Research On Iris Localization, Deformation And Feature Extraction

Biometric recognition refers to the automatic recognition of individuals based on physiological or behavioral characteristics. Physiological characteristics, such as face, iris, fingerprint, palmprint, voice etc. are connatural, whereas behavioral characteristics which are the habits of human, such as gait, handwriting, are postnatal. Compared with traditional personal recognition methods, biometrics has the following merits: (1) need not to remember intentionally; (2) have little risk to be forged; (3) available anytime anywhere.Iris recognition is performed mainly based on the intricate structure with many minute characteristics such as furrows, freckles, crypts, and coronas etc. Iris recognition is proved to be one of the most reliable biometrics in terms of identification and verification performance.In this dissertation, some of the key issues related to iris recognition system are investigated under the consideration of the iris physiological structures. The main contributions of this work are as follows:(1) Establish iris image database. The assessment of an iris recognition algorithm is somewhat dependent on the test iris database that used. By choosing the suitable test database, the algorithm can be properly evaluated. On the contrary, if the algorithm is tested on a database with too little variety, the algorithm is easy to be biased. In this dissertation, several iris image databases, such as CASIA,Ubiris,UPOL,MMU,Bath,ICE,WVU, are introduced and analyzed. Moreover, new iris database: SJTU-IDB (Iris DataBase of Shanghai JiaoTong University), is established. The iris images are captured with our self-designed iris capture devices.(2) Research on iris localization method. Iris localization is the first and the most important stage in the whole iris recognition procedure since its result is the precondition of all the subsequent processing stages. In other word, it will be nearly impossible to get a correct recognition result if mistakenly localized the iris region in this stage. In this dissertation, we analyzed and tested large numbers of iris images and make a comprehensive summary on the cases that possibly induce to wrong localization of iris. Then an iris localization method is proposed. The proposed method is characterized with the integration of an adjustment step which can revise the deviation bring about by eyelashes or eyelids or improper threshold value in pre-localization step. Then circle fitting with Least-Square Error rules is performed. Moreover, the detection of eyelashes and eyelids is considered and a new eyelashes detection method with hysteresis thresholding is proposed.(3) Propose an iris normalization method based on meshwork structure. A normalization process is implemented to compensate for iris size variations due to the possible changes in the camera-to-face distance, the pupil deformation and to facilitate the feature extraction process. In most traditional method, iris deformation due to illumination will be simplified as linear transform. But the real deformation of iris is much complex than that. The iris structure can be better represented by meshwork structure. In this dissertation, a new iris normalization method is proposed, which based on the meshwork structure of iris instead of the traditional over-simplified linear structure.(4) Propose a feature extraction method using 2-D phase congruency. There exist many local structural features in iris pattern that can be used to iris recognition. In this dissertation, a new feature extraction method which using 2-D phase congruency is proposed. Due to the characteristics of phase congruency, it can efficiently extract the local features of iris and avoid the influences of variation of contract and external illumination. Zero-Mean Normalized Cross-Correlation is computed as the similarity distance.(5) Propose a feature extraction method by region dividing according to iris physiological structures. Though many feature extraction method are applied on iris patterns, they tend to treat iris pattern the same as other normal texture images and didn't consider much about the iris physiological structures. In this dissertation, the iris structure are researched and then divided into several sub-regions. Feature extraction is performed according to the different characteristics of each sub-region. Then at last a group-of-threshold strategy is applied to further improve the recognition performance.(6) Research and test on iris pattern stability. First the controversial problem about iris recognition and iris diagnosis (Iridology) is discussed. They have different viewpoints about iris stability. Then make the iris recognition test on two cases. One is about iris recognition after vision refractive surgery. Another is iris recognition on the iris images captured between an interval of long time (two or three years).