Scribd
Upload
84 views

Eye Recognition With Mixed Convolutional and Residual Network (Micore-Net)

This document presents a new eye recognition framework called MiCoRe-Net that uses a mixed convolutional and residual neural network for biometric identification. MiCoRe-Net relaxes the data collection process, improves anti-fake quality, and outperforms other classifiers on public datasets. The framework takes whole eye images rather than segmented iris regions as input. It also uses data augmentation techniques like rotation and cropping to enrich the training data and improve the model's ability to generalize. Experiment results show MiCoRe-Net achieves over 99% and 96% accuracy on two public eye image datasets.

Uploaded by

Kathir Vel
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
84 views

Eye Recognition With Mixed Convolutional and Residual Network (Micore-Net)

This document presents a new eye recognition framework called MiCoRe-Net that uses a mixed convolutional and residual neural network for biometric identification. MiCoRe-Net relaxes the data collection process, improves anti-fake quality, and outperforms other classifiers on public datasets. The framework takes whole eye images rather than segmented iris regions as input. It also uses data augmentation techniques like rotation and cropping to enrich the training data and improve the model's ability to generalize. Experiment results show MiCoRe-Net achieves over 99% and 96% accuracy on two public eye image datasets.

Uploaded by

Kathir Vel
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 1

1

IEEE Access ( Volume: 6 ), March 2018

Eye Recognition with Mixed Convolutional and


Residual Network(MiCoRe-Net)
Zi Wang, Chengcheng Li, Huiru Shao, Jiande Sun

Abstract—Although iris recognition has achieved big successes


on biometric identification in recent years, difficulties in the
collection of iris images with high resolution and in the seg-
mentation of valid regions prevent it from applying to large-scale
practical applications. In this paper, we present an eye recognition
framework based on deep learning, which relaxes the data collec-
tion procedure, improves the anti-fake quality, and promotes the
performance of biometric identification. Specifically, we propose
and train a mixed convolutional and residual network (MiCoRe- (a) Sample with low resolution (b) Sample occluded by eyelid
Net) for the eye recognition task. Such an architecture inserts and eyelash
a convolutional layer between every two residual layers and
takes the advantages from both of convolutional networks and
residual networks. Experiment results show that the proposed
approach achieves accuracies of 99.08% and 96.12% on the
CASIA-Iris-IntervalV4 and the UBIRIS.v2 datasets, respectively,
which outperforms other classical classifiers and deep neural
networks with other architectures.
Index Terms—Eye Recognition, Iris Recognition, Deep Learn-
ing, Convolutional Neural Network, Deep Residual Network, (c) Sample interfered by eye- (d) Sample with deformation
Mixed Convolutional and Residual Network (MiCoReNet). glasses

Fig. 1: Challenging Samples in iris recognition systems.


I. I NTRODUCTION

B IOMETRIC recognition has been well studied in the


recent decades, and plays a more and more important role
in the area of information security, personal identification, etc.
pupil, iris, sclera, even with wrinkles and eyeglasses included
in the eye region, which are negative factors in iris recognition,
Among all kinds of biometric information, iris has stableness are transfered into positive factors in our proposed method
and uniqueness of features [1]. In recent years, researchers and even can provides helpful information for recognition
have achieved significant progresses in iris recognition and features. In this paper, with these concerns we employ the
the best algorithms can reach the accuracy of more than 99% powerful capability of multi-level feature learning in deep
[2] [3]. However, iris recognition has disadvantages, mainly in neural network and propose a novel recognition framework,
the following two aspects: (1) Daugman etc. suggested that if which takes the whole eye image, rather than the segmented
we want to get a satisfied accuracy of recognition, the radius and normalized iris ones, as a kind of biometrics. Such system
of iris should be at least 80-130 pixels. Moreover, internal unburdens us the demanding on segmentation and relaxes the
factors such as eyelid and eyelash, and external factors, such acquisition requirements.
as eyeglasses, can decrease the quality of the iris images. [4]. Our framework consists of two main parts: data augmenta-
(2) The iris has to be precisely segmented from the eye/face tion and mixed convolutional and residual network (MiCoRe-
image with powerful segmentation algorithms [4]. Besides the Net) as shown in Fig.2. Since the images from datasets are
interference factors mentioned above, the fact that pupil is usually collected with certain distances and perspectives, the
not strictly a circle makes the segmentation and the upcoming representation of the feature appears relatively shallow to some
procedures, such as normalization, even more difficult. extent. For example, in the UBIRIS.v2 database, images are
Considering these factors, we propose an eye recognition captured in a range of four to eight meters with not only
framework with mixed convolutional and residual network. frontal perspective, but left and right off-angle perspectives as
Compared with iris, the whole eye region consisting of the well. Data augmentation including rotation,cropping,rotation
eyelash, pupil, iris, sclera, even with wrinkles and eyeglasses after cropping,and the combination of them, can greatly enrich
which makes collecting much easier. Moreover, the eyelash, the feature of each sample/class of the training set, which
Zi Wang and Chengcheng Li are with The University of Tennessee,1520 provides the deep neural network with more information. In
Middle Dr., Knoxville,TN,USA,37996 the experiment, we will compare the performance between
Huiru Shao and Jiande Sun are with School of Information Science and different data augmentation strategies. A system can be trained
Engineering, Shandong Normal University, Jinan 250014, Shandong Province,
China to fit different kinds of testing samples by feeding such
The corresponding author: Jiande Sun, e-mail: [email protected] augmented training samples to it.

You might also like