13 Abdul Basit

Human Ear as Identity CardA. Basit

TPPD, Pakistan Institute of Nuclear Science and Technology, Islamabad, PakistanEmail: [email protected]

Abstract- In this paper human ear is used as identity card. Anew ear recognition method is proposed based on wavelet transform. Ear sub-image is cropped from the image. It is normalized to get the same sized feature vector and a very simple technique is followed for feature extraction. Various wavelet transforms are applied at different levels and matching is carried out using Euclidean distance. Correct recognition results achieved using the proposed method are up to 99.2%.

Keywords: Ear recognition, wavelet transform, biometrics.

I. INTRODUCTION

Biometrics is the branch of science which deals with automatic identification of human beings based on theirphysiological or behavioral characteristics [1]. It is well established fact that biometrics has many advantages over the traditional methods of identification such as passwords and identity cards. Passwords or pin numbers can be forgotten or shared; identity cards can be stolen, misplaced or shared etc. There are no such problems with the use of biometrics. Therefore biometrics has a big role to play in most of security systems.

In the last two to three decades biometrics has advanced much further and various biometric methods of identificationhave been developed. Currently used biometrics are keystroke, speech [2], signature [3], finger print [4], DNA, face [5], hand geometry [6], retinal scan, vascular pattern, iris [7] and Ear [8-11]. Ear is a relatively new biometrics and is becoming increasingly popular. It has certain advantages over other biometrics. For example, (a) Ear is rich in features and is a stable structure as compared to face. (b) It does not change with age. (c) Does not change with facial expressions. (d) The image of ear is much smaller than the image of other biometrics which is a distinct advantage in terms of memory and processing time.

For any new class of biometric to be acceptable, it has to be unique. The uniqueness of ear has been contested since it was proposed as biometrics by Iannarelli [12] who used manual techniques to identify ear images. Alfred Iannarelli [12]studied 10,000 ears and found that no two ears were the same. A second study carried out by Iannarelli was on twins both identical and non-identical where again the conclusion was in favor of ears being unique.

II. RELATED WORK

Victor et al. [13] and Chang et al. [14] used eigen ear for identification. The results obtained were different in both cases. Changs results show no difference in ear and face performance while victors results show that ear performance

is worse than face. According to Chang [14] views the difference in result might be due to usage of different image quality. Moreno et al. [10] used 2D intensity images of ears with three neural net approaches (Borda, Bayesian, Weighted Bayesian combination) for recognition. In his work, 6 images from 28 people were used to evaluate the recognition rate of about 93%. Chen et al. [15] studied two steps iterative closest point algorithm on 30 people with their 3D ear images that were manually extracted. The results reveal 2 incorrect matching out of 60 images.

A new ear recognition approach using coefficients of different wavelets transforms is proposed in this paper. Eight different families of wavelet are used to extract features. Euclidean distance is calculated for matching. The remainder of this paper is organized as follows: Proposed method of research is given in Section 3 and experimental results are reported in Section 4 whereas Section 5 concludes the paper.

III. PROPOSED METHOD

After image acquisition, there are three main steps in an ear recognition system; Preprocessing, Feature Extraction and Matching.

Enrollment is the first stage to identify a person, in which feature vectors are stored in the database. A certain number ofimages of the person are used for this purpose. After preprocessing, feature vectors are stored in the database for comparison at later time. This process is termed as training. Test image of the same person pass through the same process as in training phase except that instead of storing the feature vector in database, it is compared with the stored feature vectors and decision is based on minimum Euclidean distance between the test feature vector and stored feature vectors.A. Preprocessing

In preprocessing step, segmentation, normalization and conversion to grayscale are exploited. For segmentation, ear image is cropped from the original image which contains head of the person. Normalization is carried out to get the same size of feature vector for every ear image. For this purpose, cropped ear image is resized to a specific size. This process cater the changes occurred due to different sizes of the ear images. Normalized image is converted to grayscale image to make the feature extraction process fast.

The technique adopted for ear recognition in this paper is explained by Figure 1. It can be seen in figure 1 that each image has gone through the following steps before feature extraction:

Proceedings of International Bhurban Conference on Applied Sciences & TechnologyIslamabad, Pakistan, 11 14 January, 2010 192

Enrollment Image

Database

FeatureVectors

Decision

Feature Extraction

ApproximationCoefficient of

Wavelets Transform

Preprocessing

CroppingNormalization

Grayscale conversion

Matching

Euclidean Distance

Test Image

Figure 1: Steps of the proposed method

Ear image is cropped manually from the complete head image of a person.

Cropped ear image is resized. The colored ear image is converted to grayscale image.

Same number of features from each ear image is required to make consistency in feature vectors. Images are normalized bychanging their size to a specific size. Therefore, each image is resized to a fixed size of 6464 pixels. Each image was converted from RGB to grayscale and sent to feature extraction module. Figure 2 demonstrates the output at the end of preprocessing step. Figure 2(a) shows the actual image in the database and cropped image is visible in Figure 2(b) whereas Figure 2 (c) and Figure 2 (d) are the resized cropped images with RGB and grayscale respectively.

Figure 2: (a) Original image (b) Cropped ear image (c) Resized image (d) Gray scale image

B. Feature Extraction After normalizing the ear images, next step is feature

extraction. A new technique is implemented for feature extraction using various types wavelet transforms. The

wavelets transforms include, Haar, Daubechies, Symlets, Meyer, Biorthogonal, Reverse Biorthagonal and Coiflets. These wavelet transforms are applied to two sets of databases,namely the University of Notre Dame (UND) ear database [16]and ear database of Technical University of Madrid (TUM)[17] are used, both at level 1 and level 2. Wavelet decomposition of a normalized ear image is shown in Figure 3. In Figure 3 (a), upper left and right parts are the images from approximation coefficients and horizontal details respectively, whereas lower left and right parts are the images corresponding to vertical and diagonal detail of the decomposed image. The approximation coefficients are stored in a row vector instead of a matrix, which is the desired feature of the processed ear image. These feature vectors are used for training the database. Feature vectors of trained images are stored in the database. C. Matching

For matching, feature vector of test image is calculated. Euclidean distance for all the trained feature vectors in the database and the test image is calculated. The image corresponding to the minimum value of Euclidean distance matches with the image under consideration.

(a) (b)Figure 3: Wavelet decomposition of normalized image using (a) level 1, (b)

level 2

IV. EXPERIMENTAL RESULTS AND DISCUSSION

The proposed method is implemented in MATLAB 7.5 on a PC with 2.13 GHz Intel Core 2 processor and 1.0 GB RAM. In experiments, images from the UND ear database [16] and TUM [17] are used. UND Ear database [16] contains 464 images of 114 subjects with 12001600 pixels resolution with ear side view at Yaw of -90 and -75. A set of 32 people is used for experiments having six or more images each. The database from TUM [17] has also been tested using the proposed method. The results are collected on recognition rates using various types of wavelets both at level 1 and level 2. Figure 4, 5, 6 and Figure 7 show the results with number of training images versus correct recognition rate. It is clear from the tables that as the number of training image is increased, it increases the recognition rate. For two training images, the correct recognition rate achieved for TUM database is 96.08%

Proceedings of International Bhurban Conference on Applied Sciences & TechnologyIslamabad, Pakistan, 11 14 January, 2010 193

both at level 1 and level 2 for all types of wavelets used, please see Figure 4 and Figure 5for more details.

Figure 4: Results for TUM database using wavelets at level 1

Figure 5: Results for TUM database using wavelets at level 2

When the system is trained on three images the accuracy reaches as high as 99.02% at level 2 and 98.04% at level 1. Similarly, 100% accuracy is achieved when the number of training images is increased to five. The results for recognition rates obtained for UND database are given in Figure 6 and Figure 7.

It can be seen from these two tables that the accuracy for UND database is not as high as it is for TUM database. The accuracy achieved is 95.31% when the number of training images is five. The main reason for the difference in accuracy is the orientation of ear images and the angle of the camera with the ear in the UND database. In TUM database all images are properly aligned and all ears are orthogonal to the camera whereas ear images from UND database are manually cropped and are not properly aligned. Moreover each ear, captured twice, has pictures taken at three different angles with respect to the camera. Figure 8 shows the time utilized in training and

recognition using the TUM ear database with different number of training images while the applied wavelet is dmey at level 1.

Figure 6: Results for UND database using wavelets at level 1

Figure 7: Results for UND database using wavelets at level 2

The results are compared with the work done by other researcher in the same field using various techniques (see Table I). Proposed method performs better than Moreno et al. [10], Chen & Bhanu [8], Islam et al. [11], and Xie & Mu [18]. Their recognition results are less than 98% whereas our method has correct recognition rate up to 99.01% which is lower than Hurley & Nixon [9]. Moreno et al. [10] applied the Neural Network technique for training the system which is very time consuming process. Chen & Bhanu [8] quoted two incorrect matches out of 60 images (i.e. 96.6%) based on contour matching. Number of images used by Chen is very

Proceedings of International Bhurban Conference on Applied Sciences & TechnologyIslamabad, Pakistan, 11 14 January, 2010 194

TABLE ICOMPARISON OF RESULTS WITH OTHER METHODS

Sr. N

o.

Aut

hors

Tec

hniq

ue u

sed

No.

of

Subj

ects

Dat

abas

e Si

ze

Tot

al

trai

ning

im

ages

Tot

al te

st

imag

es

Acc

urac

y

1Hurley & Nixon [9]

PCA 63 254 63 63 99.2%

2Moreno et

al. [10]Neural

Net28 168 140 28 93%

3Chen &

Bhanu [8]ICP 30 60 30 30 96.6%

4Islam et al. [11]

ICP 100 200 100 100 96%

5Xie & Mu

[18]Improved LLE

79 632 553 7960.75

%

6 Proposed

Approximation coefficie

nts

32 192 96 96 82.29%

17 102 51 51 99.02%

small. Islam et al. [11] observed 3D local features on the range images using iterative closest point algorithm and reported maximum recognition rate of 96%. Xie & Mu [18] used Locally Linear Embedding (LLE), an unsupervised algorithm and found that it is not suitable for ear recognition.

02468

1012

1 2 3 4 5

T im

e(

sec

No. of Training Images

Time utilization for "dmey" wavelet using TUM Database

Training Time Rrecognition Time

Figure 8: Time utilization for complete TUM

They have improved LLE algorithm to raise recognition rate from 43.03% to 60.75%. This is the lowest recognition rate while comparing with other researchers even though the number of training images (total 553 images, 6 for each subject) is high. Main reason for such low recognition rate is that only one image per subject is used as test image. Hurley [9] used force field transformation for feature extraction based on potential channels and potential well to obtain 99.2% recognition rate. All of the mentioned methods have complex processes for obtaining feature vectors which are very time consuming whereas proposed method has the simplest feature extraction procedure and achieved comparable results in less time.

V. CONCLUSIONS

In this paper a new method of human recognition is proposed based on ear images using wavelet transforms. Proposed method is applied to two databases namely Technical University of Madrid database and University of Notre Dame ear database. Eight different types of wavelet transforms have been investigated and achieved an accuracy of up to 99.02% for one of the databases. A very simple method of feature extraction is proposed which helps to reduce the time utilized in obtaining features from the ear images and as expected, when the number of training images increases the recognition accuracy increases for all the wavelet types. The increase in accuracy is almost the same for all types of wavelet. In light of time analysis, it is recommended that probably it would be a good idea to use wavelet transforms db1 or db2 if time is going to play a role in your project as on average these two wavelet transforms have consumed lesser time during the recognition and the training process. The wavelet transform dmey turned out to be very slow but with highest recognition rate. The dmey took approximately three times more time than db1 or db2 wavelets.

ACKNOWLEDGMENT

Author would like to thanks computer vision and research laboratory at the University of Notre Dame for providing public biometric ear database Collection Set E on request.

REFERENCES

[1] A. K. Jain, A. Ross, and S. Prabhakar, "Introduction to Biometric recognition," IEEE Transaction on Circuits and Systems for Video Technology, vol. 14, pp. 4-20, 2004.

[2] W. M. Campbell, J. P. Campbell, D. A. Reynolds, E. Singer, and P. A. Torres-Carrasquillo, "Support vector machines for speaker and language recognition," Computer Speech and Language, vol. 20, pp. 210-229, 2006.

[3] A. K. Jain, F. D. Griess, and S. D. Connell, "On-line signature verification," Pattern Recognition, vol. 35, pp. 2963-2972, 2002.

[4] S. Prabhakar and A. K. Jain, "Decision-level fusion in fingerprint verification," Pattern Recognition, vol. 35, pp. 861-874, 2002.

[5] M. A. Anjum, M. Y. Javed, and A. Basit, "A New Approach to Face Recognition Using Dual Dimension Reduction," International Journal of Signal Processing vol. 2, pp. 1-6, 2005.

[6] R. Sanchez-Reillo, C. Sanchez-Avila, and A. Gonzalez-Marcos, "Biometric Identification through Hand Geometry Measurements," IEEE Trans. on Pattern Analysis & Machine Intelligence, vol. 22, pp. 1168-1171, 2000.

[7] A. Basit, M. Y. Javed, and M. A. Anjum, "Efficient iris recognition method for human identification," in International Conference on Pattern Recognition and Computer Vision (PRCV 2005), vol. 1, 2005, pp. 24-26.

[8] H. Chen and B. Bhanu, "Contour Matching for 3D Ear Recognition," in Seventh IEEE Workshop Application of Computer Vision, 2005, pp. 123-128.

[9] D. Hurley, M. Nixon, and J. Carter, "Force Field Energy Functionals for Ear Biometrics," Computer Vision and Image Understanding, vol. 98, pp. 491-512, 2005.

[10] B. Moreno, A. Sanchez, and J. F. Velez, "On the Use of Outer Ear Images for Personal Identification in Security Applications," in IEEE 33rd Annual International Carnahan Conference on Security Technology, 1999, pp. 469-476.

[11] S. M. S. Islam, R. Davies, A. S. Main, and M. Bennamoun, "A Fast and Fully Automatic Ear Recognition Approach Based on 3D Local Surface Features," in 10th International Conference on Advanced Concepts for

(

Proceedings of International Bhurban Conference on Applied Sciences & TechnologyIslamabad, Pakistan, 11 14 January, 2010 195

Intelligent Vision Systems, Lecture Notes In Computer Science, vol. 5259, 2008, pp. 1081-1092.

[12] A. Iannarelli, Ear Identification, Forensic Identification Series: Paramont Publishing, Freemont, Califoria, 1989.

[13] B. Victor, K. W. Bowyer, and S. Sarkar, "An Evaluation of Face and Ear Biometrics," in 16th Int l Conf. Pattern Recognition, 2002, pp. 429-432.

[14] K. Chang, K. Bowyer, and V. Barnabas, "Comparison and Combination of Ear and Face Images in Appearance-Based Biometrics," IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 25, pp. 1160-1165, 2003.

[15] C.-Y. Cho, H.-S. Chen, and J.-S. Wang, "Smooth Quality Streaming With Bit-Plane Labelling," Visual Communications and Image Processing, Proceedings of the SPIE, vol. 5690, pp. 2184-2195, 2005.

[16] "Ear database of Notre Dame University," http:\\www.UNDBiometricsDatabase.html, accessed 2007.

[17] M. A. Carreira-Perpinan, "Compression neural networks for feature extraction: Application to human recognition from ear images," in Faculty of Informatics, vol. MSc thesis: Technical University of Madrid, Spain, 1995.

[18] Z.-X. Xie and Z.-C. Mu, "Improved Locally Linear Embedding and Its Application on Multi-Pose Ear Recognition," in International Conference on Wavelet Analysis and Pattern Recognition, 2007, pp. 1367-1371.

Proceedings of International Bhurban Conference on Applied Sciences & TechnologyIslamabad, Pakistan, 11 14 January, 2010 196