Research On Iris Localization And Segmentation Based On Double-Circle Constraint

Iris recognition is a widely used biometric identification method,iris segmentation and iris localization,as image preprocessing,jointly determine the iris texture area for identification and matching,so accurate and efficient iris localization and segmentation methods are the key to improving iris recognition performance.In recent years,with the development of mobile device camera technology,iris recognition on images captured in user non-cooperation and visible light environment has received widespread attention.However,unlike the high-definition low-noise iris images that taken with the user cooperation and near-infrared light source,the iris images taken in the above scenarios are often accompanied by a large amount of noise,which is a great challenge for the current segmentation-based circle-fitting localization algorithm.Therefore,in order to improve the performance of iris recognition systems in the above scenarios,it is extremely necessary to study accurate and efficient iris localization and segmentation algorithms in non-cooperative and visible light environment.This project is based on the research of accurate and efficient iris segmentation and localization algorithm for the above scenarios,and the main research content and innovation points are as follows:1)This paper proposed a bicenter-based iris localization algorithm(Double circle based iris localization,DCL).At present,most of the iris localization methods use a lot of postprocessing to fit the inner and outer circles of the iris on the segmented iris mask or the predicted inner and outer circle contour mask,but the locate accurate of this segmentationbased circle fitting algorithm is not high in the iris image that captured on the user noncooperation and visible light illumination,often occurs the inner circle fitting fails or the fitted inner and outer circles do not conform to the geometric characteristics of the iris ring.The proposed DCL algorithm,directly performs the positioning of the inner and outer circles of the iris and the predict the radius of the double circle on the feature map,which extracted by the neural network.Secondly,the two-radius regression and central point localization network are designed to consistent an end-to-end iris localization framework.Finally,after analyzing the distribution characteristics of the center point on the iris image that taken in the visible light environment,a sampling strategy is innovatively proposed to assist model training.Finally,experiments on four visible light and user non-cooperation eye datasets show that the proposed DCL iris localization method is robust,and achieves competitive inner and outer circle localization performance.2)This paper proposes a dual-flow iris segmentation and localization model PSNet(Parallel iris Segmentation and localization Network,PSNet).The current mainstream multitask iris segmentation algorithm first predicts the iris mask and the inner-outer circle contour mask,and then uses the post-processing algorithm to fit the circles on the contour mask.This leads to the algorithm spending a lot of time on post-processing,and the serial design of first predicting the mask and then fitting the inner and outer circles seriously affects the real-time performance of the algorithm.However,PSNet adopts the design of dual-task parallel processing,and the network directly outputs the inner and outer circle localization results and segmentation results in an end-to-end manner without any post-processing.On the basis of the integrated DCL iris localization module,PSNet also integrates the proposed lightweight feature extraction network for iris segmentation tasks.In addition,based on the pairing and inherent geometric constraint characteristics of the inner and outer circles of the iris,this paper innovatively proposes the double circle pairwise constraint loss function CPLoss,which is used to solve the size optimization problem of circular bounding boxes.Compared with the current multi-task iris segmentation algorithm,the proposed method has the advantages of dual-task parallel design and no post-processing.While maintaining high precision,the overall task processing speed has been greatly improved.