Description

FAST PUPIL LOCATION FOR BETTER IRIS DETECTION
IMENE KHANFIR KALLEL, DORRA SELLAMI MASMOUDI, NABIL DERBEL.
RESEARCH UNIT CIELS, NATIONAL ENGINEERING SCHOOL OF SFAX, TUNISIA
BP W, 3038, SFAX, TUNISIA, +216 74 274 088
Imen.khanfir{at}betatex.com.tn
ABSTRACT
The inner edge of the iris corresponds to the pupil one.
Thus, it is enough to locate as precisely as possible the
latter to delimit the iris inner side. The distinguished
gray levels within the pupil ca

All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.

Related Documents

Share

Transcript

FAST PUPIL LOCATION FOR BETTER IRIS DETECTION
IMENE KHANFIR KALLEL, DORRA SELLAMI MASMOUDI, NABIL DERBEL.
RESEARCH UNIT CIELS, NATIONAL ENGINEERING SCHOOL OF SFAX, TUNISIABP W, 3038, SFAX, TUNISIA, +216 74 274 088
Imen.khanfir{at}betatex.com.tn
ABSTRACT
The inner edge of the iris corresponds to the pupil one.Thus, it is enough to locate as precisely as possible thelatter to delimit the iris inner side. The distinguished gray levels within the pupil can be very useful in itslocalization task. Indeed, this area often appears as thedarkest one in the image. Therefore, by studying the graylevel distribution in the image, one can locate the pupil as the whole of pixels that have the least gray levels.This can be ensured by histogram thresholding. However by using such a technique, the localized pupil isthreatened to be perforated, in case of presence of reflection points, or deformed by the addition of noisyelements such as lashes and dark textons. In this paper,a morphological cleaning technique is used to clear out the pupil binary image. This pupil localization strategymakes it possible to accurately delimit the iris by theinterior, as well as it estimates pupil center that approximately provides us the iris one. Using this strategy before applying the integro-differential operator, the Hough transform algorithm or the multi scale edge detector approach, a near 18 times faster localization of iris is achieved.
Key words:
Iris, Localization, Authentication, HistogramThresholding, Mathematical Morphologies.
I. INTRODUCTION
The iris localization in an image is a task easily carriedout by the human eye, but it becomes very complicatedonce one tries to automate it. Indeed, it is not so evidentto success the localization step as efficiently as one canhope, on any captured iris image. This returns to the factthat there often are some bothersome objects likereflection points, lashes and eyelids which come tooverlap the iris area.The traditional iris identification systems, such as theintegro-differential operator elaborated by Daugman [1,2, 4, 10, 15, 21] as well as the Hough transform basedoperator elaborated by Wildes [5, 8, 9, 11, 14], analyzeimage edges in order to find contours corresponding tothe iris-pupil and iris-sclera. Their analyses are based onthe circular approximated aspect of iris borders. Manyother researchers were based on these two techniques inorder to locate iris and to extract it from its background[1-15, 20, 21]. Following a survey of these techniques,we can deduce that in the whole of the cases, the irisdetection refers to accurately provide respective radiusof the two circles delimiting the focused area, as well astheir respective centers which are often confused.Several of these works led to good results; however suchapproaches can fail when the iris is encroached byeyelids and lashes often having circular arc aspects and arandom space distribution. Moreover, the reflection points particularly found inside pupil, often presentobstacles. A fast and efficient determination of the pupilcenter and radius could thus bring back more interest tothese approaches. It is the motivation of this work.We propose in the following a pupil localization strategy based on pixel classification based segmentationtechnique followed by a cleaning operation used toremove the undesirable elements such as lashes and brilliant points, realized owing to morphologicaloperators.This paper is articulated as follows: in the secondsection we briefly present some iris localizationtechniques. We explain in the third section the adoptedtechnique to efficiently find out the pupil center andradius. The fourth section is devoted to result presentation. We than encloses by showing our conclusions and prospects.
II. Background
Numerous methods, which are approved for irislocalization, are based on the fact that the pupil and theiris edges can be estimated by circles. Thus the problemis reduced to determine centers and radius of the twocircular shapes.
1. INTEGRODIFFERENTIAL OPERATOR
The first results of iris extraction were proposed byDaugman [1, 2, 4, 10, 15, 21]. His approach, called «coarse-to-fine strategy », aims at extracting contour details from coarsest to finest ones, in order to provide ata pre-pixel precision center coordinates as well as circleradius of the approaching internal and external iriscontours. In this context, Daugman builds anintegrodifferential called operator, acting as a circular contour detector that he applies at many space analysisscales. A large scale allows extracting edgescorresponding to very pronounced transitions, while asmall scale makes it possible to detect less perceptibleedges. The integrodifferential operator search over theimage domain (x, y) for the maximum in the blurred partial derivative, with respect to increasing radius, of the normalized contour integral of I(x, y) along a
978-1-4244-4346-8/09/$25.00 ©2009 IEEE
2009 6th International Multi-Conference on Systems, Signals and Devices
circular arc
ds
of radius
r
and center coordinates (
x
0
,
y
0
) :
0000
00,,(,,)
(,)(,,)max()*2
rxyrxy
IxyOPrxyGrdsrr
(1)
The contour integral parameterized by size and positioncoordinates (
r
,
x
0
,
y
0
), is performed on various specialanalysis scales
imposed by a Gaussianfunction
()
Gr
:
202
2
1()2
rr
Gre
(2)
Parameters of required contours (
r
c
,
x
c
,
y
c
), are those thatmaximize the integral given by (1).The contour detector iteratively seeks the maximumvalue of the contour integral derivative while graduallyincreasing the radius, on more and finer space analysisscales.The methodology suggested by Daugman in [1], isspread on two stages:
The first stage aims at searching iris sclerotic edgeon a zone delimited by two cones of 90° each one,opposed to the tops, practically covering the visibleright and left side zones of the iris.
The second stage consists in looking for the pupilcontour. For that, Daugman uses the same operator on a finer scale, and thus a weaker value of
. Thevalues of (
r
,
x
0
,
y
0
) vary in the previously localizediris zone.
2. CONTOUR DETECTION FOLLOWED BY HOUGHTRANSFORM
This technique firstly suggested by Wildes et al. [8, 14],acts in two stages:
1
st
stage:Initially, gray level information contained in the image isconverted into a binary edge map. This map is builtthanks to a gradient based directional edge detector [16,17]. This operation consists in thresholding intensitygradient amplitude in the image, given by the followingequation:
,,*,
gradIxyGxyIxy
(3)
where
,
y
(4)
and
22002
22
1(,)2
xxyy
Gxye
(5)
G(x, y) is a bidimensional Gaussian of center (x
0
, y
0
) andstandard deviation
that smoothes the image to selectspatial scale of focused edges.
2
nd
stage:In this stage, the edge points of this map, contribute tothe research of the focused edge parameters. This stageis called: vote phase.The vote process is ensured thanks to Hough transform[22, 23] defined by a parameterized expression leadingto specific edges. Particularly, for the sclerotic and papillary edges, it looks for circular ones.Thus, for a point whole picked out from the edge map (
x
j
, y
j
)
, j=1, ..., n
, the Hough transform is defined asfollows:
1
,,,,,,
nccjjcc j
Hxyrhxyxyr
(6)
where
1,,,,,0,,,,0,
jjcc jjcc
ifgxyxyr hxyxyr else
(7)
with
222
,,,,
jjccjcjc
gxyxyrxxyyr
(8)
For each edge point (
x
j
, y
j
), the expression
g
(
x
j
, y
j
, xc , yc ,
r) of the equation (8) is equal to 0, for each triplet of parameters (
xc , yc ,
r), that represents a circle goingthrough this point. So the triplet of parameters thatmaximizes H necessarily corresponds to a circlegathering the greatest number of edge points. Such atriplet presents a reasonable choice to representrequested contour. Wildes
et al.
begin by finding the iriszone inside which he looks for the pupil one.The edge detection followed by Hough transform is astandard technique of computer vision aiming atassigning simple edge models to the images [16, 17]. Itis accurate, but very greedy in term of execution time.
3. MULTI SCALE EDGE DETECTION
The multi scale edge detection can be ensured by awavelet transform as defined in the reference [14].Indeed, the wavelet transforms proved to be the mostadequate mathematical tools for singularity analysis, andespecially for edge ones, in order to efficiently detectthem. Mallat
et al.
[24, 28] showed that the localmaxima of the wavelet transform modulus, allowdetecting position of the irregular structures. Nabti
et al.
[25] profited from this survey results in order to detectiris edges. Their strategy consists in decomposing theimage according to several analysis scales in order tocontrol the type of edges to be detected. The mostsignificant edges resist more to the analysis scaleincreasing effect. Fewer edges are significant morequickly they disappear by increasing the scale.The multi scale edge detection method primarily consistsin performing the image gradient on various scales. Allfilters used on
j (j >
0) scale are up sampled by a 2
j
factor compared to those of zero scale.Being
f
(
x
y) a given image of size M×N. For each scale2
j
, with
j >
0, the wavelet transform decomposes (
S
(
j
-1)
f
)in three wavelet bands:-A low pass band (S
j
f
), giving the image approximation
f
at the 2
j
scale, and-Two high pass bands, giving the
f
image details at the2
j
scale:
WH
j
f
, corresponds to the horizontal detailsand
WV
j
f
corresponds to the vertical details ;
S
0
f
being equal to
f(x, y)
.The three wavelet bands (
S
j
f
,
WH
j
f
,
WV
j
f
) at the 2
j
scaleare
M
×
N
sized like the srcinal image.At each wavelet transform level, the
M
j
f
gradientmodulus can be performed as:
22
HV jjj
MfWfWf
(9)
At a scale 2
j
, the points corresponding to importantvariations in the image
f
, smoothed to
S
j
f
, are points (
x,
y) that coincide with local maxima of the
M
j
f
modulus.The whole of these points builds an edge map. Thus wedispose for each scale 2
j
, a vertical as well as horizontaledge maps. A Hough transform is then applied to themost adapted edge maps, in order to accurately searcheach circle corresponding respectively to the iris and the pupil.
III. FAST PUPIL RADIUS AND CENTER DETECTION
1. PUPIL LOCALIZATION
While captured in an image, the pupil area appears as thedarkest area of the human eye. Its localization is ensuredowing to the peak that it presents in the gray levelhistogram of the entire image (Cf. figure 1-b). Followingthis observation we get as relevant issue a segmentationmethod by pixel classification, that is a region basedapproach using histogram thresholding technique. In[18], we explain how we use the “Intermeans” algorithmto automatically select optimal threshold that we use toseparate the pixel whole forming pupil area from the restof the image (Cf. figure 1).
a- Original image b- Gray level histogramc- Pupil localization d- Iris inner border
Figure 1 :
Pupil localization using histogramthresholding
If there is any bothersome element risking bad pupildetection, we can have agreeable results as it is shown infigure 1-d. However, in certain cases, images presentsome reflection points on the pupil resulting from lightsources used during acquisition (Cf. figure 5-a). We arealso likely to have cases where the lashes are as dark asthe pupil, as they appear in the resulting binary imagewhereas their presence harms the ulterior treatment,namely pupil edge focusing (Cf. figure 4-a). To curethese problems, we carry out a cleaning of the resulting binary image, by a morphological operator processing based on erosions followed by dilations.
2. MORPHOLOGICAL CLEANING
Mathematical morphology is called ensemblist when it processes binary images [26]. Its basic principle is tomatch up the image, to be analyzed, to a pixel wholecharacterized by its center x and its known geometricalshape and size, called structuring element. A binaryimage usually contains some number of areas (whole of related pixels) coded with “1”, that we can define asobjects of interest, compared to the background codedwith “0”.The selected structuring element is moved so that itscenter x goes across all the positions in the binary imageto analyze. For each position x, we decide about theunion or the intersection of the structuring element withimage objects. The pixel whole corresponding to a positive answer allows building a new image.These principles, served to build the basic mathematicalmorphology operators, namely erosion and dilation [27].During dilation all objects become bigger of a partcorresponding to the structuring element size. If there areholes in the objects, they will be filled. If objects arelocated at a distance larger than the structuring elementsize, they will merge (Cf. figure 2).
Figure 2 :
Dilation progress
During erosion, the objects of size lower than that of thestructuring element will disappear; the others will betruncated of a part corresponding to the structuringelement size. If there are holes in the objects, they will be accentuated and the objects connected to each other will be separated (Cf. figure 3).
Figure 3 :
Erosion progress
Erosion followed by dilation is called “opening”. Such a procedure allows eliminating all the parts of the objectswhich cannot contain the structuring element.Dilation followed by erosion is called “closing”. Such a procedure has as property to fill all what is smaller thanthe structuring element.While following this principle, it is then possible, tolook, in the image, for a particular geometrical pixelconfigurations corresponding to that of the usedstructuring element.
In our case, in order to keep only the pupil in the binaryimage, we employ, as structuring element, a disc of radius slightly lower than the smallest pupil radius in theimage base that we have (in the CASIA V.1 iris image base, the smallest pupil radius corresponds to 28 pixels, but the largest one corresponds to 75 pixels [29].In order to eliminate lashes we apply to the pupil binaryimage an opening operation (Cf. figure 4-b-c).
a- Original image b- Histogram thresholdingresulted imagec- Opening of image (b) d- Pupil resulting edge
Figure 4 :
Example of iris image where lashes color isclose to that of the pupil. The cleaning of the binary imageis ensured by an opening.
On the other hand, to fill the reflection points we applyto the pupil binary image a closing operation (Cf. Figure5-b-c).Results of such processing are efficient in the wholecases (Cf. figure 4-d, figure 5-d).
a- Original image b- Histogram thresholdingresulted imagec- Closing of image (b) d- Pupil resulting edge
Figure 5 :
Example of iris image with reflection pointson the pupil. The cleaning of the binary image is ensured bya closing.
3. CENTER AND RADIUS SEEKING
Once the pupil is successfully localized in the image, wecan approach it with a circular formed object in order todetermine his characteristic parameters. Indeed, withsuch an approximation, its geometrical center coincideswith its mass centre [19], of which coordinates are given by:
(,)
1(,)
p pxijP
ii pxijP
(10)
(,)
1(,)
p pxijP
jj pxijP
(11)
Where
(,)
pxij
is a pixel of position
),(
ji
, and
P
is the pupil.Given that the sum of pupil pixels defines its surface
S
,its radius can be expressed by:
P
S R
(12)
IV. Results and discussion
The pupil localization referring to colorimetricresemblance and spatial neighboring of its pixels, allowsa more accurate edge deduction than that achieved froma circular approximation. Indeed, given that the pupil isnot perfectly circular; such an approximation makes uslose a part of the texture which is likely to be of asignificant discriminating value (Cf. figure 6). In fact,the texture of the iris is much richer in the zonesimmediately close to the pupil.
a- No circular pupil edge b- Circular pupil edge
Figure 6 :
Evidence of saving data by use of nocircular pupil edge.
Besides accuracy, our strategy of pupil localization andits center and radius determination, has also the goal to bring more speed and efficiency, to several irislocalization techniques.Let us take as an example the techniques presented atsection 2, and analyze the contribution added by thisstrategy. First of all, the research of papillary edge isdone according to a much less complex concept than thatof an integrodifferential operator or that of a Houghtransform. Indeed the histogram thresholding techniquesand those of mathematical morphologies are moreconvenient due to their simplicity and speed. Thus onecan always start by determining pupil center and radius

Search

Similar documents

Tags

We Need Your Support

Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks