TELK OMNIKA , V ol. 16, No . 2, Apr il 2018, pp . 739 746 ISSN: 1693-6930, accredited A b y DIKTI, Decree No: 58/DIKTI/K ep/2013 DOI: 10.12928/telk omnika.v16.i2.7418 739 Real Time F ace Recognition Based on F ace Descriptor and Its Application I Gede P asek Suta Wija y a* , Ario Y udo Husodo , and I W a y an Agus Arimba wa Depar tment of Inf or matics Engineer ing, Engineer ing F aculty , Matar am Univ ersity Jl. Majapahit 62 Matar am, Lombok, Indonesia *Corresponding A uthor , email: gpsuta wija y a@unr am.ac.id, ar io@ti.ftunr am.ac. id, ar imba w a@unr am.ac.id Abstract This pape r presents a real time f ace recognition based on f ace descr iptor and its application f or door loc king. The f ace descr iptor is represented b y both local an d global inf or mation. The local inf or mation, which is the dominant frequency content of sub-f ace , is e xtr acted b y z oned discrete cosine tr ansf or ms (DCT). While the global inf or mation, which also is the dominant frequency content and shape inf or ma tion of the whole f ace , is e xtr acted b y DCT and b y Hu-moment. Theref ore , f ace descr iptor has r ich inf or mation about a f ace image which tends to pro vide good perf or mance f or real time f ace recognition. T o decrease the dimensional siz e of f ace descr iptor , the pred ictiv e linear discr iminant analysis (PDLD A) is emplo y ed and the f ace classification is don e b y kNN. The e xper imental results sho w that the proposed real time f ace recognition pro vides good perf or mances which indicated b y 98.30%, 21.99%, and 1.8% of accur acy , FPR, and FNR respectiv ely . In addition, it also needs shor t computational time (1 second). K e yw or d: f ace recognition, real time , LD A, f ace descr iptor , f ace classification Cop yright c 2018 Univer sitas Ahmad Dahlan. All rights reser ved. 1. Intr oduction This paper presents an application of real time f ace recognitio n based on f ace descr iptor f or door loc king system. The f ace descr iptor is the dominant frequency content of sub-f ace (local) and whole f ace (global). The f ace descr iptor is e xtr acted b y z oned DCT , non-z oned DCT , and Hu-moment. The main aim of DCT coefficients based f ace descr iptor is to get r ich inf or mation of f ace image which can giv e better achie v ement than that of compact f eatures (CF) based method [1] f or real time f ace recognition. The predictiv e linear discr iminant analysis (PDLD A) is hire d to drop-off the dimensional siz e of the descr iptor and the k nearest neighborhood (kNN) is utiliz ed f or v er ification. The main aim of this w or k is to obtain strong f ace recognition against lighting v ar iation which can be applied to the secur ity system, i.e . door loc king system which is an e xtended v ersion of our pre vious w or k[2]. F ace recognition has been widely de v eloped b y man y researchers[3], such as statistical- based (ICA, PCA, and naiv e Ba y esian), globa l f eatures-based, ar tificial intelligent-based (i.e ., genetic algor ithm, ar tificial neur al netw or k, SVM and e tc ,) and an y their v ar iations-based [1, 2] f ace recognition algor ithms . The most popular algor ithm is f ace recognition based on subspace projection: LD A, eigenf ace (PCA), and their v ar iations [4, 5]. The LD A and their v ar iation become popular due to its simple implementation and less computation comple xity . In addition, their dis- cr imination po w er is higher than that of the PCA, which mak e the perf or mance of LD A and their v ar iation be better than PCA. Discrete cosine tr ansf or m (DCT) based f ace recognition [6] has been repor ted that it pro- vided good perf or mance compared to other approaches . Both PCA and LD A is possib le to be directly e x ecuted on images in JPEG standard f or mat unaccompanied b y perf or ming in v erse DCT tr ansf or m because the y can w or k in DCT domain. The DCT -based system requires cer tain nor- malization techniques to o v ercome v ar iations in f acial geometr y and illumination. Ho w e v er , both approaches e xtr acted the f ace f eatures using only b loc k-based DCT . The f ace recognition method using selection DCT coefficients from 75% to 100% DCT and setting the high frequency to z ero has been proposed to handle illumination prob lem[6]. Ho w e v er , it nee ds high computational time Receiv ed September 28, 2017; Re vised December 21, 2017 ; Accepted J an uar y 18, 2018 Evaluation Warning : The document was created with Spire.PDF for Python.
740 ISSN: 1693-6930 because in v erse DCT tr ansf or ms and Contr ast Limited Adaptiv e Histo g r am Equalization (CLAHE) is mandator y to obtain an illumination in v ar iant f ace image . Regarding real time f ace recognition algor ithms[7, 8], mostly the eigenf ace (PCA) has been successfully implemented. Ho w e v er , the PCA is lac k of discr iminant po w er , which mak e the system be lac k of accur acy . In addition, the combination of compact f eatures (CF) and LD A projection has been applied f or real time f ace recognition[1]. The CF v ector w as e xtr acted b y LBP and z oned DCT , while the classificat ion w as perf or med b y nearest neighbor r ules . The LD A w as emplo y ed f or dimensional reduction of CF v ector . Theref ore , alter nativ e real time f ace recognition using DCT coefficients based f ace de- scr iptor which consists of dominant frequency content e xtr acted b y discrete cosine tr ansf or ms (DCT), local f eatures e xtr acted b y z one DCT (b loc k-based DCT) and shape inf or mation e xtr acted b y Hu-moment. The DCT coefficients based f ace descr iptor tends to impro v e the perf or mance of CF based f ace recognition because it has r ich inf or mation. 2. Pr oposed Method In this research, there are tw o main modules: f ace recognition engine and its implemen- tation f or a door-loc king system. The f ace recognition engine pr incipally has three subsystems: f ace detection, f eature e xtr action, recognition and v er ification r ules , as sho wn in Fig. 1(a). While the door-loc king system consists of a f ace recognition engine and solenoid cont rol circuit, as presented in Fig. 1(b). 2.1. Pr oposed F ace Recognition Engine The mechanism of f ace recognition and v er ification can be descr ibed as f ollo ws: 1. Suppose , the tr aining set is giv en to the recognition engine f or finding out machine par am- eters and guiding the engine to be intelligent. Fur ther more , f ace image descr iptors that are   F ig .   1 .   Face   r ec o g n itio n   an d   v er i f icati o n   p r o ce s s .     4.1  Face D e t e ct i on   The   f a c e   de te c ti on   is  on e   im porta nt  modul   in  r e a ti me  fa c e   re c o g nit ion  e spec iall y   on   a ppli c a ti on  of   fa c e   r e c o gnit ion  for   e lec tronic k e y .   I t his  r e se a rc h,  the  h a a r - li ke   b a se d   fa c e   de te c ti on  pr ovided   b y   p rovide b y   ope n C V   li br a r y   is  im pleme nted  for   f a c e   de tec ti on.  This  a lg or it hm  ha be e re porte than  pr ovide  robust  pe rf or man c e   a mong   the   other ’s  a l g or it hm  [ 19 ] .   The   fa c e   de tec ti on  modul   star ts  fa c e   loca li z a ti on  for   de finin                     F a ce   s ig n a tu r   Fa c e   De s c r ip t o r   Ex tr a c tio n   Fa c e   De s c r ip t o r   D i m e n sio n a Re d u c tio n   E lect ro nic K ey s   Ver if ica t io n P ro ce s s   Usi ng   M et rics a nd   k NN  Rule   Query   I m age       Re giste r e d  I m age   Fa c e   De te c tio n   a n d   No r m a li z a ti o n   Fa c e   De te c tio n   a n d   No r m a li z a ti o n   ID   a n d   Re g iste r e d   K e y   (a) F a c e  D e t e c t i on & F e a t ur e s  E xt r a c t i on M odul e S ol e noi d K e y   Cont r ol  M od u l e   D e t e c t  F a c e Cl os e / O p e K e y S i gna l   C ont rol   In s t ru c t i o Fa ce D et ect i on U n i t   Sol e n oi d  C o nt r o l  U ni t (b) Figure 1. Diag r am b loc ks: (a) f ace recognition engine[2] and (b) door-loc king based on f ace image . TELK OMNIKA V ol. 16, No . 2, Apr il 2018 : 739 746 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA ISSN: 1693-6930 741 e xtr acted dur ing the tr aining process are stored in the database as registered f ace signa- tures . The f ace image descr iptor is e xtr acted b y using a f ast z oned and non-z oned DCT , and Hu-moment, then selected a sma ll par t of tr ansf or mation coefficient ha ving g reatest magni- tude . Then the chosen coefficients are quantiz ed f or shar pening the k e y f eatu res called as f ace signature . 2. In the recognition process , the quer y f ace signature is e xtr acted b y using a similar technique to the tr aining process . Ne xt, the similar ity score is deter mined b y matching quer y f ace signature and registered f ace signatures . In this case , the smallest score is concluded as the best lik eness . 3. In the v er ificat ion process , the kNN is hired to find the highest probability of quer y f ace signature , which is close to the registered f ace signature . If the probability of quer y f ace signature , which is close registered f ace signature of class B , the input quer y is v er ified as class A. The kNN is chosen because it could giv e good perf or mance (91.5% of recognition r ate and 2.66 seconds of computational time) f or f ace recognition in small and compact de vices(ARM processor)[9]. 2.1.1. F ace Acquisition F ace image acquisition is done b y using a standard USB camer a. Ne xt, histog r am equal- ization is utiliz ed to decrease the eff ect of lighting condition dur ing f ace acquisition. Finally , the Haar-lik e based f ace detection[10], which has been widely e xamined and pro vide rob ust perf or- mance among the others algor ithm, is emplo y ed f or f ace detection. Simply , the f ace detection algor ithm star ts from f ace localization to define a region of interest (R OI) of f ace , and then from the detected f ace R OI is confir med b y detecting the tw o e y es inside the R OI, finally the fir med f ace R OI is cropped and passed to f ace recognition engine f or fur ther process on real time f ace recognition. The illustr ation of f ace detection is presented in Fig. 2. fa c e   th e   re g ion  of   int e re s (RO I ) The   f a c e   R O I   w il be   c onfir med  b y   de te c ti ng   the  two   e y e c oordin a tes  that  e x is ti ng   insi de   the  fa c e   R O I F inall y the  fir me fa c e   R O I   is   c roppe a nd  sa ve fo fur ther   pr oc e ss  in  re a ti me  fa c e   r e c o g nit ion.  The   il lust ra ti on  of   fa c e   de te c ti on  is  shown  in  F ig ,   3 B a se d   on  our   re a l - ti me  e x pe rimen tal  r e sult s,  Ha a r - li ke fa c e   d e tec ti on  p rovide robust  e nou g p e rf o r manc e a g a inst   the   lar ge   il lum ination   va ria ti ons  a nd re qui re s   s hort ti me pr oc e ssi n g .     (a )   (b)   (c )   F ig .   2 .     Face   d etec tio n   p r o ce s s es: ( a)   f ac lo ca lizatio n ,   ( b )   e y es d ete ctio n ,   ( c)   cr o p p in g   f ac e.     4.2  Feat u re  E x t rac t i on   I thi re se a rc h,  a   ne a p pr oa c to  fa c e   fe a ture w hich  is  de fine d   ba se one   Diff e re nt  o f   Ga ussi a n ( D o G) .   The   Do G itself is i mpl e mente d  f or  f indi n g  out i nte re sti ng ke y  point s in   the  im a g e   c a n   be   im plem e nted  fo il lum ination  norma li z a ti on  be c a use   it   w or looks   li ke   low - pa ss  filter in g .   The   D o of   im a ge   c a b e   e x tra c ted  usin g   the   il lust ra ti on  a shown   in  F ig  3.     F ig .   3 .   The  D o G e x tar c ti on pr oc e dur e[ 10 ].   This  pr oc e du re   wor ks  f a st  a nd  e ff icie nt  b e c a use  i re plac e a   c omput a ti on a ll y   int e nsive   of   L a plac ian  o Ga ussi a pr oc e ss  with  a   sim ple   subt ra c ti on.  The re for e the  DoG   of   im a g e s is  a pp rox im a tel y   the sa me a s the  L a plac ia n of G a ussi a n.    Figure 2. F ace detection algor ithms: (a) f ace localization, (b) e y es detection, (c) cropping f ace 2.1.2. F ace Descriptor Extraction In this paper , the f ace descr iptor e xtr action process is sho wn b y using diag r am b loc k in Fig. 3. The filter ing and contr ast stretching are also emplo y ed to eliminate the lighting v ar iation eff ect dur ing f ace captur ing. In detail, the f ace descr iptor e xtr action is done b y using some steps as f ollo ws: 1. P erf or ming the local binar y patter n (LBP) and f ollo w ed b y perf or ming none-z one DCT (on entire image) to obtain the global inf or mation of f ace image . LBP and its v ar iation ha v e been successfully implemented f or f ace recognition[11]. In this case , small par t (less than 64) coefficients are selected as global inf or mation. The LBP is implemented t o get rob ust global inf or mation of f ace image against illuminations . 2. P erf or ming z one DCT (as perf or med on JPEG compression) to obtain local f eatures of the f ace image , as sho wn in Fig. 4. In this case , less than f our coefficients are selected from Real Time F ace Recognition Based on F ace Descr iptor ... (I Gede P asek Suta Wija y a) Evaluation Warning : The document was created with Spire.PDF for Python.
742 ISSN: 1693-6930 Featur E xt rac tio n     N o rmal i z a ti o n   L BP   Zo n e  DCT   S h ap e   A n alys i s   U   N on - Zo n e   DCT   Figure 3. F ace descr iptor e xtr action processes                     Featur E xt rac tio n     N o rmal i z a ti o n   L BP   Zo n e  DCT   S h ap e   A n alys i s   U   N on - Zo n e   DCT   0 50 100 150 200 250 -2 0 2 4 6 8 10 12 N o rmal i z e d   F ac e   Zo n e   DCT   S e l e c te d     c o e ff i c i e n ts   Figure 4. Local f eatures e xtr action processes each z one as local f eatures . The local f eatures represent specific inf or mation of sub f ace image which is a v ailab le in some lo w frequency components . 3. P erf or ming the shape analysis using Hu-moment to get shape inf or mation of f ace image . In this case , only f our moments (first-f our th) is considered because the fifth-se v enth moment’ s v alues are close to z ero . It means that shape inf or mation is not a v ailab le in the fifth-se v enth moments . 4. Finally , combining the global inf or mation, local f eatures , and shape inf or mation to get r ich f ace descr iptor . In this w or k, the f ace descr iptor is represented b y the dominant frequency content of whole and sub f ace images . The local f eatures represent the most inf or mation of k e y point. Similar to SIFT f eatures , this f ace descr iptor is r ich of inf or mation which tends lighting in v ar iant because the lighting v ar iation has been decreased b y filter ing and contr ast stretching. 2.1.3. Dimensional Reductions In this paper , the pre dictiv e LD A (PDLD A[1]) algor ithm is hired to drop-off f ace descr iptor siz e . The PDLD A is similar to LD A which define optim um projection mat r ix, W , b y eigen analysis of the betw een class scatter , S b , and the with-in class scatter , S w [1, 4]. The W has to satisfy the Eq. 1. J LD A ( W ) = ar g max W j W T S b W j j W T S w W j (1) This algor ithm has been estab lished that can a v oid the retr aining prob lem of LD A. It can be done b y redefining the S b and the S w using global mean, a . It means the a is estimated b y calculating it from l sub-sample data that is r andomly selected from a giv en data set. Finally , the dimensional reduction is done b y Eq. 2. y k i = W T x k i (2) where y k i is projected f ace descr iptor and x k i is an input f ace descr iptor . By using this concept, the input f ace descr iptor can be decreased more than 50% of or iginal siz e . 2.2. Door -Loc king System The door loc king hardw are system consists of fiv e subsystems named: Raspberr y mod- ule , a set of output po w er ga in system, a ser v er , a netw or k s witch, and a door solenoid system. The Raspberr y module is used to control the door solenoid to loc k ed or unloc k ed depending on the output status giv en b y the softw are recognition system in the ser v er . The ser v er pro vides a logic condition 1 (ref ers to unloc k ed) or 0 (ref ers to loc k ed). This logic condition wrote in a file which can be accessed through the netw or k. A w eb ser v er is installed on the ser v er to pro vide this f eature . A netw or k s witch is used to connect the ser v er and the Raspberr y through the computer netw or k. TELK OMNIKA V ol. 16, No . 2, Apr il 2018 : 739 746 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA ISSN: 1693-6930 743 The Raspberr y initial mode firstly sets to 0 which will loc k the door solenoid. The Rasp- berr y contin uously chec ks the ser v er output condition through the netw or k. If the ser v er status is diff erent from the initial status , then the Raspberr y processes the prog r am to command the solenoid. The ser v er status will dr iv e the prog r am to control solenoid either loc k or unloc k. The latest status , then used as the initial status , and the chec king processes will contin ue . The Raspberr y uses ser v ers logic output condition as an input and a Raspberr y GPIO (gener al pur pose input output) pin as an output. This Raspberr y GPIO output status is used as an input b y the door solenoid as a command t o loc k or unloc k the door . Since the solenoid input v oltage requirement is 12VDC and the Raspberr y GPIO output v oltage is 3.3VDC then a rela y is needed to dr iv e the solenoid. The rela y has a minim um 5V input, which is higher than the Raspberr y GPIO output (3.3V) as w ell as the current. T o dr iv es the rela y , Raspberr y will need a simple po w er gain system. A simple po w er gain system can be b uilt using a tr ansistor and some resistors as sho wn in the Fig 5. Figure 5. Circuits of door loc king hardw are system 3. Experiments and Result Discussions Both off-line and real time e xper iments w ere carr ied out to kno w the perf or mance of f ace recognition engine based on f ace descr iptor (FD). f our w ell kno wn f ace datasets: ORL[1, 12], Image Media Labo r ator y K umamoto Univ ersity (ITS) [1, 4], and India (IND)[4], and Y ale B[13] w ere chosen f or doing off-line e xper iments . The ORL dataset has 400 g r a yscale f aces that w ere tak en from 40 persons . F ace v ar iations e xample of the ORL dataset is presented in Fig. 6(a)[1]. ITS f ace database belongs to Image Media Labor ator y K umamoto Univ ersity , which is an ethnic East Asia f ace image , especially J apan and Chin ese . ITS has 90 samples and each sample has (a) ORL dataset [12] (b) ITS dataset[4]       (c) IND f ace dataset[2] Featur E xt rac tio n           (a)   Sub - Set  1   (S 1 )   (b)   Sub - Set  2   (S 2 )           (c)   Su b  se 3   (S3 )   (d)   Su b  Set  4   (S4)     Staff TI       1 st   Day       2 nd   Day       3 rd   Day       4 th   Day       5 th   Day       (d) Y ale B[13] Figure 6. The e xample of f ace v ar iations of tested datasets . Real Time F ace Recognition Based on F ace Descr iptor ... (I Gede P asek Suta Wija y a) Evaluation Warning : The document was created with Spire.PDF for Python.
744 ISSN: 1693-6930 10 to 15 f ace v ar iations . Examples of f ace v ar iations of the ITS f ace database can be presented in Fig. 6(b)[1]. Thirdly , India dataset is color f ace image dataset which has 61 persons (22 f emale and 39 male). Thee are ele v en pose v ar iations as presented in Fig. 6(c). Some f acial emotions are also included in this dataset such as smile , disgust, neutr al, and laugh[4]. The Y ele B dataset is divided into f our sets , as sho wn in Fig. 6(d). In this case , the sub-set 1 w as chosen as tr aining and the remaining sub-sets w ere selected as testing. In addition, the off-line e xper iments w ere carr ied out b y under conditions: firstly , 50% f aces of each dataset w ere arbitr ar ily elected f or tr aining data and lefto v er par t w as chosen as quer ying images; secondly 10-F old cross-v alidation w as enf orced f or perf or mance e v aluation; finally , recognition r ate and computational time w ere utiliz ed as a perf or mance indicator .         9 8 . 3 1   9 2 . 7 5   8 2 . 1 1   9 9 . 1 4   9 8 . 3 5   9 2 . 6 6   80 85 90 95 100 I T S OR L I N D R e c ogni t i on  R a t e   ( % )   Fac e   D a t a bas e s   CF FD 100   7 5 . 6 0   1 2 . 5 5   100   8 9 . 8 9   1 6 . 5 4   10 20 30 40 50 60 70 80 90 100 S 1   v s .   S 2 S 1   v s .   S 3 S 1   v s .   S 4 R e c ogni t i on  R a t e   ( % )   Y a l e   D a t a bas e   CF FD (a) On ORL, ITS , and IND datasets         9 8 . 3 1   9 2 . 7 5   8 2 . 1 1   9 9 . 1 4   9 8 . 3 5   9 2 . 6 6   80 85 90 95 100 I T S OR L I N D R e c ogni t i on  R a t e   ( % )   Fac e   D a t a bas e s   CF FD 100   7 5 . 6 0   1 2 . 5 5   100   8 9 . 8 9   1 6 . 5 4   10 20 30 40 50 60 70 80 90 100 S 1   v s .   S 2 S 1   v s .   S 3 S 1   v s .   S 4 R e c ogni t i on  R a t e   ( % )   Y a l e   D a t a bas e   CF FD (b) Y ale B dataset Figure 7. Off-line perf or mances of our f ace recognition compared to baseline (CF based method[1]) on tested datasets . The e xper imental results (see Fig. 7) sho w that f ace re cognition engine based on FD giv es better perf or mance than those baseline methods (compact f eatures (CF) based f ace recognition[1]. In a v er age , the FD based f ace recognition engine pro vides b y about 96.72% of recognition r ate on ORL, ITS , and IND datasets (see Fig. 7(a)). In other w ords , FD based f ace recognition engine can impro v e b y about 5.66% th e perf or mance of CF based f ace recognition[1]. It can be achie v ed b y our f ace descr iptor that has r ich inf or mation which is f or med b y global inf or mation, local f eatures , and shape inf or mation. The global and local inf or mation is represented b y some lo w frequency components of whole and sub-f ace images . This achie v ement is in line with the basic theor y of signal processing that the most signal inf or mation is located in the lo w-frequency element. In ter ms of rob ustness of FD to an y v ar iations of lighting condition compared with the CF method, the FD based method giv es better perf or mance than CF (see Fig. 7(b)). It pro v es th at the FD has r ich inf or mation which rob ust to lighting in v ar iant due to filter ing and contr ast stretching bef ore the e xtr action. Regarding e x ecution time , the FD based f ace recognition engine tak es less than 1 second in a v er age f or perf or ming the matching betw een quer ying f ace descr iptor among the registered f ace descr iptors of all tested datasets . It can be achie v ed because the f ace descr iptor is repre- sented b y 32 elements of or iginal siz e f ace images (128x128 pix els). F rom off-line e xper imental data, the proposed f ace recognition engine is potential to be used f or electronic k e ys f or door loc king system. In the real time e xper iments , the system w as tested using large v ar iability f ace images in ter ms of pose and captur ing time . In t his case , 1002 f ace images ha v e been collected b y using w eb camer a Logitech C300 (1.3 MP (1280 x 1024)) from 13 persons of the staff on Inf or matics Engineer ing Dept., Engineer ing F aculty , Matar am Univ ersit y , in fifth da ys . F rom this dataset, 159 f ace images captured on the first da y (almost 11 images f or each person) w ere used as the tr aining and 843 f aces w ere prepared f or testing. Examples of f aces v ar iation are sho wn in Fig. 8. The par ameters f or real time e v aluation of f ace recognition engine w ere accur acy , F alse P ositiv e Rate (FPR), F alse Negativ e Rate (FNR), and computational time . The e v aluation results affir m that the proposed f ace recognition engine using f ace descr iptor has perf or med proper ly , TELK OMNIKA V ol. 16, No . 2, Apr il 2018 : 739 746 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA ISSN: 1693-6930 745 Featur E xt rac ti o n           (a)   Sub - Set  1   (S 1 )   (b)   Sub - Set  2   (S 2 )           (c)   Su b  se 3   (S3 )   (d)   Su b  Set  4   (S4)     Staff TI       1 st   Day       2 nd   Day       3 rd   Day       4 th   Day       5 th   Day       Featur E xt rac ti o n           (a)   Sub - Set  1   (S 1 )   (b)   Sub - Set  2   (S 2 )           (c)   Su b  se 3   (S3 )   (d)   Su b  Set  4   (S4)     Staff TI       1 st   Day       2 nd   Day       3 rd   Day       4 th   Day       5 th   Day       Figure 8. Example of f ace images v ar iations f or real time e v aluation. 9 7 . 96   25 . 0 2 . 2 98 . 4 1 7. 6 6   1. 6 7   0 20 40 60 80 1 0 0 1 2 0 Ac c u r ac y F N R F P R R a t e   ( % )   Eva l u a ti o n   Pa r ame te r s   C F F D (a) Acur acy , FNR, and FPR       7 4. 9 7   73 . 5 82 . 3 8 1. 45   4 0 5 0 6 0 7 0 8 0 9 0 Sen s it iv i t y Pr e c is s io n R a t e   ( % )   Ev al u a ti o n   P ar a met er s   C F F D (b) Sensitivity and precision Figure 9. P erf or mance of real time e xper iments . which is indicated b y more than 98% of accur acy and less than 2% and 18% of FPR and FNR, respectiv ely (see Fig. 9(a)). These perf or mances can b e achie v ed because the de v eloped recog- nition engine using the f ace descr iptor of the DCT and sub-space analysis pro vides w ell data separ ation. Compared to the perf or mance of CF based f ace recognition, our proposed method significantly decreases the FNR b y about 7.37% and in another side , it is not m uch increase and decrease the accur acy and FPR, as presented in Fig. 9(a). The FD based f ace recognition engine also impro v es significantly sensitivity and precision of CF based f ace recognition b y about 7.37% and 7.95% respectiv ely , as sho wn in Fig. 9(b). It also affir ms that our proposed method can handle the f alse negativ e prob lem of the baseline method (the correct person is f alsely recogniz ed as others). Ov er all, the real time perf or mances confir m the off-line achie v ements which can impro v e the baseline perf or mances . The last e xper iment w as carr ied out to kno w the perf or mance of FD based f ace recog- nition engine f or door loc king system which w as done b y the staff of Inf or matics Engineer ing Dept., Engineer ing F aculty , Matar am Univ ersity , in one w eek. The door loc king system can w or k proper ly , which is sho wn b y the accur acy , FPR, and FNR b y about 98. 30%, 21.99%, and 1.8%, respectiv ely . The last result also re-affir m that the FD is po w erful f or real time f ace recognition engine . 4. Conc lusion and Future W ork The real time FD based f ace recognition engine giv es better perf or mances than those of baseline (CF). F rom the Off-line e v aluations , it pro vides high recognition r ate (a v er age more than 96%) f or all tested datasets , while the real time e xper iment al data pro vide high accur acy (more than 98%) and less f alse v er ification r ate (b y about 17.66 % of f alse negativ e and 1.67% of f alse positiv e r ate). Regarding the computa tional time , the proposed electronics k e y sim ulator needs less than 1 second f or the matching process . In addition, the application of FD f ace recognition engine f or the door-loc king system also w or ks proper ly , which is indicated b y 98.30%, 21.99%, and 1.8% of accur acy , FPR, and FNR respectiv ely . The door-loc king system based on f ace image has to be e v aluated in large siz e dataset Real Time F ace Recognition Based on F ace Descr iptor ... (I Gede P asek Suta Wija y a) Evaluation Warning : The document was created with Spire.PDF for Python.
746 ISSN: 1693-6930 to kno w its rob ust perf or mance against the large v ar iability of f ace images in pose , lighting, and accessor ies . I n addition, the proposed system still needs to be impro v ed b y adding some illu- mination compensation, such as Contr ast Limited Adaptiv e Histog r am Equalization (CLAHE) to decrease the f alse negativ e recognition. Ac kno wledgment W e w ould lik e to send our g reat a ppreciation to the staff of inf or matics Engineer ing Dept. on their par ticipation in the e v aluation of this system. In addition, our g reat honor is also to The Minister of Research and Higher Education Repub lic of Indonesia f or research funding under scheme competitiv e research g r ant 2015-2016. Ref erences [1] I. G. P . S . Wija y a, A. Y . Huso do , and A. H. J atmika, “Real time f ace recognition engine using compact f eatures f or electronics k e y , in Inter national Seminar on Intelligent T echnology and Its Applications (ISITIA) , Lombok, Indonesia, J uly 2016. [2] I. G. P . S . Wija y a, A. Y . Husodo , and I. W . A. Ar imba w a, “Real time f ace recognition using dct coefficients f ace descr iptor y , in Inter national Conf erence on Inf or matics and Computing (ICIC 2016) , Lombok, Indonesia, October 2016. [3] R. Chellappa, C . L. Wilson, and S . Sirohe y , “Human and machine recognition of f aces: a sur v e y , Proceedings of the IEEE , v ol. 83, no . 5, pp . 705–741, Ma y 1995. [4] I. G. P . S . Wija y a, K. Uchim ur a, and G. K outakii, “F ace recognition based on incremental predictiv e linear discr iminant analysis , IEEJ T r ansactions on Electronics , Inf or mation and Systems , v ol. 133, no . 1, pp . 74–83, 8 2013. [5] J . Zhang and D . Scholten, “A f ace recognition algor ithm based on impro v ed contour let tr ansf or m and pr inciple component analysis , TELK OMNIKA (T elecomm unication Comput- ing Electronics and Control) , v ol. 14, no . 2A, pp . 114–119, 2016. [6] A. Thamizhar asi and J . J a y asudha, “An illumination in v ar iant f ace recognition b y selection of dct coefficients , Inter national Jour nal of Image Processing (IJIP) , v ol. 10, no . 1, p . 14, 2016. [7] H. H. Lwin, A. S . Khaing, and H. M. T un, “A utomatic door access system using f ace recog- nition, Inter national Jour nal of Scientific & T echnology Research , v ol. 4, no . 6, pp . 210–221, J un 2016. [8] M. Ba ykar a and R. Da, “Real time f ace recognition and tr ac king system, in Electronics , Computer and Computation (ICECCO), 2013 Inter national Conf erence on , No v 2013, pp . 159–163. [9] E. Setia w an and A. Muttaqin, “Implementation of k-nearest neightbors f ace recognition on lo w-po w er processor , TELK OMNIKA (T elecomm unication Computing Electronics and Con- trol) , v ol. 13, no . 3, pp . 949–954, 2015. [10] P . Viola and M. Jones , “Rapid object detection using a bo osted cascade of simple f eatures , in Proceedings of the conf erence on Computer Vision and P atter n Recognition , 2001, pp . 511–518. [11] S . R. K onda, V . K umar , and V . Kr ishna, “F ace recognition using m ulti region prominent lbp representation, Inter national Jour nal of Electr ical and Computer Engineer ing , v ol. 6, no . 6, p . 2781, 2016. [12] F . S . Samar ia and A. C . Har ter , “P ar ameter isation of a stochastic model f or human f ace identification, in Applications o f Computer Vision, 1994., Proceedings of the Second IEEE W or kshop on . IEEE, 1994, pp . 138–142. [13] Y ale , “The e xtended y ale f ace database b , 2001. [Online]. A v ailab le: http://vision.ucsd.edu/ iskw ak/ExtY aleDatabase/ExtY aleB .html TELK OMNIKA V ol. 16, No . 2, Apr il 2018 : 739 746 Evaluation Warning : The document was created with Spire.PDF for Python.