Research On Several Key Problems In Iris Recognition

With the fast development of internet technology and smart mobile devices,system and information security has become a major problem in current world.Biometrics such as fingerprint,iris and other recogonition technologies are considered to be good solutions for this problem.As an accurate biometric technology,iris recognition is a hot research topoic in both image processing and pattern recognition.Although researchers have devoted great efforts to improve the performance of iris recognition in the past years.There are still some challenging problems unsolved,requiring new theories and approaches.This dissertation focuses on several key problems in iris recognition,including near-infrared iris image quality evaluation,near-infrared iris segmentation,feature extraction and classification as well as visible iris image segmentation and classification.The developed methods improve near-infrared and visible iris recognition performance significantly.The main contents are as follows:(1)In order to exclude low quality iris images,a coarse-to-fine iris image evaluation method is proposed.The method adopts IRST(Improved Radial Symmetry Transform)to localize pupil and evaluate image quality coarsely.To classify the motion-blurred iris images,circular Gabor is used to locate spot region.Then the boundary of spot region is extracted by edge detection and fitted by ellipse.Finally,the motion-blurred iris image is determined by the ratio of major and minor axis together with fitting error.Experimental result shows that the proposed method can exclude low quality iris image effectively.(2)Aiming at the problem that iris localization is easily affected by noise,an effective and efficient iris segmentation method based on SIFT(Scale-invariant Feature Transform)and SDM(Supervised Descent Method)is developed.The method first employs an optimization model for iris localization and uses SIFT feature as discriminative information for iris boundary.Then,SDM algorithm is employed to solve this model and calculate key points of iris outer boundary and eyelids.Finally,RR(Robust Regression)is used to calculate the curves of iris outer boundary,upper and lower eyelids.Experimental result indicates that the developed method can locate iris boundaries effectively and efficiently.(3)To achieve a more effective iris feature representation,an optimization model is established with unknown feature extraction filters.The object of the model is to maximize the inter-class distance and the intra-class similarity.The established objective function contains binary encoding function and Hamming distance which are both non-differentiable.To make the object function differentiable,binary encoding function is approximated by sigmoid function and Hamming distance is replaced by Euclidean distance.Finally,the approximated differentiable model is solved by SGD(Stochastic Gradient Descent)algorithm for iris feature representation The proposed model is verified by experiments on CASIA-Iris-Lamp,achieving an excellent recognition performance.(4)For the purpose of segmenting visible noisy iris image robustly,a statistical denoising iris localization method is developed.The method computes Itg-Diff(Integral Differential Operator)on iris image firstly and selects some candidate iris outer boundaries with largest Itg-Diff values.Then Pauta criterion is used to remove serious outliers for each of the selected boundaries.After that,Itg-Diff is calculated again for these denoised candidate boundaries.The boundary with the largest denoised Itg-Diff value is selected as the final localization.The developed method is used in "Betaeye" iris recognition system,which takes the second place in NICE:?(Noisy Iris Challenge Evaluation:Part ?)among all 67 research groups.(5)In order to improve the recognition performance of noisy iris under visible light,an iris recognition approach is proposed based on global and local features.The developed method trains Adaboost classifiers on a 2D Gabor-based feature set for iris verification.Firstly,irises are normalized by rubber sheet or simplified rubber sheet according to whether segmentations are accurate or not Then,the normalized irises are divided into different count of patches according to segmentation.Moreover,a feature set is constructed based on 2D-Gabor for whole iris and patches.Finally,Adaboost is trained on features for accurately and inaccurately segmented irises separately.The proposed method was evaluated by NICE:?(Noisy Iris Challenge Evaluation-Part 2).We were ranked the 2nd among all of the 67 participants from 28 different countries/districts all around the world.