Research On Iris Segmentation And Feature Extraction In Large Scale System

Iris recognition is widely used in bank access,military security,border security,identity verification,and check on work attendance in coal mines due to its natural advantages in uniqueness,stability,and defense.It is one of the most promising fields in biometrics.Recently,conventional iris recognition under homogeneous and controlled conditions has been extensively studied.More attention has been paid on uncontrolled and heterogenous iris recognition,matching iris images across different domains.In both of task patterns,it is challenging to manually design a robust encoding filter to face the complex intra-class variations of iris images due to the influence of light intensity,collection device and samples size.The non-ideal captured images intensify the difficulty of iris segmentation while parameter adaptation under cross-system makes it difficult for current iris research.Although deep learning replaces the practice of manually setting filters,it is still an urgent issue to choose the appropriate model structure based on the characteristics of the iris image and the practical problem of the actual scene.This paper proposes several solutions to solve the iris recognition issues in large scale system.1.This paper proposes a non-normalized pre-processing model based on dynamic path search for iris segmentation.It considers the preprocessing of the iris as two edge extractions,i.e.the inner and outer edge extraction of the iris.This algorithm based on the global cumulative gradient optimization search is not sensitive to light and noise.In addition,it does not use the classic rubber sheet method to normalize the model,but retains the original iris texture information directly for feature extraction.It also theoretically analyzes that the iris region normalization is not an essential key step in feature coding based on deep model.2.This paper optimizes the original deep convolutional neural network corresponding to the characteristics of the iris image itself,and proposes a multi-scale deep feature fusion model for extracting feature information,which uses partial convolution operator as the basic unit of model.The preprocessing mask information can be used to make feature extraction focus only on specific areas of the image.Two modules--attention redefine Module and feature fusion module are respectively used to optimize feature information and fuse them at different scales.Finally,the fused deep features are used to predict labels.3.This paper defines a novel loss function referred as T-Center to enhance the discriminative power of deep features.It fuses the Euclidean distance between the feature vector and its corresponding feature center.Furthermore,the L2 norm is used to constrain the effect of hard samples to maximize the intra-class variance and minimize the inter-class-to-class variance.In order to verify the effectiveness of the loss function,it performs benchmark experiments on two different task modes,i.e.1:1 and 1:N recognition tasks.The testing dataset are ND-IRIS-0405,CASIA,and IITD.This paper extends the proposed function to other existing deep models to verify its universality and cross-database recognition ability.