Iris Pattern Classification Combining Orientation Recognition

The importance of personal authentication is gradually increasing with the development of the information society. Biometrics identification technology plays an important role in cyberspace. Unlike other biometrics such as the face or fingerprints, iris recognition has high reliability for personal identification. Iris recognition methods are classified into four categories: the phase-based method (Daugman, 1993), the zero-crossing representation-based method (Boles and Boashash, 1998; Sanchez-Avila and Sanchez-Reillo, 2005), the texture-based method (Wildes, 1997; Ma et al., 2003), and local intensity variation (Ma et al., 2004, a;b). Using the internal CASIA dataset (CBSR, 2005), Ma et al. evaluates the proposed algorithm by comparing the performance of other iris recognition methods proposed by Daugman, Wildes, and Boles and Boashash (Ma et al., 2004, b). The experimental results show the equal error rates (EER) of respective algorithms (Ma, Daugman, Wildes, and Boles) are 0.07%, 0.08%, 1.76%, and 8.13%, respectively. In other studies, the Daugman’s method which is a representative algorithm of iris recognition is also evaluated using the subset of internal CASIA dataset (Sun et al., 2006) and the CASIA iris image 1.0 database (Wang et al., 2007), which is available from the CASIA web site. The EERs of Daugman’s method reported in Sun et al. and Wang et al. are 0.70% and 0.67%, respectively. These analysis indicate the high accuracy of recognition performance although the EERs of Daugman’s method described in these papers are not the same because the iris segmentation method including eyelid and eyelash detection would not be exactly the same. Iris recognition technology are applied in various fields. Especially, the iris recognition algorithm embedded on a mobile phone requires robustness to rotation changes because capturing the iris pattern by a hand using a camera built in the mobile phone causes the rotation changes. However, the iris recognition methods described above are generally fragile in rotation variation. We previously proposed a rotation spreading neural network (R-SAN net) that focused on spatial recognition/memory systems (parietal cortex(PG)) in the brain and recognized an object’s orientation and shape (Nakamura et al., 1998; Yoshikawa and Nakamura, 2000). The R-SAN net can simultaneously recognize the orientation of the object irrespective of its shape, and the shape irrespective of its orientation. The characteristics of the R-SAN net are to use a two-dimensional input pattern in a polar coordinate system converted from the Cartesian coordinate system. The R-SAN net is suitable for the shape and orientation recognition of concentric circular patterns. The orientation recognition performance of R-SAN net allows the 10