a comparison of iris localization techniques for pattern recognition analysis.pdf

6
A Comparison of Iris Localization Techniques for Pattern Recognition Analysis Nor’aini Abdul Jalil, Rohilah Sahak, Azilah Saparon Faculty of Electrical Engineering, Universiti Teknologi MARA, 40450 Shah Alam, Selangor, Malaysia [email protected] AbstractThis paper presents a comparison of iris localization techniques namely, Circular Hough Transform (CHT), Daugman's Integro Differential Operator (DIDO) and Circular Boundary Detector (CBD) for localization of iris region. The difference among these three techniques are, CHT employs a segmentation technique to highlight the edges of interest and the circular shape of the iris is detected using the equation of circle, DIDO makes use of integro differential operator for locating the iris and pupil regions while CBD is developed based on the equation of circle by first selecting two points at the iris region: one at the center and the other one at the circumference of the iris. Once this selection has been made, the computation of the outer iris boundary takes place. The same procedure is repeated for the construction of inner iris boundary. The circular iris region can be un-wrapped into rectangular form for the purpose of pattern recognition analysis. The iris localization was conducted on iris images taken from healthy women free from Human Papilloma Virus (HPV). The results show that CBD is able to localize the iris region for all tested iris images while using DIDO and CHT, not all iris images can be localized precisely. Keyword-Circular Hough transform; Daugman's Integro Differential Operator; Circular Boundary Detector; Human Pappiloma Virus. I. INTRODUCTION Localization of iris has many applications especially in biometric such as Iris recognition. Iris recognition is a method of biometric authentication that uses pattern recognition techniques based on high resolution images of the iris that able to distinguish from one person to another. It uses camera technology that able to capture detail intricate structures of the iris that provides mathematical representation which produce undeniable true identification of a person. Various methods of segmentation to localize the iris maybe applied in order to precisely locate the region of interest. Detection using Circular Hough transform (CHT) finds many applications for eye localization and eventually for facial recognition. CHT has been recognized as robust technique for curve detection whereby in can detect object even in noisy image [1]. It is also good at extracting geometrical components from any given object [2]. Iris Recognition System developed by C.H.Daouk et al uses a fusion mechanism that combines both canny edge detection and CHT to detect the iris boundaries. The iris patterns in the form of feature vectors is extracted using Haar wavelet and Hamming distance operator is used to determine the accuracy between test and train irises in the database. The results show that the system effectively able to recognize the irises [3]. Eyes Detection in Facial Images using CHT from facial images as in [4] also shows a promising result whereby 86% detection accuracy is achieved. Shylaja S.S. et al in [2] uses CHT as an input to the face recognition engine along with Feed Forward Neural Network for face recognition. The outcome of this technique has yielded very satisfactory results with an accuracy of about 98.68%. DIDO also finds many applications in iris recognition. A.E.Yahya et al utilizes DIDO for iris localization to determine the matching of iris to a specific person. A 100% success of pupil detection was achieved using CASIA database but failure of about 2% for the detection of pupil [5]. Daugman’s algorithm is also applied in [6] by Mohamed A. H. using UPOL database. The result shows that success rate is low due to incompatibility with visible light illuminated images. But a high percentage of accuracy was achieved when Daugman’s optimization algorithm is applied. Mahboubeh S. et al enhanced the use of Daugman’s method to locate the iris using average shrinking approach. Result shows that the proposed method is able to localize the iris faster than other similar method when applied on CASIA and MMU databases [7]. Another enhanced method is also carried out by Milad S. et al in [8] and result shows that the localization accuracy and speed have improved to 97% and 3.3 sec respectively compared to using original and optimized Daugman’s algorithms. The special characteristics of the circular iris and pupil allow the use of CHT to segment the pupil and the iris effectively. But due to illumination problem it may create circular patches that cause the CHT to detect false region. In addition to this, the filter effect to smoothen the texture pattern of the iris may cause the outer boundary to be detected incorrectly. Heavy complexity computation may also contribute to CHT becoming less popular and many researchers are looking for precise and simpler technique to localize the iris region. A similar problem may also occur when using DIDO. In this paper a precise and simpler technique to localize the iris region using circular boundary detector is proposed. This technique is able to precisely locate the iris region for the purpose of pattern recognition analysis and allows the user to redo the localization as desired until satisfied. II. CIRCULAR HOUGH TRANSFORM The CHT procedures consist of, generate edge image using Canny edge detection, detect the radius from the values in accumulator array, and finally detect circles of iris 2012 Sixth Asia Modelling Symposium 978-0-7695-4730-5/12 $26.00 © 2012 IEEE DOI 10.1109/AMS.2012.44 75

Upload: amr-yassin

Post on 08-Nov-2015

27 views

Category:

Documents


3 download

TRANSCRIPT

  • A Comparison of Iris Localization Techniques for Pattern Recognition Analysis

    Noraini Abdul Jalil, Rohilah Sahak, Azilah Saparon Faculty of Electrical Engineering,

    Universiti Teknologi MARA, 40450 Shah Alam, Selangor, Malaysia

    [email protected]

    AbstractThis paper presents a comparison of iris localization techniques namely, Circular Hough Transform (CHT), Daugman's Integro Differential Operator (DIDO) and Circular Boundary Detector (CBD) for localization of iris region. The difference among these three techniques are, CHT employs a segmentation technique to highlight the edges of interest and the circular shape of the iris is detected using the equation of circle, DIDO makes use of integro differential operator for locating the iris and pupil regions while CBD is developed based on the equation of circle by first selecting two points at the iris region: one at the center and the other one at the circumference of the iris. Once this selection has been made, the computation of the outer iris boundary takes place. The same procedure is repeated for the construction of inner iris boundary. The circular iris region can be un-wrapped into rectangular form for the purpose of pattern recognition analysis. The iris localization was conducted on iris images taken from healthy women free from Human Papilloma Virus (HPV). The results show that CBD is able to localize the iris region for all tested iris images while using DIDO and CHT, not all iris images can be localized precisely.

    Keyword-Circular Hough transform; Daugman's Integro Differential Operator; Circular Boundary Detector; Human Pappiloma Virus.

    I. INTRODUCTION Localization of iris has many applications especially in

    biometric such as Iris recognition. Iris recognition is a method of biometric authentication that uses pattern recognition techniques based on high resolution images of the iris that able to distinguish from one person to another. It uses camera technology that able to capture detail intricate structures of the iris that provides mathematical representation which produce undeniable true identification of a person. Various methods of segmentation to localize the iris maybe applied in order to precisely locate the region of interest.

    Detection using Circular Hough transform (CHT) finds many applications for eye localization and eventually for facial recognition. CHT has been recognized as robust technique for curve detection whereby in can detect object even in noisy image [1]. It is also good at extracting geometrical components from any given object [2]. Iris Recognition System developed by C.H.Daouk et al uses a fusion mechanism that combines both canny edge detection and CHT to detect the iris boundaries. The iris patterns in the form of feature vectors is extracted using Haar wavelet and Hamming distance operator is used to determine the

    accuracy between test and train irises in the database. The results show that the system effectively able to recognize the irises [3]. Eyes Detection in Facial Images using CHT from facial images as in [4] also shows a promising result whereby 86% detection accuracy is achieved. Shylaja S.S. et al in [2] uses CHT as an input to the face recognition engine along with Feed Forward Neural Network for face recognition. The outcome of this technique has yielded very satisfactory results with an accuracy of about 98.68%.

    DIDO also finds many applications in iris recognition. A.E.Yahya et al utilizes DIDO for iris localization to determine the matching of iris to a specific person. A 100% success of pupil detection was achieved using CASIA database but failure of about 2% for the detection of pupil [5]. Daugmans algorithm is also applied in [6] by Mohamed A. H. using UPOL database. The result shows that success rate is low due to incompatibility with visible light illuminated images. But a high percentage of accuracy was achieved when Daugmans optimization algorithm is applied. Mahboubeh S. et al enhanced the use of Daugmans method to locate the iris using average shrinking approach. Result shows that the proposed method is able to localize the iris faster than other similar method when applied on CASIA and MMU databases [7]. Another enhanced method is also carried out by Milad S. et al in [8] and result shows that the localization accuracy and speed have improved to 97% and 3.3 sec respectively compared to using original and optimized Daugmans algorithms.

    The special characteristics of the circular iris and pupil allow the use of CHT to segment the pupil and the iris effectively. But due to illumination problem it may create circular patches that cause the CHT to detect false region. In addition to this, the filter effect to smoothen the texture pattern of the iris may cause the outer boundary to be detected incorrectly. Heavy complexity computation may also contribute to CHT becoming less popular and many researchers are looking for precise and simpler technique to localize the iris region. A similar problem may also occur when using DIDO. In this paper a precise and simpler technique to localize the iris region using circular boundary detector is proposed. This technique is able to precisely locate the iris region for the purpose of pattern recognition analysis and allows the user to redo the localization as desired until satisfied.

    II. CIRCULAR HOUGH TRANSFORM The CHT procedures consist of, generate edge image

    using Canny edge detection, detect the radius from the values in accumulator array, and finally detect circles of iris

    2012 Sixth Asia Modelling Symposium

    978-0-7695-4730-5/12 $26.00 2012 IEEEDOI 10.1109/AMS.2012.44

    75

    Ali2Highlight

    Ali2Highlight

    Ali2Highlight

    Ali2Highlight

    Ali2Highlight

    Ali2Highlight

    Ali2Highlight

    Ali2Highlight

    Ali2Highlight

    Ali2Highlight

    Ali2Highlight

    Ali2Highlight

    Ali2Highlight

    Ali2Highlight

    Ali2Highlight

  • and pupil. Figure 1 shows the block diagram of the procedures using CHT for Iris localization.

    A. Canny edge detection

    Canny edge detection is used to detect edge pixels. Although it is quite old it has become one of the standard edge detection method and still being used in many researches that call for edge detection [9, 10]. The steps involve in the edge detection are as follows;

    1) Smooth image Before the image is edge detect, Gaussian filter is used to

    smoothen the image that will reduce the noise that cause false edge detection. The smoothed image produced depends on the standard deviation of the Gaussian filter. This smoothing filter will optimize the tradeoff between noise filtering and edge localization.

    2) Gradient magnitude The Gradient magnitude is computed using approximations of partial derivatives as follows [12],

    a) Gradient, a[m,n], of an image;

    yyxxy iahiahiyaix

    xaa

    (1)

    where, ix and iy are unit vectors in the horizontal and vertical direction, respectively. b) Gradient magnitude;

    2y2x ahaha (2) approximated by,

    ahaha yx ~ (3)

    3) Non-maxima suppression The thin edges are formed by applying nonmaxima

    suppression to the Gradient magnitude. The purpose of this step is to convert the blurred edges in the image of the gradient magnitudes to sharp edges. Basically this is done by preserving all local maxima in the gradient image and

    deleting everything else. The steps taken for non-maxima suppression are;

    a) Round the gradient direction to nearest 45, corresponding to the use of 8-connected neighborhood.

    b) Compare the edge strength of the current pixel with the edge strength of the pixel in the positive and negative gradient direction that is if the gradient direction is north (=90), compare the pixel to the north and south.

    c) If the strength of the current pixel is largest, preserves the value of the edge strength else suppress or remove the value [10].

    4) Double threshold Some of the detected edges after non-maxima suppression

    may not be true edges. They may be the result of noise or color variations for instance due to rough surface. Due to this, canny edge detection uses double thresholding. Edge pixel stronger than the high threshold are marked as strong while edge pixels weaker than the low threshold are suppressed and edge pixels between the two threshold are marked as weak [10].

    B. Radii detection of outer and inner circles The circular detection of the iris involves finding the

    radius of outer circle which is the outer iris boundary and the radius of the inner circle which is the inner iris boundary. After edge detection of the image each edge point is taken as a centre of a circle of radius R drawn onto an accumulator array. Many constructed circles intersect leading to a large intensity peak in the accumulator array at, or near, the centre of the circle. This statement is illustrated in Figure 3.

    Figure 3. The CHT showing the edge points and accumulator The CHT as described above may be carried out for a

    range of different radii. This leads to an accumulator array, or parameter space, consisting of three dimensional; the position of the centre and the radius, (a, R), where a = (xo,yo). The values in the array will be accumulated and increased every time a circle is drawn with the desired radii over every edge point. The accumulator keeps count of how many circles pass through the coordinates of each edge point and continue to find the highest count. The coordinates of the centre of circles are the coordinates with highest count [13].

    Radii detection of outer and inner circles (Pupil

    and Iris)

    Generate edge image using canny edge

    detection

    Detect circles of Iris and Pupil

    Circle boundar

    (ax,ay) parameter

    Edge pixel

    Figure 1. Block diagram of the procedures in CHT for iris localization

    76

    Ali2Highlight

    Ali2Highlight

    Ali2Highlight

    Ali2Highlight

    Ali2Highlight

  • C. Circular Detection of Iris using CHT Hough transform is a technique which can be used to

    isolate features of a particular shape within an image. The classical Hough transform is commonly used for the detection of regular curves such as lines, circles, ellipses, and etcetera because it requires the desired features to be specified in some parametric form [2, 12]. The main advantage of the Hough transform technique is that it is tolerant to gaps in feature boundary descriptions and is relatively unaffected by image noise. The circular shape of the iris is detected using the equation of circle as in equation (4).

    22020 ryyxx (4)

    where, 00 y,x is the coordinate of the circle centre, r is the radius of the circle.

    The detection process starts with the local maxima in the region of interest is assumed as the centre of the circle. If the linear indices among the minimum value of qualified pixel forming the circular shape, then that area is the iris/pupil region detected on the image. Every area of interest is tested with this process for it occurs as an element of the circle component which is the iris/pupil region identified in the image [12].

    III. DAUGMAN'S INTEGRO DIFFERENTIAL OPERATOR Daugman uses integro-differential operator to locate the

    circular iris and pupil regions as well as the arcs of the upper and lower eyelids using the equation as follows;

    dsr2

    )y,x(Ir

    )r(G)y,x,r(max00 y,x,r000

    (5)

    where, x0, y0, r are the center and radius of coarse circle (for pupil and iris), G0(r) is the Gaussian filter used for smoothing function and I(x,y) is the original iris image. Symbol * denotes convolution [5, 6].

    The operator searches for the image domain I(x,y) for the maximum in the blurred derivative with respect to increasing radius r, of the normalized contour integral of I(x,y) along the circular arc ds of radius r and center (x0, y0). The operator is applied iteratively with the amount of smoothing progressively reduced in order to attain precise localization. Eyelids are localized in a similar manner, with the path of contour integration changed from circular to an arc [5,14].

    Daugmans algorithm can be seen as a variation of the Hough transform, since it too makes use of first derivatives of the image and performs a search to find geometric parameters. Since it works with raw derivative information, it does not suffer from the thresholding problems of the Hough transform. However, the algorithm can fail when there is noise in the eye image, such as from reflections, since it works only on a local scale [14].

    IV. PROPOSED CIRCULAR BOUNDARY DETECTOR The proposed method is to find boundaries of outer and

    inner iris region in order to segment iris for further analysis on the eye image. The idea is to draw circular template on the eye image so that the boundaries of iris and pupil can be detected.

    To detect the iris boundary, initially the user must select two points on the iris region using a mouse click. The first point is the centre of the iris which is located at [x1, y1] and the second point is the radius of the iris which is located at [x2, y2]. Here, the pixel coordinate and radius of the circle which are defined as the iris boundary are stored. The same steps are then performed to detect pupil boundary. The circular shape is drawn using the equation of circle as in equation (6).

    r2 =(x2-x1)2 + (y2-y1)2 (6) The parametric representation of the circle is, x2=x1+ rcos() y2=y1+ rsin() (7) where, r is the radius of the circle, x1 and y1 are the x-coordinate and y-coordinate of first point or centre of circle while x2 and y2 are the x-coordinate and y-coordinate of second point at the circumference[12]. In MATLAB code the circular boundary computation uses equation (8). pp=rsmak('circle',r,[x1,y1]) (8)

    where, rsmak is the function to compute the circular boundary .

    Figure 4 shows the example of circle template obtained from equations (6) and (7).

    Radius, r

    Second point; located at [x2 y2]

    First point; located at [x1 y1]

    Figure 4. Template of circle

    Even though the coordinates of the circle centre and the

    coordinates at the circumference are selected manually but the construction of a circle is based on the computation of

    77

  • the rest of x2 and y2 coordinates as angle varies from 0 to 360. The circle is then formed automatically.

    A. Proposed Iris localization process The block diagram of the proposed iris localisation

    process is shown in Figure 5 There are three main stages involved in the localization process namely preprocessing and detection of iris and pupil boundaries. 1) Pre-processing

    In pre-processing, raw images of the eyes were cropped and resized from 3888x2592 pixels to 601x501 pixels. This is to avoid laborious computation.

    Figure 5. Block diagram of Iris localization process

    2) Detection of outer iris boundary

    In detecting the outer iris boundary, the user needs to select the first point at the centre of iris region by clicking the mouse as shown in Figure 7(a). Then the second point which is located at the circumference of the iris is selected as shown in Figure 7(b).

    (a) (b)

    Figure 7. Detection of outer iris boundary (a) Setting coordinates for the first point for outer iris boundary

    (b) Setting coordinates for the second point as well as calculating the radius of the circle for outer iris boundary

    3) Detection of inner iris boundary (pupil boundary)

    The same steps as in Section 2 are used in detecting boundary of inner iris (pupil boundary) after detecting the outer iris boundary.

    (a) (b)

    Figure 8. Detection of inner iris boundary (a) Setting coordinates for the first point for inner iris boundary

    (b) Setting coordinates for the second point for inner iris boundary

    IV. RESULTS AND DISCUSSION The database of iris images for this study is obtained by

    the research team using iridology digital camera. Fourteen eye images of healthy married women free from HP Virus that is responsible for about 70% cervical cancer are used in this study. CHT, DIDO and proposed methods of localization are applied to localize the iris region. Figure 9(a) shows circle drawn on the iris image using proposed method and this successfully detects the outer iris boundary. The complete detection of the outer and inner iris boundaries as shown in Figure 9(b) takes about 8.82 seconds computation time.

    (a) (b)

    Figure 9. Detection of outer and inner iris boundary (a) Circle on the outer iris boundary

    (b) Circle on the outer iris boundary and inner iris boundary

    Figure 10 shows the incorrect iris localization using CHT. The problem using CHT is that it may not consistently able to localize the iris region. Since the localization of iris requires precise detection, CHT may not be suitable for GUI application. Due to smoothing effect as a result of filtering the boundary may sometime not able to be detected. Also, using CHT may cause heavy and complex computation. A similar problem may occur when using DIDO for iris localization.

    EYE IMAGE

    CROP and RESIZE

    Preprocessing

    Detection of iris and pupil boundaries

    78

  • Figure 10. Incorrect iris localization using CHT

    The results from the test conducted using the three

    techniques are also presented in terms of the region of interest (ROI) detected. This was carried out by un-wrapping the iris using Daugman's rubber sheet technique [15] and the ROI was identified using Iridology chart [16]. Table 1 tabulates the results obtained using the three methods in terms of threshold values by applying Otsu threshold technique on the ROI [17] and Figure 11 shows the iris localization using the three techniques.

    Iris

    imageCircularHough

    transform

    Daugmans Integro-

    differential Operator

    Proposed

    1

    2

    3

    4

    5

    6

    7

    8

    9

    10

    11

    12

    13

    14

    Figure 11. Iris localization for right eyes using CHT, DIDO and proposed method

    TABLE I

    THRESHOLD VALUES FOR RIGHT EYES USING CHT, DIDO AND PROPOSED METHOD

    Iris image

    Houghtransform

    Daugmans Integro-differential Operator Proposed

    1 0.624 0.598 0.6712 0.773 0.694 0.6923 0.616 0.553 0.4474 0.580 0.553 0.4945 0.702 0.725 0.6866 0.682 0.706 0.6907 0.761 0.761 0.7498 0.655 0.706 0.7149 0.875 0.561 0.52710 0.733 0.749 0.78411 0.824 0.773 0.81612 0.804 0.459 0.52513 0.475 0.714 0.69414 0.808 0.765 0.863

    From Figure 11 using CHT, images 9, 12 and 13 show miss localization since the inner and outer boundaries are not on the right target. The threshold values compared to using CDB show significant difference. In the case of using DIDO, image 1, 6 and 14 show miss localization since the inner boundaries are not well targeted. The threshold values also show significant difference. CBD, the proposed

    79

  • localization technique has shown consistent localization compared to the other two techniques since this technique can be repeated as desired to acquire the right localization. Furthermore CBD technique does not require transformation to gray scale which means localization is done directly on the image.

    V. CONCLUSION The iris localization developed using CBD has shown

    consistent localization of the iris region. Applying CHT and DIDO onto the eye images may not consistently produce good result due to the effect of illumination variations and poor filtering of the eye image that may cause incorrect detection. A suitable filter is required to enhance the feature of the eye image thus able to rectify the above problems as well as increase the quality of the processed image. Besides that, the combination of other techniques can be considered to eliminate the unwanted details on the eye image. In conclusion, CBD is a relevant technique to be considered in the iris localization due to its ability to precisely localize the iris region since this technique can be repeated as desired and suitable to be used in GUI system.

    ACKNOWLEDGMENT The authors would like to extend their sincere gratitude

    to MOHE and RMI of University Teknologi MARA for awarding the Fundamental Research Grant of Science (FRGS). Without this fund the research work is impossible to be conducted.

    REFERENCES [1] M. Rizon et al, Object Detection Using Circular Hough

    Transform,American Journal of Applied Sciences 2(12), 2005, pp 1606-1609.

    [2] Shylalaja S.S. et al, Feed Forward Neural Network Based Eye localizaation and Recognition using Hough Transform, International Journal of Advanced Computer Science and Applications, Vol. 2, No.3, March 2011.

    [3] C. H. Daouk et al, Iris Recognition, IEEE ISSPIT 2002, Marakesh, pp 558-562.

    [4] W. M. K. Wan Khairosfaizal and A. J. Noraini, Eyes Detection in Facial Images using Circular Hough Transform, Colloquium of Signal Processing and its Application (CSPA 2009) , Kuala Lumpur 2009.

    [5] A. E. Yahya and M.J. Nordin, A New Technique for Localization in Iris Recognition Systems, Information Technology Journal 7, Vol. 6, 2008, pp 924-929, ISSN 1812-5638.

    [6] Mohamed A.H., Optimized Daugmans Algorithm for Iris Localization, wscg.zcu.cz.ascq2008/papers-2008/poster/All-full.pdf.

    [7] Mahboubeh S., Puteh S., Subariah I and Abdolreza R. K., Fast Algorithm for Iris Localization using Daugman Circular Integro Differential Operator, 2009 International Conference of Soft Computing and Pattern Recognition,2009, pp 393-398.

    [8] Milad S., Saeid T. Z. and Hamid R. P., Daugmans Algorithm

    Enhancement for Iris Localization, Advanced Materials Research, Vols. 403-408, 2012, pp 3959-3964.

    [9] Marcin S.and Ignacy D., Circular Object Detection Using a Modified Hough Transform, International Journal of Applied Mathematics and Computer Science, 2008, Vol. 18, No. 1, pp85-91.

    [10] Canny Edge Detection, http://www.cvmt.dk/education/teaching/f09/ VGIS8/AIP/canny-09gr820.pdf, retrieved on 26 March 2012.

    [11] Naveem S., Dilip Gandhi and Krishna P. S., Iris Recognition System using Canny Edge Detection and Circular Hough Transform, International Journal of Advances in Engineering & Technology, Vol. 1 Issue 2, May 2011, pp 221-228.

    [12] Simon J. K. P., Circular Hough Transform, Aalborg University, Vision, Graphics and Interactive Systems, November 2007.

    [13] T.J. Atherton and D.J. Kerbyson, The Coherent Circle Hough Transform, Proceedings of the British Machine vision conference, Guildford, UK, 1993, pp269-278.

    [14] Libor M., Recognition of Human Iris Patterns for Biometric Identification, Thesis, Bachelor of Engineering Degree of the School of Computer Science and Software Engineering, University of Western Australia, 2003.

    [15] Daugman, J.G., High confidence visual recognition of persons by a test of statistical independence, IEEE Trans. Pattern Anal. Machine Intell., Vol. 15, 1993, pp 1148-1161.

    [16] Bernard J. and Donald V. B., Vision of Health, AVERY a member of Penguin Putnam Inc., ISBN 0-89529-433-8, 1992.

    [17] Otsu, N., "A Threshold Selection Method from Gray-Level Histograms," IEEE Transactions on Systems, Man, and Cybernetics, Vol. 9, No. 1, 1979, pp. 62-66.

    80