KR20020088890A - Rotation Invariant Feature extraction for Iris Pattern recognition - Google Patents

Abstract

PURPOSE: A method for extracting a feature point unchangeable to a video rotation for an iris pattern recognition is provided to prevent an image from being changed caused by a rotation and prevent a luminosity of an image from being changed by calculating a feature point of a pattern using a coefficient of a discrete Fourier transform with respect to one-dimensional luminosity data of an iris pattern. CONSTITUTION: A filter having a characteristic for maintaining an edge is used for removing a noise in a video pre-processing process, and a component labeling is used for making a boundary of the pupil of the eye. Also, only the pupil of an iris video is extracted using an automatic boundary value detecting method. The pupil is detected for searching a central point of the pupil. A focus of the pupil is searched using a point being met by central values of a histogram on an X-axis of the extracted pupil and a histogram on a Y-axis of the extracted pupil. A radius of the pupil is searched from the focus and data on a 30-percent large circle more than the radius toward an outside of the pupil is searched. A radius of the circle for obtaining data becomes r(1+0.3), and the number of entire data becomes R*cos45 degree*8. A discrete Fourier transform is performed with respect to the obtained one-dimensional luminosity data.

Description

Rotation Invariant Feature Extraction for Iris Pattern Recognition

In the 21st century's highly information society, the value of information has emerged as a key determinant of personal professionalism, corporate interests and the future of the nation. The establishment of information infrastructure and the development of contents have long been a national policy goal, and companies and individuals are struggling to accumulate and disseminate high-quality information. In addition, the necessity of information accumulation is spreading more deeply thanks to the development of computers, and the effective sharing of information through the network is the biggest topic in the 21st century. However, it is also unavoidable that such a huge amount of money and information accumulated through effort is easily stolen by others' technical approach.

Therefore, efforts have been made to develop the only security device for hundreds of years before the computer was used. As a result, biometric technology using a part of the human body has been developed and has come to this day. The biggest feature of these biometrics is that there is no fear of loss, theft, forgetting, or cloning due to external factors in any case. Among these biometric techniques, the iris pattern for iris recognition does not change for life, and the left and right sides are different and even identical twins. In addition, since the iris can only be a living eye, it is impossible to falsify by scientific methods. Unlike the retina, the iris is not affected by intraocular diseases by using the pattern of the iris located around the pupil of the eye. It is convenient for use because there is no.

The human iris is located at the front of the uvea, where the anterior part of the ciliary body is extended. It is located in front of the eye, between the cornea and the lens, and the round hole in the middle of it is the pupil. The front side of the iris has irregular relief, and there is a circular pattern that is raised near the pupil. This is called a crimp wheel, and it is known that once it is born at birth, it does not change for a lifetime. . Therefore, there was an attempt to apply this pattern to personal identification and authentication by photographing it with a camera.

First, Daugman proposed a method to generate 256-byte Iris code using 2D Gabor Filter, and Wides studied iris recognition using Laplacian of Gaussian filter of Gaussian filter. Was performed. However, all of the above methods use two-dimensional data, which is complicated and computationally expensive.

On the other hand, Boles has been trying to recognize one-dimensional brightness data obtained from the iris using a wavelet transform. In this study, we tried to recognize this value by comparing the zero-crossing point data from wavelet transform data. However, in this method, when the image acquired with respect to the iris pattern of the same person is rotated due to the characteristics of the wavelet transform, the coefficient value is different so that it is not possible to provide a characteristic independent of the rotation required for image processing. In addition, the number of image data obtained is always There are constraints that must be

The following conditions are pointed out as the conditions that the biological or behavioral characteristics of human beings targeted by biometrics should have.

(1) Universality: Must be a characteristic of every individual.

(2) Uniqueness: Everyone should have different characteristics.

(3) Permanence: The characteristic must not change over time.

(4) Collectability: The characteristic should be measurable quantitatively.

However, it is difficult to obtain a biometric characteristic that satisfies all of the above conditions in a real-world biometric system. Therefore, the following conditions are considered.

(1) performance: the accuracy of personal identification aimed at the biometric system,

Robustness to the external environment of speed and perception

(2) Acceptability: The biocharacteristic of the individual who is the subject of biometrics

When you want to acquire, you should be able to accept without great objection.

(3) Safety: There is a risk of theft by unauthorized means

Should not be

Biological characteristics of human beings that are subject to bio-recognition having the above characteristics have been studied for fingerprints, faces, and irises. The fingerprints of each individual are raised by the sweat glands to form a constant flow, and their shapes are different from each other and have been used in real life since it was proved that they do not change as they are at birth. In modern times, due to the development of related technologies, the use of the system is becoming more common, and it is being widely used as an inexpensive tool for personal identification. However, since the reference coordinate axis does not exist in the fingerprint, it is difficult to process the fingerprint collected by rotating randomly. Also, the fingerprint has the flexibility as part of the body, and its shape looks different every time it is collected. Fingerprints are frequently occurring. Moreover, there is a problem of acceptability that basically gives a rejection due to the occurrence of contact with the acquisition equipment at the time of fingerprint acquisition.

On the other hand, there has been a study on the biometrics according to the face shape, which is one of the most natural life features of humans. In the problem of face recognition, despite many applications and various studies, satisfactory reliability is not guaranteed. Because in the case of a still image, it is difficult to separate the face from the image. In the case of using a moving image, time information is used, so that the face can be separated and there is no problem in using the still image. Follow. In addition, there are still many situations to be improved due to the problem of feature extraction due to the deformation of the face due to aging, the change in the length of the hair, the expression change, and the influence on the ambient light.

On the other hand, there is another study on the iris pattern as another biological feature of humans, where the anterior part of the ciliary body is elongated and is located in front of the uveal tract. It is located in front of the eye, between the cornea and the lens, and the round hole in the middle of it is the pupil. The front of the iris has an irregular relief, and there is a circular pattern that is raised near the pupil. This is called a crimp wheel, which, like a fingerprint, does not change forever once it is born. Known. Since the iris pattern can be classified into many types of patterns compared to the types of fingerprints, the iris pattern is installed in an unattended ATM or PC of a bank, mainly in the United States, and is being applied to applications such as Internet user verification and authentication.

However, the personal identification algorithm based on the iris pattern requires the following three invariants required for image processing. First, it must be immutable to change of position. This means that when the iris image is acquired, the same output value should be produced regardless of the iris pattern on the two-dimensional image plane. Second, it must be constant in size change. This means that the same output value should be produced even when the size of the image changes due to the distance between the camera and the iris. Third, it should be possible to produce the same output value even when the rotation of the image occurs due to the inclination of the eye during shooting. Finally, in addition to the above three, the recognition rate should not be reduced even when the whole image is brightened or darkened by the change of lighting value at the time of shooting. Therefore, to provide an iris pattern recognition algorithm that satisfies the above conditions in the present invention.

1 is an explanatory view for extracting one-dimensional brightness data from an imaginary circle 30% larger than a pupil radius from an extracted pupil;

FIG. 2 is a graph showing brightness data of the iris pattern obtained from FIG.

3 is a graph showing a DFT coefficient for the brightness data of FIG.

FIG. 4 is a graph of brightness data at different starting points in the iris pattern shown in FIG. 2. FIG.

FIG. 5 is a graph showing the DFT coefficients for the brightness data of FIG.

FIG. 6 is a graph of brightness data when the overall brightness is increased in the iris pattern shown in FIG.

FIG. 7 is a graph showing the DFT coefficients for the brightness data of FIG.

Problems may occur when acquiring an iris image, such as when the eyelash covers the pupil or a part that affects the iris image due to the eyelid, etc., and the image preprocessing process is performed to compensate for these problems. The median filter, which has the characteristics of maintaining the edge during the preprocessing, is used for noise reduction, and component labling, which is an eight-neighbor pixel, is used to create the pupil boundary.

In addition, only the pupil of the iris image is extracted using an automatic edge detection method. The detection of the pupil is performed to find the center point of the pupil, and the focus of the pupil can be found by using the point where the histogram on the x-axis and the center value of the histogram on the y-axis of the extracted pupil meet. From the focal point the radius of the pupil, r, can be found, from which the circular data of 30% of the radius length out of the pupil can be found. In this way the radius of the circle to acquire the data is R = r (1 + 0.3) is the number of total data is the R × cos45 ^o × 8. 1 shows a situation in which one-dimensional brightness data is obtained from the iris through this process. The method of extracting the feature points provided can obtain the feature points invariant to the above-mentioned position change. Because the position is changed on the image plane, it is irrelevant to the position change because it is referred to the center of the pupil. It is also irrelevant to the size conversion because feature points are extracted from an imaginary circle of a size relative to the size of the pupil.

Discrete Fourier Transform (DFT) is performed on the one-dimensional iris pattern brightness data obtained as described above. The DFT is applied to convert a signal in the time domain into a representation in the frequency domain, and the coefficients of the DFT for the N finite sequences are N values obtained from the following equation.

2 is a graph showing the 640 brightness data for the actual iris pattern obtained through the above process. 3 shows the DFT coefficients.

FIG. 4 is a brightness data value when the same brightness data shown in FIG. 2 is obtained by only changing the starting point, that is, when the iris image is rotated and acquired. Despite the same iris pattern, it can be seen that the shape is much different. Therefore, it can be seen that an invariant feature point is required for this rotation. Since the brightness data of the iris is acquired along the circle, the brightness data acquired along the circle becomes a periodic function having a period equal to the number of pixels around the circle. therefore About data stream of Is It is shifted by, i.e., rotated image data. The DFT coefficient for this is expressed by the following equation.

As can be seen from Equations 1 and 2 above, the following relationship is established.

Therefore, when using the magnitude of the coefficient of the DFT, it can be seen that a feature point independent of rotation can be obtained.

Meanwhile, the change in lighting About data stream of Can be displayed in the form of. Constant This causes the overall brightness to be dark or bright. This can be obtained by the following equation.

As can be seen from the above equation, only the first term of the DFT coefficient It can be seen that this is affected.

According to the proposed method, the magnitude of the DFT coefficient with respect to the brightness data of the iris shown in [FIG. 4] is [FIG. 5]. When rotation occurs for the same iris pattern, it can be seen that the presented feature points have irrelevant characteristics. FIG. 6 shows that the brightness value of the brightness data of FIG. 2 changes due to a change in illumination. The magnitude value of the DFT coefficient for this is as shown in FIG. Except for the value of, all are the same.

Claims (1)

In recognizing the iris pattern from the image of the eye, (1) finding a pupil and focusing on its focus; (2) acquiring one-dimensional brightness data from a virtual circle with a larger radius of pupil from this center; (3) A method of obtaining feature points independent of rotation on the basis of the magnitude of the DFT coefficients for the obtained one-dimensional brightness data.