IEEE 2011 Paper on Image Processing

Text-only Preview

Architecture for Precisive Face Recognition System Deploying Quantization with
Multiresolution Curvelet and Training with Support Vector Machine

Mr.K.N.Narasimha Murthy1,

Dr.YS Kumaraswamy2,
1Research Scholar, Anna University,

2Sr. Professor, Department of MCA (VTU),
Dept. of Information Science and Engineering,

Dayanadasagar College of Engineering
City Engineering College, Bangalore, India,

Bangalore, India
[email protected]

[email protected]

Dr. S T Narayanaswamy3
3Director (Academic),
City Engineering College, Bangalore, India
E-mail:[email protected]

Abstract-- The proposed research approach addresses various
While such variability and spatial nonstationarity defies any
issues in face recognition as well as in computer vision which
single statistical characterization, the multiresolution
were signified and researched by multiresolution concepts like
components are more easily handled. Not only this, it is also
wavelet transforms. But survey also shows that only wavelets
found that the Fourier transform [2] of an image is not very
are not the factors for ideal description of an image as they
informative from the perspective of object recognition. Other
have very rough directional representations and are absolutely
transforms like wavelets, curvelets etc. provide alternative
not anisotropic. With the recent developments in Curvelet
image representations. These transforms represent images in
Transform, which has the potential to overcome these flaws of
such a way that recognition is facilitated. There are five
wavelets, this proposed idea highlights an advance technique
criteria for multiresolution image transform e.g. (1)
of face recognition which is based on multiresolution curvelet.
Multiresolution. The transform should allow images to be
The application will quantize 8 bit image to 4 bit and 2 bit
representation in initial phase. In the next phase, the curvelet

successively approximated, from coarse to fine resolutions.
transform will be applied to all 3 different resolved versions of
(2) Localization. The basis elements of the transforms should
the image. In the final phase, all the 15 sets of co-efficients
be localized in both the spatial and the frequency domains.
were used to train different support vector machines. Finally,
(3) Critical sampling. For some applications (e.g.,
the accuracy of the application will be evaluated by
compression), the transforms should form a basis, or a frame
identification from different sets of facial image for the
with small redundancy. (4) Directionality. The transform
proposed robust face recognition system.
should contain basis elements oriented at a variety of
directions, compared with the few directions offered by
Keyword: Wavelet Transform, Multiresolution, Curvelet,
separable wavelets. (5) Anisotropy. When a physical
Face Recognition, Support Vector Machine.
property changes with direction, that property is said to be

anisotropic, e.g. in certain crystals the conductivity of heat is
more in horizontal direction than in vertical direction; for
such a case, the conductivity is said to be anisotropic.

The proposed work will introduce a novel face
The problems encountered in face recognition system are
recognition technique based on multiresolution curvelet
considered as one of the most important research areas in
transform where the application deploys the Fast Discrete
human visual system and image statistics. Majority of the
Curvelet Transform architecture via unequally-spaced fast
face recognition algorithms has witnessed a performance
Fourier transforms and Wrapping since Curvelets have
drop whenever face appearances are subjected to variations
ability to represent edges and other singularities along curves
by factors such as occlusion, illumination, expression, pose,
to meet 100% accuracy level in recognizing facial image.
accessories and aging. In fact, often these factors lead to
The rest of this paper is organized as follows. We discuss
intra-individual variability of face images, to the extent that
related work in Section II. The problem statement is
they can be larger than the inter-individual variability.
discussed in Section-III. Assumptions and dependency
Multiresolution techniques [1] have already been researched
considered for the research work is elaborated in Section IV
in order to mitigate the loss of classification performance
Proposed System is described in Section-V. Implementation
due to changes in facial appearance. The most popular
Scenario is discussed in Section-VI and finally conclusion
multiresolution analysis technique is found to be wavelet
and future work is described in Section-VII
transform. Wavelet decomposition is the most widely used

multiresolution technique in image processing. Images have
typically locally varying statistics that result from different
combinations of abrupt features like edges, of textured

regions and of relatively low-contrast homogeneous regions.
978-1-4244 -8679-3/11/$26.00 (c)2011 IEEE


Tanaya Mandal e.t. al [3] have introduced a new
but yet this method of correlation of unprocessed data is
feature extraction technique from still images using PCA on
frequently used by many researchers and has highest
curvelet domain which has been evaluated on two well-
probability to fail in situation where the correlation is very
known databases. This technique has been found to be robust
minimum between two images of same test-person with
against extreme expression variation as it works efficiently
different posture.
on Essex database. The subjects in this dataset make

grimaces, which form edges in the facial images and curvelet
transform captures this crucial edge information. The

proposed method also seems to work well for ORL database,
The project work assumed an image database of the same
which shows significant variation in illumination and facial
person taken from webcam under congenial illumination and
details. There has also been an extensive research to describe
lightening and converted to appropriate image file format for
an image transform [4] that may proof beneficial in image
exact input. The input image is assumed to have minimal
and in vision research.
noise, in case if it is there. The prime focus of the project is
Eero P. Simoncelli [5] presented jointly shiftable
basically the face recognition, so particular database is
transforms that are simultaneously shiftable more than one
selected for that purpose only. The term wavelet is usually
domain. Two examples of jointly shiftable transforms are
assumed to refer to an orthonormal basis set. The main
designed and implemented: a one-dimensional transform that
dependency of the project work is mainly the Matlab IDE
is jointly shiftable in position and scale, and a two-
without which the project execution fails at this stage.
dimensional transform that is jointly shiftable in position and

orientation. It was demonstrated the usefulness of these
image representations for scale-space analysis, stereo

disparity measurement, and image enhancement.
Our face recognition system is divided into two stages:
Minh N. Do [6] constructed a discrete transform that
training stage and classification stage. In training stage, the
provides a sparse expansion for typical images having
images are decomposed into its approximate and detailed
smooth contours. Szabolcs Sergyan [7] has presented a face
components using curvelet transform. These sub-images thus
detection algorithm which was used in color images, and
obtained are called curvefaces. These curvefaces greatly
some classification methods and this usage for the reduces the dimensionality of the original image. Then the
classification of image database were reviewed.
methodology is applied on selected subbands, which further
Emmanuel Candes [8] has described two digital reduces the dimension of image data. Thereafter only the
implementations of a new mathematical transform, namely,
approximate components are selected to perform further
the second generation curvelet transform in two and three
computations, as they account for maximum variance. Thus,
dimensions. The first digital transformation is based on
a representative and efficient feature set is produced. In
unequally-spaced fast Fourier transforms while the second is
classification stage, test images are subjected to the same
based on the wrapping of specially selected Fourier samples.
operations and are transformed to the same representational
Dengsheng Zhang e.t. al [9] has presented a new texture
feature based on curvelet transform.
In the initial process, recognition is done by varying bit

of quantization where digital images in black and white are
represented in 8 bits or 16 bits which result in 256 or 65536
The challenging problem consists of a data set of facial
gray levels. If we assume that the images are represented by
images and a defined set of experiments. One of the
256 gray levels, in such an image two very near regions can
impediments to developing improved face recognition is the
have different pixels values. There are lots of "edges" in a
lack of data.
gray scale image and consequently the edge information will
The problem which exists in face recognition actually
be captured by the curvelet transform. In case if there is
boils down to identification of an individual depending on
quantization of the gray levels, nearby regions which has a
array of pixel intensities. Deploying this input values and
very little differences in pixels values and formed edges in
extracted information from the other images of known
the original 256 bit image will be merged and as a result
person, it has been seen that there are several issues which
only more bold edges in face image will be represented. At
seeks to assign a name to an unidentified set of intensities of
this moment, in case the gray-level quantized images are
pixels. Another prominent issue is notifying the subjected to Curvelet transform, the transformed domain
dependencies laid between values of pixel which actually
coefficient will contain information of these bolder curves.
becomes a statistical signal processing problem.
Images of the same person from the face database, quantized
Certain more problems exist in this area if the application
to 4 bits and 2 bits from the original 8 bit representation.
is subjected to a huge database of image or a photograph.

There was an issue to pick from the database a smaller set of
records so that one of the input image records becomes
equivalent to the photograph. This problem in recognition
system is rendered difficult by great variability in tilt or
rotation of head, lighting intensity, facial expression, , etc.
Previous research has not witness any progress in this issue

Each images will
8 bit images are
previous process
be a bit quantized
converted to 4 bit
and 2 bit version
2 bit
Get 15 different
Transform is
4 bit
version of images
taken at different
Training diff
multi-cast SVM
8 bit

Fig 3. Multiresolution Curvelet
Fig 1. Data Flow Diagram

In the subsequent process, recognition is done by
The application accepts the input of image from the
multiresolution curvelet transform which permits the user to
database which is subjected to bit quantization as shown in
view an image at multiple scales. In highest magnification,
Fig 2 where the first image is original image, second image
majority of the finest details of the picture is visible, but
is 4 bit quantized image, and third one is 2 bit quantized
when magnification is reduced, a rough view of the image is
image. Immediately after this, the next step is
obtained gradually. Taking Curvelet transform at different
multiresolution curvelet transform as depicted in Fig 3. After
resolution will allow the capture of the edge information in
performing this stage, the profile can be saved by a name
facial images at varying granularities. The final processing
which can be used for understanding the proper
of our recognition system will be done by combination of the
identification of test image in this application. The next step
above two stages together to get the final output. The
of processing then assists in generation of data which might
multiresolution analysis allows for the coarse to fine
be used to train support vector machine. The process is
approximation of the images. So by combining the aforesaid
repeated for all the 10 sets of database of the training images.
steps we will have different versions of images, which will

be used to train different multi-class SVM. The final
recognition decision will be obtained by majority voting of
the classifier output as shown in data flow diagram in Fig 1.


The framework is designed on Matlab IDE. To start
working on this application, a database of image is created
by taking snaps of 10 different persons in 10 different facial
postures or angle. The application is then fed with all the
possible inputs from database and is signified by a logical

name. The input image is then subjected to bit quantization

and then multiresolution curvelet is applied. The screenshots
Fig 4. Face Recognition for input images
are as follows.

Once the training is completed, the application will
prompt the user to feed the test image for verification
purpose which is represented by fig 4. The processing is
done taking one set of input database image for training and
another set for testing the application for identification. For
each of the two pairs of training and testing set thus obtained,
two sets of experiments were performed. The application is
tested with all the 10 sets of test images for analyzing the

performance with regards to 8 bit, 4 bits, and 2 bits. The

performance of the application is also scrutinized by
Fig 2. Bit quantization
considering multiple image resolution parameters (high,

average, low, lowest), accuracy classifier, final accuracy,
rejection rate, as well as incorrect classification.

Bits Final accuracy
Rejection rate
The project work highlights one of the possible
8 98.2
utilization of multiresolution curvelet transform for better
4 99.2 0
accuracy in face recognition system. The main aim of the
project work is to design an architectural framework to
2 99.2 0
propose a multiresolution Curvelet based method for face

recognition. The research work also aims to highlight a new
Table 1. Final Accuracy Vs Rejection Rate

texture feature based on curvelet transform. The technique
The above table shows the parameters which are used for
makes use of curvelet transform which represents the latest
analyzing the accuracy of the application developed. Each
research result on multiresolution analysis where the 8 bit
experiment is checked for 8 bits, 4 bits as well as 2 bits and
image is quantized to 4 bit and 2 bit representation.
found to provide accuracy level of 98.2%, 99.2%, and 99.2%
Multiresolution methods provide powerful signal analysis
respectively. Although with a minimal rejection rate of 0.8 in
tools, which are widely used in feature extraction, image
8 bit quantization, the experiment shows absolutely zero
compression and denoising applications. Database of large
rejection rate in both 4 bits and 2 bits of quantization
classes should be experimented in future work. If small
respectively. The evaluation of this performance is shown in
database is used, it classifies as false positive results. This
Fig 5 in graph.
project can be enhanced to capture live things by integration

fast motion cameras to the server. We can enhance the same
to recognize the facial expression of a human.


[1] Hazim Kemal Ekenel, Multiresolution face recognition, Image and
Vision Computing 23 (2005) 469-477
[2] Angshul Majumdar and Rabab K. Ward, Multiresolution Methods in

Face Recognition, Recent Advances in Face Recognition, Book

edited by: Kresimir Delac, Mislav Grgic and Marian Stewart Bartlett ,
Fig 5. Graph for Table 1.
ISBN: 978-953-7619-34-3, Publisher: InTech, Croatia, Publishing

date: December 2008
The next set of performance evaluation is conducted
[3] Tanaya Mandal and Q. M. Jonathan Wu, Face Recognition using
considering distinct parameters like average accuracy
Curvelet Based PCA, Pattern Recognition, 2008. ICPR 2008. 19th
classifier and final accuracy as shown in Table 2
International Conference on 2008 IEEE

[4] Andrew B Watson, The Cortex Transform: rapid Computation of
Simulated neural Images, Computer Vision, Graphics, and Image
Bits Average
Final Accuracy
[5] Eero P. Simoncelliy, William T. Freemanz, Edward H. Adelsonx,
David J. HeegerShiftable Multi-scale TransformsMIT Media
8 94.82
Laboratory Vision and Modeling Technical ReportIEEE Trans.
Information Theory, vol. 38(2), pp. 587-607, March 1992
4 93.15
[6] Minh N. Do, Martin Vetterli, The Contourlet Transform: An Efficient
2 91.85
Directional Multiresolution Image Representation, IEEE

Table 2. Avg Accuracy Classifier Vs Final Accuracy
[7] Szabolcs Sergyan, A New Approach of Face Detection-based

Classification of Image Databases, 2009
This set of evaluation is performed with almost all
[8] Emmanuel Candes, Laurent Demanet, David Donoho] and Lexing
possible set of image resolution ranging from highest to
Ying, Fast Discrete Curvelet Transforms, March 2006
lowest. The evaluation shows the results of average accuracy
[9] Ishrat Jahan Sumana, Md. Monirul Islam, Dengsheng Zhang and
classifier as 94.82, 93.15, and 91.85 for 8 bit, 4 bit, and 2 bit
Guojun Lu, Content Based Image Retrieval Using Curvelet
Transform, Multimedia Signal Processing, 2008 IEEE 10th
respectively with all final accuracy of 100%. The graphical
Workshop on Issue Date: 8-10 Oct. 2008
representation is shown in Fig 6.

Fig 6. Graph for Table 2