A Practical Implementation of Face Detection by Using Matlab Cascade Object Detector

Transcription

1 th Internatinal Cnference n System Thery, Cntrl and Cmputing (ICSTCC), Octber 14-16, Cheile Gradistei, Rmania A Practical Implementatin f ace Detectin by Using Matlab Cascade Object Detectr Elena Alinte Department f Autmatic Cntrl and Applied Infrmatics Technical University Gherghe Asachi f Iasi Iasi, Rmania alinte.elena@ac.tuiasi.r Crneliu Lazar, Member, IEEE Department f Autmatic Cntrl and Applied Infrmatics Technical University Gherghe Asachi f Iasi Iasi, Rmania clazar@ac.tuasi.r Abstract: The detectin f faces in an image is a subject ften studied in cmputer visin literature. The algrithm which allwed face detectin, impsing new standards in this area, was the Vila Jnes algrithm. In this paper, a practical implementatin f a face detectr based n Vila-Jnes algrithm using Matlab cascade bject detectr is presented. Emplying the system type bject visin.cascadeobjectdetectr, eight face detectrs were develped using the traincascadeobjectdetectr functin and tuning the number f cascade layer and the alse Alarm Rate. r different tuning parameters, the perfrmances f the face detectrs were analyzed. Keywrds: Vila Jnes algrithm, face detectin, AdaBst, integral image, cascade bject detectr I. INTRODUCTION In many applicatins, such as driver face mnitring, face recgnitin, vide surveillance, human cmputer interface r image database management, human face detectin is an imprtant and cmplex prcess. The cmplexity f the face detectin algrithms is due t the variatins in illuminatin, backgrund, visual angle and facial expressins and the implementatin is nt easy [1]. ace detectin algrithms are usually divided int tw general categries [2]: (i) feature-based and (ii) learning-based methds. The algrithms frm the first categry are based n the assumptin that face in the image can be detected based n sme simple features, independent f ambient light, face rtatin and pse. Thus, a simple methd uses image prjectin t detect faces under the assumptin that the backgrund is unifrm and with the vertical prjectin f the gray level image is determined the face psitin [3]. Anther feature-based face detectin apprach is based n a skin clr mdel determined by using the prbability distributin in a clr space. The face is detected in image by applying a threshld n the mdeled distributin as in [4].The algrithms frm the secnd categry are mre rbust but they need a greater cmputatinal effrt. Learning-based methds use a number f training samples and benefit frm statistical mdels and machine learning algrithms. rm this categry, Vila- Jnes face detectr [5] is ne f the mst extensively used. The detectr can be extended t ther kinds f bjects. The cnvergence f the training phase f this algrithm depends a lt n the training data. The Vila-Jnes face detectr can run in real time because it is based n the fllwing main ideas [6]: - rapid cmputatin f Haar-like features using the integral image; - classifier learning with AdaBst t select the best features; - the attentinal cascade structure which rejects the majrity f the sub-windws in early layers f the detectr, making the detectin prcess extremely efficient. Due t the simplicity f extracted features prcess and selectin f the best features, Vila-Jnes face detectr is fast and rbust, being reprted many and varius implementatins fr different applicatins. An implementatin used successfully is the ne in OpenCV. A cmplete algrithmic descriptin f Vila-Jnes face detectin methd, with a learning cde and a learned face detectr is presented in [7]. Anther implementatin incrprates six different types f feature images int the Vila and Jnes' algrithm [8] t imprve its perfrmance. Cmputer Visin System Tlbx supprts several appraches t bject detectin in an image r vide, including the Vila-Jnes algrithm [9]. The Vila- Jnes algrithm uses Haar-like features and a cascade f classifiers t identify pretrained bjects, including faces, nses, eyes, and ther bdy parts. It is als pssible t train a custm classifier. In this paper is presented a practical implementatin f a frntal view face detectin algrithm based n Vila-Jnes apprach using Matlab cascade bject detectr. Emplying the Matlab system bject visin.cascadeobjectdetectr, a face detectr was develped cnfigurated t use the user classificatin mdel specified in the XMLILE input file. The file is created with the help f the traincascadeobjectdetectr functin. The attentinal cascade training is dne using a set f psitive samples (windws with faces) and a set f negative images. r btaining a mre accurate detectr, the number f cascade layers and the functin parameters were tuned. inally, fr different tuning parameters the perfrmances f the face detectr were analyzed /15/$ IEEE 785

2 This paper is rganized as fllws. In Sectin 2, is presented Vila-Jnes face detectin algrithm. Sectin 3 illustrates the implementatin f the Vila-Jnes algrithm using Matlab cascade bject detectr. In sectin 4, we tested ur prpsal face detectin system. Sectin 5 describes the cnclusin and the future wrk. II. VIOLA JONES ALGORITHM The Vila Jnes algrithm is intended fr real time detectin f faces frm an image. Its real time perfrmance is btained by using Haar type features, cmputed rapidly by using integral images, feature selectin using the AdaBst algrithm (Adaptive Bst) and face detectin with attentinal cascade. A. eature calculatin Starting frm the cmmn characteristics f the faces, such as the regin arund the eyes is darker than the cheeks r the regin f the nse is brighter than thse f the eyes, five Haar masks (ig. 1) were chsen fr determining the features, calculated at different psitins and sizes. Haar features are calculated as the difference between the sum f the pixels frm the white regin and the sum f the pixels frm the black regin. In this way, it is pssible t detect cntrast differences. btained the integral ne and hw is cmputed the sum f pixels within a rectangle regin using integral image A C B 1 D a) b) ig. 3. Integral image: a) riginal image I; b) integral image II; c) pixel cmputatin frm the regin D using integral image r the lcatin (i, j), the integral image II cntains the sum f the pixels abve and t the left f (i, j), inclusive: c) 3 4 Type 1 Type 2 Type 3 ig. 1. Haar masks used Type 4 Type 5 II ( i, j) I( s, t), 1 i N, 1 j N 1si 1t j The sum f the pixels within rectangle can be cmputed with fur array references. The value f the integral image at lcatin 1 is the sum f the pixels in rectangle A. The value at lcatin 2 is A + C, at lcatin 3 is A + B, and at lcatin 4 is A + B + C + D. The sum within D can be cmputed as (2 + 3). ig. 2. Type 2 Haar feature frm which the intensity difference between the pixels frm eyes regin and the cheek regin can be bserved If we cnsider the mask M frm ig. 2, the Haar feature assciated with the image I behind the mask is defined by: I( i, j) white I( i, j) black 1i N 1 j N The features are extracted fr windws with the dimensins f 24x24 pixels, which are mved n the image where we want t detect faces. r such a windw, Haar masks are scaled and mved, resulting 162,336 f features. T reduce the cmputatin time f the Haar features, which vary depending n the size and type f the feature, the integral image was used. In ig. 3 is illustrated hw frm an riginal image is B. eature selectin using AdaBst algrithm As the number f Haar features fr an image with 24 x 24 pixels is d = , and many f them are redundant, AdaBst algrithm was used t select a smaller number f features. The basic idea is t build a cmplex classifier (decisin rule) using a weighted linear cmbinatin f weak classifiers. Every feature f is cnsidered a weak classifier, defined by: 1, if pf ( x) p h( x, f, p, ) 0, therwise where x is a 24 x 24 pixel image,ɵ is a threshld and p is a parity. AdaBst algrithm [10] is based n a training set which cntains n pairs (x i, y i ), where x i is a psitive r a negative image, and y i is a label assigned t each image and is equal t 1 fr a psitive image and t -1 fr a negative image. Each image is weighted with wi R and AdaBst algrithm aims 786

3 t decrease lsses defined by wi 1yi f ( xi) by adjusting the values f the weights w i. AdaBst algrithm cnsists in T iteratin, frm which T weak classifiers, meaning T features, will be selected. At each stage t f bsting is perfrmed: Step 1: with data xi ( t 1) weighted frm the previus phase, train all weak classifiers (d = ) and chse the mst efficient weak classifier h t that will becme a cmpnent f the strng classifier. Step 2: Cmbine the weak classifier with ther weak classifiers declared the mst efficient in the previus phases. n i1 Step 3: calculate the weighted errr e w1 ( ) n t i ht xi yi i1 update the weights fr iteratin (t + 1) with the frmula: w() t 1 1 i i, wi ( t 1) 1ht ( xi) yi 1ht ( xi) yi 2 et 1et and In this way, the next weak classifier will fcus n harder examples frm the training set. inally, the strng classifier will be a linear cmbinatin f T weak classifiers whse decisin rule will be: T 1 1 if h ( x) t t t hx ( ) t1 2 t1 0, therwise 1 1e where the weights are ln t t. These weights will be 2 et larger fr a weak classifier that has a small errr and will be smaller fr a weak classifier with a high classificatin errr. C. Attentinal Cascade After AdaBst algrithm, a strng classifier will result that classifies the windws f NxN size well enugh. Since, n average, nly 0.01% f the windws are psitive images, meaning faces, nly ptentially psitive windws must be examined. T Instead, t achieve a higher detectin rate and a smaller misclassified images detectin rate, we shuld use anther strng classifier that classifies crrectly the befre misclassified images. This creates the attentinal cascade, as shwed in ig. 4. At the first layer f the attentinal cascade, a strng classifier with few features is used, which will filter/reject mst negative windws. A cascade f classifiers that are becming mre and mre cmplex (with mre features) will fllw and they will allw t achieve a better detectin rate. At each layer f the cascade, the negative images classified crrectly will be eliminated and the new strng classifier will have a mre difficult task than the previus step classifier. inally, the cascade f classifiers will perate as fllws: - the image will be split int multiple windws; - every windw is an input in the attentinal cascade; - at every layer, the windw is checked if it cntains a face r nt accrding t the strng classifier; - if it is negative, the windw is rejected and the steps will be repeated fr anther windw; - if it is psitive, it means that the windw is a pssible face and will mve t the next layer f the cascade; - the windw cntains a face if it passes all layers f the attentinal cascade. III. ACE DETECTOR IMPLEMENTATION USING MATLAB The Cmputer Visin Tlbx frm the Matlab envirnment cntains a cascade bject detectr (visin.cascadeobjectdetectr) which creates a system bject (detectr) capable t detect bjects using the Vila Jnes algrithm. By default, the detectr is set t detect faces in an image, but it can als detect the nse, muth, eyes r the upper part f the bdy defined by the input string MODEL (ClassificatinMdel). This paper presents a detectr based n Haar type features, that builds a system type bject, cnfigurated t use the user classificatin mdel specified in the XMLILE input file. The file is created with the help f the traincascadeobjectdetectr functin. All sub-windws Strng classifiers 1 A A A 2 3 n A ace Rejected sub - windws ig.4 Attentinal cascade 787

4 A. Training the detectr using the traincascadeobjectdetectr functin. Because the system bject frm Cmputer Visin Tlbx visin.cascadeobjectdetectr cntains nly a few pretrained classifiers, insufficient fr a face detectin applicatin, the attentinal cascade requires training fr each user, using the traincascadeobjectdetectr functin. The attentinal cascade training is dne using a set f psitive samples (windws with faces) and a set f negative images. The negative samples are autmatically generated frm the set f negative images. r btaining a mre accurate detectr the number f cascade layers, the feature type (Haar in ur case) and the functin parameters must be specified. The attentinal cascade training (ig.5) is dne layer by layer, as fllws: - Train Layer One using: Calculated number f psitive samples, which is less than the ttal number f user-prvided samples. Generated negative samples frm user-prvided negative images. - Train layer Tw: - Train layer N: Use the results frm layer ne. Classify all psitive samples and discard samples misclassified as negatives. Use the same calculated number f psitive samples f the remaining in psitive samples. Generate negative samples by prcessing negative images with sliding windw and using false-psitive classified samples. Use the results frm the previus layer. Classify all psitive samples and discard samples misclassified as negatives. The ptinal functin parameters are: - ObjectTrainingSize: the size f the bjects used fr training is in the frm f a vectr with tw elements. Befre training the attentinal cascade all negative and psitive images will be resized t the values specified by the user. - T have better results, the value f the 'ObjectTrainingSize' parameter shuld be as clse as pssible t the size f the bject detected. T reduce the training duratin the value f the 'ObjectTrainingSize' parameter must be smaller than the size f the bject. The default value f the parameter is aut the median widthlength rati will be calculated. - NegativeSamplesactr: the multiplicatin factr f the number f psitive samples fr btaining the number f negative samples used in the training f the attentinal cascade. The default value is 2. - alsealarmrate: the false psitive result rate (alsepsitiverate) accepted fr each layer f the attentinal cascade. A false psitive result represents a negative sample classified as a face. The parameter value can be a number in the interval (0, 1] and the default value is TruePsitiveRate: the psitive result rate crrectly detected fr each layer. The parameter value can be a number in the interval (0, 1] and the default value is eaturetype: the feature type used fr training. The accepted types are Haar, LBP and HOG and fr this paper, the feature type used was Haar. The functin parameters must be specified fr achieving an ptimal functining detectr. Setting the parameters is based n the fllwing cmprmises: - r a small training set - Decrease the number f layers and set a lwer false psitive rate fr each layer. - r a large training set (in the thusands) - Increase the number f layers and set a higher false psitive rate fr each layer. - r reducing the prbability f missing an bject it is recmmended t increase the true psitive rate. - r reducing the number f false detectins it is recmmended t increase the number f layers r decrease the rate f false psitive results. ig. 5 The attentinal cascade training 788

5 r training the detectr, psitive images with size f 24 x 24, which cntain nly the crpped face, are used. The psitive samples will cnsist f an array f structures where the address f the psitive sample and the regin where the face is lcated in this case, the whle image, will be retained. The pseudcde used t btain the psitive samples is: r each psitive image Put in penilename <- the address f the psitive image Put in psitiveinstances(number). ObjectBundingBxes <- [ ] Put in psitiveinstances(number).imageilename <- penilename The negative samples will be autmatically generated frm the negative images prvided by the user. The negative images are specified by the path f their flder: negativelder = 'D:\matlab2015\imagininegative\'; The executin f the traincascadeobjectdetectr functin which trains a detectr with the alse Psitive Rate equal t 0.3 and the number f layers equal t 16, will be as fllws: traincascadeobjectdetectr('haar500v2.xml',psitiveinst ances,negativelder,'numcascadelayers',16,'alsealarmrat e',0.3,'eaturetype','haar'); B. ace detectr based n Vila Jnes algrithm r creating the system bject (detectr) that detects faces frm an image, using the Vila Jnes algrithm, the fllwing cmmand is used: detectr = visin.cascadeobjectdetectr( attentinalcascade.xml ), where the nly parameter is represented by the name f the xml file in which the attentinal cascade was saved. After the creatin f the detectr, the step methd is called by the fllwing syntax: BBOX = step(detectr, I) which returns BBOX, an M by 4 matrix defining M bunding bxes cntaining the detected bjects. Each rw cntains 4 elements [x y width height], that specify in pixels, the upper-left crner and size f a bunding bx. In rder t use a detectr btained frm training, the fllwing steps are dne:i) pen the desired image; ii) create the detectr bject; iii) identify faces frm the image; iv) anntate the faces; v) shw image with anntated faces. IV. EXPERIMENTAL RESULTS The prpsed face detectin algrithm based n Vila- Jnes was implemented using Matlab cascade bject detectr with different setting parameters f the Matlab functin traincascadeobjectdetectr, resulting eight face detectrs. The perfrmances f these detectrs were analyzed using 3 different images: I1 with 6 faces (2128x1416px), I2 with 11 faces (1920x1080px) and I3 with 6 faces (1752x1360px). We have trained the detectrs, prviding 1000 psitive images and 1000 negative images. We started with the default parameters given in Sectin III, resulting the detectr D1. After that, we decreased/increased the number f layers (NL) and/r the alse Alarm Rate (AR), btaining detectrs D2- D8. The results btained with the eigth trained detectrs fr the 3 images are illustrated in Table 1, where the number f detected faces (ND) is given fr every detectr. Table 1. Results f face detectin Detectr NL AR I1-ND I2-ND I3-ND D D D D4 26/ D D D7 22/ D8 26/ There were sme detectrs that have stpped frm training because mre negative images were needed. The resulting detectrs (D4, D7, D8) had less stages than the wanted number (table 1). Starting frm the default patrameters, better results were btained by increasing the number f layers and decreasing the alse Alarm Rate. The best results were btained with detectr D8 and are shwn in ig. 6-8 fr images I1-I3. ig.6 ace detectin with D8 fr I1 rm Table 1, it is bserved that keeping cnstant AR = 0.5, the results are imprved by increasing the number f levels frm NL = 16 (D2) t NL = 26/23 (D4). A high value f the alse Alarm Rate (AR = 0.7) leads t pr results wrsen alng with reducing the number f levels (see results fr D5 789

6 and D6 in Table 1). Ntable results were btained with a lw value f the the alse Alarm Rate (AR = 0.3) and the best fr a high value f NL (see results fr D7 and D8 in Table 1). ig.7 ace detectin with D8 fr I2 ace detectin is a difficult task due t the many variatins in scale, lcatin, rientatin, pse, facial expressin, lighting cnditins, cclusins etc. r example, in the image I2 peple are nt placed in the fregrund and therefre the results btained with the 8 detectrs are nt the best, never btaining the detectin f nly 11 faces. ig.7 ace detectin with D8 fr I3 As seen in Table 2, the respnse time f the 8 face detectrs D1-D8 with different tuning parameters is satisfactry t an applicatin f real-time face detectin. V. CONCLUSIONS Using visin.cascadeobjectdetectr, a Matlab bject system, a face detectr based n Vila-Jnes algrithm has been develped. Starting frm several pretrained classifiers fr detecting frntal faces, we used the traincascadeobjectdetectr functin t train ur face detectr classifier, emplying a set f psitive samples and a set f negative images. Selecting the functin parameters, we ptimized the number f layers, the false psitive rate and the true psitive rate, resulting mre detectrs. Using three different images with 6, 11 and 6 faces, all detectrs were tested. T get better results with a higher psitive rate, mre images are required in the training prcess. REERENCES [1] Cha Zhang, Zhengyu Zhang, Bsting-based face detectin and adaptatin, Synthesis Lectures n Cmputer Visin, Vl. 2, N. 1, Pages 1-140, Mrgan & Claypl Publishers, [2] M. H. Yang, D. J. Kriegman and N. Ahuja, Detecting faces in images: a survey, IEEE Transactins n Pattern Analysis and Machine Intelligence, vl. 24, n. 1, [3]. Wang and H. Qin, A PGA based driver drwsiness detecting system, Prceedings f IEEE Internatinal Cnference n Vehicular Electrnics and Safety, Xian, Octber, [4] J. Batista, A drwsiness and pint f attentin mnitring system fr driver vigilance, Prceeding f IEEE Intelligent Transprtatin Systems Cnference,, Seattle, USA, Octber, [5] P. Vila and M. Jnes, Rapid bject detectin using a bsted cascade f simple features, Prceeding f Internatinal Cnference n Cmputer Visin and Pattern Recgnitin (CVPR), Kauai, HI, USA, [6] P.Vila and M. Jnes, Rbust real-time face detectin, Internatinal Jurnal f Cmputer Visin, 57(2), 2004, pp [7] Y. Wang, An alasysis f the Vila Jnes face detectin algrithm, Image Prcessing On Line, 4, 2014, pp [8] Qian Li, Niaz, U., Meriald, B., An imprved algrithm n Vila- Jnes ject detectr, 10 th Internatinal Wrkshp n Cntent-Based Multimedia Indexing (CBMI), Annecy, June, 2012, pp.1-6. [9] [10] R. E. Schapire, A brief intrductin t bsting Prc. f the Sixteenth Internatinal Jint Cnference n Artificial Intelligence, Having in view that the gal f face detectin is t determine whether r nt there are any faces in an image and, if present, return the image lcatin and extent f each face in real time, the respnse time f the detectr is an imprtant perfrmance. Table 2 Time respnse (sec) f face detectrs D1 D2 D3 D4 D5 D6 D7 D8 I I I