An Experimental Study of different Features for Face Recognition
M.Hanmandlu R.Bhupesh Gupta Farrukh Sayeed A.Q.Ansari
E.E Department, E.E.Department E &C Department E.E Department
I.I.T. Delhi I.I.T.Delhi P.A.College of Engineering JMI, New Delhi
New Delhi, India New Delhi, India Mangalore
mhmandlu@gmail.com rbhupeshgupta@gmail.com sayeed.farrukh@gmail.com aqansari62@gmail.com
Abstract – As a first study, the use the Gabor filter bank is
made to generate features for face recognition. The features
so obtained on the application of SVM classifier yields
accuracy rate of 96.2%. With a view to improve the
performance, two more feature types, viz., wavelet features
and wavelet-fuzzy features resulting from the application of
2D wavelet transform on the Composite detail images and
the Approximate images at 3 levels of decomposition, are
devised. The ROCs of three feature types show that wavelet-
fuzzy features have a better performance. The performance
of Gabor features is slightly inferior to that of wavelet-fuzzy
features. The algorithm was tested on ORL (Olivetti
Research Laboratory) database that has slight orientations
in face images.
Keywords: Gabor filter bank, Haar wavelets, Muti-
decomposition, wavelet-fuzzy features, SVM
I. Introduction
The face recognition is one of the most important
applications of the Biometric based authentication
systems. During the past few decades, it has received a
considerable attention, because of a large number of
application areas such as Entertainment, Smart Cards,
Information Security, Law Enforcement, Surveillance etc.
Face recognition involves a range of activities from many
aspects of human life. Humans are good at face
recognition and now machine learning is being improved
to do this task. Identification of people in different
environments requires three steps: acquisition,
normalization and recognition. Acquisition involves the
detection and tracking of face-like patches in the dynamic
scenes. Normalization is the segmentation, alignment and
standardization of the face images, and finally recognition
is concerned with the representation via modeling of face
images and association of face images with the models
learned.
The methods reported in the literature are broadly
classified into: i) Holistic matching methods in which the
whole face acts as an input to the recognition system, ii)
Feature based matching methods, in which as the name
indicates, the local features such as eyes, mouth, and nose
are first extracted and their locations and the local
statistics are fed to the recognition system, and iii) Hybrid
methods where the recognition system makes use of both
the local features as well as the whole face region to
recognize the face.
Schneiderman and Kanade [8] apply statistical likelihood
tests using feature output histograms to create the
detection scheme. Rowley and Kanade in [7] use neural
network-based filters to obtain the benchmark results. In
another early work, Papageorgiou et al. [5] propose a
general object detection scheme that uses a wavelet
representation and statistical learning techniques. Osuna et
al. [4] apply Vapnik's support vector machine for the face
detection, and Romdhani et al. [6] improve upon this work
by creating the reduced training vector sets. Fleuret and
Geman attempt a coarse-to-fine approach to face
detection, focusing on minimizing the computation
involved [3]. In perhaps the most impressive paper, Viola
and Jones use the concept of an “integral image", along
with the rectangular feature representation and Adaboost
learning algorithm to detect faces at the rate of 15 frames
per second [2]. Several basic face recognition techniques
exist suitable to the frontal faces. These include:
eigenfaces, neural networks, dynamic link architecture,
hidden Markov model and geometrical feature matching.
These approaches are analyzed in terms of the facial
representations adopted by them.
II. The Proposed Features
Here we will be exploring three types of features which
comprise Gabor, Wavelet and Wavelet-fuzzy. Our main
innovation is in the second and the third types. We will
now elaborate on these three feature types.
A. Gabor Features
Here we use the Gabor filter bank with different
orientations and frequencies. The output of each filter is
convolved with the original image I(x,y) and the
convolved outputs of the different filters are averaged to
2011 International Conference on Communication Systems and Network Technologies
978-0-7695-4437-3/11 $26.00 © 2011 IEEE
DOI 10.1109/CSNT.2011.121
567