Optimization of a Biometric System Identification by Hand ... - SEE

OPTIMIZATION OF A BIOMETRIC SYSTEM IDENTIFICATIONBY HAND GEOMETRYCarlos M. Travieso, Jesús B. Alonso, Sébastien David, Miguel A. FerrerDpto. de Señales y Comunicaciones, Universidad de Las Palmas de Gran CanariaCampus de Tafira, 35017 Las Palmas de Gran Canaria, SPAIN.Phone: +34 928 452 864 Fax: +34 928 451 243 email: ctravieso@dsc.ulpgc.esAbstractIn this paper the optimization of a very robust and novelsystem has been carried out to identify biometric by thehand geometry. Thus, a study between different type ofgeometric parameters like the distances betweenelements of the hand or the hand shape has been done.These parameters are used into two different types ofclassifiers, Neural Networks (NN) and Support VectorMachines (SVM). The best results have been foundusing the measures of distances in the hands, and whenSVM is used, achieving over a 99%, with at least threehands for the training in the identification system.It has been employed a supervised classificationsystem, utilizing therefore, two phases in the process ofidentification, one of training and another of test. For it,a database has been created with the particularity of alow resolution and a standard format.Key WordsBiometric, Hand Recognition, Security System, SVM.1. IntroductionThe problem to resolve the identity of a person can beclassified fundamentally in two types of approaches [1]:recognition (more popularly known as identification)and verification. In this paper we focused on theidentification, which is based on determining theidentity of the subject inside a set of known identities.In many situations of our daily life, we are required totest our identity, as for example when we gain access toa precinct. A good recognition of the identity of aperson would be able to dissuade the delinquency andthe fraud, put new energy into the commercialtransactions and safeguards the critical resources.To approach the identification through a possessionhas the problem that this can be stolen, lost or forgotten,and falsified in the case of the cards with a magneticband. On the other hand, the problem of the use of theknowledge as identification form rests in the difficultyto re member the PINs. The PINs that are easilyremembered, it can be guessed for the swindler [2].Therefore, the use of the biometric will enable tosolve the problem of the personal identification. Thebiometric is easy to use, nothing to remember, nothingto change, nothing to lose. Besides, it provides a higherlevel of security, since it is somewhat univocal of acharacteristic human that can hardly be guessed ordeciphered. We work on the principle that somebiological characteristics are singular and unalterable,and besides , they are impossible to lose, to transfer or toforget. This does them more dependable, friendly andsure than passwords [1].From all the possibilities of biometric parameters, ithas been chosen the hand, since it is a system very littlestudied with few references, [3][4][5][6][7], and peoplegive it credit as recognition system [1]. The use of thehand palm as measure of authentication can be a goodsolution for applications of intermediate security andwhere the convenience is considered as more importantoption than the security or the precision. At the sametime, this technique offers a good balance betweenperformance and facility of use. The fast and easyintegration in others security systems does the use of thehand geometry an obvious first step for the industry inbiometric applications. In the following figure, it can beobserved as the biometric market is currently, where itsannual mean growth should be 31.5% between 2004and 2008 [8].The approach to problem of the identification carriedout up to now has been based on two premises:• A personal possession: in our case can be anidentification card or card of magnetic band.• The knowledge of information: as it can be a PIN(Personal Identification Number).Figure 1. 2004 Comparative Market Share by Technology [8]

Inside our system of biometric recognition, somesteps have already been made as the building of thedatabase and a first approximation with system basedon Neural Networks [7]. Therefore, in this paper, thefull study of classifier according to different parametersis presented. The scheme followed in this paper isshown in the figure 2. SVM has been selected likeclassifier because of the robust and good resultsprovided in [9][10]. In such a way, in this study, wetend to optimize the identification system.The rest of paper follows with the second section andthe description of the database. Subsequently, the imageprocessing applied for the extraction of characteristics.In the fourth section, the classification is described.Experiments are in the fifth section, and finally,conclusions.2. Building of the databaseWe will restrict the feature set mainly to the contour andto some geometrical measurements. We do not usepalm-print recognition (identification of a user based inthe information extracted from the lines of the palm).The quality of the images has been fixed to 150 dpi. Tofacilitate further computation every scanned image hasbeen scaled by a factor of 20%. This scaling do notaffect the recognition rate, since all the images havebeen reduced by the same factor. The scaling processhas a two reasons:• We keep the image file size small.• The contour information is reduced making easierfurther processing.Table 1 summarizes the image databasecharacteristics.Properties of the images contained in the databaseSize80% of the original size.Quality150dpiColour256 grey levelsFile size1405KbytesMatrix size1403x1021Table 1. Images specificationsThe images have been obtained using 8 bitsresolution, which is the natural choice because of twomain reasons:• We do not need to save our image as coloured one,because no special feature as colour skin is used.• The scanner software produces a very poor binaryconversion.Most of the users are within a selective age rangefrom 23 to 30 years old. Far away from simplicities, wehave selected the users to have approximately equalhand characteristics to see how the algorithms perform.The percent of males and females are not equal. In thedatabase, 68% of the users are males.The number of repetitions selected to elaborate thedatabase is based on the references [3], [4]. In [3], theauthors took 360 images belonging to 50 users(approximately 7 images per user). In [4], they took 20users and 10 images per user. In our work, the databaseis composed by 500 images (50 users and 10 repetitionsper user) taken from the users’ rigtht hand.3. Image processing: extraction ofparametersOnce the database elaboration built, according to thescheme showed in the Figure 2, we should make a pre -processing of the data in order to simplify theparameterisation process. This stage can be split into thefollowing steps:• Binarization: the original images in our databaseare 256 grey level images. In this step, our goal isto convert the input 8-bit image into a binaryimage.• Contour and main points extraction: once the imagehas been binarized we proceed with the extractionof the principal points of the hand (geometryapproach) and of the hand-palm contour.• Computation of two significant parameters: thehand’s area and perimeter are computed.• Normalization of the hand contour: the contour ofthe hand is centred with respect to the centroid,hence achieving translation invariance.Acquisitionof imagesPreprocessingExtraction ofparametersClassificationTRAININGTESTTemplateIdentificationDecisionAcquisitionof imagesPreprocessingExtraction ofparametersFigure 2. Identification Biometric System

3.1. BinarizationAfter several experiments changing the threshold andplaying with different images extracted from thedatabase, we reach the conclusion that with a selectedbinarization threshold of 65 the results were adequatefor our purposes. We tried other binarization algorithmas the suggested by Lloyd, Ridler-Calvar and Otsu[11][12]. The two first algorithms gave us a threshold of100 and the third one a threshold of 185. The last valuewas completely discarded, since the output imageapplying this threshold of 185 is worse than the case“High Threshold” showed in the Figure 3. For a valueof 100 there were not so many differences since smallvariations in the threshold will not affect the finalresults.To complete precisely the hand-palm geometry thereare still three important points missing. These points arethe symmetric to the points located in the little-ringfinger valley, index-ring finger valley and index-thumbvalley. We can see the positions of these points inFigure 5, pointed by the labels (X1, Y1), (X2, Y2) and(X3, Y3), respectively. To locate these points we haveto look for the outer points in the contour withminimum distance to these valleys.Finally, using all the main points previouslycomputed, the geometric measurements are obtained.We take the following distances: Length of the 5fingers, distances between points (X1, Y1) and(X2,Y2), points (X2,Y2) and the valley between thethumb and index finger and the points (X3,Y3) and(X1,Y1). Figure 5 shows the final results along with thegeometric measurements taken into account.(a) Low (b) High (c) SelectedFigure 3 Different thresholds applied to the sample image3.2. Contour and main points extractionOnce the images have been binarized we have to extractthe most significant data mentioned above. Featureextraction is a very important step in building ourbiometric identification system because a wrong featureselection provides low recognition rates. In this paperwe have used two parameters, hand contour andgeometric measuments.The contour is calculated from the binarizated imageby the derivate. As one of our objectives is to recognizeusers by the hand contour, we need to obtain anumerical sequence describing the hand-palm shape.Contour following is a procedure by which we runthrough the hand silhouette by following the image’sedge. In our implementation, we take into account theradius from gravedad center to hand shape. Figure 4shows this process and the vector of parametersimplemented with boundary codification.ContourCodificationFigure 4. Contour and its parametersThe method for extraction the geometric hand-palmfeatures is quite straightforward. The first points to bedetected are the thumb and the little finger. By thesetwo points, we are able to extract most of thegeometrical information. With those two points , welocate the following main points: finger tips, valleysbetween the fingers.Figure 5. Contour and geometric measurements3.3. Computation of two significant parametersThose parameters are the area and the perimeter,calculated with the binarizated and contour images,respectively.3.4. Normalization of the hand contourThe contour vectors characterizing the silhouette ofeach hand-palm are quite large in length (about 4000samples). We should reduce these lengths as much aswe can, preserving the information needed for theclassification. The techniques used in this paper toextract and compress the information presented in acontour parameter vector are transformed descriptors[12][13] and Principal Component Analysis (PCA) [14].• Discrete Cosine Transform Descriptors [12]: it hasexcellent energy compaction for highly correlateddata. The DCT basis images are independent inputsand its information packing ability is close to theoptimal KLT. The DCT of a sequence is definedu( n),0 ≤ n ≤ N −1as:{ }N 1⎡ (2 + 1) ⎤( ) = ( ) ∑ − π n kv k α k u(n)cos⎢,0 ⎣ 2 ⎥n=N ⎦0 ≤ k ≤ N −1(1)

where12α( 0) = , α(k)= for 1 ≤ k ≤ N −1.NN(2)Signal restoration is possible by applying theinverse transform given byN 1⎡ (2 + 1) ⎤( ) = ∑ − π nu n α(k ) v(k ) cos⎢,0⎣ 2 ⎥k =N ⎦0 ≤ n ≤ N −1.(3)Choosing the appropriate v(k) coefficients we get agood reduction in the number of coefficients whilekeeping as much information as possible.• Discrete Wavelets Transform Descriptors [13]:discrete orthogonal wavelet transform is used todecompose a signal into several levels ofresolution. Multiresolution decomposition isperformed by projecting the signal onto orthogonalapproximation and detail subspaces. An efficientway of implementing discrete wavelet transform(DWT) is the use of filter banks. The filter bankimplementation of DWT is based on itsmultiresolution property.• Principal Comp onent Analysis [14]: PCA alsoknown as the Karhunen-Loève transform, can beused to reduce the feature vector dimension whileretaining the feature information by constructing alinear transformation matrix. The transformationmatrix is made up of the most significanteigenvectors of the feature vector covariancematrix. The eigenvectors are orthonormal(orthogonal and normalized) so they transform theoriginal data into independent feature informationhaving maximal variance.Finally, once the useful data has been extracted fromthe images, we store the new parameters. The geometricmeasurements is a structure that contains the 10geometric measurements extracted from the hand (5finger lengths, 3 ratio measurements, area & perimeter).On the other hand, we have to notice that the length ofthe contour parameters for each user is quite long. Forthis reason, we reduce with above techniques in size of128, 256 and 512 samples.two different classes. This decision boundary orseparate hyperplane can be designed with differentkernels, lineal, polynomial, gaussian, etc.Supposing that all the data of training satisfy thefollowing conditions:x i· w + b ≥ +1x i· w + b ≤ −1for y i = +1 (4)for y i = +1 (5)These two conditions can be transformed in a set ofinequations:yi( x · w + b) −1≥ 0ifor ∀i(6)It is observed that the points which comply with theequality of the equation 4, they are found in thehyperplane H 1 : x i· w + b = 1 with normal w and distanceperpendicular since the origin1− b.wSimilarly, the contained points in the equality of theequation 5 are found in the hyperplane H 2 :xi· w + b = −1 with normal w and distance perpendicularsince the origin −1 − b .wThe distances between H 1 and H 2 will be equal to 1 .For which the margin will be 2 .wwThe above H1 and H2 are parallels (have the samenormal) and no point of the training is located betweenthem. It will be able to find the pair of hyperplanes that2they give the most maximum margin minimizing w ,subject to the conditions of the equation 6.Therefore the solution of a typical two-dimensionalcase would have the following form:4. ClassificationTo carry out the study have been utilized two classifiersto observe the adaptation of the data to different robustforms to classify geometric parameters. Theseclassifiers are Support Vector Machines (SVM)[15][16] and Neural Networks (NN) [14][17].4.1. Support Vector MachinesSupport Vector− bwH 2HH 1wThe Support Vector Machines (SVM) resolve theclassification of geometric parameters to calculateexactly the separate hyperplane from the training data[15][16]. Besides, this solution introduces methods tobe able to work with no separable and separable lineallycases of the data. The decision boundary is decided withthe calculation of a separate hyperplane thatdiscriminates between the positive and the negativesamples; therefore a SVM treats the classification asFigure 6. Separate lineal hyperplanemarginFor the identification system in this paper, a SVM withkernel lineal and RBF has been used, as not linealseparate plane. As the samples are separated in groupsof two classes, we have established a multiple classesstructure for the 50 different hands.

4.2. Neural NetworksThe perceptron of a simple layer establishes itscorrespondence with a rule of discrimination betweenclasses based on discriminante lineal. However, it ispossible to define discriminations for not lineallyseparable classes utilizing multilayer perceptrons, thatare networks without refreshing (feed-forward) with oneor more layers of nodes between the input layer and exitlayer. These additional layers contain hidden neurons ornodes, which are directly connected to the input andoutput layer [14][17]. A neural network multilayerperceptron (NN-MLP) of three layers is shown in thefigure 7, with two layers of hidden neurals . Each neuronis associated with a weight and a bias, these weights anbiases of the connections of the network will be trainedto make their values suitable for the classificationbetween the different classes.Output layer2º hiddenlayer1º hiddenlayerInput layerFigure 7. Multilayer PerceptronParticularly, the neural network used in theexperiments is a Multilayer Perceptron (MLP) Feed-Forward with Back-Propagation training algorithm, andwith only one hidden layer.5. ExperimentsThe experiments have been based on the use of thedifferent parameters, for the two classifiers proposed.Therefore, the first step is to seek the optimum values ofthe classifier for each one of the parameters (theseresults are not shown for limitations of space in thispaper).The combinations studied for the different classifiersare;• Geometric Measures (GM-10)• Contour with DCT reduction to 128• Contour with DCT reduction to 256• Contour with DCT reduction to 512• Contour with DWT reduction to 128• Contour with DWT reduction to 256• Contour with DWT reduction to 512• Contour with PCA reduction to PCA 128The results obtained for each one of the classifierswith the optimized variables, can be observed in thefollowing tables;Maximum Recognition RateNumbers of hands 3 5 7 9DCT-128 92.21% 96.20% 98.00% 99.00%DWT-128 92.11% 96,92% 98,4% 98,8%PCA-128 84.02% 92.92% 94,66% 96,2%GM -10 99.02% 99.56% 99.73% 99.8%Table 2. Results with NNThese better results have been found for 58 neuronsin the hidden layer. From the above table 2, it can bededuced that the best parameters are the geometricmeasures, surpassing always the 99% of success as thenumber of hands for training, and the worst one by thereduction with PCA. With the transformed descriptorsparameters, the results are low in comp arison withgeometrical measurement (GM -10). Finally, it iscommented that the results obtained are the best rates ofpeak that were registered, due to that the neuralnetworks vary quite a lot, their recognition rate sincethey are randomly started.For the proofs carried out with SVM, the experimentswith PCA have been ruled out, since they have obtainedthe worst success. Thus, we focus on the tests with thetransformed descriptors and geometrical measures. Theresults are shown in the following table, where variablesare the different types of parameters and the type ofkernel of the SVM. The recognition rates have beencalculated with the optimized SVM for the lineal andRBF kernel.Mean Recognition RateNumbers of hands 4 5 6DCT-128lineal 92.27% 93.36% 94.60%RBF 85.17% 86.89% 89.00%DCT-256lineal 92.53% 93.92% 95.00%RBF 85.75% 87.21% 89.50%DCT-512lineal 92.73% 93.71% 95.50%RBF 83.25% 86.78% 91.12%DWT-128lineal 92.10% 93.45% 94.78%RBF 80.54% 84.61% 87.39%DWT-256lineal 92.67% 93.92% 95.10%RBF 82.53% 87.68% 89.37%DWT-512lineal 92.00% 94.72% 95.10%RBF 80.93% 86.08% 88.20%GM-10lineal 86.67% 89.44% 88.70%RBF 99.67% 99.84% 99.90%Table 3. Results with SVMIn the table 3, the experiments have been made with4, 5 and 6 training hands. Moreover, from the abovetables, we can deduce that the NN-MLP and the SVMwith kernel lineal have the same response fordescriptors transformed. On the other hand, the SVMwith kernel RBF responds better than the NN in the useof geometric measures. This is due to that the NN startthe training with a random beginning, move the

decision boundary until all the points of training are inthe correct side of the decision. Nevertheless, theSVM’s calculate its margin of more optimum form,between the support vector of each class, obtaining themean line (decision boundary) of both lines.Thinking about those better results, and in a possiblecommercial application, the following table shows therates averaged of the best case in this study and theirtypical deviation, after being carried out the experimentin 5 times, emphasizing the case of the training with 3hands, that of each 1000 hands is mistaken in 2.3 withtypical deviation of 0.13;NumbersMeanStandardof handsdeviation1 97.29% 1.052 99.10% 0.453 99.77% 0.134 99.67% 05 99.84% 0.226 99.90% 0.22Table 4. Best results with GM-10 and SVM with kernel RBF.6. ConclusionsA biometric identification system has been created bythe hand geometry, which is very robust and with goodresults, upper 99%, using geometric measures andclassifying them with SVM.It has been shown thanks to this study that the use ofthe SVM with the same parameters than the NN andwith an adequate training improves the efficiency of thisidentification biometric system.The next step is to increase the hand database, and tostudy if these good results remain similar.The good results indicate that this biometric systemcan be integrated for commercial applications of accessto precincts, control of timetables, etc., since turns outto be efficient and the times of test are lower to onesecond.8. Acknowledgement[4] R. Sanchez-Reillo, Hand geometry patternrecognition through Gaussian mixture modelling,Proc. 15th International Conf. on PatternRecognition, Barcelona, Spain, vol. 2, 2000, 937-940.[5] A.K. Jian & N. Duta, Deformable matching ofhand shapes for user verification, Proc. 15thInternational Conf. on Pattern Recognition,Barcelona, Spain, vol. 2, 1999, 857 -861.[6] Dong-mei Sun, Zheng-ding Qiu & Bing He,Automated identity verification based on handshapes, Proc. 6th International Conf. on SignalProcessing, vol. 2, 2002, 1596 -1599.[7] S. González, C.M. Travieso, J.B. Alonso & M.A.Ferrer, Automatic Biometric Identification systemby hand geometry, Proc. 37th IEEE InternationalCarnahan Conf. on Security Technology, Taipei,Taiwan, 2003, 39-42.[8] Biometrics Market Report 2003-2007. Visited inFebruary 2004. [www.biometricgroup.com].[9] T. Joachims, Learning to classify text usingsupport vector machines (Kluwer AcademicPublishers, 2001).[10] C.M. Travieso, J.B. Alonso & M.A. Ferrer, SignLanguage to text by SVM, Proc. 7th InternationalSymposium on Signal Processing and itsApplications, Paris, France, vol. 2, 2003, 439-442.[11] L. O’Gorman & R. Kasturi, Document ImageAnalysis (IEEE Computer Society Press, 1995).[12] Anil K. Jain, Fundamentals of Digital ImageProcessing. (Prentice Hall. 1989).[13] A. Graps, An introduction to wavelets ,IEEE Computational Science and Engineering,2(2),1995, 50-61.[14] C. M. Bishop, Neural Networks for PatternRecognition (Oxford University Press, 1995).[15] N. Cristianini & J. Shawe-Taylor, An introductionto support vector machines (Cambridge UniversityPress, Cambridge, 2000).[16] B. Schöfkopf, C. Burges & A. Smola, PairwiseClassification and Support Vector Machines (TheMIT Press, Cambridge, Massachusetts, London,England, 1999).[17] D.R. Hush & B.G. Horne, Progress in supervisedneural networks, IEEE Signal ProcessingMagazine, 10(1), 1993, 8-39.This work has been developed and agrees by researchProject UNI2003/26 from University of Las Palmas deGran Canaria (Spain).References[1] A.K. Jain, R. Bolle, & S. Pankanti, Biometrics:Personal Identification (Networked Society,Kluwer Academic Publishers, 2001).[2] J. R. Parks, Personal identification – biometrics(North Holland: Elsevier Science, 1991).[3] A.K. Jain, A. Ross & S. Pankanti, A prototypehand geometry-based verification system, Proc.2nd International Conf. on Audio and Video basedBiometric Person Authentication, 1999, 166-171.