Ear Biometrics for Machine Vision - CiteSeerX

7 downloads 105 Views 217KB Size Report
4] Alfred Iannarelli. Ear Identi cation. Forensic Identi cation Series ... 7] Steve Lawrence, C. Lee Giles, A.C. Tsoi, and A.D. Back. Face recognition: A convolutionalĀ ...
Ear Biometrics for Machine Vision M. Burge and W. Burger

Johannes Kepler University Department of Systems Science Computer Vision Laboratory A-4040 Linz, Austria [email protected]

Abstract: A new class of biometrics based upon ear features is introduced for use in the development of passive identi cation systems. The viability of the proposed biometric is shown both theoretically in terms of uniqueness and measurability over time and in practice through the implementation of a proof of concept system. Identi cation by ear biometrics is promising because it is passive like face recognition, but instead of the dicult to extract face biometrics it uses robust and simply extracted biometrics like those in ngerprinting.

The goal of automated personal identi cation that is both positive and reliable has attracted increased interest in automated biometrics. Automated biometrics can be divided into two broad classes, physiological which are based upon measurements of external physical traits, and behavioral which usually measure learned behaviors carried out by the subject. Behavioral methods include signature, voice, and keystroke biometrics, all of which vary over time and are dependent upon environmental factors. Machine vision research has concentrated on physiological, e.g. face, hand, eye and ngerprint biometrics. Section one introduces the eld of automated biometrics and contrasts the success of machine vision research on invasive methods like ngerprint identi cation against its apparent weakness on passive methods like facial biometrics. The viability of a new biometric based on ears and suitable for passive identi cation is argued in section two which examines the uniqueness and comparability over time of the ear. The short history of the ear in manual biometrics and the features developed by the forensic scientist Iannarelli are given in section three. Section four mathematically de nes identi cation and recognition while sketching our ear biometric identi cation system with emphasis on methods for ear localization. Finally two application scenarios are given followed in the last section by our current research directions.

1 Introduction to Automated Biometrics Automating face biometrics has been extensively studied in machine vision (see Chellappa [2] for a survey). Despite extensive research many problems in face recognition remain largely unsolved due to the inherent diculty of face biometrics. A wide variety of imaging problems, e.g. lighting, shadows, scale, and translation plague the attempt for unconstrained face identi cation [7]. In addition to the many imaging problems, it is inherently dicult to collect consistent features from the face as it is arguably the most changing part of the body due to e.g. facial expressions, cosmetics, facial hair and hair styling. The combination of the typical imaging problems of feature extraction in an unconstrained environment, and the changeability of the face, explains the diculty of automating face biometrics. Despite the attractiveness of face biometrics, e.g. they are easily veri able by non-experts, of biometrics, e.g. ngerprint-based, provide the basis for most commercial implementations. Unlike facial biometrics, ngerprint-based biometrics have been shown to be highly amenable to automation by machine vision techniques [1]. The automation of ngerprint biometrics began in 1971 [8] and has culminated in a number of commercial machine vision based systems. In terms of traditional image processing problems ngerprint imaging is done within a controlled environment, usually a specially designed scanner, which eliminates the problem of localization and artifacts from shadowing and lighting variations. Physical changes, a bane of facial biometrics, is a miniscule problem as the nger, baring surgery, remains comparatively constant over time. Machine vision techniques [5] have been applied successfully and have provided highly accurate and robust commercial systems which are in use worldwide. Fingerprints are not the only successful example of the application of machine vision techniques to automated biometrics, but both the three dimensional shape of the hand and retinal patterns have also been used. All of the biometrics which have been successfully automated using machine vision techniques are inherently invasive. They require the subject to participate actively in both enrolling into the system and in subsequent identi cations. The willing participation of the subject in the controlled environment of the machine vision systems is intrinsic in the success of the identi cation. One class of passive physiological biometrics are those based upon iris scans. Unlike retinal scans, which require close contact with the scanner, iris-based recognition has been reported from distances of 40 cm in controlled situations [8]. The unique collection of striations, pits, and other observable features of the iris along with the ease of segmenting the iris from the white tissue of the eye which serves as its background, make iris based biometrics attractive. The decided disadvantage is the small size of the iris which makes image acquisition from any distance greater then 40 cm problematic. To summarize the two classes of passive physiological biometrics which have been researched in machine vision up to now, face and iris-based techniques both have a number of drawbacks which

make their usage in commercial applications limited. Facial biometrics fail due to the changes in features caused by expressions, cosmetics, hair styles, and the growth of facial hair as well as the diculty of reliably extracting them in an unconstrained environment exhibiting imaging problems such as lighting and shadowing. Iris features on the other hand remain relatively consistent over time and are easy to extract, but acquisition of the image at the necessary resolution from a distance is dicult. Therefore, we propose a new class of biometrics for machine vision based upon ears which have both reliable and robust features and are localizable and segmentable from a distance for passive identi cation.

2 Viability of Ear Biometrics In proposing the ear as the basis for a new class of biometrics, we need to show that it is viable, i.e. unique to each individual, and comparable over time. In the same way that no one can prove that ngerprints are unique, we can not show that each of us has a unique ear. Instead, we will assert that this is probable and give supporting evidence by examining two studies [4]. The rst study compared over 10,000 ears drawn from a randomly selected sample in California, and the second study examined fraternal and identical twins, in which physiological features are similar. The evidence from these studies supports the hypothesis that the ear is a unique physiological features, since in both studies all examined ears were found to be unique though identical twins were found to have similar, but not identical, ear structures especially in the concha and lobe areas. Having shown uniqueness it remains to ascertain if the ear provides biometrics which are comparable over time. It is obvious that the structure of the ear does not change radically over time. The medical literature reports [4] that ear growth after the rst four months of birth is highly linear, i.e. proportional. It turns out that even though ear growth is proportional, gravity can cause the ear to undergo stretching in the vertical direction. The e ect of this stretching is most pronounced in the lobe of the ear, and measurements show that the change is non-linear. The rate of stretching is approximately ve times greater then normal during the period from four months to the age of eight, after which it is constant until around 70 when it again increases. We have shown that biometrics based upon the ear are viable in that ear anatomy is probably unique in each individual and that features based upon measurements of that anatomy are comparable over time. Given that they are viable, identi cation by ear biometrics is promising because it is passive like face recognition, but instead of the dicult to extract face biometrics it can use robust and simply extracted biometrics like those in ngerprinting.

1b 1

3

2 3

8

5 6

1c

1a 3

8

9

7

10

3

7

1d 5

4 9

11 8 6

2

4 12

Figure 1: A) The external anatomy of the ear. 1 Helix Rim, 2 Lobule, 3 Antihelix, 4 Concha, 5 Tragus, 6 Antitragus, 7 Crus of Helix, 8 Triangular Fossa, 9 Incisure Intertragica. B) The locations of the anthropometric measurements of the external ear as used in the \Iannarelli System".

3 Iannarelli's Ear Biometrics An anthropometric technique of identi cation based upon ear biometrics was developed by A. Iannarelli [4] in 1949. The \Iannarelli System" is based upon the 12 measurements illustrated (Figure 1, left). The locations shown are measured from specially aligned and size-normalized photographs of the right ear. Each photograph is aligned during development so that the lower tip of a standardized vertical guide on the development easel touches the upper esh line of the cocha area whilst the upper tip touches the outline of the anti-tragus (Figure 1, right). Since each ear is aligned and scaled during development, the resulting photographs are normalized in size and orientation, enabling the extraction of comparable measurements directly from the photographs. The distance between each of the numbered areas in Figure 1 (right) is measured in units of 3 mm and assigned an integer distance value. These twelve measurements, along with information on sex and race, are then used for identi cation. The system as stated provides for too small of a classi cation space as within each sex and race category a subject is classi ed into a single point in a 12 dimensional integer space where each unit on an axis represents a 3 mm

Figure 2:

A) Segmented outer ear, B) Segmented inner ear

measurement di erence. Assuming an average standard deviation in the population of four units, 12 mm, then the 12 measurements provide for a space with less than 17 million distinct points. Though simple remedies (e.g. the addition of more measurements or using a smaller metric) for increasing the size of the space are obvious, the method is additionally not suited for machine vision because of the dicultly of localizing the anatomical point which serves as the origin of the measurement system. All measurements are relative to this origin which if not exactly localized, results in all subsequent measurements being incorrect. In the next section we present a simple proof of concept implementation which avoids the problem of localizing anatomical points and the frailty of basing all subsequent feature measurements on a single such point.

4 Automating Ear Biometrics The goal in identi cation is to verify that the biometric extracted from the subject suciently matches the previous acquired biometric for that subject. Let be the subject at the time of identi cation and the subject at time of enrollment, further let s = ( ) represent a function which extracts some biometric from a subject as a vector s , and let (s s ) compute some previously de ned distance metric between the two biometrics s and s . Identi cation then is s

s

0

b

s

f s

b

d b ;b

b

b

0

0

Figure 3:

A) Original ear image, B) Edge map, C) Gradient

the task of determining if (s s ) , where is the given acceptance threshold. Since the subject and environment changes over time, a certain tolerance in the matching criterion must be permitted. This tolerance can be de ned in terms of the false reject rate (FRR) and the false acceptance rate (FAR) exhibited by the system. A system is usually designed to be tunable to minimize either the FAR or the FRR, i.e. in the given formulation by lowering or raising respectively, depending upon the type of security which is required. The problem of recognition is harder then that of identi cation since the system must determine if the subject's identity can be veri ed n against any o previously enrolled subject. If the system's 1    enrolled identities are the set , I = 0 1    n then recognizing some subject is equivalent to nding the least member of the set f j 2 I ^ ( ( ) ) g. We have developed a simple machine vision system as a proof of concept of the viability of ear biometrics for passive identi cation. The system implements ( ) by rst localizing and segmenting the ear of a subject from a greyscale CCD camera acquired image using deformable contours [6] on a Gaussian pyramid representation of the image gradient (Figure 2). Once segmented, the vector s consisting of a number of scale, rotation, and translation invariant areabased features [3] is computed and the distance ( ( ) s) between the stored biometric vector for is computed and if it less then the established acceptance threshold then identi cation is veri ed. We divide the identi cation system down into three general phases: localization and segmend b ;b

0

< t

t

t

b ;b ;

i i

;b

s

d f s

0

;i

< t

f s

s

0

0

0

b

d f s

s

0

;b

t

The de nition of I illustrates the point that normally the biometric bs extracted by the function f (s) from a subject s is stored by the system and not some direct representation e.g. an image, of s. 1

tation, feature extraction, and feature comparison, and will concentrate on the rst stage as a wide variety of methods developed for other machine vision based biometrics are applicable to the last two phases. We examine two ways of localizing the ear in an image. In the rst method the distinctive outline of the ear formed by the Helix rim and lobule (Figure 1) are searched for using a generalized Hough transform (GHT). Learned templates for both parts are combined into a master template and the search method attempts to optimize maxima in each template while minimizing the resulting di erences between the two templates. The result is two curves which loosely bound the ear around the top and bottom Figure 3 part B shows the computed edge map on which the edges corresponding to the marked curves Shown as think lines in Figure 3 part A show up clearly. In the second method deformable templates of the inner and outer ear based on active contour models are searched for. We are currently evaluating the two methods as well as examining a method based on Fourier descriptors. Ear biometrics can be used as a supplementary source of evidence in identi cation and recognition systems, e.g. a system designed for face recognition already includes all the necessary hardware for capturing and computing ear biometrics. A typical example is supplementary identi cation at an automated teller machine (ATM). Most ATMs contain a camera mounted so that it records the face of the user during the transaction. These cameras could be supplemented with a simple optical mirror setup to allow simultaneously recording of the face and the ear of the user. While a user identi es themselves to the ATM by inserting the bank card and keying in their personal identi cation number (PIN), the camera simultaneously records the face and ear of the user and uses ear biometrics to supplementary verify the identi cation of the user. A second scenario where passive ear biometrics are ideal is with di erent levels of security levels and are accessible to di erent groups of people. Commonly this situation is solved through the use of active badges. These badges, usually small identi cation cards containing passive transmitters, are automatically queried at each point of access, e.g. a door or elevator. Access to a restricted area e.g. unlocking of the door, is automatically allowed or disallowed according to the identity supplied by the badge. A well known drawback of such a system is that someone can illegitimately obtain and use the active badge of another to gain access to restricted areas, for this reason cameras are often installed to monitor the access points. The cameras record access and in the case that illegitimate access to an area is later discovered the taped video record can be examined to visually determine the perpetrator - but only after the fact. A prohibitively expensive solution, which defeats the purpose of automation, is to have a human analyze each request for access by comparing the identi cation reported by the active badge and the visual provided by the camera to some known visual of the subject. Ear biometrics can be used in such a scenario to allow more secure automated access. In the proposed solution, as the subject approaches the access point the active badge is queried and the identi cation is ascertained, at the same time an image of the subject's ear is acquired

and biometrics are used to verify the identi cation provided by the active badge. In the case that the the two identi cations do not match, the video of the access point including the subject is displayed at the central security counsel along with an image retrieved from the security database of the subject who should be in possession of that badge. The two images can then be visually compared by the security personnel and appropriate action taken.

5 Conclusion A simple proof of concept system for passive identi cation from ear biometrics was presented. The system lends support to the theoretical evidence that ear biometrics are a viable and promising new passive approach to automated human identi cation. One of the primary arguments against the usage of ear biometrics is that ears are often occluded, e.g. by hair or hats, rendering them unusable. In selected populations e.g. those with short hair as in the defense industry, ear biometrics are applicable and especially useful when used to supplement existing automated methods. We are currently working on the problem of ear localization as well as building both a statistical model of the ear and a larger database of ear images for retrieval and comparison experiments.

References [1] G.T. Candela and R. Chellappa. Comparative performance of classi cation methods for ngerprints. In NISTIR, 1993. [2] R. Chellappa, C.L. Wilson, and S. Sirohey. Human and machine recognition of faces: A survey. PIEEE, 83(5):705{740, May 1995. [3] M.K. Hu. Visual pattern recognition by moment invariants. IEEE Trans. Information Theory, 8:179{187, 1962. [4] Alfred Iannarelli. Ear Identi cation. Forensic Identi cation Series. Paramont Publishing Company, Fremont, California, 1989. [5] K. Karu and A.K. Jain. Fingerprint classi cation. PR, 29(3):389{404, March 1996. [6] K.F. Lai and R.T. Chin. Deformable contours: Modeling and extraction. PAMI, 17(11):1084{1090, November 1995. [7] Steve Lawrence, C. Lee Giles, A.C. Tsoi, and A.D. Back. Face recognition: A convolutional neural network approach. IEEE Transactions on Neural Networks, 8(1):98{113, 1997. [8] B. Miller. Vital signs of identity. IEEE Spectrum, 83:22{30, February 1994.