Hand Geometric and Handprint Texture based Prototype for Identity Authentication Artzai Picón Ruiz Dpto. INFOTECH / INFOTECH Dept. ROBOTIKER, Corporación Tecnológica TECNALIA. Parque Tecnológico, Edificio 202. E-48170 Zamudio (Bizkaia) SPAIN Estibaliz Garrote Contreras Dpto. INFOTECH / INFOTECH Dept. ROBOTIKER, Corporación Tecnológica TECNALIA. Parque Tecnológico, Edificio 202. E-48170 Zamudio (Bizkaia) SPAIN Begoña Suárez Fernández Dpto. INFOTECH / INFOTECH Dept. ROBOTIKER, Corporación Tecnológica TECNALIA. Parque Tecnológico, Edificio 202. E-48170 Zamudio (Bizkaia) SPAIN Silvia Renteria Bilbao Dpto. INFOTECH / INFOTECH Dept. ROBOTIKER, Corporación Tecnológica TECNALIA. Parque Tecnológico, Edificio 202. E-48170 Zamudio (Bizkaia) SPAIN Abstract: -A person’s identity verification prototype, based on a scanner or camera acquired image, offers a very good performance with very low cost. This document describes the development of a palm-print recognition prototype based on its geometrical features and texture as well as the excellent results obtained by the use of new classifiers in the hand recognition area such as neural network classifiers and texture based statistical ones which obtain better performance than the currently used algorithms. Key-Words: - Biometrics, Hand recognition, Machine vision, Neural networks, Pattern recognition, Security 1 Introduction Biometrics is the technology which can identify and obtain human features using physical characteristics or behaviour. These technologies can make a relationship between a person and his pattern in a safe and nontransferrable way. In this paper some improvements in the characteristics extraction methods are presented, as well as new approaches in the classification methods using not only geometrical characteristics but texture based characteristics too, obtaining higher performance than the present systems. 2 Developed Prototype The main difference between biometrical methods and the classical ones lies in that the very person is the key. This key can’t be stolen and its falsification is, at least very hard The verification of the person’s identity based in the measurements of the palmprint is presented as a safe and very economic method allowing its implementation in lots of Biometric applications. There are two working modes in a biometric system. Authentication and identification. Authentication consist of verifying if the person is who he pretends to be .(1vs1). Identification consist of obtaining the identity of a person searching a database so the customer has not to tell us who he is (1vsN). The prototype developed consists of a software system which takes charge of the processing and the management of the data obtained from each user using a database, and of a hardware module constituted by a PC Scanner or a CCD Camera and a compatible PC. This prototype allows camera or scanner image acquisition, users registration, information centralization in a database accessible from several ID checkpoints. To measure the performance of a biometric system False Acceptance Rate and False Rejection Rate terms are used [1][2]. FAR (False Acceptance Rate): it is defined as the point per cent of the users who are accepted by the system when shouldn’t have been accepted. FRR (False Rejection Rate): it is defined as the point per cent of the users who are rejected by the system when should have been accepted. EER (Equal Error Rate): it is defined as the threshold in which FAR and FRR are equal so a true user have the same possibilities of being rejected as a fake user of being accepted. Most of the hand recognition systems need the user’s hand to be placed in a rigid way determined by some bumpers so no possibility of relative movement between the different fingers is possible [3]. However new approaches are being developed. A characteristic extraction algorithm for computer scanner acquired hand images has been developed allowing a more flexible hand placement [4] Fig.1 Developed prototype picture The prototype is installed to allow the access into a laboratory in Robotiker and several people use it everyday. 3. Feature extraction The main problem in a hand based recognition system lies in doing the correct election of the features which are going to be used to classify a person. This algorithm tries to extract a set of noise and hand placement independent characteristics. To obtain these characteristics this method is applied - Hand image acquisition. Hand isolation. Detection of points of interest. Geometrical characteristics extraction. Texture characteristics location. 3.1 Geometric features extraction A color image is directly obtained from the scanner. Curvature Sin ( ) IA IB (1) IA IB After calculating the curvature values in the array its observed that the finger ends are corresponded to the valleys in the curvature graph and the fingers intersection are corresponded to the peaks of the graph (Fig. 4). Fig. 1 Acquired Image The method consists of obtaining the hand edges to allow the detection of finger intersections and finger ends through the hand shape curvature. Fig. 4 Curvature graph Searching for maximum and minimum points in the curvature graph allows to extract these points which we will call “main points”(Fig.5). Fig.2 Hand Obtained Borders The hand borders are extracted using classical machine vision algorithms [4]. To get the hand curvature, the whole border array is examined and outer product is used to determine the curvature of each point in the edge image (1). Fig. 3 Curvature estimation Fig. 5 Main points extracted from the captured image Using these points, and some characteristics related to the distance of a point inside the hand to the closest border point, we are able to create the feature vector of a hand image. Fig. 6 hand extracted characteristics image Finally 48 characteristics are extracted from a hand and used to create a hand feature vector. To generate the model, it is necessary to obtain a group of feature vectors from each person. The statistical model of a person’s features is named template of that person. 3.2 Texture characteristic extraction As well as the previous characteristics, a texture feature extraction is needed to perform a texture based classification. 4.1.1 Template generation To create a template, a group of samples are obtained to get their own feature vectors, For this purpose, a trapezoidal zone is extracted using the main points coordinates and a transform is applied to convert this in a rectangle (Fig. 7). Let Vi be the feature vector of a hand image of a person with m components, being m the number of features obtained from a hand and let n be the number of hand images of that person. The hand texture in the rectangle will be used as new features for hand based classification. In order to reduce the amount of information contained in the texture, a PCA (Principal component analysis) based reduction method is chosen. The mean vector and the covariance matrix is calculated using the n samples (2). n Mean V i 1 i n n Covariance Vi Mean Vi Mean (2) i 1 n A mean vector and a Covariance matrix will be used to define each person. 4.1.2 User Classification To allow the classification of a person, a new hand image is acquired, its feature vector is extracted and is compared with its corresponding template by Mahalanobis classifier (3). Fig. 7 hand texture extraction The PCA based reduction method uses the mean of the textures of the hand of each person (as described in the training method developed for facial recognition [5]) which can spectacularly decrease the intraclass differences while increases the interclass differences. Dist V Mean inv (Co var iance) V Mean ' (3) the Dist value indicates the similarity of the present vector with one template of a person. Dist is smaller if the feature vector has been obtained from a hand of the same person the template belongs. 4.2 Neural classifiers 4. Classifiers Two sort of classifiers are proposed, statistics based and neural networks based ones. 4.1 Statistical classifiers These classifiers try to get an statistical model in order to be able to estimate, under the given input features, whether the probabilities of those features belong to a particular class or not . Neural network based classifiers are proposed to perform identification and authentication tasks as an alternative to the statistical ones. Neural networks take advantage of having experience training ability and the capabilities of extracting relationships among different features unknown for us. Neural networks are trained with a part of obtained vectors and they are tested against the rest of the vectors. Several neural networks models are proposed, these will be explained in the experimental results section due to the very different structure of each model. 14 templates and 85 images not used in the template generation were used to test the system. A success is considered when the closest template of a test image belongs to the same person the test image belongs to. The obtained results are shown in Table.2. 5. Experimental Results Correct 91,76 % 5.1 Geometric features based results Table.2 5.1.1 Statistic classifier Mahalanobis classifier was used to classify the persons using their geometric features. 5.1.1.1 Authentication results Authentication tests were accomplished using 155 hand images from 14 different persons, using 5 images from each person to create the 14 templates and the rest of them were used to test the system. To determine the classifier performance, each test image was crossed against the 14 templates observing the distances produced (3) whether they belong to the same person or not. The obtained performance of the biometric system is shown in Table.1 and in Fig.8 . FAR 0.6 % FRR 14% 5.2.1 Statistic classifier 5.2.1.1 Neural network classifier for identification A multilayer perceptron neural network based classifier is considered using back propagation training. An output vector is chosen for each feature vector,.This vector consist of n values where n is the number of classes. It will have zeros if the feature vector doesn’t belong to that class and one for the class the vector belongs to. The network was trained with 70 images from 14 persons, another 43 for network verification and another 42 to perform the final test. The network assigned all the test and verification images (and, of course, the training ones too) in the 14 existing classes. EER 2.25% Table.1 Correct 100 % Incorrect 0% Table. 3. 100 FAR y FRR 90 Incorrect 8,24% 80 FAR This classifier works better than the statistical one in identification tasks. FRR 70 60 50 40 30 20 10 0 0 0 15 00 0 14 00 0 13 00 0 12 00 90 00 11 00 80 00 10 00 70 00 60 00 50 00 40 00 30 00 50 0 20 00 45 0 10 00 40 0 35 0 30 0 25 0 20 0 15 0 0 50 10 0 0 threshold Fig.8. 5.1.1.2 Identification results To test the system performance in identification process, each image is crossed against all templates but, in this case, only the closest template is assigned without taking into account the obtained distance. It was observed that although the network classified correctly all the test images, it hadn’t the ability to discriminate those images which don’t belong to any of the classes. To overcome this setback, the training images were divided in two groups, the first group with half of the persons and the second with the other half. A network was trained with each group and the two networks were tested with the whole group of test images. In this way, the behavior of the network is observed when no class belonging inputs are presented. The Euclidean distance between the output vector of the network and all the target output vectors of each class was used as a threshold. FAR 2,65% FRR 18% The network was trained using the first 595 difference vectors using another 255 for verification and 340 for the final test. Note that the training group only has the first 7 classes , the verification group the following 3 classes and the test group the last 4 classes. EER 9% Table. 4. The selected network was a Multilayer perceptron with 25 neurons in the input layer (not all the characteristics were used), an hidden layer with 12 neurons and an output layer indicating if the authentication is correct or not. All the activation functions were sigmoid. FAR-FRR Clasificador de Verificación 100 90 80 70 60 FAR 50 FRR 40 The results are shown in Table.5. 30 20 10 FAR 2% 1 0. 0 00 0 0. 1 00 0 0. 2 00 0 0. 3 00 0 0. 4 00 0 0. 5 00 06 0. 00 1 0. 01 0. 1 0. 2 0. 3 0. 4 0. 5 0. 6 0. 7 0. 8 0. 9 0 FRR 2.3% Table.5. Fig.9. 5.2.1.3 Hand texture based results The previous classifiers don’t use the hand texture information to make their decision. The use of Mahalanobis classifier is proposed to accomplish this. The obtained results are shown in Table.4. 5.2.1.2 Neural authentication network classifier for Unlike the previous classifier, which requires a different networks if the number of classes changes, or at least requires a different network for each group of persons, this classifier is pretended to be independent from the number of classes to authenticate. To perform this, the input of the network is the difference vector between the mean of the training feature vectors and the present vector. The output of the networks is 1 if the training feature vectors and the present vectors belonged to the same person, and is zero if they belong to different persons. 155 images were taken, 70 of them were used to create 14 mean vectors (one for each class) were used as template. The other 85 feature vectors were used to create 1190 difference vectors (85 feature vectors and 14 different mean vectors). These difference vectors are used as the input of the neural network. Fig.10. Hand textures from two different persons 35 hand textures were used against the previous calculated classifiers (as described in [5] for face recognition systems) obtaining a very low error rates. Mahalanobis distance was used as threshold to accept or reject a person and the results are shown in Table.6 and in figure.11. FAR 0% FRR 4.16% Table.6 EER 0.47% The texture and geometric features will be combined to get a better performance in identification and authentication tasks. % FAR y FRR textura 100 90 80 70 60 50 40 30 20 10 0 FAR FRR Due to the good results obtained with texture features, it’s planned to extend the present texture area to other hand areas such as fingers. 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5 3 Threshold Fig.11 6. Result Discussion The results obtained in geometric features based classification are very promising, similar or better than the ones obtained by other groups as is presented in Bulatov et al [4]. - Jain et.al. FAR of 2% and a false rejection rate (FRR) of 15%. - Jain and Duta FAR of 2% and FRR 3:5%. - Raul Sanchez-Reillo et. al. error rates below 10% in verification - Bulatov et al. FAR 1% y FRR 3 %. Our results using the geometric features obtain a FAR of 0.6% and a FRR of 14% and changing the threshold sensibility we obtain a result of FAR 2% and FRR 3% as shown in Fig.8. Neural network based classifiers obtain a better performance than the statistical ones in identification, but they are not able to distinguish persons who don’t belong to any class as well as the statistical methods. To correct this problem a two-stages process is proposed. First, a person is identified using the neural classifier and then a statistical classifier is used to validate the neural network classification. 7. Conclusions Neural networks are able to improve the classification in identification tasks whenever the final verification were made by a statistical classifier. Hand texture features based classification offers a very good result in authentication, better than the geometrical methods offers. A combination of both can increase a lot the system performance. The acceptance of the prototype by the users is being good enough, and no complaints are being detected. 8. Acknowledgements The authors are grateful to Basque Government for it’s partially supporting of this project though the project IODA: Gestión de la Seguridad Informática y de la Identidad Digital. References: [1] Gutierrez. J.A et al, Estado del Arte: El Reconocimiento Biométrico. Robotiker,2002. [2] Virginia Espinosa Duró, Evaluación de sistemas de reconocimiento biométrico, Departamento de Electrónica y Automática. Escuela Universitaria Politécnica de Mataró. 2001. Hand texture based classification has obtained really amazing results, better than the results obtained using geometrical features. [3] Raul Sanchez-Reillo, Carmen Sanchez-Avila, and Ana Gonzalez-Marcos, Biometric identification through hand geometry measurements, IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(10):1168–1171, 2000. The results using texture features obtain a FAR of 0% and a FRR of 4.16%, and changing the sensibility of the threshold we get a FAR of 0.47% and a FRR of 0% (as shown in Fig.11) being these results better than other groups obtained ones. [4] Yaroslav Bulatov Sachin Jambawalikary Piyush Kumarz and Saurabh Sethiay, Hand recognition using geometric classifiers, DIMACS Workshop on Computational Geometry, Rutgers University, Piscataway,2002. 7. Future Work [5] Artzai Picón, Sistema software en tiempo real para el reconocimiento de personas en base a sus características faciales, Proyecto fin de carrera sección de publicaciones ETSIIT, Bilbao , 2002 To validate the result it’s planned to repeat the tests using a huge number of images and persons.