HYBRID INTERPOLATION SUPER RESOLUTION BASED ENHANCEMENT OF IRIS IMAGES Hassan Aftab, Asif Butt, Umer Munir, Sheryar Malik, Omer Saeed National University of Sciences and Technology, Pakistan hassanaftab735@yahoo.co.uk ABSTRACT Iris recognition is a method of biometric authentication that uses pattern recognition techniques based on high-resolution images of the iris. In this paper we propose a method to obtain high resolution iris image from low resolution image for facilitating the recognition process using hybrid interpolation super-resolution technique that switches between New Covariance based interpolation and Curvature based interpolation to produce sharp and refined iris images. The results show the visually improved quality of Iris for recognition. Keywords: Hybrid Interpolation, Super Resolution, Iris, Edges. 1 INTRODUCTION Biometrics is the automated use of physiological or behavioural characteristics to determine or verify identity. A distinction may be drawn between an individual and an identity; the individual is singular, but he may have more than one identity, for example ten registered fingerprints are viewed as ten different identities. Iris are composed before birth and, except in the event of an injury to the eyeball, remain unchanged throughout an individual s lifetime. Irisscan technology has been established as one of the biometrics that is very resistant to false matching and fraud. The false acceptance rate for iris recognition systems is 1 in 1.2 million, statistically better than the average fingerprint recognition system. The real benefit is in the false-rejection rate, a measure of authenticated users who are rejected [1]. For localizing the Iris and performing segmentation to extract Iris pattern a high resolution iris image is essential. Interpolation which is sometimes called resampling is an imaging method to increase the number of pixels in a digital image. Image interpolation addresses the problem of generating a high-resolution image from its low-resolution version. There are many methods for image interpolation and commonly used ones in medical field include nearest neighbour, bilinear, bicubic and cubic-spline [2], [3]. The related work on enhancement of iris images is done by [4] which predicts the prior relation between iris feature information of different bands and incorporates this prior into the process of iris image enhancement. The hybrid image interpolation super resolution algorithm developed in Matlab was employed on the iris image taken from Chinese academy of sciences (CASIA). It yielded superior quality image for processing, thereby facilitating the use of iris recognition process. The proposed method is a fast hybrid approach for enhancement of iris images. The hybrid approach is an improvement of method proposed by [5], [6]. In previous work, the similar algorithm was employed on aerial images [10]. The present work shows the effectiveness of the algorithm to obtain high resolution iris images for the iris recognition system. Moreover cropping fused with hybrid image interpolation algorithm developed in Matlab gives the flexibility of extracting iris from the face of an individual for identification or surveillance. The results proved the potential of the image for being utilized in iris recognition systems for enhancing e-security and achieving maximum level of infallible security measure. This paper is organized as follows: Section II presents the proposed algorithm. Section III presents the experimental results of our proposed technique. The paper is concluded in Section IV. 2 PROPOSED ALGORITHM Edge detection is an important parameter in iris recognition system. The proposed algorithm differentiates between edge points and smooth areas using the four neighbors of new interpolated points. The difference of maximum and minimum values of all four points when compared with a pre-defined threshold determines the presence of an edge or smooth area. If the difference exceeds the threshold it is considered an edge otherwise the point belongs to smooth area. Edges are determined using the new covariance based method, whereas the smooth areas are handled by the iterative curvature based approach. The detail of the algorithm is explained in the ensuing paragraphs.
2.1 Interpolation Scheme The interpolation scheme of proposed approach is shown in Fig. 1. In this technique a low resolution image is taken and padded with zeros. Next bilinear interpolation is performed at the four corners of padded image. Now the four diagonal neighbors of few points are available, therefore new data points are calculated using these diagonal neighbors. In the end the remaining data points are determined using four horizontal and vertical neighbors to produce a Super Resolved image. 2.2 New Covariance Interpolation New covariance based interpolation has been used for interpolation in edges. Covariance is employed by making a circular mask around every unknown high resolution pixel and estimating the high resolution covariance coefficients from known low resolution covariance. The process is computationally complex as it keeps on looking for edges in that circular mask around the high resolution pixel until optimal MMSE (minimum mean square error) is not achieved. The new covariance based interpolation employees the geometric duality property between the lowresolution covariance of four neighbour pixels and the high resolution pixel. Geometric duality actually couples pair of pixels at different resolution but in same orientation. Since four near pixels are used we have the fourth order linear interpolation equation [6] given as: neighbouring low resolution pixels value, [a1, a2, a3, a4, b1, b2, b3, b4, d1, d2, d3, d4, e1, e2, e3, e4] are the four neighbouring pixels of low resolution pixels [C1, C2, C3, C4] and [w, x, y, z] are the high resolution coefficients. Using the concept of four equations and four unknowns the high resolution coefficients are estimated [6]. (7) where c5 is the high resolution pixel value. Fig. 2 shows the complete process of new covariance based interpolation [6]. 2.3 Curvature Interpolation For interpolation in smooth areas curvature based method [7] has been employed. Curvature based interpolation performs bilinear interpolation along the direction where second derivative is lower. It is followed by iterative refinement of the interpolated point following the isophote curve. Isophote curve is an intensity level curve which compares present value with previous and next value and change the intensity of the present value. In this way the linear curve is changed into isophote curve [8]. In case of diagonals, it finds the difference of intensity levels between diagonals at opposite direction and performs bilinear interpolation, where the difference is less. The differences V1 and V2 in two different directions are presented in [7], [8]. where Y is the required interpolated point in high resolution image, α is the high resolution covariance coefficients and Y' is neighbouring pixels to interpolated point at low resolution image. The neighbouring pixels of interpolated point are known, so the only thing required is the high resolution covariance coefficients [6]. where P, Q, P' and Q' are intensity values of the four neighbor points of interpolated pixel. The direction in which second derivative is lower, bilinear interpolation is performed [7] as given in equation 10 and 11. (4) (3) (5) (6) where [C1, C2, C3, C4] are the four where P1 and P2 are interpolated points in two different directions. Next, iterative refinements are carried out on the interpolated points to follow the isophote curve. Fig. 3 shows the complete process of iterative curvature based interpolation [7].
Figure 1: Illustration Interpolation scheme of proposed method Figure 2: New Covariance based Interpolation Scheme Figure 3: Iterative Curvature based Interpolation Scheme
3 EXPERIMENTAL RESULTS The proposed algorithm is tested on both gray scale and RGB iris images of CASIA database [9]. In Fig. 4, a comparison of Peak Signal to Noise Ratio (PSNR) of ten interpolated RGB iris images using nearest neighbour, bilinear, bicubic, inedi and proposed method is plotted using MatLab. The PSNR values of different methods have been summarized in tabular format in table 1. The average PSNR for each method is calculated which are 30.353 for new method, 30.138 for inedi method, 19.463 for Bilinear method, 22.506 for Bicubic method and 17.669 for Nearest neighbour method. The results clearly show the improved PSNR performance of proposed method as compare to conventional methods. In Fig. 5, an original iris image of size 320x280 is taken and small portion of this image of size 45x30 is cropped out of it as shown. This cropped image is then zoomed up using windows photo gallery and using new proposed method to size of 350x200. In Fig. 6, images of size 50 50 have been interpolated to size 400 x 400 using various techniques such as zooming, nearest neighbour interpolation, bilinear interpolation, bicubic interpolation, inedi method and proposed interpolation method. The result shows the visual quality improvements of new proposed method as compared to others mentioned. The result shows visually that employing the enhancement can help in identification of iris image pattern matching techniques. Figure 4: Matlab plot of PSNR of different interpolation methods of RGB iris images. Figure 5: Cropping Application of New Proposed Method Table 1: The table summarizes the PSNR in Decibels (db) of Iris images enhanced using different interpolation methods. Images 1 2 3 4 5 6 7 8 9 10 Hybrid 30.37 29.9 30.83 30 29.93 30 30.3 30.4 31 30.8 inedi 30.27 29.84 30.44 29.8 29.43 29.9 30.1 30.13 30.9 30.57 Bilinear 19.53 18.96 19.87 19.35 19 19.33 19.41 19.58 19.95 19.65 Bicubic 22.56 22.11 22.93 22.11 22.45 22.41 22.55 22.59 22.5 22.85 Nearest 17.58 17.36 17.94 17.46 17.37 17.46 17.53 17.6 18.24 18.15
(a) Original image (b) Simple zooming of image (e) Bicubic interpolation (c) Nearest neighbor interpolation (f) inedi interpolation method (d) Bilinear interpolation (g) Hybrid interpolation method Figure 6: Interpolation methods employed to enhance 50 50 size iris image taken from CASIA database to 200 200 size for visual quality comparison. a) Original image, (b) zoomed image, (c) Nearest neighbor interpolation, (d) Bilinear interpolation, (e) Bicubic interpolation, (f) inedi interpolation, (g) Hybrid interpolation 4 CONCLUSION The proposed method is a hybrid technique for enhancing iris images in order to aid the recognition process of iris recognition system. The algorithm employs new covariance based interpolation for edges and iterative curvature based interpolation for smooth areas. A threshold is selected by performing an iterative experiment. This threshold differentiates between an edge and smooth areas. The results showed improved PSNR performance and visually enhanced iris images with the propose method as compare to conventional methods. 5 REFERENCES [1] Nanavati S, T. M: Biometrics Identity Verification in A Networked World, New York: John Wiley & Sons, In (2002). [2] Rafael C. Gonzalez, R. E: Interpolation Techniques, Digital Image Processing, Pearson Prentice Hall (2007). [3] HS Hou, H.C: Cubic Splines for Image Interpolation and Digital Filtering, IEEE Transactions Acoustics, Speech, Signal
Processing., pp. 508-517 (1978). [4] Huang J. T.T. M.L. W.Y: Learning Based Resolution Enhancement of Iris Images, British Machine Vision Conference., pp. 153-162 (2003). [5] X. Li, M. T. Orchard: New edge-directed interpolation. s.l. : IEEE Trans. on Image Processing, Vols. 10, pp. 1521-1527 (October, 2001). [6] Nicola Asuni, Andrea Giachetti: Accuracy improvements and artifact removal in edge based image interpolation. s.l. : Proc. 3rd Int. Conf. Computer Vision Theory and Applications, pp. 8 (2008). [7] Andrea Giachetti, Nicola Asuni: Fast Artifacts- Free Image Interpolation. Leeds, UK : Proceedings of the British Machine Vision Conference (BMVC), pp. 10 (September, 2008). [8] B.S Morse, D.S: Isophate-Based Interpolation, IEEE International Conference on Image Processing, pp 227-231 (1998). [9] Database: Center for Biometrics and Security Research. Retrieved 2009, from Institute of Automation chinese Academy of Sciences: http://www.cbsr.ia.ac.cn/english/irisdatabase. asp [10] Aftab, Hassan Mansoor, Atif Bin Asim, Muhammad A New single image interpolation technique for super resolution, INMIC 2008, 23-24 Dec. Karachi, Pakistan 2008, page 592-596