International Journal of Video&Image Processing and Network Security IJVIPNS-IJENS Vol:13 No:01 15
138601-7575-IJVIPNS-IJENS © February 2013 IJENS
I J E N S
Abstract -- This paper deals with the analytical-statistical
method for the face recognition. Although the bases of the face
recognition are known by researchers worldwide, the statistical
tests of data obtained by measuring chosen anthropometrical
points can be found in several articles. Our aim is to show how
the data act during the various emotions of one face, which will be
helpful for deeper knowledge of how the face behaves. The focus
is also aimed at study of aging. The data were tested in the
STATISTICA software using e.g. the ANOVA, the Shapiro-Wilk
test, the k-Means clustering and the t-test. Acquired results reflect
the difficulty of describing the face and the applicability of
combination of different recognition methods (e.g. methods based
on neural networks, recognition of facial contours, distribution of
the gray scale in the image, deformation models) to get the best
results in the verification/identification of a human.
Index Term-- pattern recognition, face recognition, statistical
methods, biometrics
I. INT RODUCT ION
The face recognition is one of the most utilized technologies
for protection of assets. The roots of this technique lie in the
anthropology, respectively in the judicial anthropometry by A.
Bertillon. As can be apparent this method from the 19
th
century
evolved into criminalistic portrait identification (also called
photo-anthropometry). This attitude was utilized mainly in the
1960s with the advancement in analytical-statistical and
graphical methods for identification of a person. This paper is
focused on analytical-statistical method, which was mainly
used in the 1970s, and is a representative of one of the simplest
methods of today’s techniques . The obtained information is
characterized by the space and linear structure of the face,
which characterized the individuality. The scientific analysis
proved that only 12 basic anthropometrical points are quite
This work was supported by the Internal Grant Agency at Tomas Bata
University in Zlín, project No. IGA/FAI/2013/001, and by the European
Regional Development Fund under the Project CEBIA-Tech No.
CZ.1.05/2.1.00/03.0089.
K. Sulovská is with Tomas Bata University in Zlín, Faculty of Applied
Informatics, Department of Security Engineering, nám. T. G. Masaryka 5555,
76001 Zlín, the Czech Republic (phone: +420 57 603 5133; fax: fax: +420
57 603 2717; e-mail: sulovska@fai.utb.cz).
S. Bělašková is with Tomas Bata University in Zlín, Faculty of Applied
Informatics, Department of Mathematics, nám. T. G. Masaryka 5555, 76001
Zlín, the Czech Republic (e-mail: belaskova@fai.utb.cz).
M. Adámek is with Tomas Bata University in Zlín, Faculty of Applied
Informatics, Department of Security Engineering, nám. T. G. Masaryka 5555,
76001 Zlín, the Czech Republic (e-mail: adamek@fai.utb.cz).
adequate for this characterization (Fig. 1) [14]. As the
anthropometry is a three-dimensional measurement, the photo-
anthropometry deals with the two-dimensional photographs,
which can bring serious difficulties [9], [10], [11], [12]. One of
the limiting factors is the quality and the angulation of the
photograph [16].
Fig. 1. Extended number of anthropometrical points (marked order meets [14])
When focusing on statistical approaches to the face
recognition, especially focused on emotional changes in faces,
only the machine recognition is disposable [1], [2], [5], [6],
[7], [13], [18], [19]. This type of recognition shows the
utilization of the bases made by the photo-anthropometry in
the way enriched by the novel computational methods [8],
which raises the accuracy of verification/identification.
Although 12 points are enough, we may add extra points
adapted to the image. Such points need to be clearly visible in
the image. Features like hair features or hairlines, which can be
easily modified, should be omitted as they are unreliable and
unpredictable for computerized methods. Features like face
shape have higher rates of interobserver agreement, and
pronounced ear projection is said to be the best discriminators
[3], [4], [15], [16]. The resolution of processed image for the
use in police-court (forensic) applications is set to be
standardly 500 dpi (e.g. used by FBI also for fingerprint
Study of Face Recognition Using Statistical
Analysis
Kateřina Sulovská, Silvie Bělašková, Milan Adámek