KR20020023011A - Human iris recognition method using harr wavelet transform and lvq - Google Patents

Abstract

PURPOSE: An iris recognition method using Harr wavelet conversion and LVQ(Learning Vector Quantization) is provided to effectively represent and process an iris pattern by reduce memory space usage through a multi division of image. CONSTITUTION: The method comprises steps of executing a pre-treatment for extracting only iris region from the provided input image(S10), multiply dividing the image signal by repeatedly applying the Harr wavelet conversion to the pre-treated iris image and extracting the region including a high frequency component in order to extract a characteristic(S30), extracting the values of information constructing a characteristic vector for the extracted high frequency region and constructing the characteristic vector by quantizing the corresponding value, and executing learning and recognizing an iris characteristic vector by using an initial weight setting method of uniform dispersion and dimensional winner selection method(S60).

Description

HUMAN IRIS RECOGNITION METHOD USING HARR WAVELET TRANSFORM AND LVQ}

The present invention relates to an iris recognition method, and more particularly, extracts a feature pattern using a Harle wavelet transform, and classifies the pattern using a learning vector quantization (LVQ) algorithm. The present invention relates to an iris recognition method that enables iris recognition.

In general, passwords used in ATMs and various devices using computer systems are the most common identification method for identifying a user's identity. However, in spite of such convenience of use, there is a inconvenience that the user's attention is always required because the password includes the risk of being forgotten by the user or being misused by others. Therefore, there is a need for a method of eliminating these risks and identifying a person's identity more stably and accurately, and one of them is a method of using biological characteristics that can distinguish each person from others. Such biological characteristics include genetic traits, fingerprints, voices, veins, facial appearances, and eye irises. A system that automatically recognizes an individual's identity using the individual's biological uniqueness has high reliability of biological characteristics. Therefore, a lot of research has been conducted for a long time, and the development of the fingerprint recognition and iris recognition is outstanding.

In the case of the iris recognition system, the iris is an area between the pupil of the eye and the white part of the eye. The iris is a long band-like net (comb ligament), a red fiber like cora, Eyebrow bumps, sinusoidal relationships, ring-shaped circles, corona-shaped ligaments surrounding the pupil, iris-specific colors, and spots have different biological characteristics. Since the iris of this characteristic has its own shape within two to three years of life, and persists permanently without being affected by changes in the external environment, it is of great interest for important biological features to identify an individual. It is becoming. In other words, iris recognition is a technique of analyzing individual iris patterns for each individual and recognizing an individual's identity based on the information.

In the case of the existing iris recognition system, the method of extracting features and constructing vectors using Gabor transformation has been mainstream. Since the feature vectors generated by these methods are composed of 256 or more dimensions and occupy at least one byte per dimension, they occupy at least 256 bytes. Accordingly, there is a need for a feature configuration method capable of efficiently processing, storing, transmitting, and retrieving pattern information. Conventional methods use a simple distance measurement method such as the Hamming distance between two feature vectors (the feature vector for the input pattern and the stored reference feature vector) for pattern classification. It is not easy to construct and does not properly reflect the characteristics of the information of each dimension of the feature vector.

In addition, due to the generalization of data through learning, LVQ, which is frequently used for pattern classification, is faster than the error back-propagation learning algorithm, but has a disadvantage in that it is sensitive to an initial weight vector. Furthermore, since Euclidean distance is generally used as the winner selection method of LVQ, when the dimension of the feature vector is large, the information on each dimension is lost, so that the wrong winner can be selected, and the calculation amount increases and the processing is delayed. It is also possible.

Accordingly, an object of the present invention is to solve such a conventional problem, and an object thereof is to provide an iris that can efficiently express and process an iris pattern using a minimum information space through multiple division of an image signal using wavelet transform. It is to provide a recognition method.

In addition, another object of the present invention is to generate an initial weight only weight vector located closer to the boundary of each class in order to effectively select the initial weight in the LVQ recognizer to minimize the impact of the recognition performance by the initial weight and improve the processing performance It is to provide an iris recognition method that can.

Another object of the present invention is to effectively select the winner by comparing the information of each dimension in the process of comparing the feature vector of the input pattern and the feature vector learned to produce the final output result in the LVQ recognizer It is to provide an iris recognition method.

1 is a flowchart schematically showing the entire process of the iris recognition method according to the present invention.

2 is a detailed flowchart of the iris image preprocessing process according to the present invention.

Figure 3 shows the conversion to the polar coordinate system and the result according to the present invention.

4 is a flowchart showing a video wavelet-based image splitting method according to the present invention.

5 shows a feature extraction process through multi-division using a wavelet for implementing the present invention.

FIG. 6 (a) shows the coincidence rate with respect to the feature vector of the reference data and the comparison data through four multiplexing.

FIG. 6 (b) shows the matching rate with respect to the feature vectors of the reference data and the comparison data through five times of multi-division.

7 is a flowchart illustrating a method of constructing an iris feature vector through multiple division of a wavelet transform.

8 is a flowchart illustrating a uniform distribution weight initialization method for LVQ according to the present invention.

9 is a flowchart of a method of selecting a winner by dimension for LVQ-based recognition according to the present invention.

10 is a conceptual diagram illustrating a winner selection method for each dimension according to the present invention.

In order to achieve the above object, an iris recognition method using a wavelet transform and LVQ according to the present invention obtains an iris image, generates user registration data from the obtained iris image, and generates the registered data and authenticates the user. In the iris recognition method to perform identification recognition for the user by comparing the obtained iris image,

(a) performing preprocessing to extract only the iris region with respect to the provided input image;

(b) applying the Hare wavelet transform repeatedly a predetermined number of times to the iris image that has been pre-processed in the step (a) to divide the image signal into multiples, and extract a region containing only high frequency components for feature extraction; Wow,

(c) extracting values of information constituting the feature vector with respect to the region of the extracted high frequency component, and quantizing the corresponding values to form a feature vector;

(d) learning and recognizing the iris feature vector using a uniform distribution initial weight setting method and a dimension winner selection method.

Hereinafter, with reference to the accompanying drawings, the configuration and operation effects of the iris recognition method according to a preferred embodiment of the present invention will be described in detail.

1 is a flowchart schematically showing the entire process of the iris recognition method according to the present invention. As shown in Figure 1, in the iris recognition method according to the present invention is first subjected to a pre-processing process to extract only the iris region for the input image provided in step (s10), which is to find the center of the pupil, inside / outside the iris It consists of detailed processes such as boundary finding and conversion of iris region to polar coordinate system.

On the other hand, in the next step (s30) for the iris image that has undergone the pre-processing in step (s10), the Harah wavelet transform is repeatedly applied while satisfying a predetermined number of times or a predetermined criterion to divide the image signal into multiples. In this process, a region including only high frequency components necessary for feature extraction is separately extracted. In addition, a value of information constituting the feature vector is extracted for the region of the divided high frequency component, and the values are quantized to form the feature vector.

Finally, in the step S60, the LVQ-based iris feature vector classification process uses the uniform distribution initial weighting method and the winner selection method for each dimension to perform effective learning and recognition.

Hereinafter, a detailed method of each process will be described with reference to the accompanying drawings.

First, a black and white CCD camera is used to acquire an image for preprocessing. Since iris recognition is performed by analyzing pattern characteristics of an iris wrinkle, no additional information such as color information of an iris is used. On the other hand, when the iris image is acquired with ordinary lighting indoors, it is difficult to extract desired pattern information because the iris image is generally dark. Therefore, a separate ambient light is used. In this case, the illumination used in this case is a reflection of the light, which prevents the loss of information on the iris pattern and the deterioration of recognition performance, and also provides an appropriate illumination to obtain a clear iris pattern. Should be used. In this case, considerations include illumination of the illumination, the number of illuminations used, and the distance between the iris and the illumination.

Meanwhile, when the iris image is acquired in this way, a preprocessing process for iris recognition as shown in FIG. 2 is performed to process the obtained image. As shown in FIG. 2, first, in step S20, filtering is performed to remove noise for smooth image processing. In step s22, the center coordinates of the pupil are found, which first places a point T (x0, y0) in the pupil. By setting the initial value (120,160), which is the center of the image, the image information at the point T and the surrounding image information are compared to determine whether the point T is in the pupil. If the pupil is not located in the center of the image, while moving the point T up, down, left and right, the image information is continuously compared to set the point T located in the pupil. Once the point T is determined, the distance to the outside of the pupil is measured while increasing and decreasing the coordinate values of T in the x and y directions, respectively. In this case, if each end point is X1, X4, Y1, Y4, in step 24, X3, which is the value of the center coordinate of the x-axis of the line segments X1, X4, and Y2, which is the y-axis center coordinate of the line segments Y1, Y4, can be obtained. By the x3, y2, the center coordinates (x1, y1) of the actual pupil can be obtained.

When the center coordinate of the actual pupil is determined in step S24, in step S26, the inner boundary where the pupil and the iris contact each other and the outer boundary where the black and white parts of the eye (the outer part of the pupil) contact each other are obtained. Only the iris region can be extracted. At this time, the internal boundary can be easily obtained by using the points (X1, X4, Y1, Y4) obtained in the process of determining the center of the pupil. On the other hand, in the case of an external boundary, the coordinate of x is increased until the brightness information I (x, y) of the image of the corresponding coordinate starts from the outside portion of the internal boundary. When the brightness information of the image larger than the threshold is obtained, the distance from the x coordinate of the center of the pupil to the corresponding x coordinate becomes the radius of the external boundary.

When the iris region is judged by the separation of the inner and outer boundaries, the next step (s28) shows the distance r and the angle θ from the center of the pupil rather than the x and y coordinate system in order to express and process the characteristics of the extracted iris region. Perform the conversion to the polar coordinate system represented. Since the iris pattern is formed in a circular shape around the pupil, the iris pattern is spread out on the square plane of (r, θ) so that the change of the iris pattern according to the r and θ values can be easily obtained through image signal analysis. This is because it can be extracted.

That is, the iris region existing between two inner and outer boundaries is converted into a polar coordinate system represented by the distance r from the center of the pupil and the center angle θ using Equation 1 below.

I (x (r, θ), y (r, θ)) → I (r, θ)

x (r, θ) = (1-r) x _P (θ) + rx _S (θ)

y (r, θ) = (1-r) y _P (θ) + ry _S (θ)

(At this time, (x _P (θ), y _P (θ)): coordinate of the point rotated in the + direction by θ from the horizontal axis of the pupil, (x _S (θ), y _S (θ)): from the horizontal axis of the pupil The coordinates of the point that advances by the range of the specified feature).

3 shows an image obtained as a result of a process of converting a rectangular coordinate system to a polar coordinate system with respect to the iris image. At this time, the normalization was performed by dividing the internal / external boundary into 60 intervals so that the iris characteristics did not change due to the change in pupil size, and θ was changed by 0.8 to represent 450 information. Therefore, an iris image of θ * r = 450 * 60 is finally obtained.

4 is a flowchart illustrating an image multi-segmentation method based on wavelet transform for the iris recognition method according to the present invention, and FIG. 5 is a feature extraction process through multi-division using a wavelet for implementing the present invention. It shows a diagram for explaining.

As shown in FIG. 4 and FIG. 5, the present invention uses wavelet transform for feature extraction of an iris image, and in particular, uses a wave wavelet among various mother wavelets. First, in step S40, an iris image is divided by applying a wavelet transform. In other words, the iris image can be thought of as a two-dimensional signal in which one-dimensional signals are arranged on the x-axis and y-axis, so to analyze them, the LPF and HPF are passed through the x- and y-directions, respectively. Extract the ingredients. In this case, the signal is divided into a difference component D (difference) of a signal that passes through the high-band filter through a low wavelet transform, and an average (A) average of a signal that passes through the low-band filter. In FIG. 5 (a), when the low frequency component is represented by L and the high frequency component is represented by H, the LL portion is a component that passes LPF in both x and y directions, and HH is HPF. It means that the component passed. That is, the wavelet transform is performed on the two-dimensional video signal in the horizontal and vertical directions, respectively, so that the single wavelet is divided into four regions LL, LH, HL, and HH.

In this case, in step S40, the HH region having only high frequency components is extracted from the divided regions, and in step S44, the discrimination ratio of the pattern based on the feature value of the HH region is calculated to determine whether the discrimination ratio is equal to or less than a predetermined threshold. The process is terminated, or the process is terminated even if the number of times of repetition is more than the specified number of times.

On the other hand, if the determination result of the pattern is more than a predetermined threshold value or the repetition number is more than the specified number of times, the information of the HH area is stored as information for feature extraction in step S48.

Further, in step s50, the LL region, which is a low frequency component, is extracted for both the x-axis and the y-axis, and in the next step s52, the extracted LL portion (the image whose size is reduced to 1/4 of the previous image) is extracted. Since the main information of the image is included, it is provided as a new processing target image so that the wavelet can be applied to the portion again.

One wavelet transform is performed on the iris image to obtain the HH region of 225 * 30 whose size is reduced by half. That is, 225 * 30 pieces of information are used as the feature vector. This information can be used as it is, but the signal division process is repeated again to reduce the size of the information. Since the main information of the image is included in the LL part, when the wavelet is applied to the corresponding part again, as shown in FIG. 5 (b), the feature values having reduced sizes, such as HH 2, HH 3 and HH 4, can be obtained.

On the other hand, the number of repetitions provided as a criterion for repetition of the wavelet in step S46 should be set to an appropriate value in consideration of information loss and the size of the feature vector. That is, in the case of iris recognition, a boundary that can clearly distinguish these feature vectors in the comparison process between reference data stored as a representative feature vector of each user and feature vectors of a given user's iris image to determine an appropriate number of repetitions. To see if it exists. That is, the degree of agreement between the two vectors is calculated through one-to-one comparison between the reference data and the given feature vector, and the feature vector is extracted by repeating the multiplexing of the signal to the extent that each user can be distinguished according to the result. FIG. 6 shows the coincidence rate of the reference data and the feature vector of the comparative data for 10 persons. FIG. 6 (a) shows four cases of wavelet transform and FIG. 6 (b) shows five cases. In this case, in FIG. 6, the black point is a reference data of the user and the white point is a result of comparison with reference data of another user. In order to recognize / verify, there must be a boundary that distinguishes one's own data from another's data. In the case of FIG. 6 (a), the two data can be distinguished using a coincidence rate of about 65%, while in the case of FIG. 6 (b), it is impossible to distinguish the two data. That is, if the wavelet is performed five times or more, the number of data constituting the feature vector is reduced, so that the characteristic of the iris image cannot be properly represented, and verification using the wavelet is impossible.

Therefore, in the present invention, the HH4 region of FIG. 5 (b) obtained by performing the wavelet four times is used as the main feature region, and the values are taken as components of the feature vector. At this time, the HH4 region contains 28 * 3 = 84 pieces of information.

7 is a flowchart illustrating a method of constructing an iris feature vector through multiple division of a wavelet transform. As shown in FIG. 7, in operation S70, information on the N feature vectors extracted in the process, that is, HH1, HH2, HH3 and HH4 is received. In step s72, an average value of each region is obtained to allocate feature information for each region of HH1, HH2, and HH3 except for the information of HH4 obtained through the last wavelet transform among the N feature vectors, and then one dimension is allocated. In step S74, the respective values M present in the HH4 region are extracted as the respective features. Meanwhile, in step s76, M + N_1, that is, 84 + 3 = 87-dimensional feature vectors are generated by combining the three-dimensional values obtained by the average values of HH1, HH2, and HH3 with the feature vectors of the HH4 region. .

In step s78, all values of the previously obtained vector, that is, each of the 87-dimensional values represented by the real value are quantized to a binary value of 0 or 1, and a total of 87 bits are quantized with the quantized values in step s80. Construct a feature vector.

8 is a flowchart showing an example of the uniform distribution weight initialization method according to the present invention. As shown in FIG. 8, in the uniform distribution weight initialization method for LVQ recognition, firstly, the training data input for the first time among the training data of each class in step S90 is selected as the first weight vector of the corresponding class, and the corresponding class. All remaining weight vectors of are set to zero. This is represented by Equation 2 below.

For k = 1,2, ..., M

( the first input learning pattern of the kth class, : first weight of kth class, M: number of classes).

In step s92, different learning data is selected from each class and provided as a new input. In step s94, the feature vector of the training data provided as the new input data and each weight vector stored in each class are already learned. The Euclidean distance (d: Euclidean distance) is calculated for the similarity of.

In step S96, it is determined whether the class of the weight vector having the minimum distance d among the distances calculated for each weight vector matches the class of the training data. If it is determined that the two classes do not match, in step S98, the feature vector of the current learning data is registered as a new weight vector of the corresponding class. Then, the process proceeds to the next step (s100) to determine whether the learning of all the learning data is completed, and if it is not completed, repeats the above step (s92) to set the weight vector of each class.

9 is a flowchart illustrating an example of a method of selecting a winner by dimension for LVQ based recognition according to the present invention, and FIG. 10 is a conceptual diagram illustrating a process of selecting a winner by dimension according to the present invention. As shown in FIG. 9, the present invention applies an algorithm for selecting a winner for each dimension instead of the conventional winner determination method using Euclidean distance when learning LVQ. That is, the minimum class is selected as the winner by comparing the distance between the first dimension of the input pattern and the first dimension of each output neuron.

That is, in step S102, the similarity with the corresponding dimension of each output neuron is measured for the dimension i (i = 1,2, ..., M + N-1) of the feature vector of the input image, for example, in the 87th dimension. . In step S104, the class having the largest similarity value among the measured similarities is selected as the winner for the corresponding dimension i.

Further, in step s106, it is determined whether the measurement of the similarity for all the dimensions has been completed, and the above-mentioned step s102 or less is repeated until the measurement of the similarity for all the dimensions is completed. In addition, when the measurement of the similarity for all dimensions is completed, in step s108, the class selected as the winner for each dimension is selected as the final winner.

As shown in FIG. 10, the left side represents an arbitrary input pattern, and the right side represents each class of output neurons, and represents a dimension in which each block on the pattern forms a pattern. At this time, the similarity is measured by comparing each class of each input pattern and output neuron. Therefore, the winner determination method for each winner can reduce the processing time because each dimension can reflect the information in the winner determination process and the winner can be determined by using a smaller number of operations than the comparison operation for the total dimension. There is an advantage.

Iris recognition method according to the present invention is not limited to the above-described embodiment can be carried out in various modifications within the scope allowed by the spirit of the present invention.

As described above, in the iris recognition method using the wavelet transform and the LVQ according to the present invention, the personal identification system using the iris pattern information can be efficiently and effectively extracted and recognized in terms of processing performance or processing time. do. Through feature extraction and feature vector construction according to the present invention, since the iris pattern information can be represented with a high information expression capability while using only a small amount of space, the processing, storage, retrieval and exchange of information can be performed very effectively. Provide a technical foundation. In addition, through the LVQ-based uniform distribution initial weight setting method, the initial weight is effectively expressed, and thus, the iris pattern information can be processed with stable and high reliability.

Furthermore, the winner selection method for each dimension allows the information of each dimension to be reflected in the winner selection process even for high-dimensional feature vectors, and processing time is reduced because fewer winners can be determined than fewer comparison operations for all dimensions. A shortened effect can be obtained.

Claims (8)

In the iris recognition method for acquiring an iris image, generating the registration data of the user from the obtained iris image, and performing identification identification for the user by comparing the generated registration data and the iris image obtained at the time of authentication of the user, (a) performing preprocessing to extract only the iris region with respect to the provided input image; (b) applying the Hare wavelet transform repeatedly a predetermined number of times to the iris image that has been pre-processed in the step (a) to divide the image signal into multiples, and extract a region containing only high frequency components for feature extraction; Wow, (c) extracting values of information constituting the feature vector with respect to the region of the extracted high frequency component, and quantizing the corresponding values to form a feature vector; (d) learning and recognizing the iris feature vector using a uniform distribution initial weight setting method and a winner selection method for each dimension, comprising a wavelet transform and an iris recognition method using an elb queue. . The method of claim 1, wherein step (a) Setting arbitrary coordinates in the pupil of the obtained iris image, and comparing the image information of the arbitrary coordinates with the image information of the peripheral coordinates to find the center of the pupil; When the center of the pupil is determined, extracting the iris region by distinguishing the inner boundary between the pupil and the iris and the outer boundary of the black and white parts of the eye, And converting the iris region existing between the inner and outer boundaries into a polar coordinate system represented by a distance from the center of the pupil and a center angle. The method of claim 1, wherein step (b) Splitting the image into multiples by applying a low wavelet transform to the provided iris image; Extracting a region (HH) including only high frequency components from the divided images on the x and y axes, Calculating a discrimination rate of the pattern based on the ranging value of the extracted area HH, and storing information of the extracted area when the discrimination rate is equal to or greater than a predetermined threshold value; When the storing of the information on the high frequency region is finished, the region LL including only low frequency components is extracted from the divided regions on the x-axis and y-axis, and the multi-division of the image is repeatedly performed on the LL region. Iris recognition method using a wavelet transform and elb queue, characterized in that it comprises a step. 4. The method of claim 3, wherein step (b) Determining whether the number of repetitions of the multi-segmentation is greater than or equal to a predetermined number of times when the discrimination rate by the measured value of the extracted area HH is equal to or greater than a predetermined threshold value; If the repetition number is less than the specified number of times as a result of the determination, the multi-division of the low frequency region is performed, and if the repetition number is greater than or equal to the specified number, the multi-division is terminated. Iris recognition method. 5. The iris recognition method according to claim 4, wherein the predetermined number of times is set to four times. The method of claim 1 or 3, wherein step (c) Receiving divided images of a plurality of high frequency regions (HHi) formed by the multi-division in step (b); Obtaining an average value of information of all remaining HH regions except for the last HH region (HHn) obtained from the input high frequency region, and allocating the average values as elements of a feature vector; Allocating each value present in the last HH region (HHn) obtained from the input high frequency region as a component (m) of each feature vector; Combining the values of the multi-divided plurality of high frequency regions (HHi) to form one feature vector; And quantizing each element value of the feature vector to a binary value to generate a feature vector having a predetermined dimension. The method of claim 1, wherein the weight setting method of step (d) Selecting the first learning pattern among the learning patterns of each class as the first weight vector of the corresponding class, and setting the remaining weight vectors to 0, Selecting different learning data in each class and providing it as a new input, measuring the similarity between the feature vector of the input learning data and each weight vector selected in the above step and stored in each class; Determining whether the class of the weight vector having the largest similarity as the measurement result matches the class of training data; And when the two classes do not match, setting the feature vector of the current training data as a new weight vector to the corresponding class, and repeating the above steps until the learning of all the training data is completed. Iris recognition method using ha wavelet transform and elb queue. The method of claim 1, wherein the winner selection method of step (d) Measuring similarity with the corresponding dimension of each output neuron with respect to the dimension i of the feature vector of the input image, Selecting the class having the largest similarity value among the measured similarities as a winner for the corresponding dimension i, and repeating the similarity measurement for all dimensions, When the similarity measurement for all dimensions is completed, selecting the most winner class for each dimension as the final winner comprising the wavelet transform and iris recognition method using elb queue.