Page 1 of 8
Archives of Business Research – Vol. 9, No.1
Publication Date: January 25, 2021
DOI: 10.14738/abr.91.9612.
Lu, C., Ren, Y., & Han, L. (2021). Face and Ethnical Group Recognition with Images of Different Resolutions. Archives of Business
Research, 9(1). 140-147.
Face and Ethnical Group Recognition with Images of Different
Resolutions
Chong Lu
School of information management,
Xinjiang University of Finance and Economics, Urumqi 830012, China
Yan Ren
School of information management,
Xinjiang University of Finance and Economics, Urumqi 830012, China
Liying Han
School of information management,
Xinjiang University of Finance and Economics, Urumqi 830012, China
ABSTRACT
In this paper, a dataset for Xinjiang minority ethnical groups is
introduced, and implementation of two dimensional Linear
Discriminant Analysis (2DLDA) and two-dimensional Partial Least
Squares (2DPLS) is investigated. Two important topics for face
recognition and the ethnicity recognition are investigated for
database with different image resolutions. Experiments show that
2DLDA performances better than 2DPLS on our face database.
Key Words: face recognition, ethnical group recognition, different
resolution image
INTRODUCTION
Face recognition has been a very popular research topic in computer vision community. A
complete face recognition system includes six steps, data capture [1,2], Image preprocessing [3],
face detection [4], face feature extraction [5], face database comparison [6] and face recognition
[7,8]. The existing facial features include three categories, holistic feature description such as
PCA, LDA, ICA[9,10,11],etc. local feature description such as Gabor, LBP[12,13],etc. and some
fusion features which are obtained by combining holistic feature and local feature Holistic
features[14]. The criterion function of two-dimension LDA (2DLDA) and two-dimension PLS
(2DPLS) have many advantages that can be established with original image covariance matrix
directly instead of undergoing reshaping them to vector, and due to the importance of face
recognition, ethnicity recognition is attracting more attention with broad applications [15,16,17].
In this paper, we will consider both problems with our new dataset created in our lab.
The remainder of this paper is organized as follows: after reviewing existing techniques 2DPLA
and 2DPLS in Section 2, we briefly describe our face database in Section 3, and then discuss
Page 2 of 8
141
Archives of Business Research (ABR) Vol 9, Issue 1, January-2021
experimental results in Section 4. In Section 5 experimental evaluations of ethnic identification
has done by 2DLDA algorithm. Last but most importantly, we conclude in Section 6.
2DLDA AND 2DPLS
In this section, we will briefly outline 2DLDA and 2DPLS in order to present the contribution of
this paper clearly.
2DLDA
Let {A$}$&'
( be training images set, where A$ denotes ith an m × n training sample. N is the total
number of training samples, containing C classes, and the ith class C$ has n$ samples, with
∑ n$
0
$&' = N. 2DLDA method tries to find a projected vector x that maximize the following
objective function [17].
J(x) = x6S8x
x6S9x (1)
Where T denotes transpose, S8 and S9 are the between-class scatter matrix and the within-class
scatter matrix, respectively defined as follows.
S8 = <N$(A=$
0
$&'
− A=)6(A=$ − A=) (2)
S9 = <<(A@ − AA = )6(A@
@∈0C
− A$)
0
$&'
(3)
In witch A= and A=$ represent the global and the ith class means defined as follows.
A= = 1
N<<A@
@∈0C
0
$&'
(4)
AA = = 1
n$
<A@ (5)
@∈0C
2DLDA aims to find the optimal projection direction x in order to maximize J(x).
2DPLS
M.L. Yang et al. [18] proposed 2DPLS algorithm for face recognition and we briey describe this
technique here. Let {A$}$&'
( be training images set, where A$ denotes ith an m × n training sample.
N is the total number of training samples, containing C classes, and the ith class C$ has n$ samples,
with ∑ n$
0
$&' = N. Thus, we can obtain the mean matrix of samples matrix.
A= = 1
N<<A@
@∈0C
0
$&'
6
To finish image recognition task, sample images can be considered as a variable set in 2DPLS,
called sample matrix. Another variable set is class membership matrix, which represents the
relationship between samples and classes. It is similar to the definition in traditional CCA and in
PLS, the class membership matrix can be coded as follow [19]
Page 3 of 8
URL: http://dx.doi.org/10.14738/abr.91.9612. 142
Lu, C., Ren, Y., & Han, L. (2021). Face and Ethnical Group Recognition with Images of Different Resolutions.. Archives of Business Research, 9(1).
140-147.
Z = J
P' 0M
0' PM
... 00
... 00
⋮ ...
0' 0M
⋱ ⋮
... P0
Q
(R×0)×(S×()
(7)
Where P$ means there are n$ samples in the ith class, but each sample here is corresponding to a
matrix QR×S as large as the size of sample image. So the matrix P$ can be denoted as P$ =
[Q,Q, ... Q]R×(S×SY),
i = 1 ... . C. For obtaining the mean of class membership matrix in the sense of
two dimensional sample representation, the matrix Y is rewritten as Y = [y',', ... , y',SY], then the
covariance matrices of A and Y are denoted as
G^ = 1
N<<(A$,@ − A=)
@∈0C
(A$,@ − A=)6 (8)
0
$&'
G` = 1
N<<(y$,@ − Y=)
@∈0C
(y$,@ − Y=)6
0
$&'
(9)
G^` = G`^
6 = '
( ∑ ∑@∈0 (A$,@ − A=) C (y$,@ − Y=) 0 6
$&' (10) respectively.
FACE DATABASE CREATION
In this section, we present our face database creation. We captured face images for six different
minority nationality groups which are Kazak, Uygur, Kirgiz, Mongolian, Sibbo and Hui with
different pixel resolution cameras, i.e., Hasu, Nikon D90, Nikon S60 and Kinect. Face Database
contains eight different images of each of 48 distinct subjects. For some subjects, the images were
different with varying the lighting, facial expressions and facial details with glasses. All the images
were taken against a light homogeneous background with the subjects in an upright, frontal
position. they are shown below.
The size of Each Image in the Original Face Database captured with Hasu camera is about 41M
with12840 × 10550 pixel, Nikon D90 is about 5.4M with4280 × 2580 pixel, Nikon S60 is about
3.1M with30100 × 2100 pixel and Kinect is about 8K with200 × 160 pixel.