Research And Implementation On Key Algorithms Of Iris Recognition Based On Convolutional Neural Networks

Biometric recognition algorithms have replaced the tranditional identity authentication modes gradually due to the security,convenience,anti-counterfeiting and reliability,which are widely applied in the fields of national defense,financial,e-businesses,entrance guard and social welfare.etc.It is because uniqueness,stability,non-contact and anti-counterfeiting of iris textures that iris recognition algorithms have become one of the important branches in the biometric recognition fields.However,there are still some drawbacks in current iris recognition algorithms.This paper mainly focues on improvement of iris segmentaion algorithms and iris recognition algorithms in the whole iris reconition systems,and proposes new architectures and algorithms based on convolutional neural networks(CNNs,convnets).The main works and contributions are summarized as follows:(1)Iris segmentation algorithms are of great significance in complete iris recognition systems and directly affect the iris verification and recognition results.However,conventional iris segmentation algorithms have poor adaptability and are not sufficiently robust when applied to noisy iris databases captured under unconstrained conditions.Additionally,there are currently no large iris databases;thus,iris segmentation algorithms cannot maximize the benefits of convolutional neural networks(CNNs).The main work of this paper is as follows: first,we propose an architecture based on CNNs combined with dense blocks for iris segmentation,referred to as a dense-fully convolutional network(DFCN),and adopt some popular optimizer methods,such as batch normalization(BN)and dropout.Second,because the public ground-truth masks of the CASIA-Interval-v4 and IITD iris databases do not include the labeled eyelash regions,we label these regions that occlude the iris regions using the Labelme software package.Finally,the promising results of experiments based on the CASIA-Interval-v4,IITD and UBIRIS.V2 iris databases captured under different conditions reveal that the iris segmentation network proposed in this paper outperforms all of the conventional and most of the CNN-based iris segmentation algorithms with which we compared our algorithm's results in terms of various metrics,including the accuracy,precision,recall,f1 score,and nice1 and nice2 error scores,reflecting the robustness of our proposed network.(2)Iris recognition step contains iris features extraction and iris features classification,which is also important in complete iris recognition systems,especially the iris features extraction.On one hand,traditional iris recogniton algorithms performed badly on nosie iris images and are not robuse enough.On the other hand,the iris recognition algorithms based on CNNs did not adapt the deep architectures whose number of layers is larger than 5 only considering the convolutional layers,and have not fully taken advantage of deep CNNs.Given this two problems,works of this paper are as follows: 1)Architecture based on CNNs is proposed for iris recognition in this paper,which is named as IrisConvDeeper.2)A shallower architecture termed as IrisConvShallower is designed for comparison.IrisConvDeeper consists of structures called dense block,and have 12 or 14 convolutional layers.Meanwhile,IrisConvShallower consists of the standard convolutional layers,and have 6 or 7convolutional layers.The performance of the proposed architectures is tested on two public iris databases collected under different conditions: CASIA-Iris-V3 and IITD iris image databases.The experimental results demonstrate that the proposed IrisConvDeeper outperforms some of state-of-the-art approaches in term of correct recognition rate.