International Journal of Communication & Information Technology (CommIT) http://msi.binus.ac.id/commit/
Vol. 6 No. 1 May 2012, pp. 32-36
THE IDENTIFICATION OF EAR PRINTS USING COMPLEX GABOR FILTERS
Alexander A S Gunawan; Heni Kurniaty; Wikaria Gazali
Mathematics and Statistics Department, School of Computer Science, Bina Nusantara University Jl. KH. Syahdan No. 9, Jakarta 11480, Indonesia
aagung@binus.edu; wikaria@binus.edu
Abstract: Biometrics is a method used to recognize humans based on one or a few characteristics physical or behavioral traits that are unique such as DNA, face, fingerprints, gait, iris, palm, retina, signature and sound. Although the facts that ear prints are found in 15% of crime scenes, ear prints research has been very limited since the success of fingerprints modality. The advantage of the use of ear prints, as forensic evidence, are it relatively unchanged due to increased age and have fewer variations than faces with expression variation and orientation. In this research, complex Gabor filters is used to extract the ear prints feature based on texture segmentation. Principal component analysis (PCA) is then used for dimensionality-reduction where variation in the dataset is preserved. The classification is done in a lower dimension space defined by principal components based on Euclidean distance. In experiments, it is used left and right ear prints of ten respondents and in average, the successful recognition rate is 78%. Based on the experiment results, it is concluded that ear prints is suitable as forensic evidence mainly when combined with other biometric modalities.
Keywords: Biometrics; Ear prints; Complex Gabor filters; Principal component analysis;
Euclidean distance
INTRODUCTION
Surveillance technologies now widely used by the society. The very fast growth of security cameras is caused by the numbers of increasing crimes, and also the environmental needs that are safe but lack of labors to oversee the neighborhood.
The non-invasive biometric technology users are the important since they improve the usability of the security cameras by building automated visual surveillance systems. Non-invasive biometric technologies is the example of the improvement by building the facial recognition and the ear prints that are useful for forensic investigation based on the camera records. Therefore, nowadays, criminals are familiar with disguises to hide their identities, so the facial recognition becomes less effective for this forensic investigation needs. Here, the ear prints recognition can be the alternative way that is promising. For example, it can be seen in Fig. 1, the suspect uses the pet cap to hide his identity.
However, his ear can be clearly seen, although with a low resolution. Notice that the disguise, that is used to avoid the identification, can be socially accepted.
Fig. 1: An example of the use of ear prints for the forensic investigations
Human ears are the very stable structures, which are rich in information and easily imaged for biometric identifications. The biometric of ear prints is often compared with the biometric of face [1]. Ear has several excellences rather compared face, especially because ear is relatively unchanged due to increased age, colors of skin, and hair styles [2]. It has fewer variations than faces with expressions variations and orientations [3]. The uniqueness of the form of ear to any human being is useful for forensic investigation, because it is a
part of th ear print long tim biometric ear remai a distanc investiga identifica ear, disas face cann N organized ear and Then, it complex of ear fea it is cont compone of featu remains the lowe PCA bas described along wit
METH
Th has been However whether structure the poss exactly t small. Th identifica 1989, wh ears and In 1906, of ears, determin
he body that is has been used me [4]. If i
cs, such as iri ins greater an ce by using a ation, ear actu ation system.
ster victims c not be recogni Next, the or
d as follows:
the identific t continues w
Gabor filters atures based o tinued with th ent analysis (P ures where t
maintained. T er dimension
ed on Euclide d about the e th the conclus
HOD
he use of ear done long, at r until now, that ear is un of ear in Fig sibility of the
the same fro he most famou ations was m
hen he collec found that al Imhofer foun only four ch e the ear whic
Fig 1: The S
s rarely of hard d in the foren it is compar is, retina, and nd easier to be camera. Besi ually is also a
By using th an be identifi ized anymore.
ders of wri Firstly, the ex cation system with the exp s and its use i
on segmentati he description PCA) to reduc the variation The classifica
room that is ean distance. L experiment tha sion that is obt
as the modali t least a hundr , still a qu nique to each gure 1, it is qu e structure o om two indiv us research in made by Alfre
cted more tha ll of them we nd that from fi haracteristics ch is unique [5
Structure of ea
dly noticed. T sic science fo red with oth d fingerprint, t e apprehended ides for foren part of disas he biometric ied although t .
iting here a xplanation abo m of ear prin planation abo in the extracti ion texture. A n about princip
ce the dimensi on its ima ation is done defined by t Lastly, it will at is conducte tained.
ity in biometr red year ago [ estion emerg
individual. T uite complex, of ears, that viduals, is ve n the field of e ed Iannarelli an ten thousa ere different [ ive hundred da
are needed 5].
ar [6]
The r a her the d at sic ster of the are out nts.
out ion And pal ion age in the be ed,
ics 4].
ges The so is ery ear in and 2].
ata to
bi of fr w bi on in th bi bi bi th fo re su pr pr in pr m im By of re th do co im eq fe th on th an fe m lo ba
V V al to
A speci iometric (200 f biometrics, i
om the stabi which are: grow
iometric will nly a few per nfluence of th he biometric o
iometrics of iometric of e iometric that i hough there i ound fifteen p esearch of ear uccess of the r revious reason rints identifica The sys ncludes the m
re-processing, matching (Fig.
mage is the ea y detecting th f the form of esearch, the V he ear. The n oing a norm onversion in mprovement o qualization me eature by usin he structure of n the segment he ear-matchin nalysis (PCA) eatures where maintained. Ne ower dimensio ased on the Eu
Fig. 2: Th
Viola Jones M The firs Viola Jones m lgorithm of ob o detect objec
ialist compan 3) has compa including the lity aspect ag wth, aging, du
have a high rhaps variation he variables. B of ear is mor face, signat ar is only le is invasive lik s a fact that percent from r prints is ver
research from ns, we are doi ations.
stem of ear i main module,
, extraction 2). The dete arly step to do he ear, it will ear that fit to Viola-Jones m
next step is malization of nto grayscal of the light in ethod. Next, i ng the comple f the ear can b tation of the t ng by using th
that is used t the variation ext, a classifi on room that uclidean distan
he process of
Method st step is th method. Viola bjects detectio cts. This is be
ny of Bromb ared a differe form of ear e gainst some ung, and injury
constancy, so n, although it Based on this
re permanent ture, and vo ss permanent ke DNA and t ear prints h m the crime s ry limited be m fingerprints.
ng a research identification
as follows:
feature, a ection of the o the ear iden be obtained t o be identified method is used
the pre-proce f the image
le image, ntensity with h it is done an e x Gabor filter be made contr texture. The la he principal co o reduce the d n on its image fication is don
is defined by nce.
ear identificat
he ear detecti a Jones meth ons that are o ecause this m
ba GmbH nt variety especially variables, y. A good o there is t is in the research, t than the oice. The t than the
iris. Even have been scene, the ecause the . With the about ear generally detection, and ear-
ear in an ntification.
the image d. In this d to detect essing by e size, a
and an histogram extraction rs, so that rast based ast step is omponent dimension e remains ne in the y the PCA
tion
ion using od is the often used method has
The proc Viola Jo process, from the stages in The Haa of integr Adaptive classifica Pre-Proc W very diff image. T image is the illum (noise) i grayscale equalizat equalizat image th seem mo more or used in t image on Complex
Th distingui capability spatial or observed Dennis G simulate system, orientatio the Gab recogniti can be us based on orientatio In which is impulse Gabor fil
and can b
cess of object ones method
to obtain the e data trainin n Viola Jones ar-like extracti ral image. (3 e Boosting (A ation with the
cessing With the size o
ficult for us to Then, a proces
needed. Next mination and t in the image e image, a tion method.
tion method hat is harmon ore contrast w less almost th the overall im nly.
x Gabor Filte he ability of shing various y in identifyin rientations fro d. Gabor filte Gabor [8] is
the charac in isolating ons from imag bor filter ap on applicatio sed to conduct n the repeated ons and freque n spatial, a Ga being modul response of lter is:
be illustrated a
t detection in is done, aft classification ng. There are s method [7],
ion features.
) the learnin AdaBoost). (4) Cascade of C
of different im o compare the ss to normaliz
t, to reduce th the impacts fr is by chang and using The purpos is to produc nious, so that with the light
he same. This mage or in so
ers
the human v s textures is ng variety of om the texture r that is first one of the cteristics of g certain fr
ges. This cha ppropriate fo on in an imag t textural imag d local pattern
encies.
abor function lated by Gau
a two dimen
as the followin
an image usi ter the learni n that is form e four princip which are: ( (2) The maki g features w The compou Classifier.
mages, it will e features in t e the size of t he differences rom disturban ging it into t the histogra e of histogra ce a histogra the image w intensity that s method can ome parts of t
visual system s based on t frequencies a e which is bei t introduced filters that c human visu requencies a aracteristic ma or the textu ge. Gabor fil ge segmentati n in the certa
is a sinusoid uss function. A
nsions compl
ng:
ing ing med ple (1) ing with und
be the the of nce the am am am will t is
be the
in the and ing by can ual and ake ure lter ion ain dal, An lex
re
st Th H Fr lo co ra Th th
Th th
Th G m in ex fe fe
Fig 3: Impu
In the fr epresented as t
There a ipulated in th he six param Here are the ex requency (F) ocation of th
onstant band w ange of filters.
he variable σ
x
he spatial frequ
he variable σ
y
he angular com
he position (F Gabor filter in meticulously st nformation pro After o xtraction featu eatures, whic eature, that id
ulse Response Gabor fil
requency dom the following:
are six param he implement meters are: F, xplanations:
and orientatio he filter. Bϕ, width of frequ .
x related to the uency compon
y related to th mponent.
F, θ) and ban n the frequen
tipulated in or operly.
obtaining Gab ure can be c h can be c defined as:
e of two dimen lter
main, Gabor fil :
meters that s tation of Gab
θ, σ
x, σ
y, BF
on (θ) define t and Bθ exp uencies and th
e response of nent.
he response of
nd width (σ
x, ncy domain s rder to captur bor features, conducted. On hosen, is th
nsions
lter can be
should be bor filters.
F, and Bθ.
the central press the he angular
-6 dB for
f -6dB for
σy) from should be re textural
then the ne of the he energy
The prin and Euc
Th a method reduce d method i space tha much inf the origin not much space. By of course that shou Eu that can image of the imag weight va counted b measures [10]. The conducte
RESUL
Th identifica Studio 2 language working begins w on the c detect w frames in If the im input of be conve equalizat disturban difficultie the imag then the complex submenu can be ch
In of ear is the came through previousl
ncipal compo lidean distan he Principal C d that involve data and extra is projecting at has lower d formation as n dimension, h that lost afte
y minimizing e it will allev uld be conduct uclidean dista
be used to f ear, which i ges of ear wh
alue between by using the E s of similaritie e calculation ed based on th
LTS AND D
he applicatio ation is made 2010 softwar e C# and t mechanism when the user
camera. Next whether there
n the video th mage of the ea the image wi erted into gra tion is conduc nce (noise) in
es in the extr ge is consider process of the Gabor filter u setting, the
hosen.
n the experime taken in a p era. The image
the processi ly. And last i
nent analysis nce
Component An es a mathemat action feature
the origin sp dimension, by possible whic
are maintaine er being proje the dimensio viate the comp
ted by the syst ance is one
do the class is new (unkno hich has been the two vecto Euclidean dis es between th
of the Euclid e following fo
DISCUSSIO
on program e by using M
re, with the the library of this appli conduct an i t, Viola Jone
is an ear i hat is origin fr ar has been de ill be normali ayscale image cted. This is f n the image in raction featur red clear from e extraction f
and PCA is appropriate G ent, data traini perpendicular
e that is captu ng, as has it is given lab s
nalysis (PCA) tic procedure [9]. This PC pace into a ne y keeping that ch possessed ed and there a ected to the ne on from a spac putation proce tem.
of the metho ification of t own), to one n known. T ors i and j that tance shows t he image i an
dean distance ormula:
ON
of ear prin Microsoft Visu e programmi
EmguCV. T ication progra input by turni
es method w image in eve rom the came etected, then t zed by the siz , and histogra for reducing t nput that caus res of ear. Af
m disturbanc features with t s begun. In t Gabor parame
ing of the ima manner again ured next will
been explain ble based on t
) is to CA ew as by are ew ce, ess ods the of The t is the d j is
nts ual ing The am ing will ery era.
the ze, am the ses fter es, the the eter age nst go ned the
re fo ap
af w im th su re av co im ke in ca th th
re th do an ta
espondents’
ollowing Figu pplication prog
Fig. 4: A Scr
The pro fter the respon with camera a mage of ear is he image of uccessfully d ecognition w vailable data t onducting rec mage of ear w ept by using th n terms of alculated, is hreshold, so th he recognition
F
In tes espondents are he left and the o repetitions nd then the va aken, as showe
names who ure 4 is a
gram that is d
reen shoot of
ocess of reco ndents take th as the data t marked by a f ear from detected, th will be cond training. The ognition by c with the data tr he Euclidean the Euclide less than th he program w from the ima
Fig. 5: Image-m
sting this app e incorporated e right ears. E
as many as alue of the suc ed in the follo
are involv screen shot developed.
application pr
ognition is c he image of ea training. The
yellow square the respon hen the pro ducted based
application p comparing the
raining which distance. If it ean distance, he value of will release the ge, like in the
matching
plication prog d to take the i Each of the re
ten times ex ccessful recog wing Table 1.
ved. The from the
rogram
conducted ar directly e detected
e. Next, if ndents is ocess of d on the program is
e detected h has been t is match, , that is
a certain e result of e Figure 5.
gram, ten images of spondents periments gnitions is
.
Table 1: The result of the recognition experiment of the left and right ears
Respondent Number
The percentage of the success left ear
The percentage of the success right ear
1 60% 70%
2 80% 90%
3 80% 70%
4 90% 70%
5 80% 60%
6 70% 70%
7 90% 90%
8 90% 80%
9 80% 80%
10 80% 70%
Average 80% 75%
Based on the test result to the re- identification of the image of the left and right ears each as many as ten times experiments, it is obtained the success level of right ear recognition is seventy five percent and the left ear is eighty percent. So, in average the success level of ears recognition is seventy eight percent. This result shows the recognitions of ear prints still cannot be the principal feature for the authentication application, which needs a high accuracy and easy in the detection. But, the ear print is quite adequate, if it is used as an additional feature for the forensic and reconnaissance requirements because it can provide clear hints, besides the ear print is relatively permanent and the image of the ear is also easily obtained.
CONCLUSION
The application program of the ear identification, by using the complex Gabor filters method for the extraction feature, has a high level of success in identifying the ears of the respondents. This result shows the biometric of ear can be used for forensic investigation and visual surveillance, but it is less suitable to be used for the authentication application, which needs a higher dependability. The main problem in the application of ear identification is in the process of the detection. As for the failure of ear detection is caused by the image of the ear that is too small caught by the camera or experiencing too much deformation of perspective. This detection problem can be solved by doing an increase in the resolution from the area of the targeted image, for example with the super resolution method.
REFERENCES
[1] A. Bertillon, La Photograhie Judiciaire, avec un appendice sur la classification et l’Identification Anthropometriques. Gauthier-Villars. 1890.
[2] G. Bromba, Bioidentification Frequently Asked
Questions. Accessed from http://www.bromba.com/faq/biofaqe.html. 2003.
[3] M. Carreira-Perpinan, (1995). Compression neural networks for feature extraction: Application to human recognition from ear images. Madrid: Technical University of Madrid. 1995.
[4] K. Chang, K. Bowyer, S. Sarkar & B. Victor, Comparison and combination of ear and face images in appearance-based biometrics. Pattern Analysis and Machine Intelligence, IEEE Transactions,Vol. 25 No.
9, pp. 1160-1165, 2003.
[5] A. Eleyan, H. Demirel, PCA and LDA Based Neural Networks for Human Face Recognition. In K. D.
Grgic (Ed.), Face Recognition, 93-106, Intech, 2007.
[6] D. Gabor, Theory of communication. Journal of the Institute of Electrical Engineers, Vol. 93, pp. 429–
457, 1946.
[7] A. Hoogstrate, H. Van den Heuvel, E. Huyben, Ear Identification Based on Surveillance Camera’s Images. Accessed from: http://www.forensic- evidence.com/site/ID/IDearCamera.html, 2000, May 31.
[8] A. V. Iannarelli, Ear Identification (Forensic Identification Series). Fremount, California:
Paramount Publishing Company, 1989.
[9] H. K. Lammi, Ear Biometrics. Lappeenranta, Finland:
Lappeenranta University of Technology, Department of Information Technology, Laboratory of Information Processing, 2003.
[10] D. Pissarenko, Eigenface-based, Facial Recognition.
Pp. 1-6, 2002.
