Research On Iris Recognition Algorithms

With the development of information society and the increasing requirements of information security, the personal identification technology based on biometric has been received more and more attention. Iris recognition technology has the physiological advantages of stable, unique, non-invasive. Meanwhile, it is also one of the typical and complex computer vision and pattern recognition problems, which makes it become a research hotspot of biometrics recognition. This thesis focuses on the study iris image preprocessing and iris feature extracting. The main work is provided as following:The proposed method localizes iris in non-ideal data, which has noise issues such as specular reflections, non-uniform illuminations occlusions of eyeglasses,eyebrows. The iris is coarsely localized using light detection and statistics. The iris’ s inner boundary is detected by Canny edge operator with mask. The inner boundary location is validated by gray level value and boundary moving method, modifying the error locations. The iris’ s outer boundary is detected by fusion filters and difference in horizon. The method was tested on noisy images from CASIA-V4.0-Thousand database, of which the average accuracy is 99.65% and the average time cost is963 ms per image.Due to the regularities of iris distribution, and the different contribution rate of each sub-region, the proposed method is based on 2D-Log-Gabor filter groups and Adaboost algorithm. The complete characterizations are built with various parameters of 2D-Log-Gabor filter groups and each sub-regions of iris normalization. The training sample of Adaboost is features between every two characterizations. The iris is recognized with cascade classifiers. The method, tested on the database:CASIA-V1.0, CASIA-V4.0-Thousand, shows improvement in both accuracy and time.The traditional feature comparison method by shifting matching neither achieve satisfied result in rotate iris images which result from head tilting in iris capturing,nor apply to iris template security framework. The proposed method is adding Polar Complex Exponential Transform moments to feature extraction and generation the iris template by feature selection. The method, tested on the database:CASIA-V3.0-Interval, CASIA-V4.0-Thousand, shows a good performance in rotation invariant.