Journal Pre-proof

Automatic Face Recognition of Target Images Based on Deep

Learning Algorithms

Junxi Guo , Yuzhuo Fu , Ting Liu

PII: S0141-9331(23)00180-1
DOI: https://doi.org/10.1016/j.micpro.2023.104936
Reference: MICPRO 104936

To appear in: Microprocessors and Microsystems

Received date: 23 December 2020

Revised date: 19 January 2021
Accepted date: 21 September 2023

Please cite this article as: Junxi Guo , Yuzhuo Fu , Ting Liu , Automatic Face Recognition of Tar-
get Images Based on Deep Learning Algorithms, Microprocessors and Microsystems (2023), doi:
https://doi.org/10.1016/j.micpro.2023.104936

This is a PDF file of an article that has undergone enhancements after acceptance, such as the addition
of a cover page and metadata, and formatting for readability, but it is not yet the definitive version of
record. This version will undergo additional copyediting, typesetting and review before it is published
in its final form, but we are providing this version to give early visibility of the article. Please note that,
during the production process, errors may be discovered which could affect the content, and all legal
disclaimers that apply to the journal pertain.

© 2023 Published by Elsevier B.V.

 Automatic Face Recognition of Target Images Based on Deep

Learning Algorithms

Junxi Guo*, Yuzhuo Fu, Ting Liu

School of Electronic Information and Electrical Engineering, Shanghai Jiaotong University,
Shanghai 200240, China
*
Email of Corresponding Author: guojunxi1435@163.com
Email of Yuzhuo Fu: yzfu1234@163.com
Email of Ting Liu: louisaliu123@163.com

Abstract: The current research results of face recognition have poor recognition accuracy,
time-consuming and high information loss rate. A method of automatic face recognition based on
depth learning algorithm is proposed. The three primary RGB images are transformed into color
saturated HSV images, and the dynamic range is extended by changing the non-linearity according
to the conversion results. Based on the extended situation, the membership function is used to map
the target image to the blurred plane, and the contrast enhancement of the image is completed. The
enhanced image is substituted into the image segmentation, and the image eigenvalues are
extracted preliminarily. The seed of image segmentation is generated by the eigenvalues. The
growing seed regions are merged by the color distance and the texture distance to achieve the
target image segmentation. Based on image segmentation, LBP operator is used to extract the
local texture features of face twice, and then a deep convolution network model is constructed.
The shared weights of convolution network model and pooling down-sampling technique are used
to reduce the complexity of the model. The top layer of the model forms a feature classification
surface of face image, fuses the constraint conditions, and obtains the trained face. The deep
convolution network model is used to extract features from face images and complete face
recognition. The experimental results show that the method has good accuracy, high efficiency
and low information loss rate.
Keywords: Deep Learning Algorithm; Target Image; Face; Automatic Recognition; Convolution
Network.

1 INTRODUCTION
At present, face recognition is one of the hottest research topics in the fields of pattern
recognition, machine vision and biometric recognition. It has been widely used in video
surveillance, human-computer interaction, security access control and other fields. Different from
traditional identity authentication technology, biometric recognition technology is more
convenient and reliable, and is not easy to be stolen and lost. With the development of science and
technology, people have an urgent need for fast, reliable and convenient biometric authentication
technology [1-3]. Biometric technology refers to the authentication technology of identifying a
person or a biological identity based on the physical or behavioral characteristics of a person or an
organism. There are many characteristics of organisms, including intrinsic features such as iris,
fingerprint, face, and dynamic behavioral features such as gait, writing and voice. Although face
recognition has more prominent features, face features are easily affected by illumination, posture,
occlusion and facial expression, which may affect the recognition effect [4]. Because face
recognition is of great significance, many experts and scholars in China and abroad have begun to
study this topic.
Tang et al. proposed a fast L2-norm face recognition method based on sparse representation
[5]. In the process, the structure of dictionary is improved by extracting fusion features and
reducing the size of dictionary, and the sparsity of L2-norm is enhanced, which greatly improves
the running speed of the algorithm on the premise of guaranteeing the recognition performance.
The experimental results show that this method can significantly reduce the complexity of the
algorithm, but it has the problem of poor recognition accuracy. Li et al. proposed a face
recognition method based on Gabor wavelet and cross-covariance dimensionality reduction [6].
Firstly, Gabor features are extracted from face images, and then dimensionality reduction and
feature extraction are performed using weighted cross-covariance matrix. Finally, nearest neighbor
classifier is used to classify the processed feature images. Experiments on ORL and AR databases
show that this method has a certain recognition efficiency, but it also has the problem of poor
recognition accuracy. Yuan et al. proposed a three-dimensional face recognition method based on
the neighborhood structure of singular points [7]. Firstly, the target region is segmented on the
face texture image, the feature sub-region is divided, the two-dimensional singular points and the
neighborhood structure of singular points are extracted. Then, the three-dimensional singular
points and their neighborhood structures are marked on the geometric information of the face
space, and the three-dimensional information of the neighborhood structure of singular points is
used to represent the face features. Finally, the singular points are used to represent the face
features. Neighborhood structure nearest point method is used to recognize face identity
information. The experimental results show that the proposed method has high accuracy, but it
takes a long time. Zhao et al. proposed a multi-pose face recognition method based on single view
[8-11]. Firstly, face alignment and normalization are carried out by multi-view active appearances
model; Secondly, the relationship between face correction and side face is searched based on
linear regression algorithm, and the face image is obtained by attitude correction based on this
relationship; Finally, the parameters of support vector machine are screened by genetic algorithm,
and the corrected person is treated by support vector machine. The experimental results on
CAS-PEAL-R1 face database show that the method is simple in operation, but has the problem of
high information loss rate.
In view of the imperfect performance of current research results on face recognition, an
automatic face recognition method based on depth learning algorithm is proposed. The detailed
process is as follows:
(1) Enhance the target image to improve the efficiency of face recognition and reduce the
energy consumption of recognition.
(2) Segmenting the enhanced target image to improve the accuracy of face recognition.
(3) The feature of face image is extracted by deep learning algorithm, and the automatic face
recognition of target image is realized.
(4) The proposed method is verified by experiments and discussions. The results show that
the proposed method has strong robustness and practicability.
(5) Summarize the full text and plan the next research direction.

2 ALGORITHM DEFINITIONS
2.1 Target image enhancement
In order to reduce energy consumption and time-consuming of face recognition, it is
necessary to enhance the target image. At present, most of the images taken by camera equipment
are RGB images. RGB images are obtained by weighting the three color components of red (R),
green (G) and blue (B), but they are vulnerable to the influence of illumination changes. Moreover,
there is a high correlation between the three primary color components of RGB, which can change
the color information of a channel. Information often affects the information of other channels, so
the direct non-linear processing of each color component of the image will cause the distortion of
the image color. HSV space is not only more suitable for describing human color sensation than
RGB space, but also effectively separates chroma, saturation and brightness, which brings great
convenience to subsequent true color image enhancement [12].
In the process of illumination compensation, the RGB image is converted into HSV space,
and the brightness component is enhanced, while the hue and saturation are kept unchanged.
Finally, the generated brightness component is inversely transformed with the hue and saturation
components to generate a new image. The transformation expression from RGB space to HSV
space is as follows:


0, S =0

60  G -B , MAX  R, G , B   R & G  B
 S V

 2  B  R
H = 60  , MAX  R, G , B   G (1)
 S V
 4  R  B
60  , MAX  R, G , B   B
 S V
 6  G  B 
60  , MAX  R, G , B   R & G  B
 S V

Where,

MAX  R, G , B   MIN  R, G , B 
S (2)
MAX  R, G , B 

V  MAX  R, G , B  (3)

In the formula, R , G and B are the values of normalized RGB spaces respectively. The

value range of H component is  0, 360  and that of S and V component are  0,1 and

 0,1 , respectively.
Suppose i  H / 60 , f  H / 60  i , where i represents the divisor divided by 60 and f

represents the remainder. Suppose P  V 1  S  , Q  V 1  Sf  , T  V 1  S 1  f  , the

transformation expression from HSV space to RGB space is as follows:

  R  V , G  T , B  P,  i  0 

 R  Q, G  V , B  P, i  1
 R  P, G  Q , B  T , i  2
   (4)

 R  P , G  Q , B  V ,  i  3
 R  T , G  P, B  V ,  i  4 

 R  V , G  P, B  Q,  i  5

After the above image space transformation, the image is transformed non-linearly. Low
illumination image has the characteristics of narrow gray range, high spatial correlation of
adjacent pixels, and no obvious gray change. This makes the object, background, details, noise
and other information in the image included in a narrow gray range. Nonlinear transformation is a
smooth mapping curve, which makes the gray level change of the processed image smoother.
Considering that there is an approximate logarithmic link in the process of receiving image signals
from the human eye to form an image in the brain, the commonly used non-linear transformation
is logarithmic transformation [13, 14].
Logarithmic transformation refers to the logarithmic relationship between the gray value of
the output image and the pixels of the input image, and the gray value relationship between the
output image GG g  x , y  and the input image f  x, y  is as follows:

g  x, y   log  f  x, y  (5)

This transform can compress the contrast of the higher gray area in the original image, and at
the same time expand the lower gray value of the image. Usually, in order to make the dynamic
range of transformation more flexible, modify the change speed or starting position of the curve,
some adjustable parameters are usually added to make it become:

a  ln  f  x, y   1
g  x, y   (6)
 b ln c 
In the formula, a , b and c are adjustable parameters, which can be changed and adjusted

artificially. f  x, y   1 is to avoid logarithm with 0 and ensure that the molecular part is greater

than or equal to 0. When f  x, y   0 , ln  f  x, y   1  0 , and y  0 , a is the intercept on the

y axis, which determines the functional transformation relationship of the initial position of the
transformation function, and the change rate of the transformation function is determined by b
and c parameters. Logarithmic functions are generally applicable to over-dark images and are
used to extend low-gray areas. The logarithmic curve is shown in Fig. 1.
 g

Fig. 1 Logarithmic curve

The logarithmic function set here is:

f   lg Vd  1 (7)

d is an adjustable parameter and 1 is added to avoid logarithm of 0 when V equals 0.

After the non-linear transformation, the brightness of the image can be greatly improved, but
the contrast of the image is reduced, so that the image does not achieve the desired visual effect.
By introducing the fuzzy theory, the image is transformed from the spatial domain to the fuzzy
domain using the membership function, and the image is enhanced on the fuzzy feature plane to
increase the contrast of images.
After the target image preprocessing, the image is enhanced by using the fuzzy theory. The
whole process is divided into three parts: transformation from spatial domain to fuzzy domain,
image enhancement in fuzzy domain and inverse transformation of fuzzy image. The detailed
process is as follows:
Using sinusoidal membership function, the image is mapped from spatial domain to fuzzy
domain, and the fuzzy feature plane is obtained. The sinusoidal membership function is defined as
follows:

  f  i, j   f min   
k

uij  sin    (8)

  f max  f min 2  

Among them, UU uij is the membership degree of the gray scale xij of the pixel  i , j 

relative to f max , f max is the maximum gray level of the image to be processed, and the

corresponding f min is the minimum gray level of the image to be processed. k is an adjustable

parameter. By adjusting k , the value of uij can be changed, and different blurred feature planes

can be generated for different images, which can meet the enhancement requirements of different
images. The definition of k value is as follows:

mean  f 
k (9)
 f max  f min 
 Among them, mean  f  represents the average gray value of the image, which makes the

k value intrinsically correlated with the image. The k value of different brightness images will
change accordingly, which increases the flexibility of the algorithm and adapts to the needs of
different brightness image enhancement.

The membership function uij is transformed by formula (8) and formula (9). Finally, a new

fuzzy feature plane is obtained. A new fuzzy enhancement operator is defined. As the number of
iterations increases, the image contrast increases.
1  2 1  u 2 ,0  u  T
 ij ij
uij   (10)
2  uij  , T  uij  1
2


T , as the critical point of image enhancement, takes different values of T for different
brightness images, and the value of T is the average value of image gray level.
For the new blurred feature plane, the image can be mapped from the blurred domain to the
gray space by an inverse transformation. Finally, the enhanced image is obtained.

1/ k

f   i, j   G 1  uij    umin   umax  umin  arcsin uij   (11)
 2

G1 is the inverse transformation of the image and f   i , j  is the enhanced image.

2.2 Target image segmentation

In order to improve the accuracy of face recognition, the enhanced target image is segmented.
For each pixel in f   i , j  in the same size area, its frequency coefficient is calculated by DCT

conversion. Among them, the size n of the region is preset at 8. In practice, it can be adjusted
according to the characteristics of texture. If the texture of the image is coarse, the n value will
be larger, and then it can contain the correct features. If the texture is fine, a smaller n value will
be used, which can speed up the execution time of the algorithm. The conversion process of DCT
is as follows:

1 n 1 n 1
  2 x  1 i    2 y  1 j 
Di , j   Ci C j   h  x, y  f  i, j   cos   cos   (12)
2n x 0 y 0  2n   2n 

Among them, n represents the size of the region, h  x , y  represents the intensity

coefficient at the  x, y  position in the spatial domain, Di , j  represents the frequency

 
coefficient, Ci  takes 1/ 2,0 , and C  j  takes 1/ 2,0 .  
In practice, instead of calculating all the frequency coefficients, several representative ones
are selected, in turn D 0,1 , D1,1 and D1,0 . In addition, because the calculation of DCT takes a

lot of time, in fact, the texture features of adjacent points are not very different in theory, so we
only need to calculate the texture features of one of the pixels, and then set eight adjacent pixels to
the same eigenvalue, that is, only one DCT conversion is performed in each 3* 3 region. Nine
pixels in this region have the same texture features.
When all the eigenvalues of each pixel are ready, the next step is to generate seeds. To
become a seed, the following two conditions must be met:
(1) The seed must have a high texture similarity with the points around it. The difference

between the maximum and minimum coefficients must be less than a critical value t1 by testing

the eight directions of the central pixels in the region (vertical, horizontal and two diagonals). As
shown in Formula (13):
D  D 0,1  t1
 0,1 max min

 D1,1 max  D1,1 min  t1 (13)

D  D1,0  t1
 1,0 max min

Among them, t1 is a positive real number, the bigger the t1 , the easier it is to form seeds. If

all the eight directions satisfy the above formula, the pixel can be used as a candidate seed.
(2) The seed must have a high degree of color similarity with the points around it. The eight
directions of the central pixels in the region are tested, and the similarity is measured by Euclidean

distance. The distance must be less than a critical value t2 . As shown in Formula (14):

Y   Y  
2

L i L j
 t2 (14)

Where t2 is a positive real number. If all the points in the eight directions satisfy the above

formula, the pixel can be used as candidate seed.

A pixel can be classified as seed if it satisfies the above two conditions at the same time. In
the process of seed formation, if there is a connected situation, the connected seeds are regarded as
a seed [15, 16]. In addition, the selection of critical value in condition 1 and condition 2 should be
adjusted according to the need of segmentation. If the critical value is too large, more pixels will
be classified as seeds. Finally, some regions may be merged with other regions because the seeds
are connected together, which cannot be correctly segmented. On the contrary, if the threshold is
too large, some regions will be classified as seeds. If the threshold is too small, seeds are not easy
to generate, and some regions may have no seeds at all, resulting in regions cannot be segmented.
In the process of regional growth, using color eigenvalues as the basis for growth, there are
three main steps:
(1) Calculate the average color eigenvalues of each initial region, and set a label symbol for
each region.
(2) Look for the untreated pixels around these areas, and calculate whether these pixels can
be classified into the adjacent area.
(3) If there are still unclassified pixels, step (2) is continued.
After the completion of step (1), the Euclidean distance is used to calculate the color distance
between each pair of initial regions, and then the minimum value is found, which is set as the
variable min dr .
In step (2), the relative Euclidean distance is used as the criterion of judgment, and the
expression is as follows:

 Y   Y  C   C 
2 2
bL i j   rL i rj
Cij  (15)
Y    CrL i
2 2
bL i


Among them, YbL  and CrL  represent the eigenvalues of the pixels to be processed,
i i

and Y j and C rj represent the average values of the two adjacent regions.

In the existing algorithms, region growth starts from the only pixel with the closest similarity.
When a pixel is classified, it is necessary to find the next pixel with the highest similarity to
process. It seems reasonable on the surface. In fact, this sorting action will take a lot of time to
achieve. The effect is not obvious. A variable critical value is defined as the criterion for the
growth of seed regions as follows:
min dr
t3  A (16)
I
Among them, I is a positive constant and A is initially set to 1. With time and the
emergence of demand, A shows an increasing trend, so the critical value will become larger and
larger. As long as the similarity difference between growing pixels and regions is less than the

critical value t3 at the same time, it can be classified into the most similar adjacent region, and

then the untreated points around it can be added to the line of preparation for growth. If the

similarity difference is greater than the critical value t3 , the point will be suspended and not

classified for the time being. As the region grows toward the boundary, the similarity differences
between the calculated values become larger and larger. When the growing points are in

suspension state, the A value is raised to increase the t3 value. Then the critical values are

checked to see if the points in suspension state meet. If they are satisfied, the growth continues,
and if they are still not satisfied, they continue. Raise the A value until the action of growth
resumes. The growing process lasts until all the pixels are classified, at which time the whole
image has been roughly segmented.
Because in the process of seed production, there may be several seeds in some areas, and
after growth, there will be over-segmentation [17-19]. At this time, the same region should be
combined. In order to improve over-segmentation, two merging conditions are set:
(1) The closer the similarities between regions, the more likely they are to merge.
(2) Areas containing too few pixels should be merged.
In condition (1), the similarity distance between color and texture is used to judge whether
regions can be merged. The similarity distance calculation formulas are formula (17) and formula
(18):
The color distance is:

Y  Y    C  Crj 
2 2

c  Ri , R j  
i j ri
(17)
min  Yi 2  Cri 2 , Y j 2  Crj 2 
 The texture distance is:

D    D 
2 2
i  0,1 D j 0,1 i 1,0  D j 1,0
d R , R   (18)
min  D    D 
i j 2 2

i  0,1 D j 0,1 i 1,0  D j 1,0

Here, two real numbers with positive critical values t4 and t5 are set. When the c  Ri , R j 

between Ri and R j is less than the critical value t4 and d  Ri , R j  is less than the critical

value t5 , the two blocks can be merged. In condition (2), if a region contains fewer than a critical

value of pixels, a neighboring region with the highest color similarity is found for merging.
2.3 Automatic face recognition of target images based on deep learning algorithms
Based on 2.1 and 2.2 target image processing, depth learning algorithm is used to realize
automatic face recognition. In the process, firstly, the LBP operator is used to extract the local
texture feature of face twice, and the LBP local texture feature is used to construct the deep
convolution network model. Then, the shared weights of the convolution network model and the
pooling down-sampling technology are used to reduce the complexity of the model, and the
feature classification surface of face image is formed at the top of the model. On this basis, the
constraints are fused, and the trained deep convolution network model is obtained. The model is
used to extract features from face images and complete face recognition. The specific process is as
follows:
Generally, Local Binary Patterns (LBP) is a method for extracting texture features from
image locally. It has strong classification ability, high computational efficiency, high rotation
invariance and gray invariance, and has been widely used in image recognition, machine vision
and biometric recognition.
In the process of face recognition, local texture features are extracted by using local binary
pattern. The basic principle of LBP is to compare the gray values between the central point pixels
and their neighborhood pixels, and to express the relationship between the central point and its
neighborhood using Boolean type (0 or 1). The specific calculation formula is as follows:

 P 1

 LBPP , R  d  Ri , R j   S  gi  g c   2 c  Ri , R j 
i

 i 0
(19)

S  x   1, x  0
 
 0, x  0

gi represents the gray value of the pixels in the neighborhood with radius R  , g c

represents the gray value of the central pixel, S  gi  g c  represents the Boolean value of the gray

value of the neighborhood pixels compared with the gray value of the central pixel.
To sum up, the calculation of the pixel values of local binary modes shows that the number of
binary modes formed by an LBP operator depends on the number of samples in the neighborhood
set. The binary sequences acquired by the LBP operator are successively first and last. Assuming
that the change of the sequence of LBP operators from 0 to 1 or from 1 to 0 is less than 2 times,
the LBP operator is defined as an equivalent mode. It can be seen that the conversion of binary

pattern from the original form of classical formula (19) to P  P  1  2 can effectively reduce the

dimension of face features and reduce the loss rate of feature information.
Taking the LBP statistical histogram of the whole image as the face feature, the obtained face
feature is blurred and noisy. The idea of block processing is proposed. At the same time, the face
image is divided into blocks. First, the LBP operator is applied locally, and the obtained histogram
is connected to form a new face feature vector [20-22].
The procedure of extracting LBP histogram is as follows:
Firstly, face image is segmented on the basis of 2.2 image segmentation.
Secondly, LBP is extracted from the subgraph after block, and LBP histogram is generated by
statistics.
Finally, several LBP histograms are connected sequentially to form new local texture
features.
According to the process of LBP histogram extraction, LBP operator is used to extract local
texture features of face, which lays a foundation for the realization of face recognition process and
creates conditions for the construction of deep convolution network model.
On the basis of the local binary pattern, the local texture features of face are extracted by the
local binary pattern, and clustered. The results are used as input of each layer of the deep
convolution network, and the parameters of the deep convolution network are trained layer by
layer. Deep neural network is a multi-hidden layer structure of neural network, which is composed
of several convolution layers, pooling layers and full connection layers. The core idea of deep
network is to classify a given data sample based on a certain distance, which is the smallest
distance within a class and the largest distance between classes. In the face recognition process,
the specific steps are as follows:
(1) Select the local texture features of the front K face as the initial clustering center, and
calculating the distance between the local texture features of the remaining faces and the
clustering center, then assigning each texture feature to the nearest classes. Finally, adjusting the
clustering center of the new classes, if the two adjacent clustering centers have not changed, the
clustering center of the new classes will be adjusted, and the clustering is completed.
(2) Input the clustered local texture features into the convolution layer, assuming that they are

represented by xi , then the input xi is convoluted with different convolution kernels Kij of

j  . After each convolution, the output feature map is yi and a formula about yi can be

obtained:
 
yi  max   Ki j  xi  LBPP , R  B  (20)
 i 
In this subformula,  is a convolution operation, and the convolution weights of

convolution kernel Kij and base vector B  are training parameters. In this way, a clustered

facial local texture feature synthesis function is formed.

(3) Then, the clustered local texture features are input into the pooling layer and the spatial
resolution of the convolution layer is reduced by down-sampling, which not only reduces the
number of trained face feature weights, but also reduces the influence of face rotation, posture,
expression and illumination. Because the pooling operation is a 2*2 -size max-pooling operation,
all clustered face texture features are more representative.
(4) Finally, the clustered face texture features are input into the full-connection layer, and the

corresponding output eigenvector YY y  is obtained through the weight matrix ww W , the base

vector B  and the activation function F . Finally, a function formula about y  is formed.

y   F W  xi  B (21)

According to the synthesis function in Formula (21), a network structure diagram combining
deep convolution network and LBP is finally formed, which finally completes the construction of
face recognition model of deep convolution network.
On the basis of constructing the face recognition model of deep convolution network, the
features of face images are classified by combining deep convolution network with LBP. The
process of classification of facial image features is as follows:
In order to realize the classification of face images, two different function models should be

constructed. Firstly, it is assumed that J1 represents the similarity measure function model within

the face feature class, and JJ J 2 represents the similarity measure function model between

different face feature types. Finally, two functions about J1 and J 2 are formed.

1
J1 = y  hW , B M
2
(22)
2 m n
1
y  M 
2
J2  (23)
2 m1 m
Among them, M represents the mean of the sample, M  represents the final output value
of the network, m represents the number of sample sets, n represents the number of

categories belonging to the sample sets, and hW , B represents the membership degree of the test

samples to various types of face types. In order to classify image features accurately, reasonable
constraints need to be established.
Thus, according to the function models in formulas (22) and (23), the following formula (24)
can be used to calculate the i sample mean of the class face samples, and a constraint function
formula (24) on the mean is formed.
n 

h  J W , B 1  J2  (24)
M i 1

n
In Formula (24), in order to make the deep convolution network model more conducive to the
classification of facial image features, it is necessary to further constrain the two energy function

models of J1 and J 2 . Then a new constraint function J  on residual parameters is formed,

such as formula (25):

 J    J1   J 2 (25)

In the formula (25) of constraint function J  ,  and  represent the weight of constraints
within and between classes in constructing network model. The weight values of different face
databases are different, and the specific values need to be determined by experiments. The main
purpose of constraint function is to make the classification effect more obvious. With J 
constraints, using the above formula can make the inter-class spacing of face feature samples
larger and the inter-class spacing smaller, and make the weight parameters of each layer adjust to
the direction more conducive to face classification [23].
In the process of face recognition, the model of deep convolution network function is

obtained. For the function model J1 , which represents the constrained objective function in the

face feature class, the formula for calculating the residual of each output unit in the output layer is
as follows:
J 1
 (26)
Z
Formula (26) is a constraint formula within a class, where Z represents the output of the
last layer.

Formula J 2 for inter-class constraints is formed, and the residual formula for each output

unit in the output layer is:

J 2
 (27)
Z
According to Formula (27), the unit residuals within and between classes of facial image
features can be calculated by using the formula under the face recognition model of deep
convolution network, and the recognition value of facial image features can be calculated.
According to the deep convolution network model, face recognition can be effectively completed.
By optimizing and updating the algorithm of the deep convolution network face recognition model,
an automatic face recognition method is formed based on the local binary pattern algorithm.

3 RESULTS
In order to verify the validity of the automatic face recognition method based on depth
learning algorithm, an experimental platform is built on MATLAB. The experimental data are
from ORL and YALE face databases. There are 600 face images with different illumination
conditions, postures and expressions in ORL databases. In the YALE face database, 300 face
images are selected. The images are also face images with different illumination conditions,
different postures and different expressions. Fig. 2 shows some experimental data.
 (a)

(b)
Fig. 2 Experimental data

The experimental indicators are as follows:

(1) Accuracy of face recognition
(2) Time-consuming face recognition
(3) Information loss rate in recognition process
The experimental results are as follows:

100
Face recognition accuracy /%

90

80

70

60

50

0
0 10 20 30 40 50 60
Number of images to be identified/frame

(a) Accuracy of fast L2-norm face recognition based on sparse representation

 100

Face recognition accuracy /%

90

80

70

60

50

0
0 10 20 30 40 50 60
Number of images to be identified/frame

(b) Accuracy of face recognition based on Gabor wavelet and cross-covariance

dimensionality reduction

100
Face recognition accuracy /%

90

80

70

60

50

0
0 10 20 30 40 50 60
Number of images to be identified/frame

(c) Accuracy of automatic face recognition in target images based on deep learning
algorithms
Fig. 3 Accuracy comparisons of different face recognition methods

Analysis of Fig. 3 shows that compared with current face recognition methods, the automatic
face recognition method based on depth learning algorithm has the highest overall recognition rate.
The proposed method enhances the image before recognizing the face of the target image, which
effectively improves the accuracy of face recognition.
15

13
Face recognition time / μs

11

0
20 30 40 50 60 70 80
Number of images to be identified/frame
 (a) Time-consuming three-dimensional face recognition based on singular point
neighborhood structure

15

13

Face recognition time / μs

11

0
20 30 40 50 60 70 80
Number of images to be identified/frame

(b) Time-consuming multi-pose face recognition based on single view

15

13
Face recognition time / μs

11

0
20 30 40 50 60 70 80
Number of images to be identified/frame

(c) Time-consuming automatic face recognition in target images based on deep learning
algorithms
Fig. 4 Time-consuming comparisons of different face recognition methods

Fig. 4 shows that the automatic face recognition method based on depth learning algorithm
takes about 1 µs to 4.2 µs. Compared with other research results, it has low recognition efficiency
and poor reliability. The proposed method uses the fuzzy theory to enhance the target image,
which effectively reduces the time-consuming of face recognition.

2.5
Face feature information loss

2.0

1.5
rate /%

1.0

0.5

0
40 50 60 70 80 90
Time/s
 (a) Information loss rate in multi-pose face recognition based on single view
2.5

Face feature information loss

2.0

1.5

rate /%
1.0

0.5

0
40 50 60 70 80 90
Time/s

(b) Information loss rate in fast L2-norm face recognition based on sparse representation

2.5
Face feature information loss

2.0

1.5
rate /%

1.0

0.5

0
40 50 60 70 80 90
Time/s

(c) Information loss rate in automatic face recognition of target images based on deep
learning algorithms
Fig. 5 Comparison of information loss rates in different face recognition methods

As can be seen from Fig. 5, the method of automatic face recognition based on depth learning
algorithm has the lowest information loss rate in the process of recognition. When using deep
learning algorithm to recognize face, the binary pattern is converted from original form to

P  P  1  2 , which not only reduces the dimension of face features effectively, but also reduces

the loss rate of feature information.

4 DISCUSSIONS
In this part, the anti-interference ability of the proposed method is discussed, and the
performance of the proposed method is further verified. The discussion platform is consistent with
the experimental platform. The results are as follows:
 1.0

Anti-interference coefficient
0.9

0.8

0.7

0.6

0.5

0.4
10 20 30 40 50 60
Number of experiments

Fig. 6 Anti-jamming of an automatic face recognition method for target image based on
depth learning algorithm

Analysis of Fig. 6 shows that the proposed method is based on image enhancement and
image segmentation to realize automatic face recognition in target image. In the process, the fuzzy
theory is introduced to transform the image from the spatial domain to the fuzzy domain by using
the membership function. The image is enhanced on the fuzzy feature plane, which increases the
contrast of the image, reduces the image noise and improves the anti-interference ability of the
method.

5 CONCLUSIONS
Aiming at the problems existing in current face recognition methods, an automatic face
recognition method based on depth learning algorithm is proposed. Face recognition is
accomplished by image enhancement, image segmentation and face feature extraction of target
image, and the effectiveness of the proposed method is proved by experiments. In the next step,
the deep learning algorithm can be applied to action recognition and target detection, and broaden
the research field.

REFERENCES
[1] Li, H.; Zhang, L.; Huang, B.; Zhou, X. Sequential three-way decision and granulation for
cost-sensitive face recognition. Knowledge-Based Systems. 2016, 91, 241-251.
[2] Dewan, M.A.A.; Granger, E.; Marcialis, G.L.; Sabourin, R.; Roli, F. Adaptive appearance
model tracking for still-to-video face recognition. Pattern Recognition. 2016, 49, 129-151.
[3] Ren, C.X.; Lei, Z.; Dai, D.Q.; Li, S.Z. Enhanced local gradient order features and discriminant
analysis for face recognition. IEEE Transactions on Cybernetics. 2016, 46, 2656-2669.
[4] Xu, Y.; Li, Z.; Zhang, B.; Yang, J.; You, J. Sample diversity, representation effectiveness and
robust dictionary learning for face recognition. Information Sciences. 2017, 375, 171-182.
[5] Tang, Z.Y.; Meng, F.R.; Wang, Z.X. Fast face recognition with regularized least square via
sparse representation. Application Research of Computers. 2016, 33, 2831-2836.
[6] Li, Y.Q.; Zhang, S.W.; Li, H.B.; Zhang, W.M.; Zhang, Q. Face recognition method using gabor
wavelet and cross-covariance dimensionality reduction. Journal of Electronics & Information
Technology. 2017, 39, 2023-2027.
[7] Yuan, H.; Wang, Z.H.; Jiang, W.T. 3D face recognition approach based on singular point
 neighborhood structure. Control and Decision. 2017, 32, 1739-1748.
[8] Jin, J.; Mi, W. An aimms-based decision-making model for optimizing the intelligent stowage
of export containers in a single bay. Discrete and Continuous Dynamical Systems Series S.
2019, 12(4-5), 1101-1115.
[9] Yang, A.; Li, Y.; Kong, F.; Wang, G.; Chen, E. Security control redundancy allocation
technology and security keys based on Internet of Things. IEEE Access. 2018, 6,
50187-50196.
[10] Aldouri, T.; Hifi, M. A hybrid reactive search for solving the max-min knapsack problem with
multi-scenarios. International Journal of Computers and Applications. 2018, 40(1), 1-13.
[11] Zhao, M.H.; Mo, R.Y.; Shi, Z.H.; Zheng, F.F. A novel method for recognition of pose
invariant face with single image. Journal of Xi’an University of Technology. 2017, 33, 18-23.
[12] Weng, R.; Lu, J.; Tan, Y.P. Robust Point Set matching for partial face recognition. IEEE
Transactions on Image Processing. 2016, 25, 1163-1176.
[13] Hentschel, R.; Leyh, C.; Petznick, A. Current cloud challenges in Germany: the perspective
of cloud service providers. Journal of Cloud Computing. 2018, 7(1).
[14] Valentine, T.; Lewis, M.B; Hills, P.J. Face-space: A unifying concept in face-recognition
research. Quarterly Journal of Experimental Psychology. 2016, 69, 1996-2019.
[15] Xu, Y.; Zhang, Z.; Lu, G.; Yang, J. Approximately symmetrical face images for image
preprocessing in face recognition and sparse representation based classification. Pattern
Recognition. 2016, 54, 68-82.
[16] García-Planas, M.I.; Klymchuk, T. Perturbation analysis of a matrix differential equation ẋ =
abx. Applied Mathematics & Nonlinear Sciences. 2018, 3(1), 97-104.
[17] Murugan, S.B.; Sundar, M.L. Investigate safety and quality performance at construction site
using artificial neural network. Journal of Intelligent and Fuzzy Systems. 2017, 33(4),
2211-2222.
[18] Jamil, M.K.; Farahani, M.R.; Imran, M.; Malik, M.A. Computing eccentric version of second
zagreb index of polycyclic aromatic hydrocarbons pahkpahk. Applied Mathematics &
Nonlinear Sciences. 2016, 2(1), 247-252.
[19] Feng, J.Y.; Shen, H.Y. Research on face recognition based on artificial neural network.
Automation & Instrumentation. 2017, 24-26.
[20] Yu, H.J.; Xie, Y.H.; Zhang, T.Z.; Li, C.D. Electric vehicle identity recognition and traction
battery coding. Chinese Journal of Power Sources. 2016, 40, 113-116.
[21] Shvets, A.; Makaseyev, A. Deterministic chaos in pendulum systems with delay. Applied
Mathematics & Nonlinear Sciences. 2019, 4(1), 1-8.
[22] Liu, X.; Wang, L.X.; Lv, C.; Li, J.F. Lithium-ion battery modeling and parameter
identification. Journal of Power Supply. 2018, 16, 145-150.
[23] R.Krishna Kumar, S.Diwakaran, M.Thilagaraj, “Reactive Power Control of Modern Type
High Effective Phase Grid-Tied Photovoltaic Network Inverter”, Journal of Green
Engineering, Vol 10, No 9, pp 4874-4884, 2020.

[24] M.Thilagaraj, N.Arunkumar, S.Ramkumar, S.Hariharasitaraman, “Electrooculogram Signal

Identification for Elderly Disabled Using Elman Network”, Microprocessors and
Microsystems, https://doi.org/10.1016/j.micpro.2020.103811, 2021.
[????] Dongdong, J., Arunkumar, N., Wenyu, Z., Beibei, L., Xinlei, Z., & Guangjian, Z. (2019).
 Semantic clustering fuzzy c means spectral model based comparative analysis of cardiac
color ultrasound and electrocardiogram in patients with left ventricular heart failure and
cardiomyopathy. Future Generation Computer Systems, 92, 324-328.
[????] Bagyalakshmi, G., Rajkumar, G., Arunkumar, N., Easwaran, M., Narasimhan, K., Elamaran,
V., Mario Solarte., Iván Hernández., & Ramirez-Gonzalez, G., Network Vulnerability
Analysis on Brain Signal/Image Databases using Nmap and Wireshark Tools. IEEE Access
vol. 6, pp. 57144-57151, 2018
[27] Li, L.M.; Liu, M.H. Research on the robustness of face recognition based on machine
learning algorithms. Journal of China Academy of Electronics and Information Technology.
2017, 12, 219-224.

All Authors’ Email:

*
Email of Junxi Guo: guojunxi1435@163.com (Corresponding Author)

Email of Yuzhuo Fu: yzfu1234@163.com

Email of Ting Liu: louisaliu123@163.com

Junxi Guo, Female, 1995.4. She received a master's degree from Shanghai Jiaotong
University in 2020. She has participated in a multi-modal conference system scientific research
project and wrote a patent. Her academic research interests include face recognition, speech
recognition, and multi-modal models.

Yuzhuo Fu, Male. He received B.S. degree from Computer Engineering Department at
Changsha Institute of Technology, and M.S. and Ph.D. degrees from Computer Science and
Engineering Department, Harbin Institute of Technology. Prior to joining SJTU in 2001, he
worked as Senior Engineer for the electric engineering institute of Heilongjiang University and
Computing Center of Heilongjiang Province. He is Present Deputy Professor of undergraduate
student affair office of SJTU.

Ting Liu, Female. She obtained a master's degree in software engineering from the School of
Microelectronics, Shanghai Jiaotong University. She has participated in the research and
development of the microelectronic pyrotechnics system control bus protocol design and the
FPGA verification of the multi-core system-on-chip simulation platform. Her main research
directions include FPGA-based multi-core system-on-chip architecture design, system power
consumption model analysis under multi-core architecture, network-on-chip network architecture
and protocol design and verification, etc.

Conflict of Interest
No conflict of interest