The paper compares three different methods of iris image segmentation, namely the method using the Daugman`s integro-differential operator, the method using the Hough transform for the detection of circles, and the method based on the analysis of the distribution of edge points. The accuracy and run time of the implemented methods were estimated in the experimental study conducted using the MMU Iris Image Database. The carried out research has shown that the method using the Daugman`s integro-differential operator has the greatest accuracy, and the method based on the analysis of the distribution of boundary points has the shortest operating time.
In recent years, identifying a person by his/her biometric parameters becomes a more and more
popular field of research. The reason for this is both the increased interest in the biometric
identification by public and commercial bodies and the intensive development of technical tools of
recording, storage and processing of such information. Today, a broad range of features of human
body is used for the identification, including face images [
In fact, the iris is a circular movable diaphragm with a diameter of about 12 mm, at the core of
which there is a round hole, a pupil [
The iris color (eye color) depends on the coloration (pigmentation) of iris anterior surface visible
through the transparent cornea [
Generally, the identification systems using an iris image are built according to the classical scheme: - selection of an informative region (iris) on the image, - feature description of the selected region, - comparison of the feature descriptions.
In this case, the quality assessment of both obtained images and effectiveness of individual method
stages may be included into the above scheme [
It should be noted that the accuracy of the first of the above-mentioned stages (selection of the iris on the image) plays an important role in the achievement of high quality.
The selection of the region corresponding to the iris on the image is the segmentation-related task
that can be completed using a variety of methods. Specifically, the approaches based on the edge
detection [
When using these approaches, the internal and external iris boundaries (boundaries “iris-sclera” and “pupil-iris”) are sought, and the further approximation of detected boundaries with circles is performed. It should be noted that such approximation is only the first stage of a segmentation, because after selecting the ring (which may be non-concentric) formed by these circles, the part of the iris may remain closed and shaded by eyelids and eyelashes.
This paper describes the experimental comparison of three approaches to the selection of the iris on the image.
This paper has the following structure. Sections 2-4 briefly describe the methods considered in this report. Specifically, Section 2 describes the method based on the Daugman's integro-differential operator, Section 3 describes the method based on the circular Hough transform, and Section 4 describes the method based on the edge detection and edge points distribution analysis. The results of experimental studies carried out with MMU Iris Image Database are presented in Section 5. In the end of this paper, there are Section 6 containing the conclusion and Section 7 containing the reference list.
According to the method based on the Daugman's integro-differential operator, the possible centers of the iris should be calculated first on an input image Iin. To do this, the points on Iin with the brightness level below the predefined threshold are selected. Then, the points, which correspond to the local minima of the brightness, are selected from the above-mentioned points.
Local minimum is the minimum of the circular neighborhood with a radius equal to the minimum iris radius. In this method, the minimum and maximum radii of the iris should be set to overlap a little bit the iris radius values from the image database.
Then, the parameters of circle approximating boundary “iris-sclera” should be calculated. This is
done by using the Daugman's integro-differential operator [
Example of how the method described in this section works for figure 1 is presented in figure 2.
According to the method based on the circular Hough Transform, the Canny edge detector [
Then, the parameters of the circle, which approximates the “iris-sclera” boundary, should be
determined on the image Ied1 of edge points using the circular Hough transform [
Thereafter, image Ied2 is made up of the points of the image Ied1 that are within the circle determined
at the first stage. The parameters of the circle, which approximates the “pupil-iris” boundary, should
be determined on the image Ied2 using the circular Hough transform [
An example obtained using this method is shown in figure 3.
As a part of this work, the method based on the circular Hough transform and the analysis of the edge point distribution was proposed. This method includes the following stages. At the first stage, the iris internal boundary (boundary “pupil-iris”) should be selected on the input image, and the parameters of the circle approximating the pupil should be determined. At the second stage, the iris external boundary (boundary “iris-sclera”) is selected, and the parameters of the circle approximating the iris external boundary are determined using the least squares method. These stages are described below in more detail.
To determine the internal boundary of the iris, the Canny edge detector [
Figure 4 below provides the example of pupil boundary determination for the image depicted in figure 1.
To determine the iris external boundary, the function of the edge point distribution from the pupil
centre, which is described in [
An example of the graph of the function f(r) of the edge point distribution on the distance R to the centre of the pupil is shown in figure 5. The argument r of the function f(r) equals to the difference between R and Rpupil.
Thereafter, the direct search for the local maxima of the f(R) function is performed starting with the highest maximum. For the selected local maximum RM, the edge points, which fall into the ring defined by the radii RM – s and RM + s, are approximated by the circle using the least squares method. The circle, for which the approximation error is lower than the predefined value, is selected. It is considered that this circle approximates the iris external boundary.
An example obtained using this method is shown in figure 6.
To compare the accuracy of segmentation and operation time of the above methods, the simulation experiment was carried out. All methods were implemented in MATLAB. The experiment is carried out on a personal computer powered by Intel Core i5-4210M with a frequency 2.6 GHz and randomaccess memory of 6 GB.
When conducting this experiment, 30 iris images taken from MMU Iris Image Database [
To assess the precision of each method, the Jaccard index [
The results of the experiments are presented in Table 1 and figures 7, 8. Specifically, “A” section describes the method based on the Daugman's integro-differential operator, “B” section describes the method based on the circular Hough transform, and “C” section describes the method based on the analysis of the edge point distribution.
Method
A B C
These results demonstrate that the Daugman's integro-differential operator is the most precise, but requires the longest operation time. The method based on the circular Hough transform is a little bit less precise, but has a significantly lower operation time (in 2.5 times).
The method based on the analysis of the edge point distribution was the fastest one (almost 6 times faster than the method of Daugman's integro-differential operator). In this case, it showed the same precision with the used data set as the method based on the Hough transform.
Three methods of assessment of iris boundary parameters were evaluated in this study, namely: the method based on the Daugman's integro-differential operator, the method based on the circular Hough transform, and the method based on the analysis of the edge point distribution.
For images from MMU Iris Image Database, the reference segmentation of iris images was performed, and numerical experiments to assess the precision and operating time of implemented methods were carried out. The experiments show that the method of the Daugman's integrodifferential operator is the most precise with regard to the applied data set, and the method based on the analysis of the edge point distribution proposed in this paper is marked by the lowest operation time and has the same level of precision as the Hough transformation.
In future, it is planned to use study results to develop the method of person identification by the picture of iris.
The reported study was funded by RFBR according to the research project no. 18-01-00748 a.