IJCSN - International Journal of Computer Science and Network, Volume 9, Issue 2, April 2020

ISSN (Online) : 2277-5420

www.IJCSN.org
Impact Factor: 1.5

A Study Analysis on Face Recognition and Various

Methodologies
1
Michael Farayola; 2Aman Dureja; 3Ajay Dureja
1
Department of Computer Science and Engineering, PDM University,
124507, Haryana, India
2
Department of Computer Science and Engineering, PDM University,
124507, Haryana, India
3
Department of Computer Science and Engineering, PDM University,
124507, Haryana, India

Abstract - Face Recognition is a crucial concept in today’s world which has helped enormously in the areas of security, biometric,
surveillance, access control and many more. It deals with the identification of human faces, which serves as an individual’s identity.
Every individual can be distinguished and recognized by their faces. Numerous methods have been used extensively over the decades to
enhance and boost the technology of face recognition system accurately. These are PCA, LDA, LPP, ICA, SVM and many hybrid
combinations of these techniques. In this paper, we will study an overview of face recognition, some methods used over time and
common factors that affect the poor acceptability of face recognition system.

Keywords - Face Recognition, Feature Extraction, Accuracy.

1. Introduction improving the accuracy and efficiency of face

recognition. Apart from being a pre-processing phase,
face detection could be used for region-of-interest

C
apability of human perception system [4] and
everyday activity for humans is face recognition, detection, retargeting, video and image classification,
the capability to identify and recognize an etc.
individual, while building a similar computational
model. This model has contributed to theoretical insights 1.2 Feature Extraction
and impacted immensely to many practical applications
like automated crowd surveillance, access control, design This pre-processing phase is used for dimensionality
of human computer interface (HCI), content-based image reduction, noise cleaning, salience extraction and
database management, criminal identification etc. [7]. information packing. Human face patches are extracted
Taken a picture with a digital camera, we would want to from desired images via this phase. It is always
know if a person is inside, the position of the face and advantageous not to make use of these patches directly
the identity of the person. Generally, there are face for any face recognition system because face patches can
recognition pre-processing phase which are categorized encapsulate different factors which tamper with the
into: Face Detection, Feature Extraction. efficiency of face recognition system such as; different
light illuminations, facial expressions, camera alignment
etc. In addition, a patch can have over 1000 pixels, which
becomes cumbersome to build a robust face recognition
system. Hence, there is a need for feature extraction,
1.1 Face Detection after which the face patches are transformed into vectors
of fixed dimensions and their respective location.
This helps to identify human faces and where the faces
are located in a given image. The desire outputs of this To build a robust face recognition, there is a need for
procedure are patches which contains faces present in the database availability which is required to store every
given image. Facial alignment [1] is performed to individual image taken and features extracted. Whenever
ascertain and justify the scales on the set patches, hence an input image come, face detection is performed

46
 IJCSN - International Journal of Computer Science and Network, Volume 9, Issue 2, April 2020
ISSN (Online) : 2277-5420
www.IJCSN.org
Impact Factor: 1.5

alongside the corresponding extraction, after which every

stored features of those images are compared.

There are mainly two applications of FR: Identification

and Verification. Face identification [1] means when
there is a face image input, we expect the system to
respond by detecting and identify who the individual is;
while in face verification, on the basis of the
computational work done on the input image in the
identification, we want the system to authenticate the
identification done whether true or false. That is, if all
pattern matching is accurate with our desired output.
Every time a person is tagged in an image, the software
application stores individual's facial qualities data.
Immediately, adequate information has to been gathered,
the data is utilized to distinguish a similar face in various
photos, and hitherto recommend labeling those photos
with that person's name.

2 Face Recognition Methodology

Fig. 1 Testing Phase of LDA
2.1 Principal Component Analysis (PCA)
The above figure illustrates the procedural steps to
Human face recognition was first done by Turk and successfully have a linear discriminant analysis on
Pentland and reconstruction of human faces was done by training set. LDA is used on the training set to extract
Kirby and Sirovich. PCA can also be called Karhunen- feature which is use in comparison with the testing set.
Loeve [2] method. It is one of the popular methods for Hence, a Euclidean distance helps to classify or make a
feature selection and dimension reduction. It defines a precise decision on these data set based on the
feature space which minimizes the dimensions of the comparison made between the training set and the testing
original data space and hence, used for recognition. set.
Major problems involve in PCA method is a poor
discriminating power present in the class and large A method introduced, Gabor filter, is used to filter
computation. This limitation was overcome by Linear frontal face images and PCA is used to reduce the
Discriminant Analysis (LDA). The PCA is among the dimension of filtered feature vectors and then LDA [8] is
popular multivariate statistical techniques [3] which has used for feature extraction. The performances of
been globally used in areas of pattern recognition and appearance based statistical methods such as PCA, LDA
signal processing. Numerous methods of PCA used for and ICA are tested and compared for the recognition of
face recognition ranges from one dimensional PCA to colored faces images, hereby resulting that PCA is better
two-dimensional PCA. than LDA and ICA under different illumination
variations but LDA is better than ICA. LDA is more
2.2 Linear Discriminant Analysis (LDA) sensitive than PCA and ICA on partial occlusions while
PCA is less sensitive to partial occlusions compared to
It is the most used algorithms for feature selection in LDA and ICA. PCA is used as a dimension reduction
appearance-based methods. It is used in modeling technique and for modeling expression deformations.
different classes of data [10][12]. Although, many LDA
based face recognition system first used PCA to 2.3 Independent Component Analysis (ICA)
minimizes dimensions and then LDA is used to
maximize the discriminating power of feature selection. It is quite related to PCA. It is a powerful algorithm
It has small sample size problem in which dataset capable of finding the underlying factors or sources
selected should have larger samples per class for good when the above methods fail. It does define a generative
discriminating features extraction. Therefore, a poor model [2] to observe multivariate data, which is
extraction of discriminating features is achieved if LDA categorically given as a large database of samples which
is implemented directly. will give better results as compared with existing

47
 IJCSN - International Journal of Computer Science and Network, Volume 9, Issue 2, April 2020
ISSN (Online) : 2277-5420
www.IJCSN.org
Impact Factor: 1.5

systems. The difference between ICA and other methods 2.5 Locality Preserving Projections (LPP)
is that, it looks for component that are both statistically
independent and non-gaussian. The ICA boils down to This is a method proposed for unsupervised linear
getting a linear representation in which the components dimensionality reduction. Every local structure of the
are statistically independent. The comparison of face data is preserved by the LPP. Unlike PCA, which only
recognition using PCA and ICA on database with work towards preserving the global structure of the data,
different classifiers were discussed [2] and found that the it does helps to preserve local structure of face image
ICA is of better recognition rate as compared with PCA space which can be usually more significant than the
with statistically independent basis images and also with global structure. This shows the superiority [6][10] of
statistically independent coefficients. Every image in LPP against PCA. Locality Preserving Projections (LPP)
ICA is transformed to a vector before calculating the solves a variational problem that preserves the
independent components. It helps to reduce face neighborhood structure of the data set. It represents
recognition error and dimensions of recognition subspace linear approximation of the nonlinear Laplacian
becomes smaller. This approach relied on the eigenmaps. If a high-dimensional data lies within a low
performances of a strongly discriminating optical dimension manifold embedded in a data space, LPP
correlation method along with the robustness of the ICA works to approximate the eigenfunctions of the Laplace
model. Independent component analysis (ICA) model Beltrami operator of the manifold. It is obtained from the
had sparked interest in searching for a linear nonlinear Laplacian [5] Eigenmap. LPP is unsupervised
transformation to express a set of random variables as and a linear transformation performing method.
linear combinations of statistically independent source
variables. It does provide a more powerful data 3 Factors Affecting Face Recognition
representation than PCA as its goal was that of providing
an independent rather than uncorrelated image Many factors can affect the performance of face
decomposition and representation. This algorithm recognition system. Accuracy of any recognition system
computes the principal components of a sequence of is the level at which it can accurately extract features
image vectors incrementally without estimating the from an image. The facial patches need to be extracted
covariance matrix and at the same time transform these with a minimized pattern matching error. Some of the
principal components to the independent directions that areas to focus on are: the facial expression, the
maximize the non-Gaussian of the source. illumination or lighting in the image, the image
background, occlusion present in the image and the
2.4 Support Vector Machine (SVM) image resolution.

Following the success of SVM in signal processing 3.1 Facial Expression

pattern recognition and classification applications, we
choose to use SVM via the Lib SVM Library. SVM is The way and manner at which human can express
majorly use to tackle the FR problem [9][12]. There are a themselves is by their face. Human face can express
number of variations of SVM for handling large-scale sentiment or tempers. It has the power of expressive
classification problems, like LIBLINEAR and ALM- ability and versatile means of affecting human state of
SVM. An SVM classifier with RBF kernel and mind. More or else, it is a form of communication cue
parameters optimized by cross-validation is trained for and emotions in our daily social interaction. There are
each group of training data. Support Vector Machines majorly six basic emotions which can be expressed by
(SVM) are one of the most useful techniques in humans with their faces; happy, surprise, fear, anger, sad,
classification problems in today’s world, an example is disgust. Over the years, it has been proving that the kind
face recognition. Moreover, when there is a situation of of emotion expressed has a major effect in the accuracy
feature vectors defining samples having missing entries, of face recognition system.
SVM cannot be applied. A classification algorithm [8]
that has successfully been used in this framework is the 3.2 Illumination
all-known Support Vector Machines (SVM), which can
be applied to the original appearance space or a subspace
Illumination is one of the factors affecting face
which is obtained after applying a feature extraction
recognition and a major problem faced by many
method. The advantage of SVM classifier over
researchers [11]. It is a problem to solve whether it is an
traditional neural network is that SVMs can achieve
indoor or outdoor pattern matching. Common problem
better generalization performance.
faced in this regard is variation in lightening (contrast),

48
 IJCSN - International Journal of Computer Science and Network, Volume 9, Issue 2, April 2020
ISSN (Online) : 2277-5420
www.IJCSN.org
Impact Factor: 1.5

that is, presence of different lighting conditions when the practical application of LDA directly is not advised. At
image is taken. Hence, it poses a rise to false acceptance some point, it can achieve a worst face recognition rates
rate of face recognition. amidst all other methods. Most of the times, usage of
PCA+SVM and SVM+LDA achieves a maximum face
3.3 Image Background recognition rate than direct applications of PCA and
LDA methods. Combining LDA with SVM produces a
The environment background at which an image is taken better and consistent output than LDA. In all, the future
is also a major factor to be considered when achieving of face recognition is here as an ongoing process to
high face recognition accuracy. It can reduce face implement face recognition in subways and
acknowledgement calculation during face detection. The transportation as a whole. Also, implementing it to
environment surrounding an individual in an image adds banking sectors for ease finance transactions and use of
up to the way the face recognition system actualizes it ATM machines.
functions.
References
3.4 Occlusion
[1] A. Bansal, “A study of Factors Affecting Face
Recognition”, International Journal of Advanced in
This is one of the difficult problems affecting the face Management, Technology and Engineering Sciences,
recognition system. It is a situation whereby the face Volume 8, Issue II, FEB/2018, ISSN NO: 2249-7455.
recognition system fails to detect or identity some part of [2] V. H. Mankar, “A Review Paper on Face Recognition
the human face. Many researchers have proposed Techniques” ISSN: 2278-1323, OCT/2012.
different solutions to train the face recognition system to [3] G. S. Mandal’s, IEEE Student Member, “Evaluation
detect or observe every part of human face. Some of of Feature Extraction Techniques using Neural
these are like; dividing the face image into set of local Network as a Classifier: A Comparative Review for
regions to sophisticated statistical methods, face Recognition” 2018 IJSRST | Volume 4 | Issue 2 |
Print ISSN: 2395-6011 | Online ISSN: 2395-602X.
reconstruction of test image and making comparison.
[4] Rajesh A, “Gender Recognition from Face Image
Based on Textural Analysis and Machine Learning
3.5 Resolutions Approach”, 2018 IJSRSET | Volume 4 | Issue 1 |
Print ISSN: 2395-1990 | Online ISSN: 2394-4099.
Image resolutions has been observed to affect most real- [5] S. Arya, A. Agrawal, “Face Recognition with Partial
world applications of face images. It is noticed that these Face Recognition and Convolutional Neural
applications fail to attain good accuracy level when an Network”, International Journal of Advanced
Research in Computer Engineering & Technology
image size is reduced. Visual surveillance, smart
(IJARCET) Volume 7, Issue 1, January 2018, ISSN:
meetings, and video conferencing are some of the 2278 – 1323.
examples of these real time applications. In this regard, [6] Q. Hua, Y. Liu, L. Sun, J. F. Miazonzama, “Face
solution like novel descriptor known as Weber Local Recognition using Locality Preserving Projection on
Binary Image Cosine Transform [11] (WLBI-CT) is Wavelet Subband and Artificial Neural Network”,
introduced to cub different resolutions associated with International Journal of Computer Trends and
face images. Technology (IJCTT) – Volume 55 Number 1 January
2018.
[7] S. Shukla, A. Helonde, S. Raut, S. Salode, J. Zade,
4 Conclusions “Random Keypad and Face Recognition
Authentication Mechanism”, IRJET, e-ISSN: 2395-
Face recognition is a part of machine learning and 0056, Volume: 05 Issue: 03 | Mar-2018, p-ISSN:
pattern recognition capable of identifying a face either 2395-0072.
from a digital image or video stream. It is a fast-growing [8] S. K. Bhattacharyya, K. Rahul, “Face Recognition by
technology development in the areas of security, Linear Discriminant Analysis”, International Journal
biometric, authentication applications and access control of Communication Network Security, ISSN: 2231 –
due to its uniqueness. It holds lot of potential because it 1882, Volume-2, Issue-2, 2013.
[8] O. Faruqe, A. M. Hasan, “Face Recognition Using
is a highly effective biometric technology. This paper has
PCA and SVM”, Sep/2009
works towards an overview of face recognition, its DOI:10.1109/ICASID.2009.5276938, Source: IEEE
methodologies and common adverse effect on face Xplore.
recognition. In summary, LDA always achieve optimum [9] J. Shermina, “Application of Locality Preserving
face recognition rate, although it can be quite unstable Projections in Face Recognition”, (IJACSA)
due to high sensitivity to methodological settings. Hence,

49
 IJCSN - International Journal of Computer Science and Network, Volume 9, Issue 2, April 2020
ISSN (Online) : 2277-5420
www.IJCSN.org
Impact Factor: 1.5

International Journal of Advanced Computer Science He received his Bachelor of Technology degree in Pure and
and Applications, Vol. 1, No. 3, Sep/ 2010. Applied Mathematics in 2015 from FUTMINNA, Nigeria. His
[10] S. A. Khan, “Factors affecting the recognition research interests include data science, machine learning and
accuracy of facial expressions”, Volume 1 Issue 4 – face recognition.
2017.
[11] J. Mazanec, M. Meliˇsek, M. Oravec, J. Aman Dureja is pursuing his PhD from GGSIPU, New Delhi. He
Pavloviˇcov´a “Support Vector Machines, Pca And received his Master of Technology degree from MDU University
in year 2010 and Bachelor of Technology from Bhiwani Institute
Lda In Face Recognition”, Journal of ELECTRICAL
of Technology and Sciences in year 2007. Currently working as
ENGINEERING, VOL. 59, NO. 4, 2008, 203–209. Assistant Professor in Department of Computer Science and
[12] D. N. Parmar, B. B. Mehta, “Face Recognition Engineering, PDM University since 2010. He has published more
Methods & Applications”, Divyarajsinh N Parmar et than 20 research papers in reputed International Journals
al, Inc. Computer Technology & Applications 4 including Scopus Indexed and conferences including IEEE. His
(1),84-86, ISSN:2229-6093, March 2014. main research work focus on machine learning and deep
learning.
Authors -

Michael Farayola is pursuing his Master of Technology degree

in Computer Science and Engineering from PDM University,
India.

50