Research Of Iris Recognition Based On Bagging And Boosting

Iris recognition,as an accurate biometric identification method,has been gradually applied to the products in different conditions.In Iris Recognition System,both the segmentation and recognition are the key parts of the process.In order to solve the following problems in segmentation and recognition,this paper makes a research based on Bagging and XGBoost:(1)Due to the large range searching area,the traditional iris segmentation algorithm leads to longer segmentation time.In order to solve the problem,the paper proposes an iris localization method with multi-datasets based on Bagging.First,based on the thought of Bagging,we generate 3 sets of data from the iris datasets.In each set,we extract the SIFT features of the key points.The position transformation from current points to the expected points is linearly regressed to get the boundary key point coordinates with the SDM algorithm.And the final key points are determined by the multiple SDM models.Lastly,the key points are fitted with curves which are two parabolas for eyelids and two semicircle for iris outer boundary.The proposed algorithm was tested on the CASIA-Iris-Lamp database.The precision of the proposed iris segmentation algorithm is 94.0%,which is 3.2%higher than that of intergro-differential operator method and 0.9%higher than that of single SDM.The average positioning time is 269.0 ms,which is 11.8%shorter than intergro-differential operator algorithm.(2)In order to break the bottlenecks of the iris recognition performance of single feature,the paper proposes an algorithm based on XGBoost in Boosting Methods to fuse 22 features.Firstly,based on the 2D Gabor filters and the Multilobe Differential Fileters.22 features of different parameters are extracted from the normalized iris images.And we calculate the hamming distances of different samples to generate the hamming distance features.According to the definition that the positive samples are from the same types and negative samples are from the different types,the identification problem can be transformed to the binary classification problem.Finally,XGBoost is used to fuse the features and classify the samples for iris identification.Through many experiments,the EER of the proposed method is 4.6%lower than results based on the original best feature in CASIA-Iris-Interval Database.In CASIA-Iris-Lamp Database,the EER is 5.92%lower.In MIR and MMU databases,the EER declines 2.37%and 4.25%,respectively.