Personal Authentication Using Hand Geometry
Bahareh Aghili
Department of Electrical Engineering. Shahed University Tehran, Iran
Hamed Sadjedi
Department of Electrical Engineering. Shahed University Tehran, Iran
Abstract—this paper presents an approach to personal verification and identification using hand geometry. Having extracted, fifteen features of users' right hand, Fingers’ width, Area and circumference, are classified with two different pattern recognition systems Euclidean and Absolute Distance. The algorithm has been tested on 500 pictures of 50 users. Result shows 99.7056 identification rate and 0.3198% Equal error rate (EER) with absolute distance classifier. As the pictures are captured with a scanner without any special instrument for fixing the placement, the purposed method is an inexpensive and easy to use. The experimental result shows the system performance in identification and verification.
Keywords-Hand geometry; identification; verification; feature I. INTRODUCTION
At present, authentication based on biometric properties plays an important role in world’s society. Biometric system uses variety physical or behavioral characteristic for automatic identification and verification such as face, iris, hand geometry, voice, signature and etc. In recent years, hand geometry recognition has became a very popular biometric access control, and it is used in around a quarter of real world physical access control applications [11]. In literature, there are different techniques that have been introduced below for hand geometry identification.
Raul Sanchez-Reillo used a plate and pins for fixing the hand placement and a mirror was used for capturing the picture of the hand's side. Having been extracted, 25 features were classified with different pattern recognition system from Euclidean distance to neural network. 97% accuracy in verification was reported via GMM [1]. Marcos Faundez used 10 features of right hand. The used classifier was MLP network. 99.62% accuracy in identification was reported [2].
Ovunc Polat reported a verification/identification based on hand geometry. In this research no feature was extracted from hand. Captured images were resized to 50x70 and GRNN neural network was used for classification. The algorithm has been tested on 200 pictures of 20 people and the success rate in identification was 85% and the EER was 15% [3]. Anil K.Jain presented a biometric system based on 20 features of hand.
Pins and mirror were used for capturing the pictures. Absolute, weighted absolute, Euclidean and weighted Euclidean were used for classification. The database consists 500 pictures of 50 people. The FRR=3.5% was the reported result [4]. Javad Hashemi developed an identification system based on 31 features of the hand. Different pattern recognition techniques
like Gausssian mixture modeling (GMM), Radial basis function neural network (RBF), Multi layer perceptron (MLP), K-Nearest neighbor and Hamming distance have been used in classification. 90% success rate was reported [5]. Hui Yan has developed a verification/identification system based on hand geometry palm print and finger print. Manhattan distance has been used for classification and the algorithm has been tested on 500 pictures of 50 people. The reported EER was 1.5% and the accuracy in identification was 97% [6]. S.Selvarajan purposed the usage of 14 features for identification via hand geometry. Euclidean distance has been used for classification.
He reported 96% accuracy for the database of size greater than 50 images [7]. Xiaoqain Jiang purposed an identification system based on 15 features of hand and SVM algorithm has been purposed as a classifier. The reported success rate was 92% [8]. Anil K.Jain and Duta developed a verification system that aligns finger contours and measures the mean alignment error between them. The system has been tested on 200 pictures of 20 people. The reported EER was 15% [9].
In the reviewed algorithm there are some disadvantages.
The weakness points of reviewed algorithms are:
• In some researches peg or special demand has been used for image acquisition to fix the placement of the hand. That pegs will deform the shape of the hand and users might place their hands incorrectly is the weaknesses of using pegs.
This system is not easy to use.
• The Neural Network can be used as an identification system therefore, for verification approach system; an individual neural network should be defined for each person so this way is desirable.
• The best accuracy in reported algorithm is 99.62%
in identification .that shows hand geometry is a median security algorithm.
Overcoming these disadvantages is the purposed method's aim.
II. METHODOLOGY
Hand geometry features are extracted from an image by 3 steps as follow: Image acquisition, Image pre-processing and feature extraction.
978-1-4244-4507-3/09/$25.00 ©2009 IEEE
A. Image acquisition
The database in the link below has been used
<http://www.gpds.ulpgc.es/download>used>.
The database consists of 10 different acquisitions of 50 people by a desk scanner. The 500 images have been taken from the users’ right hand. The user in this system can place the hand palm freely over the scanning surface; pegs, templates or any other annoying method for the users didn’t use to capture their hands. The images are 256 gray levels with 120 dpi resolution.
Figure 1 shows an example of a scanned image of a hand.
Figure 1.Example of a scanned image of hand
Figure 2. Contour extracted
B. Image preprocessing
First of all the gray level image is converted to a binary image. Because of the black background, there is a clear distinct in intensity between the hand and the background therefore, the histogram of the image is bimodal. The image can be easily converted to a binary image by thresholding.
After performing several experiments changing the threshold and evaluating the results with different images extracted from the data set, conclusions were reached that, with a selected threshold of 0.25. In the next step, median and morphological, to omit objects with area less than a predefined value, Filtering is used in order to cancel the presented noise. In the last step, the hand contour is obtained using Canny edge detection algorithm. The contour is shown in Fig2.
C. Feature extraction
In this research 15 features are extracted from all fingers.
Having been located, 9 main points are used to extract the desired features. The following main points are located: valleys between the fingers and three more points (see figure 3).
Figure 3.Main Points
The first two points to be detected are the ones representing the thumb and the little finger. (See figure 4)
Figure 4.Geometric process description
From the point namely (xlittle1,ylittle1) for little finger one horizontal line is traced which cuts the hand at several points (see figure 4) in this way, points , namely (xlittle2,ylittle2), (xring1,yring1), (xring2,yring2) ,(xmiddle1,ymiddle1) , (xindex1,yindex1) and (xindex2,yindex2), are found.
Once these points are found, the location of the valleys between the fingers is extracted. For instance, in order to locate the valley between the little and ring finger, the maximum ordinate point between (xlittle2, ylittle2) and (xring1, yring1) has to be found.
Three other main points that are named little, index and Thumb2 in figure 3 should be found. For instant, to locate the little point, the outer points in the contour, with minimum distance to little-ring valley need to be found.
Once these main points are found, the fingers are extracted.
Valleys in two sides of the finger are used for extracting the fingers (see figure 5).
Finally the geometric measurements are obtained as explained below.
•
• •
•
• index
• little
• thumb2
•
•
• • • • • • •
(xlittle1,ylittle1)
(xlittle2,ylittle2)
(xring1,yring1) (xring2,yring2) (xmiddle1,ymiddle1) (xmiddle2,ymiddle2)
(xindex1,yindex1)
(xindex2,yindex2)
(xthumb,ythumb)
Figure 5.Extracted fingers
• Finger widths: The finger widths is Euclidean distancebetweentwo valleys of finger.
• Finger Areas: In this research, Finger Area is the number of pixels that are in the border of finger that is between two valleys.
• Finger circumferences: The two valley of finger is connected to each other with a line. The finger has been filled (see figure 5). Number of pixel in the finger is counted and stored as finger circumference.
III. CLASSIFICATION
Two Types of classification, identification and verification, have been used in this paper which explained completely below:
A. Verification
For a verification system, an individual claims as one of the authorized users previously registered to the system. The system confirms or denies the claimer by matching the individual’s extracted data with those of the claimed person which is stored in a database. Distance functions are used to decide whether the claimer is the claimed person
In this research, 2 distance functions are experimented as follows.
• Euclidean distance
1
• Absolute distance
∑ | | (2) Where Y = y1, y2,…,yd is the feature vector of an
unknown or a claimer.
F = f1, f2, …, fd is the feature vector with d dimension of a registered user in the database.
As mentioned before 10 pictures exist for each registered person. 3 pictures of them are used for test. Having been extracted, features from seven other pictures are saved in the database therefore, the database consist 350 feature vectors.
Calculating the distance between claimer and registered person in Euclidean and absolute is possible in two ways as below:
1-The F vector is mean between 7 feature vectors of a registered person' 7 pictures. Therefore, the distance between claimer and register user is the distance between claimer feature vector and F vector.
∑ (3) Gi = g1i, g2i, …, g7i is the feature number i of each 7 feature vectors of a registered user in the database.
2-the extracted feature of the claimer are compared with 7 feature vectors of registered person in the database. The minimum distance is the distance between claimer and registered person.
D=minimum(D1,…,D7) (4) These ways are called Euclidean_withoutmean and Absolute_withoutmean in the reported result.
The statistical measures to be used for biometrics verification are:
• FAR (False Acceptance Rate): The FAR is defined as the probability that a user making a false claim about his/her identity will be verified as that false identity.
• FRR (False Rejection Rate): The FRR is defined as the probability that a user making a true claim about his/her identity will be rejected as him/her.
• EER (Equal Error Rate): The EER is defined as the crossover point on a graph that has both the FAR and FRR curves plotted. For a verification system, the optimal performance of the system is where the FRR equals the FAR. [10]
The performance of reported system in verification has been tested with 150 pictures that consist 3 pictures of each 50 enrollment people. FRR and FAR have been calculated for different threshold. The best threshold extracted from cross point of FAR and FRR (See figure 6).
Table Ireports the percent errors from verification mode with different distance functions.
TABLE I. PERCENT ERROR FROM VERIFICATION MODE
No Classifier EER%
1 Euclidean 0.6973
2 Euclidean_withoutmean 0.4089
3 Absolute 0.4744
4 Absolute_withoutmean 0.3198
Figure 6.FAR and FRR for different threshold values
B. Identification
For identification the extracted feature vector from user is compared with the entire feature vectors in the database therefore, the Identification is a one-to-many matching process.
The person whose feature vector has minimum distance with claimer person detected. The distance is compared with the threshold that is explained in the last section. If the distance be smaller than the threshold the person identify otherwise the claimer rejected as unknown person.
Experimental results in identification mode are listed in table II.
TABLE II. PERCENT ERROR OF IDENTIFICATION
No Classifier Identification
1 Eucliden 99.1556
2 Euclidean_withoutmean 99.5889
3 Absolute 99.3500
4 Absolute_withoutmean 99.7056
As mentioned earlier 3 pictures are used for test and seven pictures are used for storing in the database. Selecting 3 pictures from 10 pictures is possible in 120 different ways therefore, the purposed algorithm is tested 120 times. The reported EER and identification rate are the mean of these 120 processes.
IV. CONCLUSION
In this paper, we have proposed a new approach for biometric authentication that is based on hand geometry .Users
can place their hands freely on the scanner surface without need of pegs to fix the placement of the hand. The features used for matching are the fingers' width, area and circumference. In purposed method, distance function is used for classification in verification and identification modes. The images used for enrolment and testing are acquired from 50 users. Absoluteψ distance gives the best performance in verification and identification, with 0.31% EER and 99.70%
accuracy in identification.
ACKNOWLEDGMENT
This work is supported by Iran Telecommunication Research Center, Project number 11528/500.
REFERENCES
[1] Raul Sanchez-Reillo, "Biometric Identification Through Hand Geometry Measurement," IEEE transaction on pattern analysis and machine intelligence, VOL. 22, NO. 10, October 2000.
[2] Marcos Faundez-Zanuy, "Authentication of Individuals Using Hand Geometry Biometrics: A Neural Network Approach", Springer Science+Business Media, LLC. 2007 Neural Process Lett (2007).
[3] Polat, Yıldırım, "Hand Geometry Identification Without Feature Extraction by General Regression Neural Network", Expert Systems with Applications (2006).
[4] A. K Jain, A. Ross and S. Pankanti, “A Prototype Hand Geometry Based Verification System”, 2nd International Conference on Audio and Video based biometric person authentication (AVBPA), Washington DC, pp.166-171, March, 22-24, 1999.
[5] Javad hashemi,Emad Fatemizadeh," Biometric Identification Through Hand Geometry" Serbia & Montenegro, Belgrade, November 22-24, IEEE 2005
[6] Hui Yan, Duo Long," A Novel Bimodal Identification Approach Based on Hand-Print", 2008 Congress on Image and Signal Processing" IEEE DOI 10.1109/CISP.2008.192.
[7] S. Selvarajan, V.Palanisamy, B.Mathivanan ," Human Identification and Recognition System using More Significant Hand Attributes"
Proceedings of the International Conference on Computer and Communication Engineering 2008 May 13-15, Kuala Lumpur, Malaysia.
[8] Xiaoqian Jiang, WanhongXu, Latanya Sweeney, Yiheng Li, Ralph Gross, Daniel Yurovsky ," New Directions in Contact Free Hand Recognition" ICIP 2007.
[9] A.K. Jain, N. Duta, "Deformable Matching of Hand for Verification", IEEE International Conference on Image Processing, Kobe, JP, 1999, pp. 857–861.
[10] Paul rate (2004) , Biometric For Network Security .
[11] Jain AK, Bolte R, Pankari S (1999) Introduction to biometrics in Biometrics Personal identification in networked society. Kluwer Academic Publishers.
0 0.5 1 1.5 2 2.5 3
0 10 20 30 40 50 60 70 80 90 100
Error(%)
Threshold Value
FAR FRR