• 検索結果がありません。

A Hybrid Global Local Adaptive Particle Swarm Optimization-Based Support Vector Machine Model for Human Facial Authentication

N/A
N/A
Protected

Academic year: 2021

シェア "A Hybrid Global Local Adaptive Particle Swarm Optimization-Based Support Vector Machine Model for Human Facial Authentication"

Copied!
10
0
0

読み込み中.... (全文を見る)

全文

(1)

西 南 交 通 大 学 学 报

第 55 卷 第 1 期

2020 年 2 月

JOURNAL OF SOUTHWEST JIAOTONG UNIVERSITY

Vol. 55 No. 1

Feb. 2020

ISSN: 0258-2724 DOI:10.35741/issn.0258-2724.55.1.23

Research Article

Computer and Information Science

A

H

YBRID

G

LOBAL

L

OCAL

A

DAPTIVE

P

ARTICLE

S

WARM

O

PTIMIZATION

-B

ASED

S

UPPORT

V

ECTOR

M

ACHINE

M

ODEL FOR

H

UMAN

F

ACIAL

A

UTHENTICATION

基于混合全局局部自适应粒子群算法的人脸认证支持向量机模型

Salam Allawi Hussein a, *, Alyaa Abduljawad Mahmood b, Mohammed Iqbal Dohan c

a

Department of Computer Science, College of Computer Science and Information Technology, University of Al-Qadisiyah

P.O. Box 88, Al Diwaniyah, AL-Qadisiyah, Iraq salam.allawi@qu.edu.iq

b

Al-Qadisiyah Education Directorate Al-Qadisiyah, Iraq

alsafat_84@yahoo.com

c Department of Multimedia, College of Computer Science and Information Technology, University of

Al-Qadisiyah

P.O. Box 88, Al Diwaniyah, Al-Qadisiyah, Iraq mohammed.iqbal@qu.edu.iq

Abstract

A new facial authentication model called global local adaptive particle swarm optimization-based support vector machine, was proposed in this paper. The proposed model aimed to solve the problem of finding the preeminent parameters of support vector machine in order to come out with a powerful human facial authentication technique. The conventional particle swarm optimization algorithm was utilized with support vector machine to explore the preeminent parameters of support vector machine. However, the particle swarm optimization support vector machine model has some limitations in selecting the velocity coefficient and inertia weight. One of the best approaches, which is used to solve the velocity coefficient problem, is adaptive acceleration particle swarm optimization. Also, the global-local best inertia weight is used efficiently for selecting the inertia weight. Therefore, the global local adaptive particle swarm optimization-based support vector machine model was proposed based on combining adaptive acceleration particle swarm optimization, global-local best inertia weight, and support vector machine. The proposed model used the principal component analysis approach for feature extraction, as well as global local adaptive particle swarm optimization for finding the preeminent parameters of support vector machine. In the experiments, two datasets (YALEB and CASIAV5) were used, and the suggested model was compared with particle swarm optimization support vector machine and adaptive acceleration particle swarm optimization support vector machine methods. The comparison was via accuracy, computational time, and optimal parameters of support vector machine. Our model can be used for security applications and apply for human facial authentication.

Keywords: Facial Authentication, Support Vector Machine, Particle Swarm Optimization, Adaptive

(2)

摘要 提出了一种新的基于全局局部自适应粒子群优化的支持向量机面部认证模型。所提出的模型 旨在解决寻找支持向量机的主要参数的问题,从而提出一种强大的人脸认证技术。将传统的粒子 群算法与支持向量机结合使用,探索支持向量机的主要参数。但是,粒子群优化支持向量机模型 在选择速度系数和惯性权重时有一定的局限性。用于解决速度系数问题的最佳方法之一是自适应 加速度粒子群优化。而且,全局局部最佳惯性权重被有效地用于选择惯性权重。因此,结合自适 应加速度粒子群算法,全局局部最优惯性权重和支持向量机,提出了基于全局局部自适应粒子群 算法的支持向量机模型。该模型使用主成分分析方法进行特征提取,并使用全局局部自适应粒子 群算法查找支持向量机的主要参数。在实验中,使用了两个数据集(耶鲁布和卡萨夫 5),并将 建议的模型与粒子群优化支持向量机和自适应加速粒子群优化支持向量机方法进行了比较。通过 准确性,计算时间和支持向量机的最佳参数进行比较。我们的模型可用于安全性应用程序并申请 人脸认证。 关键词: 面部认证,支持向量机,粒子群优化,自适应加速粒子群优化,主成分分析

I. I

NTRODUCTION

Facial authentication was one of the supreme and interesting research topics in computer vision in the last years [1], [2]. Many applications for face authentication are adopted to avoid or address the increasing threats of terrorists and illegal attacks. For example, they are used in forensic examinations, airports, social media, smartphones, etc. Numerous artificial intelligence methods have been used for this purpose, such as neural networks and machine learning methods [3]. One of the well-known methods is the support vector machine (SVM), which can provide global solutions for non-linearity problems, especially facial recognition [4]. However, the accuracy of SVM is extremely influenced by the selection of the training parameters. Several studies have utilized SVM for facial authentication as successful classifiers in this context. The researchers in [5] proposed a strategy termed as SVM + nonparametric discriminant analysis (NDA) (SVM + NDA) model, which might be seen as an augmentation to SVM by consolidating some halfway data. Nevertheless, it is computationally high. Also, there is another approach, which is called a multi-objective uniform design (MOUD). This approach is a search method adopted with SVM for facial authentication [6]. The method succeeds in reducing the cost. Furthermore, the authors in [7] presented a new facial classifier named total margin-based adaptive fuzzy support vector machines (TAF-SVM), which is based on the adaptive fuzzy concept. Of late, a swarm intelligent algorithm, named particle swarm optimization (PSO), was presented in 1995 by Kennedy and Eberhart, which can powerfully search for optimum or near-optimum solutions in

huge search spaces [8]. The PSO method is adopted in numerous domains, for example, video-based facial authentication or verification, and used to deal withvarioustypes of optimization complications [9], [10], [11]. To search for the optimum parameters in SVM, a novel approach, called PSO-SVM, has been developed by the researchers [12]. The researchers have utilized PSO to explore the ideal training parameters of SVM, and they adopted their technique for face recognition.

However, in standard PSO, there are some limitations, which could create an element of doubt about the accuracy of results. One of these limitations, utilizing a fixed numbers for velocity coefficients, is usually close to two which have a high impact on the particle’s velocity performance [13], [14]. Another limitation in conventional PSO is to choose the inertial weight coefficient, which is a user-supplied coefficient that provides the balance between local and global exploration. Normally, the value of the inertial coefficient w is between 0.8-1.2, which can reduce the inertia of the particle or fasten it to the original path [15]. In general, if the inertial coefficient is low, the value will accelerate the swarm’s convergence to optima; however, if the value is high, it will lead to investigate the whole search space. So that and to address the first drawback of PSO, an enhanced method, called AAPSO-SVM, was developed by [16]. They utilized an adaptive procedure to choose the velocity coefficients based on the fitness values of the particles. Their method achieved high accuracy performance in facial recognition. In order to deal with the inertial weight limitation, many researchers have introduced several approaches to choose the accurate weight coefficient [17], [18], [19], [20]. One of the best

(3)

3

methods, which achieved encouraging results, is global-local best inertia weight (GLbestIW) [21]. Nevertheless, there is no precise algorithm to succeed in all optimization problems. In other words, some of them provide better results for particular problems than others. As a result, enhancing the earlier optimization approaches or suggesting new approaches is an essential issue. Therefore, and by taking the advantages of adaptive acceleration particle swarm optimization (AAPSO) and GLbestIW approaches, we suggested in this paper a new hybrid approach for facial authentication called global local adaptive particle swarm optimization based support vector machine (GLAPSO-SVM). The other parts of this paper were prepared as follows: In Section 2, the problem statement was explained. Next, research methods were explained in Section 3, as well as the details of the proposed model. Section 4 shows the investigational results, whereas Section 5 concluded the paper.

II. T

HE

P

ROBLEM

S

TATEMENT

SVM is one of the common supervised learning techniques, which can attain ideal solutions for non-linear problems with lesser training trials. However, the parameters of SVM have a weighty influence on the entire development of SVM. Lately, the swarm intelligence algorithm, PSO, can professionally discover the best solutions in huge search spaces. As a result, a new facial authentication method was introduced in [12], which depended on PSO and SVM (PSO-SVM). PSO was utilized to improve the training parameters in SVM, and PSO-SVM achieved high performance in facial authentication applications, but this method lacks by the utilization fixed random number close to two in for velocity coefficients. Moreover, in choosing the inertia weight (normally between 0.8-1.2), which is a very significant parameter in PSO and may touch the convergence and investigation “trade-off” in the PSO development. The velocity coefficient problem was solved by AAPSO [16], and the inertia weight limitation was improved by [21]. To enhance standard PSO, we introduced a GLAPSO with SVM for facial authentication.

III. R

ESEARCH

M

ETHODS A. Particle Swarm Optimization

PSO involves a crowd of particles in virtual search space, and it searches the finest position concerning the equivalent finest solution for the optimization problem. Every particle travels in

the route of its plbest and glbest [8], [22]. The process of particle moving is described as:

(1) (2) where d stands for the dth repetition while g1 and g2 are velocity constants usually = 2. The is a positive random number between 0 and 1. Also, we indicated to the inertial weight coefficient (normally between 0.8 - 1.2); represents the existing location at iteration d of the particle i ; represents the velocity of a particle i; is the personal finest location of particle i; signifies the finest location of all personal finest locations in the whole particles in the swarm. Figures 1 and 2 show the flowchart of the standard PSO and pseudo code respectively:

Figure 1. The PSO flowchart Begin

Initialize particle swarm randomly

While (the number of iterations, or the ending criterion is not reach )

Calculate fitness value of particle for n = 1 to the number of particles

Find plbest Find glbest

for d = 1 to the number of particle dimension Updating the location of the particles by equation 1 and equation 2

Next d Next n

Succeeding generation until it reaches ending criterion

End End

(4)

B. SVM Classifier

The SVM classifier is a method that depends on the “statistical learning” philosophy which its goal is to separate two classes by defining an optimal hyperplane in the training dataset [4]. Suppose a train data set where represents the input vector, and is a label of class. The hyperplane is determined as , where indicates to a point on the hyperplane, The positioning of the hyperplane is determined by and is the distance of the hyperplane from the point on the origin. As displayed in Figure 3, the optimal splitting hyperplane can be established through minimizing under the constraint . Accordingly, finding the ideal hyperplane is necessary to resolve the optimization problem given by:

(3) The method can be extended by presenting “positive slack variables” in order to substitute the optimization problem, then

Figure 3. The process of SVM classification

The new optimization problem for “non-linear decision surfaces” is given as:

(4) where M denotes a penalty parameter that represents a regularization constant. It is responsible for the controls of the balance between two challenging principles of “margin maximization” and “error minimization.” Therefore, the function of the authentication problem becomes:

(5) where represents “Lagrange multipliers”; denotes the kernel function,

and through some nonlinear mapping functions , it can mapping the data to the higher dimensional space. The “radial basis function” (RBF) (defined as , is typically utilized in the literature to build SVM for image classification purposes.

C. The Facial Authentication Model via GLAPSO-SVM

The structure of the facial authentication based on GLAPSO-SVM is displayed in Figure 3. We can summarize the main steps of the proposed models as follow: Initially, we read face image from the database, then extract the face features from facial image using principal component analysis (PCA) method. The extracted facial features are utilized for the process of training and testing in GLAPSO-SVM model. The final step is facial authentication via proposed GLAPSO-SVMs.

Figure 4. The facial authentication process via GLAPSO-SVM

D. The Proposed GLAPSO

To enhance standard PSO and addresses the two limitations, which were mentioned in the problem statement, two methods were merged with conventional PSO. To address the first drawback, GlbestIW [21] adopted to choose the inertia weight parameter. It was recognized as a function of plbest and glbest values of the objective function for every particle’s generation, which did not set as linearly reducing time changing function or set to any constant value. It is given in Equation (6):

(6)

where represents the local finest value of the particles in the swarm, and denotes the best value amongst the whole local finest values in the swarm. To deal with the second

(5)

5

drawback of conventional PSO, which utilized a fixed random number close to two for velocity coefficients, we adopted an adaptive method called AAPSO proposed by [16]. In AAPSO, particle fitness values are adopted in the selection process of acceleration coefficients instead of the fixed random number (close to two).The adaptive acceleration coefficients in AAPSO were selected according to the following formulas:

+ (7)

+ (8)

where and represent the minimum values of and maximum values of . and is the particle minimum fitness valuefrom the whole generated population while and are an average and maximum fitness value of the whole generated population. Also,

represents the minimum values of and represents the maximum values of . Through applying , and in equation 9 of velocity, the formula is modified as follows:

(9) The parameters , and have a high impact on the particle’s velocity performance, and this will influence the selection of the training parameters in SVM. Overall, this will be reflected in the final classification result, specifically when SVM is used with facial authentication, which is a non-linear problem. Therefore, we developed a hybrid model by using GlbestIW and AAPSO with SVM. The pseudocode of GLAPSO is illustrated in Figure 5.

PSO pseudo code

Begin

Initialize particle swarm randomly

While (the number of iterations, or the ending criterion is not reached)

Calculate the fitness value of particle for n=1 to the number of particles

Find plbest Find glbest

for d = 1 to the number of particle dimension Updating the location of the particles by Equations (1) and (2)

Next d Next n

Updating the value of inertia weight with equation 6 Succeeding generation until it reaches the ending criterion

End End

Figure 5. The GLAPSO pseudocode

E. Parameters Optimization of SVM by GLAPSO

The RBF kernel function was used to construct SVM and particle composed of two user-determined parameters, which are M and σ. The following steps summarize the optimization process:

a) Set the PSO parameters, and reset all particles’ locations and velocity, set the velocity parameters by using Equation (9) and the inertia weight by using Equation (6), determine the maximum number of iterations.

b) Compute the fitness value for every particle by the formula ( indicates the true classification and denotes the false classification).

c) Updating the particle’s velocity and location. The velocity and location for each particle are upgraded via the Equations (9) and (2) in relation to the updated plbest and glbest.

d) Check the stop condition. If the identified number of generations is satisfied, we stop; else, we return to Step 3.

IV. E

XPERIMENTAL

R

ESULTS

The suggested facial authentication model was executed by reliable platform, which is MATLAB. The performance of the suggested model was assessed via two human facial datasets, i.e., CASIAV5 [23] and YALEB [24]. In the experiment, the face images were selected from different conditions. Also, the PCA algorithm was used to extract facial features. Moreover, the authentication process was accomplished via the suggested GLAPSO-SVM model, and the suggested model was compared with the original PSO-SVM and advanced AAPSO-SVM models. By conducting “10-fold cross-validation” over each dataset, the performance of the suggested model was analyzed. In addition, the corresponding statistical measures were specified. To accomplish cross-validation via 10-fold, training of ten folds and testing datasets were created via the folding process. From every database, a total of 100 images were used, and the chosen images were separated into 20 images for training and 80 images for testing. Figure 6 shows samples from CASIAV5 and YALEB dataset, respectively.

(a) (b)

(6)

The whole performance of the proposed GLAPSO-SVM model was evaluated by using three steps of measures: Accuracy performance, computational time, and SVM optimal training parameters.

A. The Assessment of the Proposed Approach via Accuracy Performance

The accuracy measure was used to assess the proposed GLAPSO-SVM model, and 10

experiments were performed. The results of the recognition process for the proposed GLAPSO-SVM, PSO-GLAPSO-SVM, and AAPSO-SVM models are illustrated in Table 1 and Figure 6. The proposed GLAPSO-SVM model was executed in 10 experiments on each dataset. For CASIAV5, the proposed model attained higher accuracy than the other models (8 experiments from 10 experiments) and (10 experiments from 10 experiments) for YALEB dataset.

Table 1.

The accuracy result for the GLAPSO-SVM, PSO-SVM, and AAPSO-SVM models from the two datasets

10 fold cross validation

GLAPSO-SVM

PSO-SVM AAPSO-SVM

GLAPSO-SVM PSO-SVM AAPSO-SVM CASIA V5 CASIA V5 CASIA V5 YALE B YALE B YALE B 1 92 88 90 92 88 90 2 95 90 92 95 85 92 3 98 90 93 92 86 92 4 80 92 94 92 80 90 5 95 89 90 94 88 92 6 94 91 92 91 85 91 7 94 90 94 93 81 91 8 95 91 90 95 80 90 9 89 92 94 94 86 93 10 96 90 93 92 81 90 (a) (b)

Figure 6. The accuracy performance for the proposed model, (a) CASIAV5 and (b) YALEB

B. The Assessment of the Proposed Method via Computational Time

The second assessment measure is the computational time; Figure 7 shows the computational time in the original PSO-SVM, AAPSO-SVM, and suggested GLAPSO-SVM model. The proposed model needed lesser

computational time compared with the other methods in most cases.

Figure 7. The computational time for the proposed GLAPSO-SVM model against the other models C. The Assessment of the Proposed Method

via Optimal Training Parameters

The third assessment is to find the optimal training parameters of SVM. Tables 2-4 show samples from the optimal training parameters for the proposed GLAPSO-SVM, PSO-SVM, and AAPSO-SVM models, respectively.

Table 2.

Training parameters of GLAPSO-SVM

Method/Image C σ GLAPSO-SVM1 102.23 1.18 GLAPSO-SVM2 94.39 0.84 GLAPSO-SVM3 85.12 1.01 GLAPSO-SVM4 33.86 0.83 GLAPSO-SVM5 62.17 1.44 GLAPSO-SVM6 68.83 1.58 GLAPSO-SVM7 34.72 0.87 GLAPSO-SVM8 56.42 1.23

(7)

7

GLAPSO-SVM9 61.22 1.42

GLAPSO-SVM10 48.63 1.97

Table 3.

Training parameters of PSO-SVM

Method/Image C σ PSO-SVM1 110.12 1.26 PSO-SVM2 105.22 1.07 PSO-SVM3 92.16 1.24 PSO-SVM4 45.32 1.24 PSO-SVM5 73.35 1.96 PSO-SVM6 74.85 1.84 PSO-SVM7 48.64 1.04 PSO-SVM8 63.44 1.58 PSO-SVM9 72.32 1.89 PSO-SVM10 51.86 1.99 Table 4.

Training parameters of AAPSO-SVM

Method /Image C σ AAPSO-SVM1 106.31 1.12 AAPSO-SVM2 99.44 0.91 AAPSO-SVM3 90.14 1.73 AAPSO-SVM4 41.93 0.98 AAPSO-SVM5 71.15 1.76 AAPSO-SVM6 72.34 1.66 AAPSO-SVM7 42.72 1.02 AAPSO-SVM8 61.18 1.63 AAPSO-SVM9 68.33 1.75 AAPSO-SVM10 49.17 1.89

V. C

ONCLUSION

In this paper, a facial authentication model was proposed based on GLAPSO and SVM. To find the training parameters of SVM, GLAPSO was used, and the PCA approach was utilized for feature extraction. The optimized SVM approach via GLAPSO was utilized for facial authentication. The performance of the proposed GLAPSO-SVM model was investigated by two datasets, CASIAV5 and YALEB. The investigational results demonstrated that the suggested GLAPSO-SVM model provided high performance compared with the original PSO-SVM and AAPSO-PSO-SVM models in accuracy, computational time, and optimizing SVM parameters. As a future work, we can use the proposed GLAPSO-SVM model for iris recognition.

A

CKNOWLEDGMENT

We would like to thank the University of Al-Qadisiyah for their assist in labs and their support, as well as our staff for their guidelines.

R

EFERENCES

[1] DENG, J., GUO, J., XUE, N., and

ZAFEIRIOU, S. (2019) ArcFace: Additive

angular margin loss for deep face recognition.

In:

Proceedings

of

the

IEEE/CVF

Conference on Computer Vision and Pattern

Recognition, Long Beach, California, June

2019. Computer Vision Foundation, pp.

4690-4699.

[2] LI, P., PRIETO, L., MERY, D., and

FLYNN, P. (2018) Face recognition in low

quality images: a survey. Available from

https://arxiv.org/pdf/1805.11519.pdf.

[3] JAFRI, R. and ARABNIA, H.R. (2009)

A survey of face recognition techniques.

Journal of Information Processing Systems,

5 (2), pp. 41-68.

[4] FAUZI, M.A. (2018) Indonesian review

rating prediction using support vector

machines and N-gram. Journal of Southwest

Jiaotong University, 53 (5). Available from

http://jsju.org/index.php/journal/article/view/

237.

[5] KHAN, N.M., KSANTINI, R., AHMAD,

I.S., and BOUFAMA, B. (2012) A novel

SVM + NDA model for classification with

an application to face recognition. Pattern

Recognition, 45 (1), pp. 66-79.

[6] LI, W., LIU, L., and GONG, W. (2011)

Multi-objective uniform design as a SVM

model selection tool for face recognition.

Expert Systems with Applications, 38 (6), pp.

6689-6695.

[7] LIU, Y.H. and CHEN, Y.T. (2007) Face

recognition

using

total

margin-based

adaptive fuzzy support vector machines.

IEEE Transactions on Neural Networks, 18

(1), pp. 178-192.

[8] AHMED, K., AL-KHATEEB, B., and

MAHMOOD, M. (2019) A Multi-Objective

Particle Swarm Optimization for Pavement

Maintenance with Chaos and Discrete.

Journal of Southwest Jiaotong University, 54

(3).

Available

from

http://jsju.org/index.php/journal/article/view/

283.

[9] ZOU, F., LI, J., and MIN, W. (2019)

Distributed Face Recognition Based on Load

Balancing and Dynamic Prediction. Applied

Sciences, 9 (4), 794.

[10] CONNOLLY, J.F., GRANGER, E., and

SABOURIN,

R.

(2012)

An

adaptive

classification system for video-based face

recognition. Information Sciences, 192, pp.

50-70.

(8)

[11] WEIRONG, C.H., WEIYING, W.A.,

YIBIN, Z.H., YONGKANG, Z.H., and QI,

L.I. (2018) MPPT Control of Partial Shadow

Photovoltaic Generation System Based on

Improved PSO Algorithm. Journal of

Southwest Jiaotong University, 53 (6).

Available

from

http://jsju.org/index.php/journal/article/view/

240.

[12] WEI, J., JIAN-QI, Z., and XIANG, Z.

(2011) Face recognition method based on

support vector machine and particle swarm

optimization.

Expert

Systems

with

Applications, 38 (4), pp. 4390-4393.

[13] LIU, H., XU, G., DING, G.Y., and SUN,

Y.B. (2014) Human behavior-based particle

swarm optimization. The Scientific World

Journal, 2014, 194706.

[14] ABDULAMEER, M.H., MOHAMMED,

D., MOHAMMED, S.A., AL-AZAWI, M.,

AL-MAYALI, Y.M., and ALAMERI, I.A.

(2018) Face recognition technique based on

adaptive-opposition

particle

swarm

optimization (AOPSO) and support vector

machine

(SVM).

ARPN

Journal

of

Engineering and Applied Sciences, 13, pp.

2259-2266.

[15]

BANSAL,

J.C.,

SINGH,

P.K.,

SARASWAT, M., VERMA, A., JADON,

S.S., and ABRAHAM, A. (2011) Inertia

weight

strategies

in

particle

swarm

optimization. In: Proceedings of the 3rd

World Congress on Nature and Biologically

Inspired Computing, Salamanca, October

2011. Piscataway, New Jersey: Institute of

Electrical and Electronics Engineers, pp.

633-640.

[16] ABDULAMEER, M.H., ABDULLAH,

S., HUDA, S.N., and OTHMAN, Z.A.

(2014) Support vector machine based on

adaptive

acceleration

particle

swarm

optimization. The Scientific World Journal,

2014, 835607.

[17] FENG, Y., TENG, G.F., WANG, A.X.,

and YAO, Y.M. (2007) Chaotic inertia

weight in particle swarm optimization. In:

Proceedings of the 2nd International

Conference

on

Innovative

Computing,

Information

and

Control,

Kumamoto,

September 2007. Piscataway, New Jersey:

Institute of Electrical and Electronics

Engineers, pp. 475-475.

[18] AL-HASSAN, W., FAYEK, M.B., and

SHAHEEN,

S.I.

(2006)

PSOSA:

An

optimized particle swarm technique for

solving the urban planning problem. In:

Proceedings of the International Conference

on Computer Engineering and Systems,

Cairo, November 2006. Piscataway, New

Jersey: Institute of Electrical and Electronics

Engineers, pp. 401-405.

[19] LI, H.R. and GAO, Y.L. (2009) Particle

swarm optimization algorithm with exponent

decreasing inertia weight and stochastic

mutation. In: Proceedings of the 2nd

International Conference on Information and

Computing Science, Manchester, May 2009.

Piscataway,

New

Jersey:

Institute

of

Electrical and Electronics Engineers, pp.

66-69.

[20] CHEN, G., HUANG, X., JIA, J., and

MIN, Z. (2006) Natural exponential inertia

weight

strategy

in

particle

swarm

optimization. In: Proceedings of the 6th

World Congress on Intelligent Control and

Automation, Dalian, June 2006. Piscataway,

New Jersey: Institute of Electrical and

Electronics Engineers, pp. 3672-3675.

[21] ARUMUGAM, M.S. and RAO, M.V.

(2006) On the performance of the particle

swarm optimization algorithm with various

inertia weight variants for computing optimal

control of a class of hybrid systems. Discrete

Dynamics in Nature and Society, 2006,

79295.

[22] RONGRONG, S.O. and ZILI, C.H.

(2015) Optimization and Design of Maglev

System PID Controller Based on Particle

Swarm Optimization Algorithm. Journal of

Southwest Jiaotong University, 50 (1).

Available

from

http://www.jsju.org/index.php/journal/article/

view/187.

[23] CHINESE ACADEMY OF SCIENCES

(2007) CASIA face database. [Online]

Available

from:

http://biometrics.idealtest.org/

[Accessed

25/10/19].

[24] GEORGHIADES, A.S., BELHUMEUR,

P.N., and KRIEGMAN, D.J. (2001) From

few to many: Illumination cone models for

face recognition under variable lighting and

pose. IEEE Transactions on Pattern Analysis

(9)

9

and Machine Intelligence, 23 (6), pp.

643-660.

参考文

:

[1] DENG,J., GUO, J. , XUE, N. 和

ZAFEIRIOU,S.(2019)弧面:用于深脸

识别的加法角余量损失。在:电气工程师

学会/ CVF 计算机视觉和模式识别会议论

文集,加利福尼亚州长滩,2019 年 6 月。

计算机视觉基金会,第 4690-4699 页。

[2] LI,P.,PRIETO,L.,MERY,D. 和

FLYNN,P.(2018)低质量图像中的人脸

识 别 : 一 项 调 查 。 可 从

https://arxiv.org/pdf/1805.11519.pdf 获得。

[3] R. JAFRI 和 H.R. ARABNIA(2009)

人脸识别技术调查。信息处理系统杂志,

5(2),第 41-68 页。

[4] FAUZI,M.A.(2018)使用支持向量

机和克的印尼评论评级预测。西南交通大

学 学 报 , 53 ( 5 ) 。 可 从

http://jsju.org/index.php/journal/article/view/

237 获得。

[5] KHAN , N.M. , KSANTINI , R. ,

AHMAD , I.S 。 和 BOUFAMA , B 。

(2012)一种用于分类的新型支持向量机

+ 保密协议模型,应用于人脸识别。模式

识别,45(1),第 66-79 页。

[6] LI , W. , LIU , L. 和 GONG , W.

(2011)多目标统一设计作为用于人脸识

别的支持向量机模型选择工具。专家系统

的应用,38(6),第 6689-6695 页。

[7] 刘永辉。和 陈永泰(2007)使用基于

总余量的自适应模糊支持向量机进行人脸

识别。电气工程师学会神经网络交易,18

(1),第 178-192 页。

[8] AHMED,K.,AL-KHATEEB,B. 和

MAHMOOD,M.(2019)一种多目标粒

子群算法,用于混沌和离散路面养护。西

南 交 通 大 学 学 报 , 54 ( 3 ) 。 可 从

http://jsju.org/index.php/journal/article/view/

283 获得。

[9] ZOU , F. , LI , J. 和 MIN , W.

(2019)基于负载平衡和动态预测的分布

式人脸识别。应用科学,9(4),794。

[10] CONNOLLY ,J.F.,GRANGER,E.

和 SABOURIN,R.(2012)一种基于视

频的面部识别的自适应分类系统。情报科

学,192,第 50-70 页。

[11] 魏荣昌,WEIYING W.A.,宜宾 Z.H.,

YONGKANG Z.H. 和 QI L.I.。(2018)基

于改进粒子群算法算法的部分阴影光伏发

电系统的 MPPT 控制。西南交通大学学报,

53

6

http://jsju.org/index.php/journal/article/view/

240 获得。

[12] 魏健,詹建奇,和 项忠(2011)基于

支持向量机和粒子群优化的人脸识别方法。

专家系统的应用,38(4),第 4390-4393

页。

[13] LIU H.,X。G.,DING.G.Y。和 SUN,

Y.B。(2014)基于人类行为的粒子群优

化。科学世界杂志,2014,194706。

[14]

M.H.

ABDULAMEER

D.MOHAMMED,S.A。MO。,M。AL-AZAWI , Y.M 。 AL-MAYALI 和 I.A.

ALAMERI。(2018)基于自适应对立粒

子群优化(奥普索)和支持向量机(支持

向量机)的人脸识别技术。ARPN 工程与

应用科学杂志,13,第 2259-2266 页。

[15] 班萨尔·J.C。,辛格·P·K,萨拉斯瓦

特·M,维尔马·A,贾顿·S.S。和 亚伯拉罕

·A。(2011)粒子群优化中的惯性权重策

略。于:2011 年 10 月在萨拉曼卡召开的

第三届世界自然与生物启发计算世界大会

论文集。新泽西州皮斯卡塔维:电气与电

子工程师协会,第 633-640 页。

[16]

M.H.

ABDULAMEER

S.ABDULLAH , S.N 。 HUDA 和 Z.A.

OTHMAN。(2014)基于自适应加速粒

子群算法的支持向量机。科学世界杂志,

2014,835607。

[17] 冯 Y,滕 G 发,王爱霞,姚 Y 民。

(2007)粒子群优化中的混沌惯性权重。

于:2007 年 9 月,第二届国际创新计算,

信息与控制国际会议论文集,熊本。新泽

西州皮斯卡塔维:电气与电子工程师协会,

第 475-475 页。

[18] AL-HASSAN , W. , FAYEK , M.B.

和 SHAHEEN,S.I.(2006)PSOSA:一

(10)

种用于解决城市规划问题的优化粒子群技

术。于:2006 年 11 月在开罗举行的国际

计算机工程与系统会议论文集。新泽西州

皮斯卡塔维:电气与电子工程师协会,第

401-405 页。

[19] LI。H.R. 和 GAO,Y.L.(2009)具有

减少惯性权重和随机突变的指数的粒子群

优化算法。于:2009 年 5 月在曼彻斯特举

行的第二届信息与计算科学国际会议论文

集。新泽西州皮斯卡塔维:电气与电子工

程师协会,第 66-69 页。

[20] 陈 国强,黄 兴国, 贾嘉 和 闵 忠民

(2006)粒子群优化中的自然指数惯性权

重策略。于:2006 年 6 月,大连,第六届

世界智能控制与自动化大会论文集。新泽

西州,皮斯卡塔维:电气与电子工程师协

会,第 3672-3675 页。

[21] ARUMUGAM,M.S。和 M.V. RAO

(2006)的性能与各种惯性权重变量的粒

子群优化算法的性能计算一类混合系统的

最优控制。自然与社会的离散动力学,

2006,79295。

[22] RONGRONG , S.O 。和 ZILI, C.H.

(2015)基于粒子群算法的磁悬浮系统

PID 控制器的优化设计。 西南交通大学学

50

1

http://www.jsju.org/index.php/journal/article/

view/187 获得。

[23] 中国科学院(2007)卡西亚人脸数据

库 。 [ 在 线 ] 可 从 以 下 网 站 获 得 :

http://biometrics.idealtest.org/ [ 访问时间:

19/10/25]。

[24] A.S. GEORGHIADES , P.N 。

BELHUMEUR 和

D.J. KRIEGMAN 。

(2001)从稀有到很多:可变光照和姿势

下用于面部识别的照明锥模型。电气工程

师 学 会 模 式 分 析 与 机 器 智 能 交 易 , 23

(6),第 643-660 页。

Figure 1. The PSO flowchart
Figure 3. The process of SVM classification
Figure 6. The accuracy performance for the proposed model,  (a) CASIAV5 and (b) YALEB

参照

関連したドキュメント

q-series, which are also called basic hypergeometric series, plays a very important role in many fields, such as affine root systems, Lie algebras and groups, number theory,

Evolution by natural selection results in changes in the density of phenotypes in the adaptive space.. An adaptive trait is a set of values (say height, weight) that a

5 used an improved version of particle swarm optimization algorithm in order to solve the economic emissions load dispatch problem for a test system of 6 power generators, for

In this paper we study BSDEs with two reflecting barriers driven by a Brownian motion and an independent Poisson process.. We show the existence and uniqueness of local and

In this paper, motivated by Yamada’s hybrid steepest-descent and Lehdili and Moudafi’s algorithms, a generalized hybrid steepest-descent algorithm for computing the solutions of

In this paper, we establish a Stroock-Varadhan support theorem for the global mild solution to a d (d ≤ 3)-dimensional stochastic Cahn-Hilliard partial differential equation driven by

(In a very recent preprint, Niethammer and Vel´azquez [9] have obtained a remarkable estimate for the effective potential of a single particle in the supercritical case by taking

In order to demonstrate that the CAB algorithm provides a better performance, it has been compared to other optimization approaches such as metaheuristic algorithms Section 4.2