Research On Iris Localization And Recognition Algorithms

Because iris recognition has the least false recognition rate among all kinds of biometrics, and it hasthe merits of stability, uniqueness, collectability, non-invasive, antifalsification and liveness detection, irisrecognition has the broad application prospect, huge economic benefits and research value. Iris recognitiontechnology has been more than20years since it was officially proposed by Dr. Daugman in1993. However,because of the complexity of its application environment, its key technologies still need to be improved.This paper mainly focuses on improvement and innovation of the technologies of the iris image localization,local iris image quality evaluation, iris feature extraction, feature dimension reduction, the fusion ofdifferent sub-regions of iris. The experimental results demonstrate good performance.The main work and contributions are provided as follows:(1) The shape of pupil and iris are neither absolute circle nor ellipse and iris may be affected by manyfactors, so it is difficult to build the shape model. In this paper, we propose a self-adaptive CV (SACV)level set iris localization model according to the nature property of the iris images. A coarse-to-finelocalization strategy is adopted by SACV model. Firstly, the coarse boundaries of pupil and iris arelocalized. Then, interference factors in pupil and iris areas are detected and evaluated. Finally, thecorresponding parameters of SACV are adaptively set according to interference values. The experimentalresults, based on three public (CASIA-V3Interval, Lamp, MMU-V1) iris image databases and one private(JLUBRIRIS-V1) iris image database, show that the SACV model has strong robustness, whichoutperforms some of the existing methods in terms of segmentation error rate. The proposed model is notonly suitable for ideal iris image but also for noideal iris image. Most of important, the segmentation resultof SACV only keep the iris real region.(2) As iris regions have very rich texture information, only adopting single feature extraction methodcannot effectively express iris texture information. This paper puts forward the joint feature extractionmethod to extract iris feature. That is to combine iris features extracted by two-dimension Gabor methodwith iris features extracted by gray level co-occurrence matrix (GLCM) method to form the joint irisfeatures. The former method represents the frequency spectrum feature extraction methods and the latterone on behalf of the statistical feature extraction methods. As classifier plays a crucial role, we optimize theparameters of radial based function of support vector machine (SVM) to get enhanced SVM classifier.Experimental results, based on one private JLUBRIRIS-V1iris image database, show that joint featuresand enhanced classifier can improve the overall performance of the iris recognition system.(3) The selections of the optimal sub-feature have become an important issue for iris recognition system. However, traditional SIFT features contain redundant features. This paper proposes three featureselection strategies to select feature with more discriminative information, which are orientation probabilitydistribution function (OPDF) based strategy, magnitude probability distribution function (MPDF) basedstrategy, and compounded strategy based on combined OPDF and MPDF. The experimental results, basedon three public iris image databases，demonstrate that:1) The first strategy can reduce the number ofkeypoints;2) the second strategy can reduce dimension of feature elements; and3) the third strategy canreduce the number of keypoints and dimension of feature elements.(4) In view of the traditional SIFT technology does not consider the location of the features, whichmay further cause two features corresponding to the minimum distance that could not be related to the samesub-part image. We propose an iris recognition system based on sub-region feature matching and weightedfeature fusion. Firstly, we directly divide anular iris into three sub-regions in no overlapping way. Then,feature matching works in three sub-regions, respectively. Thirdly, assign corresponding weights for threesub-regions in the form of training. Finally, weighted different sub-regions’ matching scores to generate thefinal decision. The experimental results, on three public iris image databases, demonstrate that the proposedsystem has three advantages:1) the first advantage is that local matching within a grid constrains the SIFTfeatures to match features from nearby regions only;2) the second one is increase of speed in matchingsince the number of features decrease; and3) the major one is that weighted approach assure that higherinformation carrying regions of the image are associated.(5) To analyze the characteristics of different sub-regions of iris, and to study whether parts of the irisimage can replace the whole iris image or not, we study the influence of different sub-regions of iris on irisrecognition system. Firstly, the normalized iris image is divided into different sub-regions. Secondly, bothtwo dimension Gabor filters and GLCM are utilized to extract iris features. Thirdly, SVM and k nearestneighbor methods are applied to classify iris features. Finally, recognition correct rate is adopted to analyzethe iris feature information of different sub-regions. The experimental results, based on the four iris imagedatabases, demonstrate that:1) the middle sub-region of iris near the pupil contains more stable information;2) the partial iris image cannot completely replace the entire iris image for iris recognition system.(6) According to the proved truth that different sub-regions of iris contains different textureinformation, this paper proposes the strategy of the local feature fusion. Firstly, to evaluate local quality foreach sub-region of iris and to assign corresponding weighted coefficient based on evaluation score. Then,sub-region feature fusion to regenerate new feature template. Finally, the new feature is used for irisrecognition. The experimental results, based on four iris image databases, demonstrate that:1) the proposedquality evaluation algorithm is a self-adaptive algorithm and it can automatically optimize the parametersaccording to iris image samples’ own characteristics;2) our proposed feature information fusion strategycan effectively improve the performance of iris recognition system, as well as further effectively restrain or weaken the fragile bits of different tracks on the basis of retaining the consistent information.