PERSONAL IDENTIFICATION USING RETINA 1. INTRODUCTION

Transcription

1 JOURNAL OF MEDICAL INFORMATICS & TECHNOLOGIES Vol. 13/29, ISSN Rszard S. CHORAŚ PERSONAL IDENTIFICATION USING RETINA retina, retina biometrics, vessel pattern, feature etraction, Gabor transform, feature vector, matching This paper proposes a biometric sstem for authentication that uses the retina blood vessel pattern. The retina biometric analzes the laer of blood vessels located at the back of the ee. The blood vessels at the back of the ee have a unique pattern, from ee to ee and person to person. The retina, a laer of blood vessels located at the back of the ee, forms an identit card for the individual under investigation. In particular retinal recognition creates an ee signature from its vascular configuration and its artificial duplication is thought to be virtuall impossible. 1. INTRODUCTION A biometric sstem is a pattern recognition sstem that recognizes a person on the basis of a feature vector derived from a specific phsiological or behavioural characteristic that the person possesses. The problem of resolving the identit of a person can be categorized into two fundamentall distinct tpes of problems with different inherent compleities: (i) verification and (ii) identification. Verification (also called authentication) refers to the problem of confirming or dening a person s claimed identit (Am I who I claim to be?). Identification (Who am I?) refers to the problem of establishing a subjects identit. The personal attributes used in a biometric identification sstem can be phsiological, such as facial features, fingerprints, iris, retina scans, hand and finger geometr; or behavioural, the traits idiosncratic of the individual, such as voice print, gait, signature, and kestroking. One technolog has emerged in the biometric area: retinal recognition. This paper proposes a biometric sstem for authentication that uses the retina blood vessel pattern. The retina is a thin laer of cells at the back of the eeball of vertebrates. It is the part of the ee which converts light into nervous signals. The unique structure of the blood vessels in the retina has been used for biometric identification. Retina scanning is quite accurate and ver unique to each individual and tpicall requires the user to look into a receptacle and focus on a given point for the user's retina to be scanned. Retina scans require that the users (person) removes their glasses and positions his ee close to the unit s embedded lens, with the ee socket resting on the sight. In order for a retinal image to be acquired, the user must gaze directl into the lens and remain still, movement defeats the acquisition process requiring another attempt. Retina recognition technolog captures and analzes the patterns of blood vessels on the thin nerve on the back of the eeball that processes light entering through the pupil. Retinal patterns are highl distinctive traits. The retina biometric analzes the laer of blood vessels located at the back of the ee. The blood vessels at the back of the ee have a unique pattern, from ee to ee and person to person. The retina, a laer of blood vessels located at the back of the ee, forms an identit card for the individual under investigation. In particular retinal recognition creates an ee signature from its vascular configuration and its artificial duplication is thought to be virtuall impossible. Ever ee has its own totall unique pattern of blood vessels; even the ees of identical twins are distinct. Although each pattern normall remains stable over a person s lifetime, it can be affected b disease such as glaucoma, diabetes, high blood pressure, and autoimmune deficienc sndrome. Fig. 1. Tpical retina image analsis (biometrics) sstem Uniwerstet Technologiczno-Przrodnicz, Wdział Telekomunikacji i Elektrotechniki, ul. Kaliskiego 7, Bdgoszcz, choras@utp.edu.pl

2 In this paper we implementation such a sstem using the following steps: 1. Image retina acquisition, 2. Image preprocessing (colour transformation, edge detection, etc.), 3. Etraction of geometrical features, 4. Etraction of teture features, 5. Integration of geometrical and teture features. 2. PREPROCESSING Most digital images are stored in RGB colour space. RGB colour space is represented with red (R), green (G), and blue (B) primaries and is an additive sstem. RGB colour space is not perceptuall uniform, which implies that two colours with larger distance can be perceptuall more similar than another two colours with smaller distance, or simpl put, the colour distance in RGB space does not represent perceptual colour distance. Fundus images contain full colour information. The first step is to separate channel to a RGB channels and from the three colour channels the green (G) component of RGB colour space for blood vessel recognition is chosen (Fig.1). Additional to represent retinal characteristic we using luminance component (Y) from YC b C r (YIQ) colour space (Fig 2). YCrCb is an encoded nonlinear RGB signal for image compression work. Colour is represented b luminance, computed from nonlinear RGB, and two chrominance components. These colour spaces separate RGB into luminance and chrominance information. Y, 299, 587,114 R C r,5,419,81 G = C b,169, 331, 5 B (1) a b c d Fig. 2. Original retina image in RGB colour space (a) and Red (b), Green (c) Blue (d) channels. a) b) c) Fig. 3. Retina image in YC b C r colour space: a) Y component and components C b (b), C r (c) respectivel GEOMETRICAL FEATURE EXTRACTION In the geometrical feature etraction steps retina image with edge binar vessels is processed. The centroid of binar image is also the centre of concentric circles of the radius r i. The algorithm uses the surrounding circle of retina vessel line for partitioning it to o radial partitions (Fig. 4). For each contour-line we specif following characteristic points: contour ending points, contour bifurcations, points of contour intersections with the concentric circles.

3 Let basis of the coefficient Fig. 4. Radial partitions retina image binar vessel g denote the current point of the contour line, so that ( ) 8 N c (Table 1): g = g i, j = 1. Contour line points are classified on the where S = (1,3,5,7). N ( g ) = ( g g g g ) (2) 8 c k k k + 1 k + 2 k = S Table 1. Classification of the vessel line points g 8 Value of N Classification of the point g c Interior (or isolated) point 1 Line ending 2 Normal point 3 Bifurcation point 4 Crossing point The feature vector corresponding to vessel topolog and consecutivel the number of endings, bifurcations and intersection points with the concentric circles are stored in the feature vector. Moreover, the coordinates of all the etracted characteristic points are stored. The feature vector for each contour consists of the following parts: l, N, N corresponding to the number of intersection points, ending points and bifurcation points in each 3 numbers ( ) l E B c contour, subvector in which the coordinates of the intersection points are stored, subvector in which the coordinates of the ending points are stored, subvector in which the coordinates of the bifurcation points are stored. The first part of the final feature vector has alwas the same length, while the net parts of the vectors for contours c depend on the number of the etracted characteristic points. Such division of the feature vector allows faster classification. 4. TEXTURE FEATURE FROM GABOR WAVELET TRANSFORM Gabor wavelet based teture is robust to orientation and illumination change. It is a powerful tool to etract teture features. Gabor functions are Gaussians modulated b comple sinusoids. In two dimensions the take the form (Fig. 5): 1 1 g(, ) = ( )ep[ ( + ) + 2 π jw] (3) 2πσ σ 2 σ σ 1 ( u W ) v G( u, v) = ep{ [ + ]} (4) 2 σ u σ v 1 1 where σ u = v =. 2πσ 2πσ Gabor wavelets can be obtained b appropriate dilations and rotations of g(,) through the generating function: m g (, ) = a g(, ); a > 1 m =,1,, S 1 (5) mn here 55

4 m m = a ( cosθ + sin θ ), = a ( sinθ + cos θ ) and θ ( θ [, π ]) specifies the orientation of the Gabor wavelets. Fig. 5. Real (a) and imaginar (b) parts of Gabor wavelets and Gabor kernels with different orientations (c) The normalized retinal image (Y components) are divided into blocks (Fig. 6). The size of each block in our application is k l ( k = l = 2). Each block is filtered b (Fig. 7) k k + + Gab(,, α) = I(, ) g(, ) (6) k k iπ The orientation angles of this set of Gabor filters are αi αi =, i =,1, 2,3. 4 The magnitudes of the Gabor filters responses are represented b three moments X Y 1 µ ( α ) = G(,, α) (7) XY = 1 = 1 X Y 2 = G (8) = 1 = 1 std( α ) (,, α ) µ ( α ) 3 X Y 1 G(,, α ) µ ( α ) Skew = XY (9) = 1 = 1 std ( α ) The feature vector is constructed using µ ( α ), std( α ) and Skew as feature components. Fig. 6. Original block retinal images (Y component) 56

5 a) b) c) d) e) Fig. 7. Original block retina image (a) and real part of Gab(; ; θ i) for θ = (b), θ = 45 (c), θ = 9 (d), θ = 135 (e) Table 2. Mean, standard deviation and skewness Gabor power some block (Fig. 7.) retina image Parameters Retinal image 1 Retinal image 2 µ ( α ) std( α ) Skew µ ( α ) std( α ) Skew θ = θ = θ = θ = Parameters Retinal image3 Retinal image4 µ ( α ) std( α ) Skew µ ( α ) std( α ) Skew θ = θ = θ = θ = MATCHING AND CONCLUSION The matching algorithm finds the proimit of two retina calculating the Average Euclidean Distance of the two features vectors. A new method has been presented for retina recognition based on geometrical and Gabor features. This paper analses the details of the proposed method. Eperimental results have demonstrated that this approach is promising to improve retina recognition for person identification. 57

6 BIBLIOGRAPHY [1] GOH K.G., LEE M.L., HSU W. WANG H., ADRIS: An Automatic Diabetic Retinal Image Screening Sstem, Medical Data Mining and Knowledge Discover, Springer-Verlag, 2. [2] HSU W., PALLAWALA P.M.D.S., LEE M.L. KAH-GUAN A.E., The Role of Domain Knowledge in the Detection of Retinal Hard Eudates, IEEE Computer Vision and Pattern Recognition, Hawaii, Dec 21. [3] LI H., CHUTATAPE O., Automated feature etraction in color retinal images b a model based approach, in IEEE Trans. Biomed. Eng., 24, vol. 51, pp [4] SALEM N.M., NANDI A.K., Novel and adaptive contribution of the red channel in pre-processing of colour fundus images, in Journal of the Franklin Institute, 27, p [5] KIRBAS C., QUEK K., Vessel etraction techniques and algorithm: a surve, Proceedings of the 3rd IEEE Smposium on BioInfomratics and Bioengineering (BIBE 3), 23. [6] CHANG S, SHIM D., Sub-piel Retinal Vessel Tracking and Measurement Using Modified Cann Edge Detection Method, Journal of Imaging Science and Technolog March-April 28 issue. [7] CHANWIMALUANG T, FAN G., An efficient algorithm for etraction of anatomical structures in retinal images, in Proc. IEEE International Conference on Image Processing, pp , (Barcelona, Spain), [8] FARZIN H, ABRISHAMI-MOGHADDAM H, MOIN M.S., A novel retinal identification sstem, EURASIP Journal on Advances in Signal Processing, vol. 28, Article ID 28635, 1 pages, 28. [9] CHORAS R.S., Image Feature Etraction Techniques and Their Applications for CBIR and Biometrics Sstems - International Journal of Biolog and Biomedical Engineering, Issue 1, vol.1, pp. 6-16, 27 [1] CHORAS R.S., Iris Recognition - M. Kurznski, M.Wozniak (Eds.): Computer Recognition Sstems 3, AISC 57, pp , Springer-Verlag Berlin Heidelberg 29 [11] CHORAS R.S., Iris-based person identification using Gabor wavelets and moments Proceedings 29 International Conference on Biometrics and Kansei Engineering ICBAKE 29, pp , CPS IEEE Computer Societ, Los Alamitos, California, Washington, Toko 58