A Robust Iris Segmentation Scheme Based on Improved U-Net

Wei Zhang,Yu Gu,Jing Li,Xiaoqi Lu,Xianjing Meng,Yang Liu

doi:10.1109/access.2019.2924464

Abstract

Iris segmentation plays an important role in the iris recognition system, and the accurate segmentation of iris can lay a good foundation for the follow-up work of iris recognition and can improve greatly the efficiency of iris recognition. We proposed four new feasible network schemes, and the best network model fully dilated convolution combining U-Net (FD-UNet) is obtained by training and testing on the same datasets. The FD-UNet uses dilated convolution instead of original convolution to extract more global features so that the details of images can be processed better. The proposed method is tested in the near-infrared illumination iris datasets (CASIA-iris-interval-v4.0 and ND-IRIS-0405) and the visible light illumination iris dataset (UBIRIS.v2). The f1 scores of our model on the CASIA-iris-interval-v4.0, ND-IRIS-0405, and UBIRIS.v2 datasets reached 97.36%, 96.74%, and 94.81%, respectively. The experimental results show that our network model improves the accuracy and reduces the error rate, which performs well on both near-infrared illumination and visible light illumination iris datasets with good robustness.

Highlights

With the development of information technology, identity recognition is becoming increasingly difficult and important
The results show that fully dilated convolution combining U-Net (FD-UNet) has higher accuracy than that of Part Dilated Convolution combining U-Net (PD-UNet) in iris segmentation and can get more accurate segmentation results
The results show that our method can effectively improve the iris segmentation performance on the near-infrared illumination iris dataset

Summary

INTRODUCTION

With the development of information technology, identity recognition is becoming increasingly difficult and important. Our main innovations and contributions are as follows: 1) A novel network model combining U-Net and dilated convolution is proposed for iris image segmentation. The rest of this paper is organized as follows: section 2 introduces the related work of iris segmentation, including traditional algorithms and deep-learning-based algorithms. B. CONVOLUTIONAL NEURAL NETWORK (CNN) FOR IRIS SEGMENTATION In recent years, with the appearance of deep learning theory, more and more researchers apply it to iris image segmentation [9]–[13]. Yang et al [16] proposed a network model combining FCN with dilated convolution to segment iris, and trained and tested it on CASIA-iris-interval-v4.0, UBIRIS.v2 and VOLUME 7, 2019. Bazrafkan et al [17] proposed an end-to-end convolutional neural network for low-quality iris image segmentation with good results. Hofbauer et al [19] marked iris images for iris segmentation, and their public data will be used in our work

MODELS AND METHODS

EXPERIMENT

Findings

DISCUSSION

CONCLUSION