Inter national J our nal of Electrical and Computer Engineering (IJECE) V ol. 10, No. 4, August 2020, pp. 4080 4092 ISSN: 2088-8708, DOI: 10.11591/ijece.v10i4.pp4080-4092 r 4080 Local featur e extraction based facial emotion r ecognition: A sur v ey Khadija Slimani 1 , Mohamed Kas 2 , Y oussef El merabet 3 , Y assine Ruichek 4 , Rochdi Messoussi 5 1,2,3,5 Laboratory Systems of T elecommunication and Decision Engineering (LASTID), F aculty of Sciences, Ibn T of ail Uni v ersity , Morocco 2,4 CIAD UMR 7533, Uni v ersit ´ e Bour gogne Franche-Comt ´ e, UTBM, France Article Inf o Article history: Recei v ed Oct 2, 2019 Re vised Feb 25, 2020 Accepted Mar 4, 2020 K eyw ords: Basic emotion Features e xtraction Image processing Machine learning ABSTRA CT Notwithstanding the recent technological adv ancement, the identification of f acial and emotional e xpressions is s till one of the greatest challenges scientists ha v e e v er f aced. Generally , the human f ace is identified as a composition made up of te xture s arranged in micro-patterns. Currently , there has bee n a tremendous increase in the us e of Lo- cal Binary P attern based te xture algorithms which ha v e in v ariably been identified to being essential in the completion of a v ariety of tasks and in the e xtraction of essen- tial attrib utes from an image. Ov er the years, lots of LBP v ariants ha v e been literally re vie wed. Ho we v er , what is left is a thorough and comprehensi v e analysis of their independent performance. This research w ork aims at filling this g ap by performing a lar ge-scale performance e v aluation of 46 recent state-of-the -art LBP v ariants for f acial e xpression r ecognition. Extensi v e e xperim ental results on the well-kno wn challenging and benchmark KDEF , J AFFE, CK and MUG databases tak en under dif ferent f acial e xpression conditions, indicate that a number of e v aluated state-of-the-art LBP-lik e methods achie v e promising results, which are better or competiti v e than se v eral re- cent state-of-the-art f acial recognition systems. Recognition rates of 100%, 98.57%, 95.92% and 100% ha v e been reache d for CK, J AFFE, KDEF and MUG databases, respecti v ely . Copyright © 2020 Insitute of Advanced Engineeering and Science . All rights r eserved. Corresponding A uthor: Slimani Khadija, Laboratory Systems of T elecommunication and Decision Engineering (LASTID), F aculty of Sciences, Ibn T of ail Uni v ersity , K enitra, Morocco. Email: slimani.khadija@uit.ac.ma 1. INTR ODUCTION W ith the de v elopment of artificial intelligence and pattern recognition, computer based f acial e xpres- sion recognition has attracted man y researchers in the domain of computer vision. Se v eral studies ha v e sho wn that the f acial e xpression contrib utes to better understand the con v ersations [1, 2], and it helps to e xpress the indi vidual’ s internal emotions, also, it is considered as the main modality for human communication. Recent progresses in psychology and neuroscience fields gi v e a more positi v e interpretation of the emotions role in human beha vior [3]. The f acial emotion recognition system resides of three i mportant steps; f ace de- tection, feature e xtraction and classification. By taking image or series of images as input, the most important step is feature e xtract ion that all o ws to descri be the input images and calculate their characteristic v ector using a gi v en operator . Indeed, e xtracting poor features in v olv es producing poor recognition quality e v en with the use of best classifiers. Because of the e xceptional e xhibition of LBP based techniques, the y ha v e de v eloped as one of the most unmistakable local image descriptors. Although initially intended for te xture analysis [4], the LBP descriptor has gi v en e xcellent outcomes in dif ferent applications because of its in v ariance to monotonic global grayle v el changes, furthermore, its better resistance ag ainst brightening changes property in real-w orld J ournal homepage: http://ijece .iaescor e .com/inde x.php/IJECE Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 r 4081 applications including f ace recognition. Another equally important property is its computational ef fortlessness and the lo w length of its histogram v ector , which mak e it ready to e xamine images in challenging real-time settings. The achie v ement of the LBP in numerous applications concei v ed an of fspring of an immense number of LBP v a riations, which ha v e been proposed and k eep on being proposed. W ithout a doubt, since Ojala’ s w ork [4] and because of its adaptability and ef fecti v eness, the general LBP-lik e w ay of thinking has demonstrated e xtremely well kno wn, and an e xtraordinary assortment of LBP v ariations ha v e been proposed in the writing to impro v e discriminati v e po wer , rob ustness, and appropriateness of LBP . The main o bj ecti v e of this study is to perform a lar ge scale performance e v aluation for f acial emotion recognition, assessing 46 recent state-of-the- art te xture features, on four widely-used benchmark databases. Performance of the adopted f acial e xpression recognition system coupled with the best e v aluated te xture descriptor on each dataset is compared ag ainst those of state-of-the-art approaches. W e disclose in the e xperimental section the f act that some descriptors originally proposed for applications other than f acial emotional recognition allo w outperforming se v eral recent state-of- the-art systems. The remaining sections of this research w ork are arranged in the follo wing w ay: Section 2. re vie ws the traditional LBP operator as well as some of its recent and popular v ariants. Section 3. re vie ws the fe w e xisting surv e ys on te xture descriptor based classification a nd recognition as well as the e v aluated state- of-the art LBP-lik e methods. Section 4. pro vides detailed e xplanation on the results of t he e xperiments while comparing the performances of the best performing descriptors on each tested datasets with those of recent state-of-the-art f acial emotional recognition systems. Finally , section 5. dra w this paper to a close by proposing some future research perspecti v es. 2. BRIEF REVIEW OF EXISTING METHODS The original Local Binary P atter n (LBP) operator proposed by Ojala et al [4] , which consists in coding the pix el-wi se information in an image, is a po werful te xture analysis descriptor . It aims to search micro-te xtons in local re gions. The v alue I p of the pix els in a 3 3 grayscale image patch around the central pix el I c are turned into binary v alues (0 or 1) by comparing them with I c (v alue of the central pix el). The obtained binary numbers are encoded to characterize a local structure pattern and then the code is transformed into decimal number . Once a LBP code of each pix el is obtained, a histogram is b uilt to represent the te xture image. F or a 3 3 neighborhood, the definition of the k ernel function of LBP operator is gi v en in (cf. Eq (1)), where I p (p 2 f 1, 2, ..., P g ) signifies the gray le v els of the peripheral pix els, P corresponds to the number of neighboring pix els (P=8) and ' ( ) is the Hea viside step function (cf. Eq (1)). LBP ( I c ) = P = 8 X p = 1 ' ( I p I c ) 2 p 1 ; ' ( z ) = 1 ; z 0 0 ; z < 0 (1) Local binary patterns by neighborhoods (nLBPd) operator [5] consists in encoding the relationship between each pair of the peripheral pix els I 0 , I 1 , I 2 , ..., I 7 around the central pix el I c in a 3 3 square neighborhood. The pairs of pix els are compared with sequential neighbors or within neighbors possesing a distance length d. The k ernel function of nLBPd code is defined by (cf. Eq. (2)). When d=1, the binary code of the central pix el I c is gotten as belo w (Eq. (3)): nLBP d ( I c ) = P 1 X p =0 ' ( I p ; I ( p + d mod P ) ) 2 p (2) I c = ' ( I 0 > I 1 ) ; ' ( I 1 > I 2 ) ; ' ( I 2 > I 3 ) ; ' ( I 3 > I 4 ) ; ' ( I 4 > I 5 ) ; ' ( I 5 > I 6 ) ; ' ( I 6 > I 7 ) ; ' ( I 7 > I 8 ) (3) The pr ocedur e of Local Gr aph Structur e (LGS) descriptor introduced by Ab usham et al. [6] is to e xploit the d om inant graph process in order to encode the spatial data for an y pix el in the image. LGS is based on local graph structures in local graph neighborhood. The graph structure of LGS represents more left-handed neighbor pix els than right-handed ones. T o o v ercome this defect, Extended Local Graph Structure (ELGS) operator is proposed [7]. The procedure for ELGS is based on using the LGS te xture descriptor to b uild tw o descriptions (horizontally and v ertically) and then combine them into a global description. Local featur e e xtr action based facial... (Slimani khadija) Evaluation Warning : The document was created with Spire.PDF for Python.
4082 r ISSN: 2088-8708 3. EV ALU A TED ST A TE-OF-THE-AR T LBP V ARIANTS The pioneering LBP w ork [4] and its success in numerous computer vision problems and a p pl ications has inspired the de v elopment of great number of ne w po werful LBP v ariants. LBP descriptor is adaptable to suit in man y dif ferent applications requirements. Indeed, after Ojala’ s w ork, e.g., Heikkila et al [8], se v eral modifications and e x t ensions of LBP ha v e been de v eloped with the aim to increase its rob ustness and discrimi- nati v e po wer . These e xtensions and modifications of LBP , de v eloped usually i n conjunction with their intended applications (see T able 1), focus on se v eral aspects of the LBP method such as, Quantization to multiple le v el via thresholding; sampeling local feature v ectors and pix el patterns with some neighborhood topology; com- bining multiple complementary features within LBP-lik e and with non-LBP descriptors for both images and videos and finally , re grouping and mer ging patterns to increase distincti v eness. T able 1. Summary of te xture descriptors tested. Ref Y ear Complete name Abbre viation Application [4] 2002 Local Binary P attern LBP T e xture classification [9] 2003 Simplified T e xture Unit + STU+ T e xture classification [10] 2004 Gradient te xture unit coding GTUC T e xture classification [11] 2005 Dif ference Symmetric Local Graph Structure DSLGS Finger v ein recognition [8] 2006 Center -Symmetric Local Binary P atterns CSLBP T e xture classification [12] 2008 Centralized Binary P attern CBP F acial e xpression recognition [13] 2010 Local T ernary P atterns L TP F ace recognition [14] 2010 Directional Binary Code DBC F ace recognition [15] 2010 Impro v ed Local T ernary P atterns IL TP Medical image analysis [16] 2010 Local Directional P attern LDP F ace recognition [17] 2011 Binary Gradient Contours (1) BGC1 T e xture classification [17] 2011 Binary Gradient Contours (2) BGC2 T e xture classification [17] 2011 Binary Gradient Contours (3) BGC3 T e xture classification [18] 2011 Center -Symmetric Local T ernary P atterns CSL TP Feature description [18] 2011 Extended Center -Symmetric Local T ernary P atterns eCSL TP Image retrie v al [19] 2011 Impro v ed Local Binary P atterns ILBP F ace detection [6] 2011 Local Graph Structure LGS F ace recognition [20] 2012 Local Maximum Edge Binary P atterns LMEBP Image retrie v al [16] 2013 Impro v ed binary gradient contours (1) IBGC1 T e xture classification [21] 2013 Local Directional Number P attern LDN F ace e xpression analysis [22] 2013 Local Gray Code P attern LGCP F ace e xpression analysis [23] 2013 Rotated Local Binary P attern RLBP T e xture classification [24] 2015 Adapti v e Extended Local T ernary P attern AEL TP T e xture classification [5] 2015 Directional Local Binary P atterns dLBP T e xture classification [5] 2015 Local Binary P atterns by neighborhoods nLBPd T e xture classification [25] 2015 Maximum Edge Position Octal P attern MMEPOP Image retrie v al [26] 2015 Multi-Orientation W eighted Symmetric Local Graph Structure MO W -SLGS Finger v ein recognition [27] 2015 Orthogonal Symmetric Local T ernary P attern OSL TP Image re gion description [26] 2015 Symmetric Local Graph Structure SLGS Finger v ein recognition [28] 2015 eXtended Center -Symmetric Local Binary P attern XCS LBP T e xture classification [29] 2016 Adapti v e Local T ernary P atterns AL TP F ace recognition [29] 2016 Center -Symmetric AL TP CSAL TP F ace recognition [30] 2016 Diagonal Direction Binary P attern DDBP F ace recognition [7] 2016 Extended Local Graph Structure ELGS T e xture classification [31] 2016 Local Extreme Sign T rio P atterns LESTP Image retrie v al [32] 2016 Quad Binary P attern QBP T ar get tracking [31] 2016 Sign Maximum Edge Position Octal P attern SMEPOP Image retrie v al [33] 2016 Complete Eight Local Directional P atterns CELDP F ace recognition [34] 2017 Centre Symmetric Quadruple P attern CSQP F acial image recognition and retrie v al [35] 2017 Local Directional Binary P atterns LDBP T e xture classification [36] 2017 Local neighborhood dif ference pattern LNDP Natural and te xture image retrie v al [37] 2017 Local Quadruple P attern LQP A T F acial image recognition and retrie v al [38] 2018 Local Diagonal Extrema Number P attern LDENP F ace recognition [39] 2018 Local Conca v e-and-Con v e x Micro-Structure P atterns LCCMSP T e xture classification [40] 2018 Local Directional T ernary P attern LDTP T e xture classification [41] 2018 Repulsi v e-and-Attracti v e Local Binary Gradient Contours RALBGC T e xture classification Int J Elec & Comp Eng, V ol. 10, No. 4, August 2020 : 4080 4092 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 r 4083 There are se v eral researches reported in the literature that are de v oted to surv e ying LBP and its v ari - ants. One can cite: (a) Hadid et al . [42] re vie wed 13 LBP v ariants and pro vided a comparati v e analysis on tw o dif ferent problems which are gender and te xture classification. (b) The w ork of Fernandez et al. [43] attempted to b uild a general frame w ork for te xture e xamination that the authors refer to as histograms of equi v alent patterns (HEP). A set of 38 LBP v ariants and non LBP strate gies are e x ecuted and e xperimentally assessed on ele v en te xture datasets. (c) Huang et al. [44] displayed a surv e y of LBP v ariants in the application re gion of f acial image processing. Ho we v er , there is no e xperimental study of the LBP strate gies themselv es. (d) Nanni et al. [45] e xamined the performance of LBP based te xture descriptors in a f airly specific and narro w application, which consists in classifying cell and tissue images of v e datasets. (e) Michael Bereta et al. [46] highlighted man y types of local descriptors including local binary patterns and their combination with Gabor filters. The y e xamined only 14 LBP v ariants on FERET database. (f) Lumini et al. [47] e v aluated the ef fecti v eness of LBP , HOG, POEM, MBC, HASC, GOLD, RICLBP , and CLBP descriptors. Each of these feature e xtraction methods is carried out only on tw o datasets: FERET and the Labeled F aces in the W ild (LFW). (g) Liu et al. [48] pro vided a systematic re vie w of LBP v ariants while re grouping them into dif ferent cat- e gories. 40 te xture features including thirty tw o LBP-lik e descriptors and eight non-LBP methods are e v aluated and compared on thirteen te xture datasets. (h) Slimani et al. [49] re vie wed the performance of 22 state-of-the-art LBP-lik e descriptors and some of its recent v ariations and pro vides a comparati v e analysis on f acial e x pr ession recognition problem using tw o benchmark databases. It can be infer red that there is a limited number of state-of-the-art published w orks which are de v oted to surv e y LBP-lik e methods in te xture and f ace recognition and in particular f acial emotion recognition which is practically none xistent. Note that, most of these w orks remain limited in terms of num b e r of LBP-lik e de- scriptors re vie wed and tested datasets, suf fer from lack of recent LBP v ariants and some of them do not include e xperimental e v aluation. Since no broad assessment has been performed on an incredible number of LBP v ari- ations, and considering recent rapid increase in the number of publications on LBP-lik e descriptors, this paper aims to pro vide such a comparati v e study in f acial emotion recognition problem and of fers a more up-to-date introduction to the area. F or that, 46 recent state-of-the-art LBP v ariants are e v aluated and compared o v er four challenging representati v e widely-used f acial e xpression databases. The performance of the best te xture de- scriptor on each dataset is also composed to those of state-of-the-art f acial emotion recognition systems. Note that for the descriptors, we utilized the original source code if it is freely accessible; otherwise we ha v e b uilt up our o wn implementation. The e v aluated state-of-the-art te xture descriptors and their intended applications are summarized in T able 1. 4. EXPERIMENT AL RESUL TS AND DISCUSSION In t his section, the state-of-the-art LBP v ariants summarized in T able 1 are e xtensi v ely e v aluated and compared o v er four publicly a v ailable f acia l e xpression datasets (see section 4.2.). In addition, performance of the best performing method on each dataset has been compared ag ainst those of recent state-of-the-art f acial emotion recognition systems. The follo wing subsections describe: 1) the e xperimental configuration; 2) the datasets considered in the e xperiments, 3) the obtained results and 4) comparisons with other e xisting approaches. 4.1. Experimental configuration In order to systematical ly e v aluate the performance of the tested m ethods, we setup a comparati v e analysis through a supervised image classification task. Similar to most state-of-the-art f acial e xpression recog- nition systems, the adopted system, sho wn in Figure 1, in v olv es se v eral s teps including 1) image processing to alter and resize f aces to ha v e a common resolution; 2) feature e xtraction using the e v aluated LBP v ari- ants; 3) histogram v ector calculation. In this step, in order to incorporate more spatial information into the final feature v ectors, the obtained feature images were spatially di vided into multiple non-o v erlapping re gions and histograms were e xtracted from each re gion. F or e xample, the LBP code map is di vided into m n non- o v erlapping sub-re gions, from each of which a sub-histogram feature is e xtracted and is normalized to sum Local featur e e xtr action based facial... (Slimani khadija) Evaluation Warning : The document was created with Spire.PDF for Python.
4084 r ISSN: 2088-8708 one. By concatenating these re gional sub-histograms into a single v ector , a final LBP based f ac ial emotion representation is obtained; and 4) image class ification using the SVM classifier . In this step, the images of each dataset are preliminarily di vided into a random split containing tw o sub-sets, one for the training and the other for testing. In the e xperiments, we tackled the 7-e xpression classification problems and o v erall results are computed as the a v erage of the per -class accuracies and not the a v erage accurac y of all samples, which a v oids biasing to w ard e xpressions with more samples in the databases. Figure 1. Outline of the adopted f acial emotion recognition system. 4.2. T ested datasets In our e xperiments, we used four benchmark databases; the Cohn Kanade (CK), the Japanese Fe male F acial Expression (J AFFE), the Karolinska Directed Emotional F aces (KDEF) and the Multimedia Understand- ing Group (MUG) databases. The main characteristics of each database are described herein belo w . The four datasets include f acial e xpressions of six basic emotions; Anger , Disgust, Fear , Happiness, Sadness, Surprise and the neutral f acial e xpression. (a) The J AFFE database [50] contains 213 f acial e xpression images from 10 Japanese females where e v ery subject e xpres ses three times the se v en f acial e xpressions. The images ha v e a resoluti on of 256x256 pix els. (b) The CK database [51] includes 2105 digitized image sequences (video) from 182 adults ranging from 18 to 30 years old. Each image has a resolution of 640x490 pix els with eight-bit accurac y for gray scale v alues. (c) The KDEF dataset [52] contains tw o sessions of multi-vi e w posed f acial e xpression images from 70 am- ateur actors, with age ranging from 20 to 30 years old. The database has totally 4900 2D images of se v en human f acial e xpressions of emotions. The images ha v e a resolution of 562x762 pix els, and each of the se v en f acial e xpressions is acquired from v e dif ferent angles -90 , -45 , 0 , 45 , 90 . (d) The MUG Database [53] contains 86 subjects, where 51 are males and 35 are females. All subjects are between 20 and 35 years old. Only 52 subject images are a v ailable for usage with this database. F or each e xpression, a total of 50 to 160 images are e xisting. The images ha v e a resolution of 896x896 pix els. 4.3. Results and analysis T ables 2 and 3 report the a v erage accurac y of each tested descript or obtained on CK, J AFFE, KDEF and MUG Databases. The first column consists of the name of the descriptor along with the parameter used if that concerns a parametric descriptor . The other columns c o nc ern the abbre viation of em otion cate gories that we tested and the accurac y obtained; NE: NEUTRAL, HA : HAPPY , FE : FEAR, SA: SAD, AN: ANGR Y , DI: DISGUST , SU: SURPRISE, Acc: Accurac y . 4.3.1. P erf ormance analysis on Cohn-Kanade (CK) Database F or this database, we used a subset of 10 sequences that reflect only the samples e xpressing the se v en cate gories of emotions, and then we selected the four latest frames of each sequence that ha v e the highest e xpression intensity . The optimal number of non-o v erlapping sub-re gions to compute the histogram features is 14x14 for al l the tested descriptors. F or each emotion e xpression, tw o images are used as training set and the tw o others are used as test set. T able 2 illustrates the obtained e xperimental results for the basic emotion Int J Elec & Comp Eng, V ol. 10, No. 4, August 2020 : 4080 4092 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 r 4085 recognition recorded on CK dataset using the 46 e v aluated state-of-the-art te xture descriptors. It can be inferred that almost all the tested descriptors produce good results on CK dataset where their a v erage accurac y is abo v e 96%. T weenty-se v en LBP-met hods lik e RALBGC, BGC1, BGC2, BGC3, dLBP , ELGS manage successfully to dif ferentiate all classes perfectly (a v erage accurac y equal to 100%), lea ving then, essentially , no room for impro v ement. Note that, all the e v aluated descriptors reached a score of 100% for ”Happ y” and ”Surprise” classes. 4.3.2. P erf ormance analysis on J AFFE Database In this second e xperiment, each emotion in J AFFE database is desi gnated i nto 10 femal es with three samples. One image is tak en for each person and for each emotion e xpression in the test, making a total of 70 samples in the testing phase while the remaining 140 samples depict the training set. All f aces are preprocessed to align them into a canonical images with a resolution of 128x128. The histograms are produced on the feature images spatially di vided into 12x12 non-o v erlapping sub-re gions. It is apparent from T able 2 that DSLGS, ELGS and SLGS operators yield the highest a v erage rate as the y reached a score of 98.57%. Then, come the eight descriptors: BGC2, CSLBP , dLBP , ILBP , LCCMSP , LDENP , LGCP and OS L TP which reached a recognition rate of 97.14%. It can be noticed that se v eral tested LBP-lik e descriptors ha v e perf ectly recognized some classes by getting the accurac y of 100%. Note that there is a significant performance drop for all the tested descriptors on the class of ”sadness” where the reached accurac y is in the range [50%, 90%]. It also emer ges from T able 2 that some methods lik e CSAL TP , GTUC and LMEBP produce the w orst performance on almost all the classes where their accurac y is sometime belo w 70%. W e w ould also point out that although parametric methods lik e eCS L TP , IL TP , GTUC, AEL TP are re g arded as ”optimized” since their parameter v alues are tuned during the e xperiment, their performance is mark edly weak er than the non-parametric ones. 4.3.3. P erf ormance analysis on KDEF database W e choose the images of both sessions for each subject and only the vie w angle 0 is considered. The subset contains 70 subjects, each one e xpresses tw o times the se v en emotion cate gories. Thus , in total we use 980 images. W e altered the sizes of all the f aces of KDEF database into a steady sized template, which ha v e the same resolution of 256x256 and the f aces were then split into 14x14 blocks for re gion-based feature e xtraction. Each subject e xpress tw o times the se v en cate gories, so we selected one f acial image per subject for training phase and the other one for test phase. It is apparent from T able 3 that the LGS operator is rank ed as the top 1 descriptor in KDEF database as it achie v es a recognition rate of 95.92%, with perfect recognition (100%) of happ y and neutral cate gories, follo wed by DSLGS, SLGS and LBP descriptors which reached a score of 95.31%. Then, come se v en de- scriptors lik e BGC2, BGC3, CSLBP , dLBP , ELGS, ILBP and LQP A T which allo wed to achie v e accuracies between [94.08% - 94.90%]. Then tweenty-six LBP-methods attained accuracies between [90.20% - 93.88%] where three descriptors RLBP , BGC1 and SMEPOP reached 93.88% and tw o descriptors MMEPOP and DBC attained 90.20% and 90.41%, respecti v ely . Accuracies between [80.61% - 86.53%] were achie v ed by eight LBP-lik e me thods in which 80.61% w as achie v ed by AL TP and 86.53% by XCS LBP . W e can o bs erv e from T able 3 that the w orst performance of 59.39% w as attained by CSAL TP descriptor . 4.3.4. P erf ormance analysis on Multimedia Understanding Gr oup (MUG) Database W e ha v e used 924 f acial e xpression im ages, i.e., 132 images for each f acial e xpression. All f a ces were altered and resized to ha v e a common resolution of 256x256. Then, the y were split into 18x18 blocks for re gion-based feature e xtraction. F or this e xperiment, in each emotion cate gory , we used four images per subject, tw o for training phase and tw o for test phase. T able 3 g athers the obtained e xperimental results. Clearly , it can be observ ed that eight of the test ed descriptors ELGS, LDTP , LDENP , LGCP , LNDP , L TP , LQP A T and SMEPOP manage to dif ferentiate all classes perfectly 100% in accurac y lea ving then, no room for impro v ement. In addition, thirty-one LBP-lik e methods gi v e accuracies between [99.03% - 99.68%], LBP attained 98.73%, DBC reached 98.05%, XCS LBP got 97.40% and finally , GTUC attained an accurac y of 97.08%. As we can observ e, all tested methods obtain v ery promising results on the MUG dataset, e xcpect three state-of-the-art methods AEL TP , LMEBP and CSAL TP attained the lo west accuracies comparing with the other methods t ested. The undermost accurac y of 71.43% w as achie v ed by CSAL TP . Then an accurac y of 84.09% w as attained by AEL TP and finally 89.94% w as obtained when testing LMEBP method. Local featur e e xtr action based facial... (Slimani khadija) Evaluation Warning : The document was created with Spire.PDF for Python.
4086 r ISSN: 2088-8708 T able 2. Experiments Results on CK and J AFFE Databases Cohn Canade Database J AFFE Database NE HA FE DI AN SA SU Acc NE HA FE DI AN SA SU Acc LDTP 100 100 100 95 95 100 100 98.57 90 90 80 90 60 70 100 82.86 RALBGC 100 100 100 100 100 100 100 100 90 100 80 90 80 80 100 88.57 RLBP 100 100 100 100 100 100 100 100 90 90 90 80 90 80 100 88.57 CELDP 100 100 100 100 100 100 100 100 90 80 80 80 100 80 100 87.14 AEL TP f 1 g 95 100 100 100 95 95 100 97.86 80 80 90 80 90 80 100 85.71 AL TP f 0.006 g 90 100 100 100 95 95 100 97.14 100 100 90 90 100 80 100 94.29 BGC1 100 100 100 100 100 100 100 100 90 90 80 100 100 80 100 91.43 BGC2 100 100 100 100 100 100 100 100 100 100 100 100 100 80 100 97.14 BGC3 100 100 100 100 100 100 100 100 100 90 90 100 100 80 100 94.29 CBP 1 100 100 100 90 100 90 100 97.14 100 90 90 100 100 90 100 95.71 CSAL TP f 0.006 g 100 100 100 100 100 95 100 99.29 70 90 80 80 50 60 100 75.71 CSLBP f 1 g 100 100 100 100 100 100 100 100 100 100 100 100 100 80 100 97.14 CSL TP f 1 g 100 100 100 100 100 100 100 100 100 100 90 100 90 80 100 94.29 CSQP 100 100 100 100 100 100 100 100 100 90 100 90 100 80 100 94.29 DBC f 45 g 100 100 100 95 95 100 100 98.57 100 100 90 90 90 90 100 94.29 DDBP 100 100 100 100 95 100 100 99.29 90 90 100 100 100 80 100 94.29 dLBP f 45 g 100 100 100 100 100 100 100 100 100 100 100 90 100 90 100 97.14 DSLGS 100 100 100 100 100 100 100 100 100 100 100 100 100 90 100 98.57 eCS L TP f 1 g 100 100 100 90 95 100 100 97.86 100 100 80 90 90 90 100 92.86 ELGS 100 100 100 100 100 100 100 100 100 100 100 100 100 90 100 98.57 GTUC f 2 g 100 100 95 95 100 100 100 98.57 100 90 60 70 80 50 80 75.71 IBGC1 100 100 100 100 100 100 100 100 90 90 70 90 90 70 100 85.71 ILBP f 1 g 100 100 100 100 100 100 100 100 100 90 100 100 100 90 100 97.14 IL TP f 1 g 95 100 100 100 95 95 100 97.86 90 100 80 90 80 80 100 88.57 LBP 100 100 100 100 95 100 100 99.29 100 100 100 90 90 80 100 94.29 nLBPd f 1 g 100 100 100 100 100 100 100 100 100 90 80 100 100 80 100 92.86 LCCMSP 100 100 95 90 95 95 100 96.43 100 90 100 100 100 90 100 97.14 LDBP 100 100 100 100 100 100 100 100 100 90 80 100 100 80 100 92.86 LDENP 100 100 100 100 100 100 100 100 100 100 100 100 100 80 100 97.14 LDN 100 100 100 100 100 100 100 100 100 100 90 100 100 80 100 95.71 LDP f 1 g 100 100 100 100 100 100 100 100 100 100 90 100 100 70 100 94.29 LESTP 10 100 100 100 100 95 100 100 99.29 90 100 90 90 100 80 100 92.86 LGCP 100 100 100 100 100 100 100 100 100 100 100 100 100 80 100 97.14 LGS 100 100 100 100 100 100 100 100 100 100 90 100 100 80 100 95.71 LMEBP 100 100 100 90 95 100 100 97.86 60 90 70 90 50 60 80 71.43 LNDP 100 100 100 100 95 100 100 99.29 90 100 100 100 100 80 100 95.71 L TP f 1 g 90 100 100 100 95 95 100 97.14 90 100 90 90 100 80 100 92.86 LQP A T 100 100 100 100 100 100 100 100 90 100 100 100 100 80 100 95.71 MMEPOP 100 100 100 100 100 100 100 100 100 100 90 90 100 80 100 94.29 MO W SLGS 100 100 100 100 100 100 100 100 100 90 90 100 100 80 100 94.29 OS L TP f 1 g 100 100 100 100 100 100 100 100 100 100 100 100 100 80 100 97.14 QBP f 1 g 100 100 100 95 100 100 100 99.29 100 100 90 100 100 70 100 94.29 SLGS 100 100 100 100 100 100 100 100 100 100 100 100 100 90 100 98.57 SMEPOP 100 100 100 100 100 100 100 100 90 100 100 90 100 80 100 94.29 STU+ f 1 g 100 100 100 95 100 100 100 99.29 100 100 80 100 100 70 100 92.86 XCS LBP 100 100 100 100 95 100 100 99.29 90 100 90 70 90 80 100 88.57 Int J Elec & Comp Eng, V ol. 10, No. 4, August 2020 : 4080 4092 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 r 4087 T able 3. Experiments Results on KDEF and MUG Databases KDEF Database MUG Database AN DI FE HA NE SA SU Acc AN DI FE HA NE SA SU Acc LDTP 82.86 91.43 90 95.71 95.71 91.43 95.71 91.84 100 100 100 100 100 100 100 100 RALBGC 84.29 87.14 91.43 100 97.14 91.43 98.57 92.86 100 100 100 100 100 100 97.73 99.68 RLBP 88.57 92.86 91.43 97.14 98.57 92.86 95.71 93.88 100 100 100 100 97.73 100 100 99.68 CELDP 87.14 87.14 87.14 97.14 94.29 92.86 95.71 91.63 100 100 100 100 97.73 100 100 99.68 AEL TP f 10 g 68.57 80 80 95.71 94.29 82.86 87.14 84.08 93.18 65.91 68.18 68.18 100 93.18 100 84.09 AL TP f 0.006 g 60 74.29 74.29 97.14 91.43 75.71 91.43 80.61 100 100 97.73 100 100 100 100 99.68 BGC1 85.71 92.86 90 100 98.57 92.86 97.14 93.88 100 97.73 100 100 100 100 97.73 99.35 BGC2 90 91.43 90 100 100 94.29 95.71 94.49 100 95.45 100 100 97.73 100 100 99.03 BGC3 94.29 92.86 90 98.57 97.14 91.43 94.29 94.08 100 100 97.73 100 100 100 100 99.68 CBP f 10 g 87.14 88.57 88.57 97.14 95.71 88.57 92.86 91.22 100 97.73 97.73 97.73 100 100 100 99.03 CSAL TP f 5 g 40 41.43 34.29 65.71 82.86 55.71 95.71 59.39 65.91 68.18 68.18 63.64 65.91 68.18 100 71.43 CSLBP f 1 g 92.86 91.43 88.57 100 98.57 92.86 94.29 94.08 100 97.73 100 100 100 100 100 99.68 CSL TP f 1 g 91.43 88.57 87.14 100 98.57 92.86 94.29 93.27 100 95.45 100 100 100 100 100 99.35 CSQP 90 91.43 88.57 98.57 95.71 92.86 94.29 93.06 100 100 97.73 100 97.73 100 100 99.35 DBC f 45 g 85.71 87.14 87.14 100 94.29 87.14 91.43 90.41 100 95.45 97.73 97.73 97.73 100 97.73 98.05 DDBP 85.71 90 91.43 98.57 97.14 91.43 95.71 92.86 100 95.45 100 100 97.73 100 100 99.03 dLBP f 135 g 87.14 90 91.43 98.57 98.57 95.71 97.14 94.08 100 100 100 100 100 100 97.73 99.68 DSLGS 87.14 94.29 92.86 100 98.57 95.71 98.57 95.31 100 97.73 97.73 100 97.73 100 100 99.03 eCS L TP f 1 g 91.43 91.43 91.43 94.29 97.14 88.57 92.86 92.45 100 97.73 100 97.73 100 100 100 99.35 ELGS 85.71 94.29 92.86 100 100 92.86 98.57 94.90 100 100 100 100 100 100 100 100 GTUC f 1 g 78.57 81.43 87.14 95.71 90 85.71 85.71 86.33 97.73 95.45 95.45 93.18 100 100 97.73 97.08 IBGC1 82.86 88.57 90 100 95.71 91.43 97.14 92.24 100 97.73 97.73 100 100 100 97.73 99.03 ILBP 87.14 94.29 90 100 100 92.86 95.71 94.29 100 95.45 100 100 97.73 100 100 99.03 IL TP f 1 g 62.86 75.71 75.71 97.14 90 80 91.43 81.84 100 100 100 100 97.73 100 100 99.68 LBP 88.57 94.29 91.43 100 100 94.29 98.57 95.31 97.73 95.45 97.73 100 97.73 100 100 98.73 nLBP d f 1 g 81.43 90 91.43 100 98.57 92.86 95.71 92.86 100 100 100 100 97.73 100 100 99.68 LCCMSP 82.86 87.14 87.14 98.57 97.14 92.86 98.57 92.04 100 100 100 100 97.73 100 100 99.68 LDBP 81.43 88.57 87.14 100 98.57 91.43 98.57 92.24 100 97.73 100 97.73 100 100 97.73 99.03 LDENP 90 90 87.14 100 100 91.43 97.14 93.67 100 100 100 100 100 100 100 100 LDN 87.14 88.57 90 98.57 97.14 92.86 95.71 92.86 97.73 95.45 100 100 100 100 100 99.03 LDP f 1 g 88.57 90 91.43 97.14 97.14 91.43 94.29 92.86 100 97.73 100 100 100 100 100 99.68 LESTP f 10 g 64.29 78.57 77.14 97.14 91.43 80 91.43 82.86 100 100 100 100 100 100 97.73 99.68 LGCP 88.57 92.86 84.29 100 100 92.86 95.71 93.47 100 100 100 100 100 100 100 100 LGS 90 95.71 92.86 100 100 94.29 98.57 95.92 100 95.45 100 100 97.73 100 100 99.03 LMEBP 75.71 77.14 90 94.29 84.29 81.43 91.43 84.90 81.82 95.45 88.64 93.18 90.91 90.91 88.64 89.94 LNDP 77.14 87.14 90 100 97.14 91.43 97.14 91.43 100 100 100 100 100 100 100 100 L TP f 10 g 65.71 80 77.14 95.71 94.29 81.43 90 83.47 100 100 100 100 100 100 100 100 LQP A T 84.29 88.57 95.71 100 98.57 94.29 97.14 94.08 100 100 100 100 100 100 100 100 MMEPOP 74.29 87.14 90 98.57 95.71 91.43 94.29 90.20 100 100 100 100 97.73 100 100 99.68 MO W SLGS 84.29 94.29 87.14 100 95.71 94.29 95.71 93.06 100 97.73 97.73 100 100 100 100 99.35 OS L TP f 1 g 91.43 91.43 87.14 100 98.57 91.43 94.29 93.47 100 97.73 100 100 100 100 100 99.68 QBP f 1 g 91.43 90 88.57 97.14 97.14 92.86 87.14 92.04 100 95.45 100 100 100 100 100 99.35 SLGS 87.14 94.29 92.86 100 98.57 95.71 98.57 95.31 100 97.73 97.73 100 97.73 100 100 99.03 SMEPOP 87.14 94.29 90 100 97.14 94.29 94.29 93.88 100 100 100 100 100 100 100 100 STU+ f 1 g 88.57 88.57 92.86 98.57 98.57 94.29 90 93.06 100 100 100 93.18 100 100 100 99.03 XCS LBP 82.86 84.29 80 98.57 94.29 78.57 87.14 86.53 100 93.18 95.45 93.18 100 100 100 97.40 Local featur e e xtr action based facial... (Slimani khadija) Evaluation Warning : The document was created with Spire.PDF for Python.
4088 r ISSN: 2088-8708 4.4. Comparison with state-of-the-art methods In this section, we compare the performance of the best performi ng descriptors on each database with those of e xisting state-of-the-art methods. W e should note that the performance e v aluation with other state- of-the-art approaches may not be directly comparable due to the dif ferences in partitioning the dataset into training and testing sets, number of classes, number of subjects and features used. Ho we v er , distincti v e results of e v ery approach still can be indicated. The e xtracted results from the re vie we d state-of-the-art papers as well as the recognition rates reached by the best performing e v aluated LBP-v ariants on each database are arranged in T able 4. It can be observ ed from T able 4 that, e xcept for both J AFFE and KDEF databases, where t he number of the used samples is relati v el y the same for almost all the e xisting systems, the used number of samples on CK and MUG databases v aries from one e xisti ng approach to another . Gi v en tw o dif ferent systems to compare on a gi v en database, tw o cases are possible to pro vide a f air and accurate comparison of their results. In the first one, the used number of samples and the configuration into train/test sets should be the same, whereas in the second case, the system using a less number of samples, must at least be tested with a delicate configuration into train/test sets compared to the other which uses a higher number of samples. W e used the second case in our e v aluation for comparing the state-of-the-art methods with the adopted syst em, which uses the most dif ficult configuration in terms of train/test sets. Indeed, almost all the e xisting state-of-the-art systems us e a partition where the number of training images is superior to that of test images (e.g., 10-fold), while in this study , the half-half configuration is adopted. T able 4. Comparison with state-of-the-art methods Database Ref (Y ear) Method Samples Classifier (Measure train-test) Classes Accuracy KDEF [54] (2016) Local dominant binary pattern 1168 SVM (10-fold) 7 class 83.51 [55] (2017) F acial landmarks + Center of Gra vity (COG) 980 SVM (70%-30%) 6 class 90.82 [56] (2017) LBP + HOG - K-means + self-or g anizing map 6 class 85.8 [57] (2017) Lo w-Rank Sparse Error dictionary (LRSE) 980 CRC (lea v e one-subject-out 10-fold) 7 class 79.39 [58] (2017) L TP+HOG 280 SVM (10-fold) 7 class 93.34 This paper LGS 980 SVM (half-half) 7 class 95.92 MUG [59] (2013) Local Fisher Discriminant Analysis 567 1NN (lea v e-one-out) 7 class 95.24 [60] (2014) ASM 1260 LD A (2/3-1/3) 7 class 99.71 [61] (2015) Geometric features 324 SVM (fi v e-fold) 6 class 95.50 [62] (2017) MRDTP+GSDRS 567 ELM (10-fold) 7 class 95.7 [63] (2017) GLBP - Random F orest (10-fold) 7 class 92.60 This paper Se v eral LBP v ariants including ELGS, LDTP , LDENP 924 SVM (half-half) 7 class 100 J AFFE [59] (2013) Local Fisher Discriminant Analysis 213 1NN (lea v e-one-out) 7 class 94.37 [64] (2016) Curv elet transform 213 OSELM-SC 7 class 94.65 [65] (2017) HOG 182 SVM (70%-30%) 7 class 92.75 [66] (2017) DDL + CRC LBP 213 CRC (10-fold) 7 class 97.3 [62] (2017) MRDTP+GSDRS 213 ELM (10-fold) 7 class 94.3 [67] (2017) HOG + U-L TP 213 SVM (64%-36%) 7 class 97.14 This paper DSLGS, ELGS and SLGS 213 SVM (half-half) 7 class 98.57 Ck [68] (2015) IMF1 + KLFD A 404 SVM (10-fold) 7 class 99.75 [69] (2015) LGBP 150 SVM (57.2%-42.8%) 7 class 97.4 [65] (2017) HOG 1478 SVM (70%-30%) 7 class 98.37 [58] (2017) L TP+HOG 610 SVM (10-fold) 7 class 96.06 [66] (2017) DDL + CRC LBP - CRC (10-fold) 7 class 98.8 This paper 27 LBP v ariants including RALBGC, ELGS, DSLGS 280 SVM (half-half) 7 class 100% Examining T able 4, we could mak e the follo wing findings : (a) KDEF database: It can be easil y observ ed that the LGS operator is the best performing method which achie v ed the higher performance o v er the recent state-of-the-art systems with a recognition rate reaching 95.92%. (b) J AFFE database: It is easily found that the accurac y recorded by three LBP-lik e v ariants outperformed those obtained by the state-of-the-art approaches. Indeed, it emer ges from T able 4that the top rank ed method on J AFFE database is that presented in [66] as it reached a score of 97.3% which is lo wer than that obtained by DSLGS, ELGS and SLGS operators (98.57%). (c) CK database: It is apparent from T able 4 that the highest score achie v ed on CK database is 99.75% obtained by the method presented in [68] while T able 2indicates that 27 LBP v ariants reached a score of 100%. (d) MUG database: As for CK database, se v eral e v aluated LBP v ariants lik e ELGS, LDTP , LDENP , LGCP , LNDP LQP and SMEPOP descriptors reached a score of 100% outperforming the best performing state- of-the-art approach presented in [60] which reached a score of 99.71%. The LGS, DSLGS and ELGS descriptors, which are based on the graph concept, manage to achie v e remarkable accuracies o v er all the tested benchmarks. This f act is clearly highlighted on KDEF e xperiment where we find that fe w descriptors succeeded to record abo v e 94% a v erage accurac y . Then, the dominant graph Int J Elec & Comp Eng, V ol. 10, No. 4, August 2020 : 4080 4092 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Elec & Comp Eng ISSN: 2088-8708 r 4089 encoding process justifies the rob ustness and ef fecti v eness of LGS, DSLGS and ELGS des criptors. On the other hand, we remark that CSAL TP descriptor suf fers on KEDF e xperiment reaching just 59.39% also on J AFFE and MUG e xperiments, on which the results were v ery high by the majority of the tested descriptors, the reason behind is the user specified threshold used in this operator , which needs to be identified on each e xperiment based on testing man y v alues requiring man y computations. Rather than this, all the other descriptors record good performances pro ving the discriminati v e po wer of the local description concept. 5. CONCLUSION AND FUTURE W ORKS W e reported in this present w ork a comprehensi v e comparati v e e xperimental analysis of a great num- ber of recent state-of-the-art LBP-lik e descriptors on f acial e xpression recognition. It is note w orth y that the choice of an appropriate descriptor is crucial and genera lly depends on the intended application and man y f ac- tors, such as computational ef ficienc y , discriminati v e po wer , rob ustness to illumination and imaging system used. The e xperiments presented herein significantly constitute a good reference model when trying to find an appropriate method for a gi v en application. Our e xperiments on f acial e xpression recognition included a detailed and comprehensi v e performance study of 46 te xture descriptors of the literature co v ering numerous application areas lik e te xture classifica tion, image retrie v al, finger v ein recognition, medical image analysis, f ace recognition, f ace e xpression analysis, etc. T o sho w descriptors performance o v er se v eral challenging situ- ations, the test ed descriptors were applied on four f a mous and widely used datasets such as J AFFE, CK, KDEF and MUG databases. The main finding that can be dra wn from the analysis of the o v erall performance from the e xperiments is that although some LBP-lik e features ha v e been originally concei v ed and proposed for te xture classification, the y sho w considerable performance in f acial e xpression recognition. Indeed, e v en though the y were not specifically designed for f acial e xpression recognition, some LBP v ariants outperform all state-of- the-art approaches o v er the tested databases. It is of great importance to note that the descriptors based on dominating set and graph present a significant performance stability ag ainst the other e v aluated state-of-the-art descriptors as the y are often foun d among the best performing LBP v ariants on the four tested databases. F or KDEF database, LGS opera tor , which is based on dominating set and graph theory , is the best performing de- scriptor reaching a score of 95.92% outperforming the recent state-of-the-art systems. F or J AFEE database, the better recognition rate which w as 98.57% has been achie v ed by three descriptors based also on dominating set and graph theory such as DSLGS, ELGS and SLGS. 27 LBP v ariants including ag ain those based on dominat- ing set and graph theory reached a score of 100% on CK database. Finally , man y e v aluated LBP v ariants lik e LDTP , LDENP , LGCP , LNDP LQP and SMEPOP descriptors as well as the ELGS operator reached a score of 100% o v er MUG database. As future w orks, we look forw ard to e xtend this study to include the e v aluation of deep features and deep classifiers. Furthermore, we wish to further e xplore the po wer of te xture descriptors in other applications such as compound emotion recognit ion, gender classification, f ace recognition, te xture classification, etc., in order to assess their ability to w ork with v arious classification problems. A CKNO WLEDGMENTS The authors are thankful to the National Center for Scientific and T echnological Research for funding this research through the scholarship of e xcellence No 757UIT and No 7UIT2017 for the first and second authors. Our w ork w as also part of the V olubilis AI 33/SI/14 program. REFERENCES [1] S.-J. W ang, W .-J. Y an et al. , “Micro-e xpression recognition using rob ust principal component analysis and local spatiotemporal directional features, in W orkshop at the Eur opean confer ence on computer vision . Springer , 2014, pp. 325–338. [2] K. Slimani, R. Messoussi et al. , An i n t elligent system solution for impro ving the distance collaborati v e w ork, in Intellig ent Systems and Computer V ision (ISCV) . IEEE, 2017, pp. 1–4. [3] S. Bourekkadi, S. Khoulji et al. , “The design of a psychotherap y remote intelligent system, J ournal of Theor etical and Applied Information T ec hnolo gy , v ol. 93, no. 1, p. 116, 2016. [4] T . Ojala, M. Pietik ¨ ainen et al. , “Multiresolution gray-scale and rotation in v ariant t e xt ure classification with local binary patterns, IEEE T r ansactions on pattern analysis and mac hine intellig ence , v ol. 24, no. 7, pp. 971–987, 2002. Local featur e e xtr action based facial... (Slimani khadija) Evaluation Warning : The document was created with Spire.PDF for Python.