TELK OMNIKA T elecommunication, Computing, Electr onics and Contr ol V ol. 19, No. 5, October 2021, pp. 1622 1629 ISSN: 1693-6930, accredited First Grade by K emenristekdikti, Decree No: 21/E/KPT/2018 DOI: 10.12928/TELK OMNIKA.v19i5.19566 r 1622 PCA-based dimensionality r eduction f or face r ecognition Md. Ab u Marjan 1 , Md. Rashedul Islam 2 , Md. P alash Uddin 3 , Masud Ibn Afjal 4 , Md. Al Mamun 5 1,3,4 Department of Computer Science and Engineering, Hajee Mohammad Danesh Science and T echnology Uni v ersity , Bangladesh 2 Information T echnology Cell, Hajee Mohammad Danesh Science and T echnology Uni v ersity , Bangladesh 5 Department of Computer Science and Engineering, Rajshahi Uni v ersity of Engineering and T echnology , Bangladesh Article Inf o Article history: Recei v ed Jan 6, 2021 Re vised Mar 19, 2021 Accepted Mar 30, 2021 K eyw ords: Data reduction Dimensionality reduction Eigen analysis F ace recognition Principal component analysis ABSTRA CT In this paper , we conduct a comprehensi v e study on dimensionality reduction (DR) techniques and discuss the mostly used statistical DR technique called principal com- ponent analysis (PCA) in detail with a vie w to addressing the classical f ace recognition problem. Therefore, we, more de v otedly , propose a solution to either a typical f ace or indi vidual f ace recognition based on the principal components, which are constructed using PCA on the f ace images. W e simulate the proposed solution with se v eral train- ing and test sets of ma nually captured f ace images and also with the popular Oli v etti Research Laboratory (ORL) and Y ale f ace databases. The performance measure of the proposed f ace recognizer signifies its superiority . This is an open access article under the CC BY -SA license . Corresponding A uthor: Md. Ab u Marjan Department of Computer Science and Engineering Hajee Mohammad Danesh Science and T echnology Uni v ersity Dinajpur -5200, Bangladesh Email: marjan@hstu.ac.bd 1. INTR ODUCTION Data mining is a w ay for e xtracting or mining kno wledge from lar ge amounts of data [1]-[4]. In de v eloping data mining application, the amount data ta k e n from v arious repositories such as databases, data w arehouse, and W orld W ide W eb (WWW). is typically huge to be either stored or processed. Long time may be required for analyzing comple x data and mining on huge amounts of data. Therefore, it mak es such analysis sometimes impractical or infeasible. Data reduction techniques are traditionally applied to find a reduced representation of the dataset, which is much smaller i n size ensuring the close inte grity of the original data. T o what follo ws, mining on the reduced dataset should be more ef ficient producing the same or almost the same analytical results. The common strate gies for data reduction incl u de data cube aggre g ation, attrib ute subset selection, dimensionality reduction (DR) and numerosity reduction [1]. Recently , the dataset size in terms of number of records and attrib utes i s e xploring v ery rapidly , which prompts the de v elopment of a number of big-data pl atforms, parallel data analytics algorithms and the usage of data DR procedures ef ficiently . In order to handle the real-w orld data ef fecti v ely , the respecti v e dimensionality needs to be reduced in an ef fecti v e (more economic) amount. DR is the study of methods of transformations for reducing the number of dimensions describing the object of high-dimensional data into a meaningful rep- resentation of reduced dimensionality . Theoretically , the reduced representation of dataset should ha v e such a dimensionality that corresponds to the intrinsic dimensionality of the dataset. The intrinsic dimensionality of J ournal homepage: http://journal.uad.ac.id/inde x.php/TELK OMNIKA Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA T elecommun Comput El Control r 1623 dataset means the minimum number of ar gument s needed to account for the observ ed properties of the data. The general objecti v es of DR are to remo v e irrele v ant and redundant data for reducing the manipulation cost and a v oiding data o v er -fitting, and increasing the quality of data for ef ficient data-intensi v e processing tasks, such as pattern recognition, data mining, visualization, database na vig ation, and compression of high-dimensional data. As such, DR of fers an ef fecti v e solution to the di v erse problem of “curse of dimensionality” and fix es other undesired properties of high-dimensional spaces [5]. Mathematically , the DR techniques can be defined as to con v ert a gi v en dataset represented in a n D matrix X consisting of n data v ectors x i ; i = 1 ; 2 ; :::; n with dimensionality D into another dataset Y that has an intrinsic dimensionality d , where d < D , and often d << D . The intrinsic dimensionality of data signifies that the points in dataset X are belonging to or near a manifold with dimensionality d that is implanted in the D dimensional space. In another w ords, the DR methods encode the gi v en dataset X ha ving dimensionality D into a ne w datas et Y with dimensionality d retaining the geometry of the data as much as possible. In general, neither the intrinsic dimensionality d of the dataset X nor the geometry of the data manifold is completely kno wn. Therefore, DR of a dataset is an ill-posed problem that can only be solv ed by assuming certain properties of the data such as its intrinsic dimensionality [5]. There are some DR techniques for the purpose of taking a smaller image and compression and there are some other DR techniques for machine learning purpose (e.g., for better data analysis, classification, statis- tics, and visualization) [6]. In machine learning, dimension reduction is usually concerning with the feature v ectors. In this case, DR techniques can be di vided into tw o cate gories: feature e xtraction and feature selection methods. Feature e xtraction can furt her be di vided into l inear and non-linear methods. The main goal of some methods is to preserv e fidelity with respect to the original data using a certain metric such as mean squared error , and the goal of some other methods is to impro v e the performance of a typical task, such as classifica- tion, prediction, and visualization [7]. Linear feature e xtraction methods include principal component analysis (PCA), f actor analysis, i n de p e nd e nt component analysis (ICA), and linear discriminant analysis (LD A). Non- linear feature e xtraction methods include the front-rank ed techniques such as multidimensional scaling (MDS), Isomap, maximum v ariance unfolding, k ernel PCA etc [5]. Feature selection is di vided into feature ranking and feature subset selection. Feature ranking commonly uses tw o scoring function, such as Eucli d e an distance and correlation and information g ain ratio. On the other hand, the feature subset selection methods are di vided into filter method, wrapper method and embedded method. The filter methods do not use an y learning algorithm [8]. In this paper , after conducting a comprehensi v e study on the DR techniques, we present a f ace recog- nition approach using PCA transformation. W e perform e xperiment using Oli v etti Research Laboratory (ORL) and Y ale f ace databases. The e xperimental results manifest the superiority of the proposed method. The main contrib ution of this paper is listed: i) comprehensi v e study on the DR techniques; ii) technical and mathematical intuitions behind the PCA approach; iii) tw o f ace recognition proposals using PCA data; and i v) performance e v aluation on ORL and Y ale f ace databases. The remainder of this paper is or g anized as follo ws. W e pro vide the technical detail of the PCA method in se ction 2. Then, we discuss the related w orks to ours in section 3. After that, we e xplain the proposed f ace recognition approach in section 4. The e xperiments and results are pro vided in section 5. At last, we summarize and conclude the findings and observ ations in section 6. 2. PRINCIP AL COMPONENT AN AL YSIS The constituent attrib utes of real-w orld dataset re v eal relationships among them. The relati onships are often linear or approximately linear . This mak es the attrib utes amenable to common analysis techniques. One of such techniques is PCA, which rotates the original data to ne w coordinates with a vie w to making the data as flat as possible. PCA is a statistical transformation that identifies patterns in data through detecting the correlation between attrib utes [9]. If there e xists a strong correlation between attrib utes, the attempt to reduce the dimensionality only mak es sense. PCA finds the directions of maximum v ariance in high-dimensional data and then projects it onto a reduced dimensional subspace while retaining most of the information of the original dataset [10]. Mathematically , gi v en a matrix of tw o or more attrib utes, PCA produces a ne w matrix with the same number of attrib utes, called the principal components. Each generated principal component is a linear transformation of the entire original dataset. The measurements of the principal components are calculated in such a w ay that the first principal component holds the maximum v ariance, which can tentati v ely PCA-based dimensionality r eduction for face r eco gnition (Md. Ab u Marjan) Evaluation Warning : The document was created with Spire.PDF for Python.
1624 r ISSN: 1693-6930 be thought as the maximum information. The second principal component is calculated to ha v e the second most v ariance, and, significantly , in a linear sense is uncorrelated with the first principal component. The further principal components, if there are an y , e xhibit decreasing v ariance and are uncorrelated with all other principal components. The steps for the implementation of PCA are illustrated [11]: Step 1: T ak e the whole dataset consisting of d -dimensional samples ignoring the class labels. Step 2: Compute the d -dimensional mean v ector . The mean v ector consists of the means of each v ariable. The mean is the sum of the data points di vided by the number of data points. That is, = A = P n i =1 A i n . The mean is that v alue that is most commonly referred to as the a v erage. The mean v ector is often referred to as the centroid. The v ariance is roughly the arithmetic a v erage of the squared distance from the mean. The v ariance is defined as 2 = s 2 = v ar ( A ) = P n i =1 ( A i A ) 2 n 1 , where A is the mean of the data. Note that the standard de viation ( ) is the square root of the v ariance. Step 3: Compute the co v ariance matrix, alternati v ely , the scatter matrix of the whole dataset. a. Co v ariance matrix: The v ariance-co v ariance matrix consists of the v ariances of the v ariables along the main diagonal and the co v ariances between each pair of v ariables in the other matrix posi- tions. The formula for computing the co v ariance of the v ariables S and T is cov ar ( S ; T ) = P n i =1 ( S i S )( T i T ) n 1 , where S and T denote the means of S and T , respecti v ely . The co v ariance matrix is defined as X = 2 6 6 4 2 11 2 12 ::: 2 1 n 2 21 2 22 ::: 2 2 n ::: ::: ::: ::: 2 n 1 2 n 2 ::: 2 nn 3 7 7 5 Here, 2 ii is the v ariance of each v ariable A i in A, 2 j k is the co v ariance between A i and A k in A. b . Scatter matrix: The scatter matrix is computed as P n i =1 ( A i m )( A i m ) , where m is the mean v ector and it is defined as m = P n i =1 A i n Step 4: Perform eigendecomposition i.e., compute eigen v ectors ( e 1 ; e 2 ; :::; e d ) and corresponding eigen- v alues ( 1 ; 2 ; :::; d ) . The eigen v ectors or principal components determine the directions of the ne w feature space, and the eigen v alues determine their magnitude. Step 5: Sort the eigen v ectors by decreasing eigen v alues and choose k eigen v ectors with the lar gest eigen v alues to form a d k dimensional matrix W , where e v ery column represents an eigen v ector and k is the number of dimensions of the ne w feature subspace with k 6 d . Step 6: Use the d k eigen v ector projection matrix, W to transform the original samples onto the ne w subspace. This can be summarized by the mathematical equation: y = W x , where x is a 1 d - dimensional v ector representing one sample, and y is the transformed 1 k -dimensional sample in the ne w subspace. Alternati v ely , this can be performed as Y = A W (or Y = W A ) , where Y is the transformed n k -dimensional samples in the ne w subspace. 3. RELA TED W ORK Dash et al. [12] presented a PCA based entrop y measure for ranking features and compares with a similar feature ranking method (Relief) in [12]. Maaten, Postma, and Herik ha v e in v estig ated the performances of the nonlinear techniques on artificial and natural tasks, also conduct re vie w and systematic comparison of DR techniques [5]. Spectr al DR methods ha v e e xplained with a short tutorial in the follo wing paper [13]. In re vie w w ork [14], the authors cate gorized the plethora of a v ailable DR methods and illustrated the mathem atical insight behind them. Loog a, Ginnek en, and Duin ha v e proposed a DR technique for image features using the canonical conte xtual correlation projection in [15]. In [16] article, the authors pro vide a comprehensi v e re vie w and comparison of the performance of the principal methods of dimension reduction proposed in the approximate Bayesian computation literature. Silipo, Adae, and Berthold ha v e discussed se v en techniques for DR which are missing v alues, lo w v ariance filter , high correlation filter , PCA, random forests, backw ard feature elimination, and forw ard feature construction in [17]. Joshi and Machchhar [18] conduct a comprehensi v e surv e y on DR methods and proposed a DR met hod that depends upon the gi v en set of parameters and v arying conditions [18]. The authors in v estig ate that recursi v e feature elimination, and genetic and e v olutionary feature weighting and selection gi v e better classification result than PCA [19]. TELK OMNIKA T elecommun Comput El Control, V ol. 19, No. 5, October 2021 : 1622 1629 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA T elecommun Comput El Control r 1625 Se v eral w orks ha v e also b e en conducted on recognition problem based on PCA in v arious w ays. Huang and Y in [20] compare and in v estig ate linear PCA and v arious nonlinear techniques for f ace recognition. Alkandari and Aljaber [21] ha v e presented the importance of PCA to identify the f acial image without human interv ention [21]. Da n dpat and Meher proposed a f ace recognition for impro ving performance using PCA and tw o-dimensional PCA in [22]. PCA in linear discriminant analysis space for f ace recognition has been proposed by Su and W ang [23]. The follo wing paper in v estig ates the performance when tw o DR methods such as self-or g anizing map (SOM) and PCA ha v e been combined [24]. 4. PR OPOSED APPR O A CH T O F A CE RECOGNITION In this paper , after dis cussing the w orking principle of PCA in detail, we propose a solution for f ace recognition problem based the principal components of the trai ning grayscale f ace image matrices. The pro- posal is a customization of v arious principal components-based e xisting classifiers. The main customization is made in case of deri ving the training and test sets, where the images are placed as matrices rather than as v ectors of the traditional approaches and introducing the transpose of the main sets as discussed later . T o implement the proposal, the f ace recognition problem is di vided into tw o cate gories. 4.1. Pr oblem statement-1: Recognition of a typical face Gi v en a ne w image, classify it to “f ace” or “non-f ace” from a set of N original peoples’ f ace images, each image is R pix els high by C pix els wide i.e., the pix el resolution is R C . T o solv e this, we mer ge N training image matrices into a single big matrix by placing one after another . Then, we also place the input image matrix N times one after another to form another big matrix. After that, we tak e the transpose of both big matrices. Subsequently , we apply PCA on the four big matrices and select k eigen v ectors for each. W e then determine the similarity of the normal input big matrix with the normal training big matrix, and transposed input big matrix with the t ransposed training big matrix using selected k features (eigen v ectors). Finally , the decision is tak en based the similarity result. The solution is illustrated with the follo wing steps: Step 1: Input the N original images of size R C . Step 2: F or each of the N images, con v ert the image to a matrix of length (dimension) R C a. Step 2.1: Put all the matrices together in one big image-matrix, T rain1 lik e this: T rain1 = 2 6 6 6 6 4 ImageMatrix1 ImageMatrix2 ::: ::: ImageMatrixN 3 7 7 7 7 5 b . Step 2.2: T ak e the transpose of T rain1 and assign it to another matrix, T rain2 . T rain2 = T r anspose ( T rain1 ) Step 3: F or the ne w image to be classified, a. Step 3.1: Con v ert the image to a matrix of length R C and put it N times together in another big image-matrix, T est1 lik e this: T est1 = 2 6 6 6 6 4 NewImageMatrix NewImageMatrix ::: ::: NewImageMatrix 3 7 7 7 7 5 b . Step 3.2: T ak e the transpose of T est1 and assign it to another matrix, T est2 . T est2 = T r anspose ( T est1 ) [label=.] PCA-based dimensionality r eduction for face r eco gnition (Md. Ab u Marjan) Evaluation Warning : The document was created with Spire.PDF for Python.
1626 r ISSN: 1693-6930 Step 4: F or both big image matrices, a. Step 4.1: Apply PCA b . Step 4.2: Select k eigen v ectors with the highest eigen v alues Step 5: Determine the similarity of the ne w image with the e xisting images using the k e xtracted features i.e., determi ne the similarity of T est1 with T rain1 and T est2 with T rain2 using the k e xtracted features. Step 6: Classify the ne w input image either to “f ace” if the similarity is highest, or to “non-f ace”, other - wise. 4.2. Pr oblem statement-2: Recognition of indi vidual face Gi v en a ne w image, classify it to most similar ima g e(s) from a set of N original f ace images for e ach of the m peoples, each i mage is R pix els high by C pix els wide i.e., the size is R C . T o solv e this, we mer ge N training image matrices for each of the m people into a separate single big matrix by placing one after another . Then, we also place the input image matrix N times one after another to form another big matrix. After that, we tak e the transpose of all big matrices. Subsequently , we apply PCA on all big matrices and select k eigen v ectors for each. W e then determine the similarity of the normal input big matrix with all normal training big matrices, and transposed input big matrix with all transposed training big matrices using selected k features. Finally , the decision is tak en based the similarity result. The solution is illustrated with the follo wing steps: Step 1: Input the N original images of size R C for each of the m peoples. Step 2: F or each of the N images of each of the m peoples, con v ert it into a matrix of length R C Step 2.1: Put the matrices together in a separate big image-matrix, T rain3 lik e this: T rain3 = 2 6 6 6 6 4 ImageMatrix1 ImageMatrix2 ::: ::: ImageMatrixN 3 7 7 7 7 5 Step 2.2: T ak e the transpose of T rain3 and assign it to another matrix, T rain4 . T rain4 = T r anspose ( T rain3 ) Step 3: F or the ne w image to be classified, Step 3.1: Con v ert the image to a matrix of length R C and put it N times together in another big image-matrix, T est3 lik e this: T est3 = 2 6 6 6 6 4 NewImageMatrix NewImageMatrix ::: ::: NewImageMatrix 3 7 7 7 7 5 Step 3.2: T ak e the transpose of T est4 and assign it to another matrix, T est3 . T est4 = T r anspose ( T est3 ) Step 4: F or both big image matrices, Step 4.1: Apply PCA Step 4.2: Select k eigen v ectors with the highest eigen v alues Step 5: Determine the s imilarity of the ne w image with all the e xisting i mages of m peoples using the k e xtracted features i.e., determine the similarity of T est3 with T rain3 and T est4 with T rain4 using the k e xtracted features. TELK OMNIKA T elecommun Comput El Control, V ol. 19, No. 5, October 2021 : 1622 1629 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA T elecommun Comput El Control r 1627 Step 6: Classify the ne w input image to the most probable ima g e(s) with the highest similarity . T o determine the similarity for both problem statement s, first, each eigen v ector in a training set i s subtracted with its corresponding eigen v ector in the testing set. Then the result of each eigen v ector is a v eraged. No w , the ne w instance is classified as “yes”, if the a v erage v alues are near to a threshold v alue, say , that w ould be ideally around zero (0). 5. RESUL TS The proposed method for f ace recognition based on principal components has been implemented in MA TLAB simulation platform. The implemented code has been tested for some common f ace images captured manually . In addition, it has been tested for the tw o popular f ace image databases: ORL and Y ale. In ORL database, there are 10 dif ferent grayscale images of each of 40 distinct subjects. F or some of the subjects, the images were tak en at dif ferent times, and with the v ariation of lighting and f acial e xpressions. All images were captured ag ainst a dark homogeneous background with the subjects in an upright, frontal position. In Y ale database, there are 11 dif ferent grayscale images of each of 15 distinct subjects/indi viduals, one per dif ferent f acial e xpression or configuration. Y ale has e xtensions called Extended Y ale F ace Database A and B. Extended Y ale F ace Database B has 38 subjects/indi viduals and around 64 near frontal images under dif ferent illuminations per subject. F or both databases, there are tw o types of pix el resolution for the images a v ailable: 32 32 and 64 64 . Some images from ORL and Y ale and e xtended Y ale f ace database B are sho wn in Figures 1 (a-c) respecti v ely [25] while T able 1 sho ws the results on dif ferent data distrib utions. (a) (b) (c) Figure 1. F ace databases: (a) sample images from the ORL database, (b) sample images from the Y ale database, and (c) sample images from the e xtended Y ale f ace database B F or the database, the training and testing sets are created in the same manner mentioned abo v e. F or the first problem statement, a random subset of images from e v ery subject w as tak en to form the training set, T rain1 and thus T rain2 . The other images were considered to be the testing set, T est1 and thus T est2 . F or the second problem statement, a random subset of images per e v ery subject w as tak en to form the training set, T rain3 and thus T rain4 . An y of the rest image(s) of the respecti v e subject, upon which the training sets are formed, w as considered to be the testing set, T est3 and thus T est4 . The recognition result of the PCA-based dimensionality r eduction for face r eco gnition (Md. Ab u Marjan) Evaluation Warning : The document was created with Spire.PDF for Python.
1628 r ISSN: 1693-6930 T able 1. Databases and results T ask Database T otal number of samples Samples of indi vidual subject Recognition of a typical f ace ORL 400 40 Theoretically: 0; Practically: around 0 Y ale 2432 38 Theoretically: 0; Practically: around 0 Recognition of Indi vidual F ace ORL 400 40 Theoretically: 0; Practically: around 0 Y ale 2432 38 Theoretically: 0; Practically: around 0 proposed method w as quite acceptable because of, especially , the training sets, T rain2 and T rain4 , which are the transpose of the original training sets, T rain1 and T rain3 respecti v ely . The recognition accurac y can significantly be decreased with the inconsistent images in the training sets. 6. CONCLUSION AND FUTURE W ORK The discussed comprehensi v e o v ervie w of DR techniques and the w orking principle of PCA can be the ingredients for de v eloping a typical image-data mining a p pl ication. The proposed method for f ace recognition based on principal components can, mostly , be used in those applications where a fe w images are enough to train. The proposed a pp r oach can be used for not only f ace recognition b ut also for other kind of objects recognition in the same manner . In future, the proposed technique will be applied on ORL and Y ale databases completely along with other f ace databases and its performance will be compared with the e xisting classifiers based on either machine learning algorithms or other statisti cal approaches. In addition, an adapti v e range of the threshold, to recognize an instance will be determined. REFERENCES [1] J. Han, J. Pei, and M. Kamber , ”Data mining: concepts and techniques, Else vier , 2011. [2] I. H. W itten and E. Frank, ”Data mining: practical machi ne learning tools and techniques with Ja v a implementations, A CM SIGMOD Recor d , v ol. 31, no. 1, pp. 76-77, 2002, doi: 10.1145/507338.507355. [3] M. F . Rabbi et al. , ”Performance Ev aluation of Data Mini ng Classification T echniques for Heart Disease Prediction, American J ournal of Engineering Resear c h (AJER) , v ol. 7, no. 2, pp. 278-283, 2002. [4] S. M. M. Hasan, M. A. Mamun, M. P . Uddin and M. A. Hossain, ”Comparati v e Analysis of Classification Approaches for Heart Disease Prediction, 2018 International Confer ence on Computer , Communication, Chemical, Material and Electr onic Engineering (IC4ME2) , 2018, pp. 1-4, doi: 10.1109/IC4ME2.2018.8465594. [5] L. Maaten, E. Postma, and J. Heri, ”Dimensionality Reduction: A Comparati v e Re vie w , J ournal of Mac hine Learn- ing Resear c h , v ol. 10, no. 1, 2009. [6] A. W . Altaher and S. K. Abbas, “Image processing analysis of sigmoidal Hadamard w a v elet with PCA todetect hidden object, TELK OMNIKA T elecommunicati on Computing Electr onics and Contr ol , v ol. 18, no. 3, pp. 12161223, Jun. 2020, doi: 10.12928/telk omnika.v18i3.13541. [7] S. A. Bak er , H. H. Mohammed, and H. A. Aldabagh, “Impro ving f ace recognition by artificial neural netw ork using principal component analysis, TELK OMNIKA T elecommunication Computing Electr onics and Contr ol , v ol. 18, no. 6, pp. 3357–3364, doi: 10.12928/telk omnika.v18i6.16335. [8] R. Ka vitha and E. Kannan, ”An ef ficient frame w ork for heart disease classification using feature e xtraction and feature selection technique in data mining, 2016 International Confer ence on Emer ging T r ends in Engineering , T ec hnolo gy and Science (ICETETS) , 2016, pp. 1-5, doi: 10.1109/ICETETS.2016.7603000. [9] R. N. Rohmah, B. Handag a, N. Nurokhim, and I. Soesanti, ”A statistical approach on pulmonary tuberculosis detec- tion system based on X-ray image, TELK OMNIKA T elecommunication Computing Electr onics and Contr ol , v ol 17, no. 9, pp. 1474–1482, Jun. 2019, doi: 10.12928/telk omnika.v17i3.10546. [10] O. A. Ade gbola, I. A. Ade yemo, F . A. Semire, S. I. Popoola, and A. A. Atayero, ”A principal component analysis- based feature dimensionality reduction scheme for content-based image retrie v al system, TELK OMNIKA T elecom- munication Computing Electr onics and Contr ol , v ol. 18, no. 4, pp. 1892–1896,Aug. 2020, doi: 10.12928/telk om- nika.v18i4.11176. [11] S. Raschka, “Principal component analysis in 3 simple steps, ”2015. [Online]. A v ail- able:http://sebastianraschka.com/Articles/2015pcain3steps.html [12] M. Dash, H. Liu and J. Y ao, ”Dimensionality reduction of unsupervised data, Pr oceedings Ninth IEEE International Confer ence on T ools with Artificial Intellig ence , 1997, pp. 532-539, doi: 10.1109/T AI.1997.632300. [13] A. Ghodsi, “Dimensionality reduction a short tutorial, Department of Statist ics and Actuarial Science, Uni v ersity of W aterloo W aterloo, Ontario, Canada, pp. 1–25, 2006. TELK OMNIKA T elecommun Comput El Control, V ol. 19, No. 5, October 2021 : 1622 1629 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA T elecommun Comput El Control r 1629 [14] C. O. S. Sorzano, J. V ar g as, and A. P ascual, A surv e y of dimensionality reduction techniques, Mac hine Learning , pp. 1–35, 2014. [15] M. Loog, B. v an Ginnek en, and R. P . W . Duin, “Dimensionality reduction of image features using the canonical conte xtual correlation projection, P attern Reco gnition , v ol. 38, no. 12, pp. 2409-2418, 2005, doi:10.1016/j.patcog.2005.04.011. [16] M. G. B. Blum, M. A. Nunes, D. Prangle, and S. A. Sisson, A Comparati v e Re vie w of Dimension Reduction Methods in Approximate Bayesian Computation, Statistical Science , v ol. 28, no. 2, pp. 189–208, May2013, doi: 10.1214/12-STS406. [17] R. Silipo, I. Adae, and A. H. M. Berthold, “Se v en techniques for dimensionality reduction, Open for Inno v ation KNIME, 2014. [Online]. A v ailable: https://www .knime.com/sit es/def ault/files/inline- images/knime se v entechniquesdatadimreduction.pdf [18] S. K. Joshi and S. Machchhar , ”An e v olution and e v aluation of dimensionality reduction techniques A comparati v e study , 2014 IEEE International Conf er ence on Com putational Intellig ence and Computing Resear c h , 2014, pp. 1-5, doi: 10.1109/ICCIC.2014.7238538. [19] W . Nick, J. Shelton, G. Bullock, A. Esterline and K. Asamene, ”Comparing dimensionality reduction techniques, SoutheastCon 2015 , 2015, pp. 1-2, doi: 10.1109/SECON.2015.7132997. [20] W . Huang and H. Y in, ”Linear and nonlinear dimensionality reduction for f ace recognition, 2009 16th IEEE Inter - national Confer ence on Ima g e Pr ocessing (ICIP) , 2009, pp. 3337-3340, doi: 10.1109/ICIP .2009.5413898. [21] A. Alkandari and S. J. Aljaber , ”Principle Component Analysis algorithm (PCA) for image recognition, 2015 Second International Confer ence on Computing T ec hnolo gy and Information Mana g ement (ICCTIM) , 2015, pp. 76-80, doi: 10.1109/ICCTIM.2015.7224596. [22] S. K. Dandpat and S. Meher , ”Performance impro v ement for f ace recognition using PCA and tw o-dimensional PCA, 2013 International Confer ence on Computer Communication and Informatics , 2013, pp. 1-5, doi: 10.1109/IC- CCI.2013.6466291. [23] H. Su and X. W ang, ”Principal Component Analysis in Linear Discriminant Analysis Space for F ace Recognition, 2014 5th International Confer ence on Digital Home , 2014, pp. 30-34, doi: 10.1109/ICDH.2014.13. [24] D. K umar , C. S. Rai and S. K umar , ”F ace Recognition using Self-Or g anizing Map and Principal Component Analysis, 2005 Internati onal Conference on Neural Netw orks and Brain, 2005, pp. 1469-1473, doi: 10.1109/IC- NNB.2005.1614908. [25] D. Cai, “F our f ace databases in matlab format. [Online]. A v ailable: http://www .cad.zju.edu.cn/home/dengcai/Data/F aceData.html PCA-based dimensionality r eduction for face r eco gnition (Md. Ab u Marjan) Evaluation Warning : The document was created with Spire.PDF for Python.