Ear Biometrics in Passive Human

Identiﬁcation Systems

Michał Chora

Image Processing Group, Institute of Telecommunications

University of Technology & Agriculture

S. Kaliskiego 7, 85-796 Bydgoszcz

Abstract. The article discusses various issues concerning ear biometrics in hu-

man identiﬁcation systems. The major advantage of ear as the source of data for

human identiﬁcation, is the ease of image acquisition, which can be performed

even without examined person’s knowledge. Moreover, user’s acceptability and

easy interaction with the system make ear biometrics a perfect solution for secure

authentication e.g. in access-control applications. In the article the focus is on the

ear biometrics motivation, ear identiﬁcation system design and interaction with

the user. Feature extraction methods from ear images are also discussed.

1 Motivation for Passive Human Identiﬁcation Systems

Biometrics methods of human identiﬁcation have gained much attention recently, mainly

because they easily deal with most problems of traditional identiﬁcation, since users are

identiﬁed by who they are, not by something they have to remember or carry with them.

The passive methods of biometrics do not require any action from users. Some systems

can even verify the identity of humans without their cooperation and knowledge, which

is actually the future of biometrics. Crowd-surveillance, monitoring of public places

like airports or sports arenas are the most important applications that need such solu-

tions. Possible passive methods include popular and well-examined face recognition,

but one of the most interesting novel approaches to human passive identiﬁcation is the

use of ear as the source of data.

The most interesting human anatomical parts for passive, physiological biometrics sys-

tems based on images acquired from cameras are face and ear. Both of those body

parts contain large volume of unique features that allow to distinctively identify many

users and can be implemented into efﬁcient biometrics systems for many applications.

However, still the automated system of ear recognition has not been commercially im-

plemented, even though there are many advantages of using ear as a source of data for

person identiﬁcation (small size, stable features, uniform colours). Furthermore, ear is

one of our sensors, therefore it is usually visible (not hidden underneath anything) to

enable good hearing. In the process of acquisition, in contrast to face identiﬁcation sys-

tems, ear images cannot be disturbed by glasses, beard or make-up. However, occlusion

by hair or earrings is possible, but in access control applications, making ear visible is

not a problem for user and takes just single seconds (Fig. 1).

Chora

s M. (2006).

Ear Biometrics in Passive Human Identiﬁcation Systems.

In 6th International Workshop on Pattern Recognition in Information Systems, pages 169-174

DOI: 10.5220/0002474901690174

 SciTePress

Fig. 1. Ear visibility can be easily achieved in access control systems.

2 Ear Biometrics Systems: Feature Extraction Methods

The ﬁrst, manual method used by Iannarelli was based on measuring the distances be-

tween speciﬁc points of ear [1]. Another well-known method by Burge and Burger [2]

was based on building neighbourhood graph from Voronoi diagrams of the detected

edges. Hurley et al. [3] introduced a method based on energy features of the image.

Another method used by Victor et al. [4], in the experiment comparing ear and face

properties in order to identify humans was based on PCA. Their work proved that ear

images are a very suitable source of data for identiﬁcation and their results for ear im-

ages were not signiﬁcantly different from those achieved for face images. The method,

however, was not fully automated, since the reference points had to be manually inserted

into images. Another approach presented by Moreno et al. [5] was based on macrofea-

tures extracted by compression networks. Recently, various approaches towards 3D ear

biometrics has been developed and published [6][7].

2.1 Geometrical Feature Extraction

Our methods based on geometrical feature extraction are motivated by actual proce-

dures used in police and forensic evidence search applications. Nowadays, human ears

and earprints are standard features of identity taken into account by forensic specialists

and criminal policemen (so called ear otoscopy). In reality, well-established proce-

dures of handling ear evidence are based on geometrical features such as size, width,

height and earlobe topology [8].

Therefore we decided to compute geometrical parameters of ear contours extracted

from ear images. Such approach gives information about local parts of the image, which

is more suitable for ear biometrics than global approach to image feature extraction.

Moreover, geometrical features of extracted contours are more adequate for ear identi-

ﬁcation than colour or texture information, which is not distinctive enough for various

ear images. On the other hand, contours corresponding to earlobes are very diversiﬁed

and contain enormous amount of information allowing ear identiﬁcation.

In the proposed method of feature extraction from ear images in order to perform human

identiﬁcation we use the geometrical parameters and properties of ear contour images.

The ﬁrst step of the method is the extraction of contours from ear images in such way,

that the extracted contours contain distinctive information about shape and geometrical

properties of given ear. Then for each of the extracted contours we construct the feature

170

vector on the basis of geometrical parameters.

Our method of human identiﬁcation based on ear image analysis consists of the follow-

ing stages:

– ear image preprocessing - we perform such operations as contrast enhancement,

ﬁltration and histogram equalization,

– contour detection - we use local-based method based on pixel illumination, mean

and variance changes in 3 × 3 window,

– contour processing - the aim of contour image processing is the selection of con-

tours containing the most distinctive information characterizing human ear images.

For each extracted contour, we calculate its length, and then on the basis of the se-

lection algorithm we eliminate contours which are classiﬁed as short. Usually we

obtain binary ear contour images with 7-10 longest contours.

Fig. 2. Longest and numbered contours for the test images ’macﬁr’ and ’szysob’.

– image normalization - image size is normalized and invariance to rotation, transla-

tion and scaling is achieved,

– geometrical feature extraction algorithms - we developed 5 novel algorithms,

– classiﬁcation algorithms based on feature vector distances in feature space.

On the basis of the extracted and selected contours we proposed 5 methods of feature

extraction, described in our previous work [9][10][11]:

– concentric circles based method - CCM ,

– contour tracing method - CT M ,

– angle-based contour representation method - ABM,

– geometrical parameters method GP M which is divided into:

• triangle ratio method - GP M − T RM,

• shape ratio method - GP M − SRM.

In the feature extraction methods binary contour images with the selected number of

contours and normalized coordinates are processed. Feature extraction methods are

based on concentric circles centred in the centroid point obtained from binary contour

image. Contour tracing method is also based on extracting characteristic contour points

analogically to ﬁngerprint minutiae search [12]. Moreover, methods based on the de-

veloped geometrical parameters, calculated for selected contours, were developed.

After experiments we came into conclusion that the proposed geometrical methods,

which were motivated by the manual process of feature extraction used in criminology,

171

allow effective person identiﬁcation on the basis of features extracted from ear images.

The most effective methods were GP M and CT M and the achieved results are compa-

rable with face recognition systems. All the tests were performed on our own ear image

database, which is secured and used only for research in our laboratories. According

to the Polish law (Personal Data Protection Law), human features which are analysed

and processed in the biometrics systems are a subject of protection. In the identiﬁcation

systems the problem of overcast ears (e.g. by hair) is marginal, since there is always

the possibility of a proper ear image acquisition (with the user’s cooperation) (Fig. 1).

Therefore in the experiments we focused on images of the visible ears. The cumulative

results in the ﬁrst scenario for all the methods are presented in the Table 1.

Table 1. The cumulative results of the developed identiﬁcation methods for the ﬁrst scenario. The

presented parameter is the standard false rejection ratio F RR.

method number of tests number of acceptances number of rejections FRR

CCM 40 36 4 10

CT M 40 39 1 2.5

ABM 40 36 4 10

GP M 40 40 0 0

3 Applications and Test Scenarios

In order to verify the effectiveness of the presented feature extraction methods we pro-

posed two experimental scenarios. The ﬁrst one involves the ﬁnite ear images database.

One of the users, who took part in the enrolment process and his ear image is surely

stored in the database, is chosen randomly. The acquisition of the user’s test ear image

is performed. Next, we compute the feature vectors for the test user and we search for

the corresponding image from the database. In result of such scenario we obtain one

ear image for which the computed feature vectors are the closest to test image feature

vectors in terms of distance in the feature space [13]. The ﬁrst scenario reﬂects such ap-

plications of the biometrics identiﬁcation systems as access control to controlled places

(company’s buildings, rooms). For one test user only the identiﬁcation decision (yes/no)

is expected.

In the second scenario the ﬁnite ear image database is also considered, but it is not

known if the test user’s ear image is also in the database (the user didn’t have to take

part in the enrolment step and therefore his ear image might not be stored in the data-

base). In the result of the second scenario we obtain H images with the most similar

feature vectors in terms of distance in the feature space.

Such scenario is similar to Content Based Image Retrieval Systems, in which for the

user’s input query (test image), a number of the most similar images are found. It is not

a popular situation in real-time biometrics identiﬁcation systems, but it might be useful

in criminal applications. In such case (where there is no place for mistake) the trained

172

policeman/forensic expert gets H images with the most similar feature vectors to the

input query and, on the basis of his knowledge, he veriﬁes the result.

4 Ear Biometrics System Design and Studies on Interaction with

Users

Besides veriﬁcation of the effectiveness of the proposed methods, we also studied inter-

action with the users aspects. Those aspects are always crucial for biometrics and other

HCI (Human Computer Interaction) systems development.

In order to perform the experiments our own ear image database of over 100 users’ ears

was used. In the process of enrolment, by analogy with face recognition methods, we

store 10 images for each person (perpendicular to the camera 0

◦

, 30

◦

and −30

◦

, 60

◦

and −60

◦

) for 2 values of illumination in the room.

We mainly worked with students who agreed to take part in the enrolment process. All

of the students had little time to make decision whether to take part in experiments or

not, and all of them were astonished by our request (it is due to the fact that none of

them had heard about ear biometrics before).

During ear image enrolment we came to important conclusion that potential users are

not afraid to interact with ear biometrics system, which means that ear biometrics is

less stressful than dactyloscopy. Moreover, our test users admitted that they would feel

less comfortable while taking part in face images enrolment (people tend to care how

they look on photographs). Furthermore, in ear biometrics there is no need to touch

any devices and therefore there are no problems with hygiene. Therefore ear biomet-

rics seems to be a perfect solution for passive identiﬁcation systems. After user studies

we concluded that ear biometrics is more favourable than ﬁngerprint, iris or even face

recognition.

Since security of biometrics features is a hot topic nowadays, it is worth mentioning

that ear images are more secure than face images, mainly because it is very difﬁcult to

associate ear image with a given person (in fact, most of users are not able to recognize

their own ear image). Ear image databases do not have to be as much secured as face

databases since the risk of attacks is much lower.

Therefore, storing and processing ear images in a biometric identiﬁcation system is

more secure that in the case of face databases. It is more difﬁcult to perform identity-

stealing on the basis of ear images. It is an important issue for biometrics system devel-

opers since the number of such crimes is tripled each year.

5 Conclusion

Human ear is a perfect source of data for passive person identiﬁcation in many appli-

cations. In a growing need for security in various public places, ear biometrics seems

to be a good solution, since ears are visible and their images can be easily taken, even

without the examined person’s knowledge.

Moreover users’ acceptance of ear biometrics is very high. We examined user interac-

tion in the enrolment step and we concluded that ear images acquisition is more user-

friendly and less stressful than other methods, even face recognition.

173

The proposed geometrical feature extraction methods can be used to determine per-

sonality of some individuals, for instance terrorists at the airports and stations. Access

control to various buildings and crowd surveillance are among other possible applica-

tions of ear biometrics.

References

1. A. Iannarelli, Ear Identiﬁcation, Forensic Identiﬁcation Series, Paramont Publishing Com-

pany, 1989.

2. M. Burge, W. Burger Ear Biometrics, in: Biometrics: Personal Identiﬁcation in Networked

Society (Eds: A.K. Jain, R. Bolle, S. Pankanti, 273-286, 1998.

3. D.J. Hurley , M.S. Nixon, J.N. Carter, ”Force Field Energy Functionals for Image Feature

Extraction,” Image and Vision Computing Journal, vol. 20, no. 5-6, pp. 311-318, 2002.

4. K. Chang, B. Victor B., K.W. Bowyer, S. Sarkar, ”Comparison and Combination of Ear and

Face Images in Appearance-Based Biometrics,” IEEE Trans. on PAMI, vol. 25, no. 8, pp.

1160-1165, 2003.

5. B. Moreno, A. Sanchez, J.F. Velez, ”On the Use of Outer Ear Images for Personal Identiﬁ-

cation in Security Applications,” Proc. of IEEE Conf. On Security Technology, pp. 469-476,

1999.

6. H. Chen, B. Bhanu, ”Contour matching for 3D ear recognition,” Proc. of Workshop on Ap-

plications of Computer Vision (WACV), 123-128, 2005.

7. P. Yan, K. W. Bowyer, ”ICP-based approaches for 3D ear recognition,” Proc. of SPIE Bio-

metric Technology for Human Identiﬁcation, 282291, 2005.

8. J. Kasprzak, Forensic Otoscopy (in Polish), UWM Olsztyn, 2003.

9. M. Chora

s, ”Ear Biometrics Based on Geometrical Method of Feature Extraction”, in: F.J

Perales and B.A. Draper (Eds.): Articulated Motion and Deformable Objects, LNCS 3179,

Springer-Verlag, pp. 51-61, 2004.

10. M. Chora

s, ”Ear Biometrics Based on Geometrical Feature Extraction,” Journal ELCVIA,

vol. 5, no. 3, pp. 84-95, 2005.

11. M. Chora

s, ”Human Identiﬁcation Based on Ear Image Analysis” (in Polish), Ph.D. Thesis,

ATR Bydgoszcz, 2005.

12. D. Maio, D. Maltoni, ”Minutiae Extraction and Filtering from Gray-Scale Images,” in: L.C.

Jain et al. (Eds): Intelligent Biometric Techniques in Fingerprint and Face Recognition, CRC

Press, pp. 1-34, 1999.

13. A. Kapczy

nski, ”The Valuation of Biometrics Methods in the Process of User Authentication

(in Polish),” Studia Informatica, vol. 22, no. 1(43), 2001.

174