You are on page 1of 7

Journal of Physics: Conference Series

PAPER • OPEN ACCESS You may also like


- An Improved Artificial Neural Network
The Research and Analysis of Different Face Design for Face Recognition utilizing
Harmony Search Algorithm
Recognition Algorithms Maryam Mahmood Hussein, Ammar
Hussein Mutlag and Hussain Shareef

- Research on Face Recognition Based on


To cite this article: Wang Xiang 2022 J. Phys.: Conf. Ser. 2386 012036 CNN
Jie Wang and Zihao Li

- Age Group Classification using


Convolutional Neural Network (CNN)
Muhammad Firdaus Mustapha, Nur
View the article online for updates and enhancements. Maisarah Mohamad, Ghazali Osman et al.

This content was downloaded from IP address 49.43.250.132 on 03/11/2023 at 21:35


CONF-CIAP 2022 IOP Publishing
Journal of Physics: Conference Series 2386 (2022) 012036 doi:10.1088/1742-6596/2386/1/012036

The Research and Analysis of Different Face Recognition


Algorithms

Wang Xiang
WLSA Shanghai Academy, Shanghai, China

xwhs20050815@163.com

Abstract. With the background of fast developing artificial intelligence and Internet technology,
face recognition, one of the product technologies under the theory of machine learning, has
become widely applied on safety tests, identity confirmation, public security photos, and access
control system. Based on the trend of its development speed, this technology will have well
developed circumstances and the field of application will expand. One of the most essential
reasons why it has become so popular is that the algorithms used for face recognition are
advanced. In this paper, the author mainly researches five algorithms to realize face recognition
based on relevant literature. They are separately CNN, PCA, SVM, MTCNN, and Facenet. After
comparison and analysis, it can be found that CNN is one of the most suitable algorithms to
recognize human faces, and PCA and SVM have problems of calculating speed. The use of
complex neural networks can intensify the accuracy of face recognition just like MTCNN and
Facenet. If combining different algorithms together and using their advantages in face
recognition in different steps, the prediction of the algorithms will be better.

Keywords: Face recognition, algorithm, CNN, PCA, SVM.

1. Introduction
Nowadays, with the development of information technology, the face recognition system, as a technique
contains many professional techniques like artificial intelligence and machine learning, widely applied
in different realms[1]. Face recognition is a technique to extract a human face features information from
an image and match it to the specific face that is in the image dataset the computer stores[1]. There are
already a lot of algorithms used in this field. However, the algorithm still has space to improve since the
recognition sometimes fails due to a change in hair style or the wearing of glasses. The academic world
has researched the algorithms of face recognition and divided them into four types: feature-based
recognition algorithms; appearance-based recognition algorithms; template-based recognition
algorithms; and recognition algorithms using neural networks[2]. The author uses the research from the
previous scholar to summarize the main five algorithms of face recognition. In this paper, the author
describes the theory of CNN, PCA, SVM, MTCNN, and Facenet through a literature review and
compares their advantages and disadvantages. This research will help people who want to learn face
recognition understand the basic principles of algorithms.

Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution
of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd 1
CONF-CIAP 2022 IOP Publishing
Journal of Physics: Conference Series 2386 (2022) 012036 doi:10.1088/1742-6596/2386/1/012036

2. CNN

2.1. Definition and principle


Convolutional Neural Network is an algorithm that contains convolution calculations based on the
structure of neural networks. It uses the process of convolution to supersede the calculation of neural
networks. It is a typical deep learning algorithm. The principle of CNN is that it is like a function that
will give outputs based on inputs. For example, in the field of face recognition, it will state the name of
a person based on his picture.

2.2. Process
The process of convolutional neural network can be divided into four steps: filtering, pooling, activation,
and fully connected layer[3]. CNN uses convolution kernels to complete the filtering conduction.One
image has many characteristics, which means that there are many convolution kernels. These
convolution kernels could be understood as matrixes since we convert the pattern information to
numbers. As shown in Figure 1, when detecting a new image, the computer will convert this image into
several matrixes with numbers and line up the convolution kernel and the image patch. Then it will
multiply the numbers within the two matrixes correspondingly and add them up. After dividing the result
number by the number of pixels in the matrix, the final number will replace the initial number of this
image. This process will be repeated several times until the total image is changed into a big new matrix.
Since there are many convolution kernels, the replacement of the image will be executed multiple times,
depending on how many kernels it has.

Figure 1. The process of filtering.


Then it comes to the step of pooling. To make it easy to understand, the pooling process is to simplify
the matrix into low dimensional matrix. As shown in Figure 2, it will change a part of a region into one
number, whether by getting the average number or by getting the maximum number. Through this
process, the speed of calculation will be fastened, and it will also prevent the possibility of overfitting.

Figure 2. The process of pooling.

2
CONF-CIAP 2022 IOP Publishing
Journal of Physics: Conference Series 2386 (2022) 012036 doi:10.1088/1742-6596/2386/1/012036

The next step is activation, which is the nonlinear reflection of the pooling layer[3]. The pooling layers
pass through the activation function, shown in Figure 3, giving the multilayer networks actual meaning.
In CNN, this function is often ReLU.

Figure 3. Activation function.


The ultimate step is fully connection. After finishing the process of filtering, pooling, and activation, the
computer will put the layers together to finally give an answer through fully connected neural networks.
It will drop the layer into one dimension. In the neural networks, different data has different weights
that depends on the training times. In the end, after the calculation, the computer will give a number
representing the possibility of this image’s meaning.

2.3. Discussion
Convolution neural network is a robust algorithm in dealing with picture information since the
backpropagations will refine the weights and bias, which will make the judgement more accurate. Since
it is machine learning, the election of convolution kernels will be more reasonable. For example, after
thousands of training sessions, the computer will recognize noses, eyes, and mouths as convolution
kernels without human intervention[4]. The only thing people should do is to give enough image data.
Although this algorithm cannot deal with the same data information after changing the row and column,
it is an excellent algorithm for face recognition.

3. PCA-SVM
Face recognition based on PCA and SVM algorithms uses PCA to reduce the dimension of images,
which accelerate the calculation process. Then through the classier of SVM, the computer can judge
who the person is in the image[5].

3.1. Principal Component Analysis (PCA)


Principal Component Analysis is an algorithm to make dimensionality reduction of a group of data
through orthogonal transformations[6]. Through the dimension reduction process, the work of
calculation decreases with the maintaining of the information accuracy. In the field of face recognition,
since the image of people consists of thousands of pixels, Principal Component Analysis can help reduce
the dimension into small numbers, which are called eigenfaces. This algorithm contains two steps:
decentration and finding a coordinate system. Decentration is a process of putting the origin of
coordinates in the middle of the data. It will help the subsequent calculation since the origin is zero.
Finding coordinates system is to find a direction that has the maximum variance of these data. PCA uses
the covariance matrix to complete this. The covariance matrix reflects the correlation between two
variables. Hence, through calculating the eigenvectors and eigenvalues, the computer will figure out
what the degree the coordinate rotates and what the variance of the coordinate in that direction.
According to the principle of PCA, the larger the value of eigenvalues, the more accurate the information

3
CONF-CIAP 2022 IOP Publishing
Journal of Physics: Conference Series 2386 (2022) 012036 doi:10.1088/1742-6596/2386/1/012036

remained. Thus, if someone wants to reduce the dimension into k-dimension, the computer remains the
greatest k eigenvalues, using them to create several eigenfaces.

Figure 4. PCA model.

3.2. Support Vector Machine (SVM)


Support Vector Machine belongs to supervised learning and is an algorithm that can both conduct
classification and regression[7]. Its principle is to find a maximum-margin hyperplane. The maximum-
margin hyperplane will divide two classes into two parts with the maximum distance to the nearest data.
If we give enough data to create an SVM model, the prediction of new data will be very precise. Due to
the Support Vector Machine being only suitable for binary classification, if we want to realize multi-
classification, we need to construct n*(n-1)/2 vector machines and combine these together. Finally,
through one-to-one voting, we find out the class of the data.

Figure 5. The model of SVM.


In the field of face recognition, after the process of dimensionality reduction, we put the eigenfaces into
the SVM model to train. Then this classifier will give us the answer to who this person is in the image.
However, whether the eigenfaces or the SVM model, they all have some shortcomings. The predictions
of eigenfaces will have significant differences if the images only have illumination differences and the
SVM model only supports tiny amounts of data, if we want to make face recognition based on big data,
the calculation of SVM needs a large amount of time.

4
CONF-CIAP 2022 IOP Publishing
Journal of Physics: Conference Series 2386 (2022) 012036 doi:10.1088/1742-6596/2386/1/012036

4. MTCNN and facenet


The process of face recognition can be divided into two parts. The first part is face detection and image
extraction, and the second part is feature extraction. MTCNN is used to complete the first step, and
Facenet is used to complete the second step.

4.1. MTCNN(Multi-Task Convolutional Network)


MTCNN consists of three convolutional networks: PNet(Proposal Network), RNet(Refine Network),
and Onet(Output Network)[9]. To satisfy the detection of different sized images, before putting the
image into PNet, we need to change the size of the image to create an image pyramid, as shown in Figure
6. Then the PNet will obtain the candidate windows of face region and perform regression to refine the
candidate window. After that, we put the image into RNet, which has one more fully connected layer
than PNet, and makes high accuracy filtering of candidate windows. Finally, the Onet is used to make
ultimate optimizing of the model performance, completing the regression of face region and label of
feature points[8].

Figure 6. The process of MTCNN.

4.2. Facenet
Facenet reflects the face feature of image into multi-dimension and calculates the Euclidean space
distance to make judgement of human face[9]. Since a same person’s image will have shorter Euclidean
space distance, Facenet use this to make face recognition. Finally, Facenet will give a 128-dimension
eigenvector of the human face. Both MTCNN and Facenet use the principle of neural networks.
Compared to the single CNN algorithm, MTCNN complicates the algorithm and makes the processing
of images more precise. In addition, by dividing the process of face recognition into two steps, the
prediction of this new system of algorithms works better than others.

5. Discussion
Different algorithms have their advantages in different situations. The methods of face recognition based
on geometrical features have less sensitivity to illumination, but its accuracy is low when people have
intense facial expressions. The methods based on template matching, like PCA, can use dynamic
templates to solve the problem of intense facial expression. However, at the same time, the complexity
of the calculation makes the period and time of calculation long[1]. According to Face Recognition
Research Based on Transform Domain and Local Direction Pattern[10], the researcher combined the
Non-subsampled Contourlet Transform(NSCT) algorithm with the Absolute Double Center Symmetry

5
CONF-CIAP 2022 IOP Publishing
Journal of Physics: Conference Series 2386 (2022) 012036 doi:10.1088/1742-6596/2386/1/012036

Local Pattern(ADCSLP), which solves the effect of illumination intensity and the problem of feature
extraction at the same time. The research shows that through algorithm combination, the effect of face
recognition is better than a single algorithm like LDP. Face recognition algorithms are becoming
increasingly advanced. The convolution neural network perfectly handles figure information processing,
and the accuracy of prediction is improving with the combination of different algorithms. From face
extraction to face recognition, the popular algorithms nowadays have already solved some of the
problems of deviation caused by makeup and illumination[2]. Although multiparameter algorithms still
add pressure on both software and hardware, making the cost of face recognition technology expensive,
the future of face recognition is to be expected. The trend of this technology development is to combine
it with the network. As it solves the problem of daily life identity recognition, face recognition
technology will be applied to more fields and share information through the Internet. With the purpose
of ensuring information safety, this technology will combine with other biological recognition
technologies to further confirm a person's identity.

6. Conclusion
This paper discussed the principles of five algorithms used in the face recognition field and the merits
and weaknesses of each algorithm. The research shows that convolution neural network is good at
dealing with image information but will be affected by illumination the same as PCA. The SVM will
spend too much time on calculation, which may not be easy to realize face recognition with large
amounts of information. The combination algorithms like MTCNN and Facenet have more accurate
judgments of human faces than the single algorithm. Face recognition technology will be used in the
Internet field in the future. However, this paper lacks the establishment of face recognition models.
Future research will use experiments and modeling to test the accuracy of different algorithms.

Acknowledgement
I would like to show my gratitude to my paper teacher and professor during this scientific research. My
teacher helped me with my essay topic and format amendment, and my professor explained some of the
problems with the algorithms I mentioned in this paper. Further, I want to thank my parents for
supporting me in completing this research. Without their encouragement, I could not finish my essay
topic.

References
[1] He Yanqin, Wu Lian, Guo Qingfen & Zeng Guinan. (2021). Current situation and future trend of
face recognition technology. Wireless Internet Technology (13), 80-82.
[2] Sun Peng. (2019). Overview of research and application of face recognition technology.
Propagation of science and technology (24), 130-131.
[3] S. Albawi, T. A. Mohammed and S. Al-Zawi, "Understanding of a convolutional neural network,"
2017 International Conference on Engineering and Technology(ICET), 2017, pp. 1-6.
[4] Duan Jincheng, Zhang Fengxia, Zhu Xiaoqing. (2022). Research on Face Recognition Algorithm
Based on Convolutional Neural Network. Scientific and Technology Innovation (10), 73-76.
[5] Meng Yifan & Liu Yijun. (2021). Research on Face Recognition Algorithm Based on PCA-SVM.
Horizon of Science and Technology (07), 77-79.
[6] Wang Zhiyang, Liu Jinlong & Tang Zixian. (2016). Application Research on PCA algorithm in
Face Recognition. Horizon of Science and Technology (01), 19-20.
[7] Liu Ming & Wu Zhaoxia. (2018). Theory and application of support vector machine. Horizon of
Science and Technology (23), 68-69.
[8] Li Zhihua, Zhang Jianyu & Wei Zhongcheng. (2022). Design on face recognition system based
on MTCNN and Facenet. Modern Electronic Technology (04), 139-143.
[9] Zhang Ying. (2022). Study on improved face recognition algorithm based on Facenet. Electronic
Technology (03), 25-27.
[10] Wei Wei. (2019). Face Recognition Research Based on Transform Domain and Local Direction
Pattern. Master thesis, Xiangtan University.

You might also like