Iris Recognition Algorithm Based On Kernel Methods Research

Due to the high reliability, personal identification based on the human iris is one of the most promising recognition techniques. An iris recognition system always consists of the following modules: image auto acquiring, image preprocessing, feature extraction and classification. Based on analyzing original methods, algorithms of iris location, feature extraction and feature matching are respectively researched in this dissertation. Main works are as follows:(1) Snake algorithm combined with Hough algorithm, and two-step locating algorithm named cursory locating and accurate locating are respectively adopted in iris inner and outer location.(2) Kernel-based nonlinear feature extraction and classification algorithms are a new popular research direction in machine-learning and widely used in many fields. Because of the disadvantage inherently owned by traditional transform (such as Gabor and Wavelets), such as that radix vectors are fixed and independent of other data, leading to complexity on parameters selecting (e.g., spatial location, orientation, and frequency), in this dissertation, kernel-based feature extraction methods (Kernel Principal Component Analysis (KPCA), Kernel Independent Component Analysis (KICA), Kernel Discriminant Analysis (KDA), and Kernel Springy Discriminant Analysis (KSDA)) are adopted, which firstly map iris data to higher dimensional feature space using kernel function, and then train the certain number of samples to obtain features. Three distance measures are used to evaluate the performance of kernel-based feature extraction methods.(3) Compared with simple distance measures, more effective classifiers (Kernel-based Fisher Discriminant Analysis and support vector machine), are adopted in classifier-designing section, where the data used to be classified is extracted by PCA method. However, KFD and SVM are binary classifiers which just can separate data between two classes but helpless for multi-class data. Therefore, One-Against-All method and One-Against-one method are adopted to make KFD and SVM work in iris recognition. The performance of these encoding algorithms and classifiers are analyzed on CASIA II database. A series of experiments indicate that compared with corresponding linear feature extraction methods, kernel-based feature extraction methods have an encouraging performance. Furthermore, KFD and SVM adopted in this dissertation outperform the existing classifiers for iris recognition.