. There are total of 272 training and 504 testing
images. The configuration of the system used for image capturing is: Pentium
133MHz, 64 MB RAM and 2GB HD, colour camera: SONY 3CCDs, a 12mm
optics and frame grabber: Matrox Meteor RGB.
5.4.2 Image pre-processing
Preprocessing of images prior to wavelet transformation is essential in this
implementation to maintain the size of the training and test images as same. The
RGB image was converted into grayscale image and then cropped to size of 128
x128 pixels. Elliptical masking was utilized to mostly remove the background.
Also images were originally transformed to obtain the eyes at the fixed points and
histogram equalization was done to have better background intensity.
5.4.3 Pixel domain experiments
In these experiments, Canonical Correlation Analysis based face
recognition was performed on the original uncompressed 128x128 images and the
results for all the three image Datasets of AR Face database were noted which is
represented by original images column in Table 5.1. The images were divided into
training and testing image Dataset where training data are formed the whole AR
database images and they are stored as vectors. Test image and train images were
transformed as vectors using the equations given in section 5.3. Face Recognition
was done by measuring canonical correlation coefficient between the test face and
train faces. The correlation between two face images has maximum value when
the correlation coefficient is 1 or -1. Then the faces are identical.
5.4.4 Compressed domain experiments
All the preprocessed images used in the experiments were compressed
according to JPEG2000 compression scheme with various compression ratios of
98
1 bpp, 0.5 bpp, 0.3 bpp and 0.2 bpp. The proposed approach eliminates major part
of the decompression phase as the decoding was interrupted after entropy
decoding and all entropy points obtained were used as input to the feature
projection method based on CCA. The coefficients after entropy decoding phase is
enough to carry out the face recognition process and hence the recognition system
was positioned after the entropy decoding phase in the proposed work.
The training and test images were transformed as vectors using the
equations in section 5.3. Face recognition was performed by measuring the
canonical correlation coefficient between the test face and all training images. The
algorithm for the proposed CCA based face recognition is given in figure 5.2. The
Mode based matching has been used for image classification which is explained in
section 5.4.5. The results for face recognition of the various experiments
conducted on the image Datasets at different compression ratios are shown in the
figures given in section 5.5.
5.4.5 Mode Based matching of images
The Euclidean distance method for matching of images possess only 50%
of the result and another matching algorithms like Kd-tree matching is not suitable
for this feature projection technique based on CCA. Hence mapping of the image
data was done by means of finding Mode method. The differences between two
square roots of canonical correlation were taken as „M‟. If there are 20 image data
in a database, then 20 different „M‟ values will be obtained. The square of each
„M‟ value will be calculated and mode of this squared value is obtained for
mapping purpose.
99
Figure 5.2 Algorithm of the proposed CCA approach
5.5 Experimental results and analysis
Numerous experiments have been conducted using the proposed CCA
based method in compressed domain. The results obtained confirms that the
proposed method based on CCA is superior in terms of total Recognition Rate
(RR) and Normalized Recognition Rate (NRR) with a computational time when
compared to the previous approaches explained in chapter 4.
CCA based Feature projection:
1. Get the image data from the compressed domain after entropy
decoding.
2. Separating single coloured plane from RGB planes of the true
colour image in order to reduce the processing time.
3. Generating identity matrix in the same size of image data from
previous step.
4. Apply Canonical Correlation Analysis for the image data and
identity matrix.
5. This CCA analysis should be taken for both query image and all
the images in Database.
6. Differences between two square root of canonical correlation are
taken as (M).
7. If there are 20 image data in a database then we will obtain 20
different „M‟.
8. Square each „M‟.
9. Mode of this squared value is obtained for mapping purpose.
100
The Recognition rates for various image Datasets obtained by using
entropy points as input to the proposed new classification method based on CCA
are given in tables given below. The table 5.1 gives the recognition rates for
images with varying expressions at different compression ratios. The table 5.2 and
5.3 give the recognition rates for images with varying illumination conditions and
minor occlusions respectively. A comparison of the recognition accuracy of the
CCA based method with existing approaches is given in Table 5.4. Comparatively,
better results were observed when evaluated against approaches based on standard
face recognition systems with Euclidean matching. Also the proposed method
shows significantly better results for the probe sets fc and fd when compared to
approaches based on PCA, ICA and KPCA with Kd-tree matching. For image
Datasets with varying expression (fb), a marginal drop in recognition rate is noted
when compared to the previous approaches based on PCA, KPCA and ICA with
Kd-tree matching.
Images with varying expression (fb probe set)
The fb probe set has shown 0.5% increase at 1 bpp, 0.35 % to 10.28%
decrease at higher compression levels when compared to PCA with Euclidean
matching & 4.85% to 13.18% decrease when compared to ICA with Euclidean
matching.
Images with varying Illumination (fc probe set)
The image Dataset with varying illumination conditions has shown 16.51%
(at .2 bpp) to 33% (at 1 bpp) increase in Recognition Rate when compared to
PCA with Euclidean matching & 4.8%(at .3 bpp) to 13.94% (at .5 bpp) increase
in Recognition Rate when evaluated against ICA with Euclidean matching.
Images with partial occlusion (fd probe set)
Partially occluded image set. al.so has shown 12.14% (at .2 bpp) to
18.04% (at 1 bpp) increase in Recognition Rate (RR) when compared to PCA
with Euclidean matching & 6.17% (at .3 bpp) to 9.26% (at 0.5 bpp) increase in
Recognition Rate when matched to ICA with Euclidean matching.
101
Table 5.1 Recognition Rate (%) of fb probe set for proposed methodology
fb probe set (images with varying expression)
Results at Rank 1 Recognition (%)
Feature
Projection
Technique
Classification
method
applied
Compressed Domain
Original
Image
(Pixel
Domain)
Compressed Images
1 bpp 0.5 bpp 0.3 bpp 0.2 bpp
Canonical
Correlation
Analysis
(CCA)
Mode based
matching 77.76 77.90 78.65 71.52 69.92
Table 5.2 Recognition Rate (%) of fc probe set for proposed methodology
fc probe set (images with varying illumination conditions)
Results at Rank 1 Recognition (%)
Feature
Projection
Technique
Classification
method
applied
Compressed Domain
Original
Image
(Pixel
Domain)
Compressed Images
1 bpp 0.5 bpp 0.3 bpp 0.2 bpp
Canonical
Correlation
Analysis
(CCA)
Mode based
matching 79.32 81.20 81.74 72.20 68.36
102
Table 5.3 Recognition Rate (%) of fd probe set for proposed methodology
fd probe set (Partially occluded images)
Results at Rank 1 Recognition (%)
Feature
Projection
Technique
Classification
method
applied
Compressed Domain
Original
Image
(Pixel
Domain)
Compressed Images
1 bpp 0.5 bpp 0.3 bpp 0.2 bpp
Canonical
Correlation
Analysis
(CCA)
Mode based
matching 53.89 55.24 51.76 48.07 45.14
Table 5.4 Comparison of Results at Rank 1 Recognition (%) for the existing
and new methodologies
The NRR evaluation of the proposed method for the fb probe set (images
with varying expression) at various compression levels is detailed in figure 5.3. It
103
is clearly observed from the values that the proposed approach outperforms the
recognition after decompression. The image Datasets with varying illumination
conditions and partial occlusions also justify the same conclusion (Figure 5.4 and
5.5 respectively).
Figure 5.3 NRR evaluation of fb probe set – CCA Mode based matching
Figure 5.4 NRR evaluation of fc probe set - CCA Mode based matching
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1 0.5 0.3 0.2
No
rmal
ize
d R
eco
gnit
ion
Rat
e (
NR
R)
Compression Ratio ( bpp)
NRR evaluation - fb probe set CCA Mode based matching
NRR afterdecompression
NRR compresseddomain
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1 0.5 0.3 0.2
No
rmal
ize
d R
eco
gnit
ion
Rat
e (
NR
R)
Compression Ratio ( bpp)
NRR evaluation - fc probe set CCA Mode based matching
NRR afterdecompression
NRR compresseddomain
104
Figure 5.5 NRR evaluation of fd probe set - CCA Mode based matching
Computational Time and Storage requirements
The computational time of the proposed CCA approach was compared with
the existing recognition system which uses Euclidean distance method (B. Li.
2002) and Kd-tree matching (section 4.4) respectively for matching of images.
The proposed method took only 2.87 seconds for recognition with a significant
improvement in recognition rate.
Figure 5.6 Comparison of computational time
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1 0.5 0.3 0.2
No
rmal
ize
d R
eco
gnit
ion
Rat
e (
NR
R)
Compression Ratio ( bpp)
NRR evaluation - fd probe set CCA Mode based matching
NRR afterdecompression
NRR compresseddomain
105
5.6 Conclusion
A new approach for performing face recognition systems directly in
JPEG2000 compressed domain based on Canonical Correlation Analysis has been
applied and assessed. The effect of compression on recognition accuracy was
observed and also issues like achieving computational time saving was examined
in the new method. The results obtained confirms that the proposed method based
on CCA is better in terms of total Recognition Rate (RR) and Normalized
Recognition Rate (NRR) with a less computational time when compared to the
previous approaches explained in the previous chapter. The suggestion for future
enhancement is to develop a new method for extracting feature vector from
entropy coded image thus completely eliminating the decompression phase in face
recognition and also a technique to enhance the proposed method by applying
artificial neural networks for better recognition of images with varying
expressions. The researcher has tried to address both the suggestions successfully
in the successive chapters 6 and 7.