Biometrics Technology: Hand Geometry References: [H1] Gonzeilez, S., Travieso, C.M., Alonso, J.B., and M.A. Ferrer, Automatic biometric identification system by hand geometry, Proceedings of IEEE the 37th Annual International Conference on Security Technology, Carnahan, Oct. 14-16, 2003 Pages:281-284. [H2] R. Sanchez-Reillo, "Smart card information and operations using biometrics ", IEEE Aerospace and Electronics Systems Magazine, Vol. 16, No. 4, pp. 3-6, Apr. 2001. [H3] Sanchez-Reillo, R.; Gonzalez-Marcos, A., "Access control system with hand geometry verification and smart cards", IEEE Aerospace and Electronics Systems Magazine, Vol. 15, No. 2, pp. 45-48, Feb. 2000. [H4] Sanchez-Reillo, R.; Sanchez-Avila, C.; Gonzalez-Marcos, A., "Biometric identification through hand geometry measurements ", IEEE Transactions on Pattern Analysis and Machine Intelligence, Volume: 22 Issue: 10, Oct. 2000 Page(s): 1168-1171. [H5] Sanchez-Reillo, R., "Hand geometry pattern recognition through Gaussian mixture modelling ", Proceedings. 15th International Conference on Pattern Recognition, Volume 2, pp.937-940, 2000. [H6] Jain, A.K., N. Duta, "Deformable Matching of Hand Shapes for Verification" Proc. of IEEE Int. Conf. on Image Processing, Kobe, Japan, Oct. 25-28, 1999. [H7] D.C.M. Wong, Personal Identification/Authentication by using Hand Geometry, MPhil. Thesis, Dept. of Computer Science, HKUST, 2003. [H8] A. Kumar, D.C.M. Wong, H.C. Shen, A.K. Jain, "Personal Verification using Palmprint and Hand Geometry Biometrics," Proceedings of the fourth International Conference on audio- and video-based biometric personal authentication, June 2003. w02-hand Geometry Biometrics - Summer 2006 1
Hand Geometry We shall discuss these papers concentrating on the following aspects: Acquisition what kind of device is used and the type of raw information acquired. Feature extraction representational information Matching Scheme the invariant/discriminatory information. Which similarity/distance measure are used as matching score. Performance of proposed system in verification mode or identification mode. Advantages and limitations of the proposed approach. w02-hand Geometry Biometrics - Summer 2006 2
Hand Geometry Acquisition: Ref.[H1-H6] Device: desktop scanner, CCD camera. 6 pegs to fix the fingers onto the platform Image captured: half of the back of the hand and the thickness of the hand. w02-hand Geometry Biometrics - Summer 2006 3
Hand Geometry Acquisition: Ref. [H7] Device: CCD camera No pegs no contact Image captured: inside of the whole hand (including the palm) w02-hand Geometry Biometrics - Summer 2006 4
Image preprocessing: Hand Geometry Binarization (convert color or gray level image to binary image) Compute contour of the hand. w02-hand Geometry Biometrics - Summer 2006 5
Feature extraction: Geometric Hand Geometry Widths, height (thickness), deviations, angles [H4] 31 features Widths, lengths, area [H7] 29 features. Contour information Shape of the fingers [H6] set of ordered points in the Euclidean plane (between 120 to 350 points) w02-hand Geometry Biometrics - Summer 2006 6
Hand Geometry 29 Feature extracted [H7] Finger length (FL) x 4 Finger sub-length (SL) x 12 Finger width (FW) x 8 Palm length (PL) x 1 Palm width (PW) x 1 Hand length (HL) x 1 Hand contour length (HCL) x 1 Hand area (HA) x1 w02-hand Geometry Biometrics - Summer 2006 7
Hand Geometry Feature Extraction [H7] Detection of finger-tip and gap-betweenfingers points Finger-tip point Finger Tip Gap-between-finger Gap-between-fingers point w02-hand Geometry Biometrics - Summer 2006 8
Hand Geometry Feature extraction [H7] Detection of side-points and palm-widthreference points Side Points Palm width reference w02-hand Geometry Biometrics - Summer 2006 9
Hand Geometry Feature Extraction [H7] Detection of the four finger-end points Finger End Side Points Gap-between-Fingers Points w02-hand Geometry Biometrics - Summer 2006 10
Hand Geometry Feature Extraction [H7] Detection of the wrist point Hand Langth Wrist point Wrist Point Palm Length w02-hand Geometry Biometrics - Summer 2006 11
Hand Geometry Feature Extraction [H7] Detection of 2 finger-line points for each of the four fingers Line image Lines on finger Finger line points w02-hand Geometry Biometrics - Summer 2006 12
Hand Geometry Feature Extraction [H7] Detection of finger-width reference points Finger width reference w02-hand Geometry Biometrics - Summer 2006 13
Hand Geometry Feature Extraction [H7] Example: FL:320,378,361,298;FW:113,89,109,90,113,93,105,90; HL:926;PL:469;PW:590; Finger sub-length Finger length Finger width Palm length Hand length Palm width w02-hand Geometry Biometrics - Summer 2006 14
Hand Geometry Feature Extraction [H7] Example: HC:4,478; HA:426,774 Hand contour Hand Area w02-hand Geometry Biometrics - Summer 2006 15
Hand Geometry Feature Selection Some of the extracted features may not be useful for hand recognition Ref[H4] based on statistical analysis reduces 31 features to 25. The ratio F indicates the discriminability of the feature. The higher this ratio, the more discriminant the feature is. F = where inter intra ν - - is class class the = ν N i = 1 standard N i = 1 ν f i f i ; deviation, N is the no. of users, f i is the mean of the th feature of the i th user. w02-hand Geometry Biometrics - Summer 2006 16
Hand Geometry Feature Selection Ref[H7] based on information gain and the gain ratio reduces 29 features to 17. Information Gain and Gain Ratio High information gain of the feature implies that the feature can provide more information in classifying an identity Information gain tends to give higher value for variables that have higher probabilities Gain ratio is the normalized information gain by the entropy of the testing variable InfoGain GainRatio ( f ( f C ) ) = = Entropy ( f InfoGain ( f Entropy ( f ) Entropy C ). ) ( f C ); w02-hand Geometry Biometrics - Summer 2006 17
Hand Geometry Matching Schemes Euclidean Distance Hamming Distance [H1-H5] This measures the number of components that differ in value. D ( X Q, T I ) = L = 1 d ( x, t I ) = L = 1 ( x t I > t v ); where t I is the mean and t v is the standard deviation of the th feature of user I, resp. w02-hand Geometry Biometrics - Summer 2006 18
Hand Geometry Matching Schemes Gaussian Mixture Model (GMM) [H1-H5] On modelling the pattern with a determined number of Gaussian distributions Each user is assumed to be modelled as a Gaussian distribution. Requires training!! Radial Basis Function Neural Networks (RBF) [H1-H5] RBF is a two layer neural network: first layer is based on a radial basis function (Gaussian) and the second is linear. Each class needs to be trained. For verification mode, samples belonging to other classes are required to be trained. Thus, this is not used for verification mode in the paper. w02-hand Geometry Biometrics - Summer 2006 19
Hand Geometry Matching Schemes Jain&Duta, [H6] proposed a deformable matching scheme: Feature extracted: Hand shape which is a set of ordered points in the Euclidean plane Shape distance Alignment - fingers are aligned separately Basis: D is a distance function between two sets of points A and B, the point set B is aligned to the point set A with respect to a transformation G, if D(A,B) cannot be decreased by applying to B a transformation from G. A least square type of distance measure is used. Mean Alignment Error (MAE) between two hand shape is the average distance between the corresponding points. The pair of hand shapes are said to belong to the same hand if their MAE is smaller than a threshold. w02-hand Geometry Biometrics - Summer 2006 20
Hand Geometry Performance [H4] Identification: Ref.[H4] Data base 20 users, 10 samples each comparison of the matching schemes, size of training samples, size of feature vectors. best result GMM, 21 features, 5 samples for training gives 97% success rate. (refer to Table 1 in [H4]) w02-hand Geometry Biometrics - Summer 2006 21
Hand Geometry Performance [H7] Identification: Data base 100 users, 10 samples from each hand, total of 2000 samples. comparison of the left hand and right hand samples, the matching schemes, size of feature vectors and size of the training samples. Totally eight datasets used in our experiment. 69 outliers are removed. Four matching schemes knn, Naïve Bayes, Neural Network, decision tree Evaluation methods 10-fold and 2-fold (holdout) cross-validation w02-hand Geometry Biometrics - Summer 2006 22
Hand Geometry Performance [H7] Identification Datasets \ Classifiers KNN (k=1) KNN (k=3) KNN (k=5) Naïve Bayes Neural Network Decision Tree 10-F 2-F 10-F 2-F 10-F 2-F 10-F 2-F 10-F 2-F 10-F 2-F L-1000-29 94.5 91.9 91.4 86.6 88.7 82 89.8 86 94.2 90.5 80.1 70.4 L-931-29 96.5 92 94.8 86.9 90.7 80.8 93.2 84.9 94.5 90.9 79.5 70.7 R-1000-29 91.6 86.9 85.7 81.1 84 76.6 91.1 83.3 92.1 64.4 76.9 64.1 R-931-29 93.9 90 90.7 82.5 87.7 77.3 91.8 84.7 94.3 90.5 76.7 69.1 L-1000-17 97.6 96.3 95.7 93.8 94.4 91.9 95 89.4 94.2 85.8 79.8 71.8 L-931-17 98 96.5 96.9 93.3 96 90.5 96.4 88.6 95.4 91.1 80 69.7 R-1000-17 95.8 94.8 93.9 90 92.2 86.2 94.2 90.4 92 48.9 78.5 69.1 R-931-17 98.7 97.2 96.8 94.5 95.6 90.9 96.2 91.2 95 89.9 79.5 72.8 w02-hand Geometry Biometrics - Summer 2006 23
Hand Geometry Performance [H7] Identification Training time No training required for knn More than 20 minutes for Neural Network Testing time Less than one second for all classifiers KNN (k=1) KNN (k=3) KNN (k=5) Naïve Bayes Neural Network Decision Tree Training time NA NA NA <1s 20 mins 6s Testing time <1s <1s <1s <1s <1s <1s w02-hand Geometry Biometrics - Summer 2006 24
Hand Geometry Performance Verification: Ref.[H4] GMM gives the best results EER remains the same for different feature vector sizes Variation of FMR (FAR) and FNMR (FRR) is more distinct for different feature sizes. (refer to figure 3 of the paper[h4]) Ref.[H6] Data base: 353 images from 53 persons, no. of samples per person varied between 2 to 15) 3002 genuine scores and 3992 (randomly chosen) of imposter scores were used. (refer to Fig.4 of paper [H6]) ROC (note: genuine accept rate = 1- FRR on the y-axis) (Fig. 5 of [H6] w02-hand Geometry Biometrics - Summer 2006 25
Hand Geometry Performance [H7] Verification Half of the dataset for training Another half for testing 1 2 Genuine score Impostor score 1 2 Score calculated by Euclidean distances 3 3 Test set Training set w02-hand Geometry Biometrics - Summer 2006 26
Hand Geometry Performance [H7] Verification Score distribution for L-1000-29 18% 16% 14% Overlap area Percentage 12% 10% 8% 6% 4% 2% 0% Genuine Imposter 0 1 2 3 4 5 6 7 8 9 10 Score w02-hand Geometry Biometrics - Summer 2006 27
Hand Geometry Performance [H7] Verification FAR_FRR for L-1000-29 100% 90% 80% 70% Error rate 60% 50% 40% 30% FAR FRR 20% 10% 0% 0.1 0.8 1.5 2.2 2.9 3.6 4.3 5 5.7 6.4 7.1 7.8 8.5 9.2 9.9 EER Threshold value w02-hand Geometry Biometrics - Summer 2006 28
Hand Geometry Performance [H7] Worst ones: R-1000-29 (----) L-1000-29 (----) Best ones: L-931-17 (- - - -) R-931-17 (- - - -) w02-hand Geometry Biometrics - Summer 2006 29
Hand Geometry Advantages of these approaches: (?) Limitations of these approaches: (?) w02-hand Geometry Biometrics - Summer 2006 30