KR102449839B1 - Authentication method based on the ecg, authentication apparatus, learning method for ecg based authentication and learning apparatus - Google Patents

Abstract

Authentication and learning based on the electrocardiogram are initiated. The authentication method according to an embodiment includes the steps of generating a feature vector from an electrocardiogram obtained from an object using a dictionary, classifying the electrocardiogram through a classifier based on the feature vector, and performing authentication based on the classification result. .

Description

AUTHENTICATION METHOD BASED ON THE ECG, AUTHENTICATION APPARATUS, LEARNING METHOD FOR ECG BASED AUTHENTICATION AND LEARNING APPARATUS

Disclosed with respect to signal processing technology, and specifically disclosed with respect to the technical field of pattern identification. More specifically, it discloses authentication based on an electrocardiogram.

With the development of science and technology, methods of authenticating identity through biometric characteristics are diversifying. In authentication based on biometric characteristics, the identity of an object is verified using unique biological characteristics or behavioral characteristics (eg, iris, fingerprint, human face, voice, and gait, etc.) possessed by each person. Compared to traditional authentication, biometric feature-based authentication has the advantage of high authentication accuracy, but there are aspects that require some improvement. For example, in the case of some biometric features such as a fingerprint, a human face, or a voice, the security of the authentication system may be weakened by using the forged finger, a human face, or a pre-recorded voice.

According to an embodiment, an electrocardiogram-based authentication method includes: receiving an electrocardiogram; generating a feature vector for the electrocardiogram based on a pre-learned dictionary; generating a classification result for the electrocardiogram based on the feature vector and a pre-learned classifier; and performing authentication of the electrocardiogram based on the classification result.

In the electrocardiogram-based authentication method according to an embodiment, the dictionary and the classifier are learned synchronously with each other.

In the electrocardiogram-based authentication method according to an embodiment, the dictionary includes category dictionaries corresponding to a plurality of categories, and the classifier includes category classifiers corresponding to the categories.

In the electrocardiogram-based authentication method according to an embodiment, the dictionary and the classifier generate a plurality of feature vectors for training electrocardiograms using the category dictionaries, and perform the training based on the plurality of feature vectors. Prediction categories of the electrocardiograms are determined, and the category dictionaries and the category classifiers are updated alternately based on the categories and the prediction categories, so that they can be learned synchronously with each other.

In the electrocardiogram-based authentication method according to an embodiment, the generating of the feature vector for the electrocardiogram comprises: expressing the electrocardiogram as a linear combination of a plurality of category dictionaries included in the dictionary. determining coefficients; and generating the feature vector based on the coefficients.

In the electrocardiogram-based authentication method according to an embodiment, the generating of the feature vector includes: dividing the electrocardiogram into at least one signal segment; generating at least one feature representation corresponding to the at least one signal segment; generating at least one second characteristic indication by performing group sparse coding based on the dictionary on the at least one characteristic indication; and generating the feature vector based on the at least one second feature indication.

In the electrocardiogram-based authentication method according to an embodiment, the characteristic indication is expressed as a linear combination of some of the plurality of category dictionaries included in the dictionary, and the second characteristic indication includes coefficients of the partial category dictionaries. include

In the electrocardiogram-based authentication method according to an embodiment, the generating of the classification result for the electrocardiogram may include a classification result for at least one signal segment of the electrocardiogram based on the feature vector and the classifier. generating at least one label; and generating a classification result for the electrocardiogram based on the at least one label.

In an electrocardiogram-based authentication method according to an embodiment, the performing of the authentication on the electrocardiogram includes: acquiring a classification result for a reference electrocardiogram; and performing the authentication based on a voting mechanism for the classification result for the electrocardiogram and the classification result for the reference electrocardiogram.

In the electrocardiogram-based authentication method according to an embodiment, the reference electrocardiogram includes at least one reference signal segment, and performing the authentication based on the voting mechanism includes: at least one label for the electrocardiogram based on the method, generating a first distribution comprising a number of signal segments belonging to each of the plurality of categories; obtaining a second distribution including the number of reference signal segments belonging to each of the plurality of categories; and calculating a degree of similarity between the first distribution and the second distribution. and performing the authentication by comparing the similarity and a predefined threshold value.

In the ECG-based authentication method according to an embodiment, the classification result for the reference ECG may be obtained by applying the dictionary and the classifier to the reference ECG.

A learning method for electrocardiogram-based authentication according to an embodiment includes generating a plurality of feature vectors for training ECGs using category dictionaries corresponding to a plurality of categories; determining prediction categories of the training electrocardiograms based on the feature vectors; and alternately updating the category dictionaries and category classifiers based on the categories and the prediction categories.

In the learning method for electrocardiogram-based authentication according to an embodiment, the generating of the feature vectors includes: dividing each training electrocardiogram of the training electrocardiograms into at least one signal segment; generating at least one feature representation corresponding to the at least one signal segment; generating at least one second feature mark by performing group sparse coding based on the category dictionaries on the at least one feature mark; and generating the feature vectors based on the at least one second feature indication corresponding to the at least one signal segment.

In the learning method for electrocardiogram-based authentication according to an embodiment, the step of alternately updating the category dictionaries and the category classifiers includes fixing the category classifiers and minimizing errors between the prediction categories and the categories. optimizing the category dictionaries; and fixing the optimized category dictionaries, and optimizing the category classifiers to minimize the prediction categories and errors of the categories, wherein the optimization of the category dictionaries and the optimization of the category classifiers are repeated synchronously with each other. can be

In the learning method for electrocardiogram-based authentication according to an embodiment, the optimizing the category dictionaries includes optimizing the category dictionaries to minimize the error using a gradient descent algorithm, and Optimizing includes optimizing the category classifiers to minimize the error using the gradient descent algorithm.

A learning method for electrocardiogram-based authentication according to an embodiment may include terminating the update when the optimizations repeated synchronously with each other satisfy a predefined condition; and outputting the updated category dictionaries and category classifiers, wherein the predefined condition is that when the order of the iteration reaches a predefined order, the error is a predefined first value. It includes at least one of a smaller case and a case where the amount of change of the error is smaller than the second predefined value.

The learning method for ECG-based authentication according to an embodiment further includes initializing the category dictionaries based on the training ECGs corresponding to the categories.

In the learning method for ECG-based authentication according to an embodiment, the step of initializing the category dictionaries includes: preprocessing the training ECGs; generating a plurality of clusters by clustering the pre-processed training ECGs according to the categories; and generating the initialized category dictionaries by performing normalization on each centroid of the clusters.

In the learning method for electrocardiogram-based authentication according to an embodiment, the pre-processing includes: dividing each training electrocardiogram of the training electrocardiograms into at least one signal segment; extracting each feature of the at least one signal segment; and performing dimensionality reduction processing on each of the features to generate each feature representation for the at least one signal segment.

In the learning method for electrocardiogram-based authentication according to an embodiment, the dividing into at least one signal segment comprises: filtering each training electrocardiogram; detecting QRS waves for each of the filtered training electrocardiograms; and generating the at least one signal segment based on the detection result.

The learning method for electrocardiogram-based authentication according to an embodiment further includes initializing the category classifiers based on the feature vectors and the category dictionaries.

In the learning method for electrocardiogram-based authentication according to an embodiment, the step of initializing the category classifiers comprises: calculating a probability distribution of the categories for the at least one signal segment using the feature vectors and the category dictionaries. generating; calculating a prediction error of the at least one signal segment based on the probability distribution and the actual category of the at least one signal segment; obtaining, based on the prediction error, the prediction categories and category classifiers that minimize the errors of the categories; and generating the initialized category classifiers as the obtained category classifiers.

A learning method for electrocardiogram-based authentication according to an embodiment includes generating a dictionary by merging the updated category dictionaries; and generating a classifier by merging the updated category classifiers.

A program for executing the ECG-based authentication method and the ECG-based authentication learning method according to an embodiment may be recorded in a computer-readable recording medium.

An electrocardiogram-based authentication apparatus according to an embodiment generates a feature vector for the electrocardiogram based on a pre-learned dictionary, and generates a classification result for the electrocardiogram based on the feature vector and the pre-learned classifier, and a processor for authenticating the electrocardiogram based on the classification result.

A learning apparatus for electrocardiogram-based authentication according to an embodiment generates a plurality of feature vectors for training electrocardiograms using category dictionaries corresponding to a plurality of categories, and based on the feature vectors, the training electrocardiogram and a processor that determines prediction categories of , and alternately updates the category dictionaries and category classifiers based on the categories and the prediction categories.

1 is a flowchart illustrating an electrocardiogram-based authentication method according to an exemplary embodiment.
2 is an example of a waveform of an electrocardiogram.
3 is a flowchart illustrating a learning method for ECG-based authentication according to an embodiment.
4 is a flowchart illustrating an embodiment of alternately updating category dictionaries and category classifiers.
5 is a flowchart illustrating an embodiment of optimizing category dictionaries and category classifiers.
6 is a flowchart illustrating an embodiment of generating a feature vector corresponding to a test electrocardiogram.
7 is a flowchart illustrating an embodiment of authentication for an electrocardiogram.
8 is a diagram illustrating a configuration of an authentication device and a learning device according to an embodiment.

Specific structural or functional descriptions of the embodiments are disclosed for purposes of illustration only, and may be changed and implemented in various forms. Accordingly, the embodiments are not limited to a specific disclosure form, and the scope of the present specification includes changes, equivalents, or substitutes included in the technical spirit.

Although terms such as first or second may be used to describe various elements, these terms should be interpreted only for the purpose of distinguishing one element from another. For example, a first component may be termed a second component, and similarly, a second component may also be termed a first component.

When a component is referred to as being “connected” to another component, it may be directly connected or connected to the other component, but it should be understood that another component may exist in between.

The singular expression includes the plural expression unless the context clearly dictates otherwise. In this specification, terms such as "comprise" or "have" are intended to designate that the described feature, number, step, operation, component, part, or combination thereof exists, and includes one or more other features or numbers, It should be understood that the possibility of the presence or addition of steps, operations, components, parts or combinations thereof is not precluded in advance.

Unless defined otherwise, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art. Terms such as those defined in a commonly used dictionary should be interpreted as having a meaning consistent with the meaning in the context of the related art, and should not be interpreted in an ideal or excessively formal meaning unless explicitly defined in the present specification. does not

The embodiments may be implemented in various types of products, such as personal computers, laptop computers, tablet computers, smart phones, televisions, smart home appliances, intelligent cars, kiosks, wearable devices, and the like. For example, the embodiments may be applied to recognize a user in a smart phone, a mobile device, a smart home system, and the like. Embodiments may be applied to a payment service through user recognition. Also, the embodiments may be applied to an intelligent vehicle system that automatically starts the engine by recognizing a user. Hereinafter, embodiments will be described in detail with reference to the accompanying drawings. Like reference numerals in each figure indicate like elements.

Authentication based on an electrocardiogram (ECG) performs identification and authentication of an object using an electrocardiogram, which is known to have relatively high stability among authentication based on biometric characteristics. Electrocardiography has received widespread attention in the field of security technology because it is difficult to forge, has a low acquisition path cost, and has a high degree of discrimination.

Since most of the authentication based on ECG is used in the security area where stability is important, a high level of accuracy is required. For this, it is necessary to extract features with high level of discrimination. Electrocardiograms of the same object in different physiological states (eg, maintaining a stationary state and after exercise) show different states. Even with the ECG of the same object, changes in the ECG may occur in different time periods. Electrocardiograms of the same object may indicate different states, thereby causing a difference in categories for classifying electrocardiograms. Therefore, in the authentication based on the electrocardiogram, it is required to learn and authenticate good characteristics with discrimination power for signals collected in different states and different periods of the same object.

Authentication based on an electrocardiogram according to an embodiment may be applied to various types of electronic devices, and such electronic devices may include an electrocardiogram sensor and may be used to collect an electrocardiogram of a living body. The electronic device may include, but is not limited to, a desktop computer, a notebook computer, a smart phone, a tablet PC, a PDA, a smart security device, a door security device, a smart building device, and an automobile electronic device.

If the authentication method based on the electrocardiogram according to an embodiment is applied to the electronic device, the identity can be identified, for example, whether the user has the right to use or whether the user has visited through an appropriate route can be determined based on the user's electrocardiogram. . According to an embodiment, the user's identity may be determined based on the user's electrocardiogram, and information related to the user may be additionally searched for or recorded based on the user's identity. The authentication method based on an electrocardiogram according to an embodiment can be variously applied to identify an identity, for example, it can be used in terms of visit control, financial transaction, airport boarding, patient privacy protection and border security, etc. However, the present invention is not limited thereto.

The reference ECG is a pre-stored ECG. The reference electrocardiogram may be obtained from each user and stored in advance, and may be used as a target for comparison with the test electrocardiogram in order to perform authentication for the user. The test electrocardiogram is an electrocardiogram to be tested for user authentication, and may be an electrocardiogram input from a user.

According to an embodiment, a reference ECG for each user may be obtained through a method of collecting one ECG of each user in advance and corresponding to the user. According to another embodiment, a reference ECG for each user is obtained by collecting a plurality of ECGs of each user in advance, and matching one representative ECG obtained through selection/synthesis processing for the plurality of ECGs to the user. can be obtained.

A test electrocardiogram, which is a signal to be authenticated, is obtained from a user for performing authentication. When authentication is required for the identity of an unknown user, it is necessary to obtain a test electrocardiogram corresponding to the user to be authenticated. According to an embodiment, by acquiring one ECG of a user to be authenticated, a test ECG corresponding to the user may be acquired. According to another embodiment, one representative ECG obtained through selection/synthesis of a plurality of ECGs of a user to be authenticated may be used as a test ECG corresponding to the user. A method of obtaining a test electrocardiogram for a user to be authenticated is not limited to the above-described embodiment, and various types of methods may be employed.

An electrocardiogram (ECG)-based authentication method may include four steps: data preprocessing, valid feature display (or feature extraction), robust authentication mechanism, and rational decision mechanism. The feature used in this authentication method is extracted from a primitive ECG signal segment and represents information in the time domain or frequency domain. can In addition, in the authentication method based on ECG, since the steps of the two feature extraction and authentication mechanisms are separated and the label information of the signal is not used continuously, the mutual compensation action of the feature extraction and authentication mechanism is not realized. Due to this, the accuracy of the authentication method based on the ECG may be deteriorated.

According to an embodiment, the electrocardiogram-based authentication method may acquire a classifier with high accuracy by synchronously learning the dictionary and the classifier to mutually promote the acquisition of high-level features having discriminant power.

Hereinafter, an embodiment will be described with reference to FIGS. 1 to 8 . Mainly, the ECG-based authentication will be described with reference to FIGS. 1, 6 and 7, and the learning for the ECG-based authentication will be described with reference to FIGS. 4 to 6 .

1 is a flowchart 100 for explaining an electrocardiogram-based authentication method according to an exemplary embodiment.

According to an embodiment, the electrocardiogram-based authentication apparatus operates according to an electrocardiogram-based authentication method, and performs authentication based on a pre-learned dictionary and a pre-learned classifier. The dictionary and the classifier used by the authentication device may be synchronously and previously learned from each other according to the learning method for electrocardiogram-based authentication according to an embodiment. According to an embodiment, the pre-trained dictionary includes category dictionaries corresponding to a plurality of categories, and the pre-learned classifier includes category classifiers corresponding to the plurality of categories.

According to an embodiment, a plurality of feature vectors for the training ECGs are generated using category dictionaries, prediction categories of the training ECGs are determined based on the plurality of feature vectors, and based on the categories and prediction categories By alternately updating the category dictionaries and the category classifier, the dictionary and the classifier are learned synchronously with each other.

According to an embodiment, in order to express the electrocardiogram as a linear combination of a plurality of category dictionaries included in advance, coefficients of the plurality of category dictionaries are determined, and a feature vector is generated based on the determined coefficients. According to an embodiment, the electrocardiogram is divided into at least one signal segment, and at least one feature mark corresponding to the at least one signal segment is generated, and group sparse coding is performed on the at least one feature mark in advance. Thus, at least one second feature mark is generated, and a feature vector is generated based on the at least one second feature mark. According to an embodiment, the characteristic indication may be expressed as a linear combination of some of a plurality of category dictionaries included in the dictionary, and the second characteristic indication may include coefficients of some category dictionaries.

Referring to FIG. 1 , an authentication apparatus according to an exemplary embodiment receives a test electrocardiogram ( S101 ). The test electrocardiogram is an electrocardiogram to be authenticated, and may be obtained from the object through a sensor in order to authenticate the object.

According to an embodiment, the authentication apparatus may perform pre-processing on the input test ECG to extract a feature having discrimination power from the test ECG. Specifically, the preprocessing performed by the authentication device on the test electrocardiogram divides the test electrocardiogram into at least one signal segment, extracts individual features from the segmented signal segments, and performs dimension reduction processing on each extracted feature and generating each feature representation corresponding to the at least one signal segment.

The pre-processing performed by the authentication device according to an embodiment is to perform data filtering processing on the collected test ECG to remove a noise signal generated from a device or object, and to detect a QRS wave for the filtered test ECG includes action. The pre-processing performed by the authentication device according to an embodiment divides the test electrocardiogram into at least one signal segment based on the detection result of the QRS wave for the test electrocardiogram, and displays each characteristic corresponding to the divided signal segments. action to create

2 is an example 200 of a waveform of an electrocardiogram.

The electrocardiogram is a signal of the heart's electrical activity with periodicity. 2 is an example of a waveform of a typical electrocardiogram. Referring to FIG. 2 , a typical electrocardiogram includes a P wave 201 , a QRS wave 202 , and a T wave 203 . The wave that appears preferentially on the electrocardiogram is called the P wave. The P wave 201 represents a left and right atrial depolarization process. The QRS wave 202 represents the potential variation of the depolarization process of the left and right conscience. The QRS wave 202 generally includes three successive displacement waves. The first downward wave is called the Q wave, the first upward wave after the Q wave is called the R wave, and the single downward wave connected with the descending of the R wave is called the S wave. The three waves of a typical QRS are interconnected, and the total width (time) generally does not exceed 0.12 s, which is collectively called a QRS wave group. The T wave 203 has a relatively low amplitude and a relatively long occupancy time after the ST segment, and represents a change in potential of ventricular repolarization.

The ratio of the energy of the QRS wave in the electrocardiogram is relatively large, and includes important characteristic parameters of the electrocardiogram. In the case of dividing the electrocardiogram according to an embodiment, the detection of QRS is usefully used, and through this, valid information on the group of QRS waves can be obtained. An embodiment of the electrocardiogram filtering and detecting operation performed in the electrocardiogram preprocessing step is not limited to a specific technical field, and may be implemented by any technology developed in the future.

According to an embodiment, the authentication device may divide the electrocardiogram into at least one signal segment based on the detection result of the QRS wave, and each divided signal segment may correspond to one or a plurality of heartbeats.

According to an embodiment, the authentication device extracts a characteristic of each signal segment for each divided signal segment, for example, extracts at least one or a combination of a time domain characteristic, a frequency domain characteristic, and a statistical characteristic of the signal segment. can do. Examples of time-domain characteristics include, but are not limited to, a waveform of an electrocardiogram, a moment of occurrence of a specific waveform (eg, a breaking moment of a T-wave), a moment of change, and a continuous time. The frequency domain characteristic includes, for example, a frequency or a spectral distribution of an electrocardiogram, but is not limited thereto.

According to an embodiment, the authentication device performs dimensionality reduction processing using Principle Component Analysis (PCA) or Linear Discriminant Analysis (LDA) on the extracted features as a pre-processing step to perform at least one signal It is possible to obtain each characteristic indication for a segment. Since the dimension of features extracted from each signal segment can be relatively high, in order to save calculation time and improve calculation accuracy, the authentication device generates a feature representation through dimensionality reduction processing for the extracted features.

Referring to FIG. 1 , the authentication device generates a feature vector for a test electrocardiogram based on a pre-learned dictionary ( 102 ).

According to an embodiment, the authentication apparatus generates at least one second characteristic indication from at least one characteristic indication corresponding to at least one signal segment of the test electrocardiogram using the classifier and the dictionary learned synchronously with each other. According to an embodiment, the authentication apparatus performs group sparse coding based on a dictionary on each feature indication corresponding to at least one signal segment obtained through preprocessing, and at least one second corresponding to each divided signal segment. 2 Create a feature representation. The authentication device generates a feature vector for the test electrocardiogram based on the at least one second feature indication corresponding to the at least one signal segment.

이고, D의 각 열 d_j는 하나의 기초함수를 나타내며, z는 D를 이용하여 획득된 x의 계수를 나타내고, 이 중에서

일 수 있다. 여기서, 코딩 계수 z를 데이터 x의 새로운 특징표시라고 한다. 일실시예에 따르면, 인증 장치는 사전을 이용한 그룹 스파스 코딩에 기초하여 심전도의 적어도 하나의 신호 세그먼트에 대응하는 적어도 하나의 특징표시로부터 새로운 특징표시를 생성할 수 있고, 여기서 새로운 특징표시는 제2 특징표시에 해당한다.In group sparse coding for generating feature representations, data is expressed as a linear combination of one group of fundamental functions. Since the coefficients used in the linear combination are assumed to be sparse, the number of non-zero coefficients compared to the total number of coefficients is much less Group sparse coding can be performed as a feature representation of a small portion of the elementary function data x having the representativeness found in one group of elementary functions. Here, the basic function of this group is called a dictionary, each column of the dictionary is called a code book, and a coding coefficient is a new characteristic representation of data x. For example, x is one d-dimensional data, dictionary D is a dx J matrix, where

, and each column d _j of D represents one fundamental function, and z represents a coefficient of x obtained using D, among which

can be Here, the coding coefficient z is referred to as a new feature representation of data x. According to an embodiment, the authentication apparatus may generate a new characteristic indication from at least one characteristic indication corresponding to at least one signal segment of the electrocardiogram based on the group sparse coding using the dictionary, wherein the new characteristic indication is the second characteristic indication. 2 Corresponds to the characteristic indication.

Learning about a dictionary based on group sparse coding can be applied to various fields. Classification means determining which category the unknown data belongs to. Assuming that a dictionary for each category is pre-configured, if there is a category associated with one new data, the data is a dictionary for the corresponding category. can be expressed as the optimal sparseness of Learning on a dictionary allows the dictionary to learn a group of fundamental functions that better represent or code a signal given as a training sample.

According to an embodiment, the authentication device may generate a new characteristic indication of the preprocessed electrocardiogram using the dictionary. The higher the level of discrimination against the features of the learned dictionary, the more accurate the authentication result.

Referring to FIG. 1 , the authentication apparatus generates a classification result for a test electrocardiogram based on a generated feature vector and a pre-learned classifier ( S103 ).

The classifier may identify which category the electrocardiogram belongs to by using the feature vector based on the new feature indication extracted through the dictionary. 3 to 5, an embodiment in which the dictionary and the classifier are learned synchronously will be described. According to an embodiment, the dictionary and the classifier may be learned in the authentication device, and may be loaded by the authentication device after being learned in a separate device, but the device for performing the learning is not limited thereto.

Referring to FIG. 1 , the authentication apparatus authenticates the test electrocardiogram based on the generated classification result ( S104 ).

According to an embodiment, the authentication apparatus may determine whether the test ECG has passed authentication based on the classification result of the test ECG, wherein a plurality of methods may be used. For example, there is a method of calculating the average value of the feature display of the signal segmentation segment and inputting it to a classifier to generate a classification result. This method reduces accuracy because it is vulnerable to the influence of noise. According to an embodiment, the authentication device may determine whether authentication has been passed through a decision based on a voting mechanism. According to the embodiment based on the voting mechanism, the authentication device determines whether or not the authentication has passed based on the voting of each category of all divided signal segments of the test electrocardiogram, thereby reducing the effect of noise. More specifically, the authentication apparatus may calculate a similarity between the classification result of the test ECG and the classification result of the reference ECG, and determine whether the test ECG has passed authentication by using the calculation result. An embodiment based on a decision based on a voting mechanism will be described with reference to FIG. 7 .

According to an embodiment, the process of processing the reference ECG (preprocessing the reference ECG and obtaining a classification result for the reference ECG preprocessed using a dictionary and a classifier) may be performed simultaneously with the processing of the test ECG described above. . Alternatively, by performing processing on the reference ECG in advance and storing the classification result generated according to the processing in advance, the authentication device may obtain the classification result stored in advance and use it in the process of authenticating the test ECG, but processing the reference ECG The embodiment regarding the viewpoint is not limited thereto.

The authentication method based on an electrocardiogram according to an embodiment uses a dictionary and a classifier that are learned synchronously with each other. , the classifier makes it possible to generate high-accuracy classification results.

3 is a flowchart 300 illustrating a learning method for ECG-based authentication according to an embodiment.

Referring to FIG. 3 , the learning apparatus generates a plurality of feature vectors for training ECGs using category dictionaries corresponding to a plurality of categories ( 301 ). According to an embodiment, the learning apparatus performs preprocessing on a plurality of training ECGs belonging to different categories, and divides each training ECG into at least one signal segment.

According to one embodiment, the training electrocardiograms are obtained from electrocardiograms of a plurality of individuals (persons). According to an embodiment of classifying the ECG, each individual may correspond to one category, where the classification of the ECG may be used for identification or authentication of identity. In addition, a plurality of individuals having the same characteristics may correspond to one category, where the classification of the electrocardiogram may be applied to identify a heart disease. However, the embodiment of classifying the electrocardiogram according to the category is not limited to what has been described herein and may be applied in various ways. Hereinafter, in a classification in which one individual corresponds to one category, an embodiment in which the classification of the electrocardiogram is used for identification or authentication of identity will be described.

According to an embodiment, the learning apparatus acquires a plurality of training electrocardiograms. As described with reference to FIG. 1 , the learning apparatus may perform pre-processing on the plurality of acquired training ECGs to divide each training ECG into at least one signal segment. Since the above-described embodiment is applied to the pre-processing process, detailed details thereof will be omitted. According to an embodiment, the learning apparatus divides each training electrocardiogram into at least one signal segment, and generates characteristic indications for each segmented signal segment.

로 표시될 수 있다. 여기서,

이고, d는 분할된 각 신호 세그먼트의 특징표시의 특징차원(feature dimension)이며,

는 제c번째 개체의 신호 세그먼트에 속하는 특징표시의 개수이다. 모든 C개 개체의 신호 세그먼트들을 순차적으로 배열하면 모든 트레이닝 심전도는

의 특징표시들로 나타낼 수 있고, 여기서

이고 N은 분할된 모든 신호 세그먼트들의 총 개수이다. For example, it is assumed that the learning apparatus acquires the ECGs of C objects as training ECGs. When the learning device pre-processes the training ECGs (including segmentation into a plurality of signal segments, feature extraction and dimensionality reduction processing, etc.), each individual

can be displayed as here,

, d is a feature dimension of the feature display of each segmented signal segment,

is the number of feature marks belonging to the signal segment of the c-th entity. If the signal segments of all C objects are sequentially arranged, all training ECGs are

It can be expressed by the characteristic marks of

and N is the total number of all segmented signal segments.

The learning apparatus synchronously performs iterative learning on the category dictionary and the category classifier of each category based on the plurality of pre-processed training ECGs. According to an embodiment, the learning apparatus sets a category dictionary and a category classifier corresponding to each category, and acquires the dictionary and the classifier based on the category. The learning apparatus uses the label information of the divided signal segment (the most important medium for discriminating a feature as a classification result for the signal segment) sufficiently, A category dictionary and a category classifier can be obtained.

If a method for synchronously learning the category dictionary and the category classifier is not used, since there are relatively few codebooks used for reconstructing an apparatus for identifying features, the ability to discriminate features may be reduced. In addition, when a dictionary is learned for all training data, most of the training data belonging to other categories may become unnecessary information for determining a sample of any specific category. In addition, when the label of the training data is not used for learning, the ability to discriminate the feature may be reduced.

According to an embodiment, since a dictionary related to a category is used, a feature display with higher discrimination power can be obtained. According to an embodiment, since label information, which is a medium that provides a high level of discrimination power of a feature, is utilized, strong identification of an entity may be provided.

로 표시될 수 있다When the learning device pre-processes the training ECGs (including segmentation into a plurality of signal segments, feature extraction and dimensionality reduction processing, etc.), each individual

can be displayed as

If the dictionary according to an embodiment is D, the dictionary may be expressed as follows.

는 제c번째 카테고리(제c번째 개체)와 대응하는 카테고리 사전(간략히 제c 카테고리 사전이라 할 수 있음)을 나타내는데,

는 아래와 같이 표현된다. Among them,

denotes a category dictionary (which may be briefly referred to as a c-th category dictionary) corresponding to a c-th category (c-th entity),

is expressed as below.

로 표시한다면,

에 기반한 그룹 스파스 코딩을 통해

로부터 새로운 특징표시가 추출될 수 있고,

는 아래와 같은 식으로 나타낼 수 있다.Based on the feature representation of the segmented signal segment, the c-th object is

If indicated as

Through group sparse coding based on

A new feature indication can be extracted from

can be expressed in the following way.

는 제c번째 카테고리의 하나의 코드 북을 나타내고,

는 코드계수를 나타내며

의 새로운 특징표시(제2 특징표시)라 할 수 있다.

는 제c 카테고리 사전의 코드 북의 개수를 나타내고, J는 사전 D의 총 코드 북의 개수를 나타낸다.here,

represents one code book of the c-th category,

represents the code coefficient

It can be called a new characteristic display (second characteristic display) of

denotes the number of codebooks in the c-th category dictionary, and J denotes the total number of codebooks in the dictionary D.

The learning apparatus generates a new feature representation of X by using group sparse coding, and when a dictionary D is given, z derived by performing group sparse coding on X can be expressed by the following equation.

(1)

(One)

,

,

는 z중에서 제c 카테고리 사전과 대응하는 코드계수를 나타내고,

는 가중 파라미터(weight parameter)를 의미한다.

의 범위는 재구성에 사용되는 코드 북이 해당하는 카테고리를 구속하므로, 적을수록 좋다. 예를 들면, 제 c카테고리에 속하는 신호 세그먼트의 특징표시에 기초한

에 대하여 그룹 스파스 코딩을 통해 도출된 z의 0이 아닌 요소는, 가급적이면 제c 카테고리 사전과 대응하는 차원에 속할수록 좋은데, 이러한 경우 식별에 관계가 없는 트레이닝 데이터가 가져다 주는 계산오차를 줄일 수 있다. 사전 D가 주어졌을 때, 그룹 스파스 코딩의 공식(1)을 통하여 분할된 각 신호 세그먼트에 대응하는 z를 구할 수 있다. 일실시예에 따르면, 학습 장치는 트레이닝 심전도들에 대하여 카테고리 사전들에 기초한 그룹 스파스 코딩을 수행하여 제2 특징표시들을 생성하고, 분할된 신호 세그먼트들에 대응하는 제2 특징표시들에 기초하여 특징벡터들을 생성할 수 있다.here,

,

,

represents the code coefficient corresponding to the c-th category dictionary among z,

denotes a weight parameter.

Since the range of the codebook used for reconstruction constrains the corresponding category, the smaller the number, the better. For example, based on the characterization of the signal segment belonging to the c-th

For , the non-zero element of z derived through group sparse coding is better if it belongs to the dimension corresponding to the c-th category dictionary. have. Given a dictionary D, z corresponding to each segmented signal segment can be obtained through formula (1) of group sparse coding. According to an embodiment, the learning apparatus generates second characteristic marks by performing group sparse coding based on category dictionaries on the training electrocardiograms, and based on the second characteristic marks corresponding to the segmented signal segments. You can create feature vectors.

According to an embodiment, the learning apparatus determines predictive categories of the training ECGs based on a plurality of feature vectors for the training ECGs ( 302 ). According to an embodiment, in order to sufficiently use the label information of the signal segment, the target function used in the iterative process of synchronous learning for the dictionary and the classifier may reflect the category and label information of the training ECG. The target function may be designed through various application methods, and the target function is not limited to the embodiments described herein.

According to an embodiment, minimizing the error function of the predicted category and the actual category of the training ECG may be set as a target function, and this error function may be expressed as follows.

(2)

(2)

는 선형회귀(regression) 분류기를 나타내며, W의 임의의 한 개 열인

는 대응하는 카테고리의 분류기를 나타낸다. G는 분할된 모든 신호 세그먼트에 대응하는 실제 카테고리를 나타내는데, 이는 라벨 행렬로 표시될 수 있으며, 그 중에서,

은 제i번째 신호 세그먼트가 제c번째 카테고리(제c번째 개체에 속함)에 대응하는 것을 나타내고, 그렇지 않은 경우는

이다. 식(2)의 제1항

는 모든 신호 세그먼트의 예측오차를 나타내는데,

는 하나의 NxC차원 행렬이면서 모든 라벨에 대한 모든 신호 세그먼트의 확률분포를 나타낸다.

을 최소화한다는 것은 예측 카테고리와 실제 카테고리 사이의 오차를 될수록 작게 하는 것을 의미한다. 식(2)의 제2항

는 과적합(over-fitting)을 방지하기 위함이다.Here, α means the weighting parameter,

represents a linear regression classifier, and is any one column of W,

denotes a classifier of the corresponding category. G denotes the actual category corresponding to all segmented signal segments, which can be represented by a label matrix, among which:

indicates that the i-th signal segment corresponds to the c-th category (belonging to the c-th entity), otherwise

to be. Paragraph 1 of Equation (2)

represents the prediction error of all signal segments,

is one NxC-dimensional matrix and represents the probability distribution of all signal segments for all labels.

Minimizing the error means making the error between the predicted category and the actual category as small as possible. Paragraph 2 of Equation (2)

is to prevent over-fitting.

은 식(1)에서 획득한 것이다. 따라서, 식(2)중의 변화하는 요소는 D와 W이다. 사전 D가 주어진 경우, 학습 장치는 그룹 스파스 코딩에 기반한 식(1)을 통하여 트레이닝 심전도들에 대한 특징벡터들을 획득할 수 있다. 그 다음에, 학습 장치는 식(2)를 통하여 분류기 W를 획득할 수 있고, 다시 분류기 W를 고정하고 식(2)를 통하여 사전 D를 최적화하고, 최적화된 사전 D를 고정하고 식(2)를 통하여 분류기 W를 최적화할 수 있다. 이러한 방식으로 최적화를 반복적으로 계속 수행하여, 학습 장치는 사전 D 와 분류기 W가 동시에 로컬 최적화에 도달하게 할 수 있고, 각 개체에 대응하는 카테고리 사전

와 카테고리 분류기

를 획득할 수 있다. 특징을 추출하고 카테고리를 분류하는 사전과 분류기의 동기 학습을 반복적으로 하여, 학습 장치는 사전 및 분류기의 양자 간에 학습을 서로 촉진시킬 수 있다.According to an embodiment, the learning apparatus alternately updates category dictionaries and category classifiers based on actual categories and predicted categories ( 303 ). Referring to Equation (2), Z is the implicit function of the dictionary D,

is obtained from Equation (1). Therefore, the changing factors in equation (2) are D and W. Given the dictionary D, the learning apparatus may obtain feature vectors for the training ECGs through Equation (1) based on group sparse coding. Then, the learning apparatus can obtain a classifier W through Equation (2), fix the classifier W again, optimize the dictionary D through Equation (2), fix the optimized dictionary D, and use Equation (2) Classifier W can be optimized through By iteratively continuing to perform optimization in this way, the learning device can make dictionary D and classifier W reach local optimization at the same time, and the category dictionary corresponding to each object

and category classifier

can be obtained. By repeatedly performing synchronous learning of the dictionary and the classifier for extracting features and classifying categories, the learning apparatus can promote learning between both the dictionary and the classifier.

, c=1, ... , C를 획득하고, 이러한 카테고리 사전이 배열되는 방식으로 합병된 사전 D를 생성하고,

로 표현된다.According to an embodiment, the learning apparatus creates a dictionary by merging the updated category dictionaries ( 340 ). The learning device is a category dictionary for C categories updated through repeated learning

, c=1, ... , C, and create a merged dictionary D in such a way that these category dictionaries are arranged,

is expressed as

According to an embodiment, the learning apparatus generates a classifier by merging the updated category classifiers ( 350 ). The learning device may acquire the classifier in a similar manner to merging category dictionaries.

4 is a flowchart 400 illustrating an embodiment of alternately updating category dictionaries and category classifiers.

According to an embodiment, the learning apparatus fixes the category classifiers and optimizes the category dictionaries to minimize the error between the predicted categories and the actual categories ( 401 ).

According to an embodiment, the learning apparatus initializes a category dictionary corresponding to each category. In order to obtain category dictionaries and category classifiers corresponding to a plurality of categories through Equations (1) and (2) above, since it is necessary to obtain a classifier by fixing the dictionary, it is required to initialize each category dictionary. .

Various methods may be applied to a method for initializing the category dictionary. According to one embodiment, the learning apparatus generates a plurality of clusters by clustering the preprocessed training ECGs according to categories. According to an embodiment, a k-means clustering method may be used, and since there may be many types of clustering methods, the method to which clustering is applied does not limit the embodiment. The learning apparatus may generate initialized category dictionaries by performing normalization on each center of a plurality of clusters generated through clustering.

According to an embodiment, the learning apparatus fixes the optimized category dictionaries, and optimizes the category classifiers to minimize the errors of predicted categories and actual categories ( 402 ).

According to an embodiment, the learning apparatus may synchronously update each of the category dictionaries and the category classifiers by using the initialized category dictionaries to minimize the error function of the predicted categories of the training ECGs and the actual categories. . First, the learning apparatus generates an initialized dictionary by arranging initialized category dictionaries corresponding to all categories. Then, the learning apparatus generates a feature vector for each training ECG through the group sparse coding of Equation (1) using the initialized dictionary. Then, the learning apparatus obtains a classifier W based on the generated feature vector and Equation (2). Then, the learning apparatus optimizes the dictionary D based on the obtained classifier W and equation (2). Then, the learning apparatus optimizes the classifier W based on the optimized dictionary D and equation (2). By repeating this calculation process, the dictionary D and the classifier W can be optimized. Here, Equation (2) is used to minimize the error function between the predicted category and the actual category of the training ECG.

According to an embodiment, the learning apparatus terminates the update when optimizations that are synchronously repeated with respect to the dictionary and the classifier satisfy a predefined condition, and outputs the updated category dictionaries and category classifiers. The higher the iteration order of update, the better the performance of the obtained dictionary and classifier. In addition, since the accuracy of the updated dictionary and classifier may exceed the required level, the iteration of the update is terminated when an appropriate condition is satisfied, and the update of the dictionary and the classifier should be output.

According to an embodiment, the predefined condition is that when the order of repetition reaches a preset order, when the value of the error function between the predicted category of the training electrocardiogram and the actual category is smaller than a predefined first value, and when the error function is At least one of a case in which the change amount of the value is smaller than the predefined second value may be included. When a predefined condition is satisfied, the learning apparatus may end repetition of synchronous learning and output a category dictionary and a category classifier corresponding to each category.

5 is a flow chart 500 illustrating an embodiment of optimizing category dictionaries and category classifiers.

In order to derive an optimized dictionary and classifier, various kinds of calculation methods may be applied. According to an embodiment, the learning apparatus may synchronously and iteratively calculate the dictionary and the classifier using a gradient descent algorithm. Other optimization algorithms may be used for optimizing the dictionary and the classifier, and an embodiment in which a gradient descent algorithm is used will be described with reference to FIG. 5 , but this does not limit the embodiment to the method of calculating the optimization.

According to an embodiment, the learning apparatus may optimize the category dictionaries to minimize the error between the predicted categories of the training ECGs and the actual categories using the gradient descent algorithm ( 501 ).

As described above, the learning apparatus generates a feature vector for the training ECG through group sparse coding for each signal segment of each training ECG using the current dictionary. When optimization is performed for the first time, the current dictionary corresponds to the initialized dictionary. In the next update, the current dictionary corresponds to the dictionary optimized through previous synchronous learning.

According to an embodiment, the learning apparatus applies the current dictionary to Equation (1) to perform group sparse coding for each training ECG, and generates a feature vector corresponding to each training ECG. Here, a second characteristic indication for each signal segment is generated using the current dictionary, and the coding coefficient through group sparse coding is the second characteristic indication of each signal segment. As described above, data corresponding to a category that is not related to the training electrocardiogram introduces a calculation error. When group sparse coding is used, the calculation error is reduced through the reconstructed codebook.

According to an embodiment, the learning apparatus may optimize the category classifiers to minimize the error using a gradient descent algorithm ( 502 ). The learning apparatus may update each category classifier using a gradient descent algorithm based on a feature vector and an error function.

로 획득된다. 이 식에서, I는 사전 D의 총 차원과 같은 차원의 단위행렬을 가리키고,

와

는 모두 상수의 파라미터를 나타낸다. Y는 분할된 각 신호 세그먼트의 실제 라벨을 의미하는데, 이는 이미 알고 있는 값이다. Z와 Y를 모두 알고 있는 상태에서, W가 계산될 수 있다. 이후에 반복되는 동기적 학습과정에서 도출된 W는 초기 값이 된다.The initial value of the classifier W can be calculated through Equation (2). First, the learning apparatus obtains a derivative of W from Equation (2) using the feature vector Z generated based on the initialized dictionary. Here, if we set the derivative to 0, W is

is obtained with In this expression, I denotes the identity matrix of the same dimension as the total dimension of dictionary D,

Wow

are all constant parameters. Y denotes the actual label of each segmented signal segment, which is a known value. With both Z and Y known, W can be computed. Then, W derived from the repeated synchronous learning process becomes the initial value.

를 획득한다. 그 다음, 경사 하강 알고리즘의 공식

에 기초하여, W가 갱신된다. 여기서,

는 경사 방향으로의 검색보폭(step size in search)이다. 일실시예에 따르면, 학습 장치는 갱신된 카테고리 분류기와 오차함수에 기초한 경사 하강 알고리즘을 이용하여, 각 카테고리 사전을 갱신한다.According to an embodiment, the learning apparatus fixes the current dictionary D based on Equation (2) and calculates the derivative for the classifier W,

to acquire Then, the formula of the gradient descent algorithm

Based on , W is updated. here,

is the step size in search in the oblique direction. According to an embodiment, the learning apparatus updates each category dictionary by using the updated category classifier and a gradient descent algorithm based on the error function.

가 계산된다. 이러한 연산이 수행될 때, 학습 장치는 각 신호 세그먼트를 랜덤으로 교란하고, 교란 후의 순서에 따라 경사를 계산할 수 있다.According to an embodiment, the learning apparatus fixes the updated classifier W, and obtains a derivative with respect to the dictionary D. In equation (2), since there is no explicit expression for D, only an explicit expression for Z, Z is a function of D, and Z contains a plurality of z, each z is all a function of D. Therefore, using the fixed point implicit difference method, the derivative of D with respect to the error function of the nth training ECG

is calculated When such an operation is performed, the learning apparatus may randomly perturb each signal segment and calculate a gradient according to the order after perturbation.

를 구할 때, 학습 장치는 연쇄법칙(chain rule)을 이용하여 먼저 도함수

과

를 구하고, 그 다음 양자를 곱하여

를 획득한다. 학습 장치는

를 구한 후, 경사 디센트 알고리즘의 공식

에 기반하여 D를 갱신한다. 여기서,

는 경사 방향으로의 검색보폭이다. According to one embodiment,

When finding , the learning device first uses the chain rule to

class

, and then multiply by both

to acquire learning device

After finding , the formula of the gradient descent algorithm

D is updated based on here,

is the search stride in the oblique direction.

만 구하고, 모든 트레이닝 샘플들의 오차함수의 값에 대한 D의 도함수

를 구하지 못하는 것을 고려하여, 학습 장치는 동기학습을 반복할 때, 복수의 트레이닝 심전도에 대해 랜덤배열을 하고, 매번의 교란 후에 순서에 따라 경사를 계산하여 오차를 줄일 수 있다.In the process of applying the gradient descent algorithm to the dictionary according to an embodiment, the derivative of the dictionary D with respect to the value of the error function of the nth training sample

, and the derivative of D with respect to the value of the error function of all training samples

Considering that , when synchronous learning is repeated, the learning apparatus may randomly arrange a plurality of training electrocardiograms, calculate the gradient in sequence after each disturbance, and reduce the error.

An embodiment of synchronously learning the dictionary and the classifier has been described above. As described above, since the category dictionary is learned for each category, calculation errors caused by data not related to training can be reduced. Specifically, when the group sparse coding is performed to determine the coding coefficient z, the number of categories of codebooks used for reconstruction should be smaller. This allows the data to be presented as a dictionary of categories related to it.

According to an embodiment, the learning apparatus utilizes label information for prior learning, has a goal function to minimize the error function of the predicted category and the actual category of the training ECG, and synchronously learns the dictionary and the classifier to obtain a high level It is possible to obtain a feature having the discriminant power of .

6 is a flowchart 600 illustrating an embodiment of generating a feature vector corresponding to a test electrocardiogram.

According to an embodiment, the dictionary and classifier learned in the manner described above are used to extract features from the test electrocardiogram in the test or authentication phase and classify the object. The authentication device divides the test electrocardiogram into at least one signal segment ( 601 ).

The authentication apparatus according to an embodiment generates ( 602 ) at least one characteristic indication corresponding to the at least one signal segment. Here, the pre-processing for the electrocardiogram described above may be applied to the specific embodiment of generating the feature display.

The authentication apparatus according to an embodiment generates at least one second characteristic indication by performing group sparse coding based on a dictionary on at least one characteristic indication ( S603 ). The above-described embodiment is applied to the dictionary and group sparse coding used here, and the second characteristic indication can be generated through Equation (1).

로 표현된다고 가정하면, 인증 장치는 식 (1)에 기초한 S에 대한 그룹 스파스 코딩을 통해 특징벡터

를 생성할 수 잇다. 여기서, P는 테스트 심전도와 대응하는 신호 세그먼트의 개수이다.The authentication apparatus according to an embodiment generates a feature vector based on at least one second feature indication corresponding to the at least one signal segment ( 604 ). For example, if the preprocessed test ECG and corresponding signal segments are characterized

Assuming that it is expressed as , the authentication device performs a feature vector through group sparse coding for

can create Here, P is the number of signal segments corresponding to the test electrocardiogram.

The authentication device classifies the test ECG based on the feature vector and the classifier W. According to an embodiment, the authentication device may calculate the label (classification result for each signal segment of the test electrocardiogram) of each segmented signal segment based on the following equation and the classifier W.

는 테스트 심전도와 대응하는 라벨이다.here,

is a label corresponding to the test electrocardiogram.

7 is a flowchart 700 illustrating an embodiment of authentication for an electrocardiogram.

According to an embodiment, the authentication device determines whether the test ECG has passed authentication based on the classification result of the test ECG. The authentication device obtains a classification result for the reference ECG based on at least one label of the reference ECG (operation 701).

로 가정하면 인증 장치는 식(1)을 통하여 Y에 대한 그룹 스파스 코딩을 통해 새로운 특징벡터

를 생성할 수 있다. 여기서, Q는 참조 심전도의 신호 세그먼트의 개수이다.According to an embodiment, the classification result of each signal segment obtained by using a dictionary and a classifier with respect to the reference electrocardiogram may be obtained from pre-stored data and a corresponding classification result, and at the same time as processing for the test electrocardiogram in the authentication step It may be obtained by using a dictionary and a classifier with respect to the reference ECG. According to one embodiment, the characterization of the signal segment of the preprocessed reference electrocardiogram is

Assuming that , the authentication device performs a new feature vector through group sparse coding for Y through Equation (1).

can create where Q is the number of signal segments of the reference electrocardiogram.

The authentication device calculates a label of each signal segment (ie, a classification result for each signal segment of the reference electrocardiogram) based on the following equation and the classifier W.

는 참조 심전도와 대응하는 라벨이다. here,

is a label corresponding to the reference electrocardiogram.

The authentication device may perform authentication based on a voting mechanism for a classification result for the test ECG and a classification result for the reference ECG, and may determine whether the test ECG has passed authentication.

The authentication device generates ( 702 ) a first distribution in all categories for the at least one signal segment of the test electrocardiogram, based on the at least one label for the test electrocardiogram. According to an embodiment, the first distribution comprises a number of signal segments belonging to each of the plurality of categories.

로 계산된다. 인증 장치는 각 카테고리에 대한 투표수에 기반하여, 테스트 심전도가 대응하는 제1 분포

또는 히스토그램을 획득할 수 있다.Specifically, the authentication device calculates the number of votes (distribution) for each category of each test ECG based on the test ECG and the corresponding label. For example, if the test electrocardiogram includes a total of P signal segments, and k signal segments are predicted as the c th category (ie, the c th individual), the number of votes corresponding to the c th category is

is calculated as The authentication device is based on the number of votes for each category, and the first distribution to which the test electrocardiogram corresponds

Alternatively, a histogram may be obtained.

(분포)에 기초하여 참조 심전도가 대응하는 제2 분포

가 획득될 수 있다.The authentication device acquires a second distribution in all categories for at least one signal segment of the reference electrocardiogram ( 703 ). According to an embodiment, the second distribution includes the number of reference signal segments belonging to each of the plurality of categories. Similar to the way described above, the number of votes for each category of reference ECG

a second distribution to which the reference electrocardiogram corresponds based on (distribution)

can be obtained.

The authentication device calculates a degree of similarity between the first distribution and the second distribution ( 704 ). Various methods may be applied to the similarity calculation. According to an embodiment, the authentication device may calculate the similarity between two distributions or histograms by using a method of crossing the histograms. For example, the degree of similarity may be calculated using the following equation.

The authentication device performs authentication by comparing the degree of similarity and a predefined threshold ( 705 ). According to an embodiment, the authentication device may compare whether the similarity is greater than or equal to a predefined threshold, and may determine whether authentication has been passed according to the comparison result.

When the similarity is greater than or equal to a predefined threshold, the authentication device may determine that the test ECG and the reference ECG are sufficiently similar, and determine that the test ECG has passed the authentication. Conversely, when the similarity is less than a predefined threshold, the authentication apparatus may determine that the difference between the test ECG and the reference ECG is relatively large, and determine that authentication of the test ECG is rejected.

Although the operations according to the embodiments have been described in the order specified in the accompanying drawings, the embodiments are not limited to that the operations are necessarily executed in the above-described order or all operations are performed. For example, in FIG. 3 , the operation of merging the category classifiers may precede the operation of merging the category dictionaries. Alternatively, these operations may be processed in parallel. In FIG. 7 , the operation of obtaining the second distribution corresponding to the reference electrocardiogram may be performed preferentially in the generation of the first distribution corresponding to the test electrocardiogram. In addition, an embodiment may additionally or alternatively omit some steps and merge a plurality of steps to execute as one step and/or decompose one step into a plurality of steps to execute.

As described above, the authentication method based on an electrocardiogram can be applied to identify an identity. With respect to different ways of identifying identity, there may be a plurality of specific and different types of embodiments. According to an embodiment, a reference electrocardiogram corresponding to a user having a preset authority (eg, a user having a right to visit or use an arbitrary resource) may be obtained in advance. The obtained reference electrocardiogram may be stored and used as a basis for judging whether a pre-set authority for the visiting user is provided.

When a user to be authenticated visits, a test ECG corresponding thereto may be acquired, and authentication of the test ECG may be performed through the authentication method according to an embodiment. If the test electrocardiogram passes the authentication, it is determined that the user has a preset authority. For example, the smart phone may acquire and store a reference electrocardiogram corresponding to the device owner, and when the user intends to use the smart phone, authentication of the test electrocardiogram may be performed through the authentication method according to an embodiment. If the user corresponds to the owner of the smart phone, authentication is passed, and the resource of the smart phone or the smart phone can be used.

There may be a plurality of users with preset privileges. In this case, it is necessary to store reference ECGs corresponding to a plurality of users. When a user to be authenticated visits, a test electrocardiogram corresponding thereto is acquired, and authentication of the test electrocardiogram based on each reference electrocardiogram stored in advance is performed through the authentication method according to an embodiment. If authentication is passed only once for the test electrocardiogram, it is determined that the user has a preset authority. For example, the door guard device of a certain complex acquires and stores reference ECGs corresponding to all residents of the complex, and when a visitor who is a visitor tries to enter the complex, obtains and executes a test ECG corresponding to the visitor Through the authentication method according to the example, each reference ECG stored in advance is searched and the test ECG is authenticated. If the visitor is just a resident, the authentication must be passed once, so the door guard equipment is released.

According to another embodiment, reference electrocardiograms corresponding to some specific users (eg, all employees of a certain company or all students of a certain school) may be acquired in advance. The user's related information corresponding to the reference electrocardiogram is merged with the reference electrocardiogram and stored, used as a basis for confirming the user's identity, and used to retrieve or record the user's related information. When the user visits, a corresponding test ECG is obtained, and authentication is performed on the test ECG using each reference ECG stored in advance through the authentication method according to an embodiment. When the test electrocardiogram passes the authentication for any reference electrocardiogram, the user's identity is confirmed based on the reference electrocardiogram-related related information, the user's related information can also be obtained, and the user's new information is merged with the identity. More can be recorded. For example, the attendance equipment of any company can acquire the reference electrocardiogram corresponding to all the employees of the company, merge and store the names and IDs of the corresponding employees, and when the employee records the attendance of the employee, the correspondence with the employee A test electrocardiogram may be acquired, and authentication may be performed on the test electrocardiogram using each reference electrocardiogram stored in advance through the authentication method according to an embodiment. When the test ECG passes the authentication of the parent reference ECG, an employee name and ID associated with the reference ECG may be obtained and displayed on the screen. The time and the identity of the employee going to work or leaving the office may be merged and recorded, and this may be used as the employee's attendance information.

8 is a diagram illustrating a configuration of an authentication device and a learning device according to an embodiment.

The authentication device 801 according to an embodiment includes a sensor 802 , a processor 803 , and a memory 804 . The sensor 802 may detect the test ECG from the object, and the processor 803 may acquire the test ECG of the object through the sensor 802 . The processor 803 generates a feature vector for the test electrocardiogram based on the pre-learned dictionary, generates a classification result for the test electrocardiogram based on the feature vector and the pre-learned classifier, and generates a classification result for the test electrocardiogram based on the classification result. authentication can be performed. According to an embodiment, the memory 804 may record the pre-trained dictionary and the pre-trained classifier, and the processor 803 loads the information stored in the memory 804 to be used to perform authentication for the test electrocardiogram. can The embodiment described with reference to FIGS. 1 to 7 may be applied to the authentication device 801 , and thus the detailed description thereof will be omitted.

The learning device 805 according to an embodiment includes a sensor 806 , a processor 807 , and a memory 808 . The sensor 806 may detect the plurality of training ECGs, and the processor 807 may obtain the plurality of training ECGs through the sensor 806 . The processor 807 is configured to generate a plurality of feature vectors for the training ECGs by using category dictionaries corresponding to the plurality of categories, determine prediction categories of the training ECGs based on the feature vectors, the categories and Based on the prediction categories, category dictionaries and category classifiers may be updated alternately. According to one embodiment, the memory 808 may record category dictionaries and category classifiers, and the processor 807 alternately updates the category dictionaries and category classifiers recorded in the memory 808, so that the dictionary and the classifier can learn The embodiments described with reference to FIGS. 1 to 7 may be applied to the learning apparatus 803 , and thus detailed details thereof will be omitted.

Referring to FIG. 8 , the authentication device 801 and the learning device 805 may be distinguished as independent devices and may perform separate operations. In addition, the authentication device 801 and the learning device 805 may be integrated into one device 809 and applied to a smart device or the like. Although the authentication device 801 and the learning device 805 may mutually operate or operate separately as an independent subject, embodiments related thereto may be applied in various ways.

The embodiments described above may be implemented by a hardware component, a software component, and/or a combination of a hardware component and a software component. For example, the apparatus, methods and components described in the embodiments may include, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate (FPGA). array), a programmable logic unit (PLU), a microprocessor, or any other device capable of executing and responding to instructions, may be implemented using one or more general purpose or special purpose computers. The processing device may execute an operating system (OS) and one or more software applications running on the operating system. The processing device may also access, store, manipulate, process, and generate data in response to execution of the software. For convenience of understanding, although one processing device is sometimes described as being used, one of ordinary skill in the art will recognize that the processing device includes a plurality of processing elements and/or a plurality of types of processing elements. It can be seen that may include For example, the processing device may include a plurality of processors or one processor and one controller. Other processing configurations are also possible, such as parallel processors.

The software may comprise a computer program, code, instructions, or a combination of one or more thereof, which configures a processing device to operate as desired or is independently or collectively processed You can command the device. The software and/or data may be any kind of machine, component, physical device, virtual equipment, computer storage medium or device, to be interpreted by or to provide instructions or data to the processing device. , or may be permanently or temporarily embody in a transmitted signal wave. The software may be distributed over networked computer systems and stored or executed in a distributed manner. Software and data may be stored in one or more computer-readable recording media.

The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded in a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc. alone or in combination. The program instructions recorded on the medium may be specially designed and configured for the embodiment, or may be known and available to those skilled in the art of computer software. Examples of the computer-readable recording medium include magnetic media such as hard disks, floppy disks and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic such as floppy disks. - includes magneto-optical media, and hardware devices specially configured to store and execute program instructions, such as ROM, RAM, flash memory, and the like. Examples of program instructions include not only machine language codes such as those generated by a compiler, but also high-level language codes that can be executed by a computer using an interpreter or the like. The hardware devices described above may be configured to operate as one or more software modules to perform the operations of the embodiments, and vice versa.

As described above, although the embodiments have been described with reference to the limited drawings, those skilled in the art may apply various technical modifications and variations based on the above. For example, the described techniques are performed in a different order than the described method, and/or the described components of the system, structure, apparatus, circuit, etc. are combined or combined in a different form than the described method, or other components Or substituted or substituted by equivalents may achieve an appropriate result.

Therefore, other implementations, other embodiments, and equivalents to the claims are also within the scope of the following claims.

Claims (26)

receiving an electrocardiogram;
generating a feature vector for the electrocardiogram based on a pre-learned dictionary;
generating a classification result for the electrocardiogram based on the feature vector and a pre-learned classifier;
obtaining a classification result for the reference electrocardiogram; and
performing authentication of the electrocardiogram based on a voting mechanism for the classification result for the electrocardiogram and the classification result for the reference electrocardiogram;
including,
the reference electrocardiogram comprises at least one reference signal segment,
The step of performing the authentication based on the voting mechanism comprises:
generating, based on the at least one label for the electrocardiogram, a first distribution including a number of signal segments belonging to each of a plurality of categories;
obtaining a second distribution including the number of reference signal segments belonging to each of the plurality of categories;
calculating a similarity between the first distribution and the second distribution; and
performing the authentication by comparing the similarity and a predefined threshold value
containing,
Authentication method based on electrocardiogram.
According to claim 1,
The dictionary and the classifier are learned synchronously with each other,
Authentication method based on electrocardiogram.
According to claim 1,
the dictionary includes category dictionaries corresponding to a plurality of categories;
the classifier comprises category classifiers corresponding to the categories;
Authentication method based on electrocardiogram.
4. The method of claim 3,
The dictionary and the classifier are
A plurality of feature vectors for the training ECGs are generated using the category dictionaries, prediction categories of the training ECGs are determined based on the plurality of feature vectors, and the categories and the prediction categories By alternately updating the category dictionaries and the category classifier,
learned synchronously with each other,
Authentication method based on electrocardiogram.
According to claim 1,
The step of generating a feature vector for the electrocardiogram comprises:
determining coefficients of the plurality of category dictionaries to express the electrocardiogram as a linear combination of the plurality of category dictionaries included in the dictionary; and
generating the feature vector based on the coefficients
including,
Authentication method based on electrocardiogram.
According to claim 1,
The step of generating the feature vector comprises:
dividing the electrocardiogram into at least one signal segment;
generating at least one feature representation corresponding to the at least one signal segment;
generating at least one second characteristic indication by performing group sparse coding based on the dictionary on the at least one characteristic indication; and
generating the feature vector based on the at least one second feature indication;
containing,
Authentication method based on electrocardiogram.
7. The method of claim 6,
The characteristic indication is expressed as a linear combination of some of the plurality of category dictionaries included in the dictionary,
wherein the second characteristic indication includes coefficients of the some category dictionaries;
Authentication method based on electrocardiogram.
According to claim 1,
The step of generating a classification result for the electrocardiogram comprises:
generating at least one label corresponding to a classification result of at least one signal segment of the electrocardiogram based on the feature vector and the classifier; and
generating a classification result for the electrocardiogram based on the at least one label
containing,
Authentication method based on electrocardiogram.
delete delete According to claim 1,
The classification result for the reference ECG is obtained by applying the dictionary and the classifier to the reference ECG,
Authentication method based on electrocardiogram.
generating a plurality of feature vectors for the training ECGs by using category dictionaries corresponding to the plurality of categories;
determining prediction categories of the training electrocardiograms based on the feature vectors; and
Alternately updating the category dictionaries and category classifiers based on the categories and the prediction categories;
including,
Alternately updating the category dictionaries and the category classifiers comprises:
fixing the category classifiers and optimizing the category dictionaries to minimize the prediction categories and errors of the categories; and
fixing the optimized category dictionaries and optimizing the category classifiers to minimize the prediction categories and the errors of the categories;
including,
The optimization of the category dictionaries and the optimization of the category classifiers are repeated synchronously with each other,
Learning method for ECG-based authentication.
13. The method of claim 12,
The step of generating the feature vectors comprises:
dividing each training electrocardiogram of the training electrocardiograms into at least one signal segment;
generating at least one feature representation corresponding to the at least one signal segment;
generating at least one second feature mark by performing group sparse coding based on the category dictionaries on the at least one feature mark; and
generating the feature vectors based on the at least one second feature indication corresponding to the at least one signal segment;
containing,
Learning method for ECG-based authentication.
delete 13. The method of claim 12,
optimizing the category dictionaries includes optimizing the category dictionaries to minimize the error using a gradient descent algorithm,
wherein optimizing the category classifiers includes optimizing the category classifiers to minimize the error using the gradient descent algorithm.
Learning method for ECG-based authentication.
13. The method of claim 12,
terminating the update when the synchronously repeated optimizations satisfy a predefined condition; and
outputting the updated category dictionaries and category classifiers
further comprising,
The predefined condition is
at least one of a case in which the degree of iteration reaches a predefined order, a case in which the error is less than a predefined first value, and a case in which the amount of change of the error is smaller than a predefined second value,
Learning method for ECG-based authentication.
13. The method of claim 12,
initializing the category dictionaries based on the training electrocardiograms corresponding to the categories;
further comprising,
Learning method for ECG-based authentication.
18. The method of claim 17,
Initializing the category dictionaries includes:
preprocessing the training electrocardiograms;
generating a plurality of clusters by clustering the pre-processed training ECGs according to the categories; and
Generating the initialized category dictionaries by performing normalization on each centroid of the clusters
containing,
Learning method for ECG-based authentication.
19. The method of claim 18,
The pre-processing step is
dividing each training electrocardiogram of the training electrocardiograms into at least one signal segment;
extracting each feature of the at least one signal segment; and
performing dimensionality reduction processing on the respective features to generate respective feature representations for the at least one signal segment;
containing,
Learning method for ECG-based authentication.
20. The method of claim 19,
The dividing into at least one signal segment comprises:
filtering each training electrocardiogram;
detecting QRS waves for each of the filtered training electrocardiograms; and
generating the at least one signal segment based on the detection result;
containing,
Learning method for ECG-based authentication.
13. The method of claim 12,
initializing the category classifiers based on the feature vectors and the category dictionaries.
further comprising,
Learning method for ECG-based authentication.
22. The method of claim 21,
Initializing the category classifiers comprises:
generating a probability distribution of the categories for at least one signal segment using the feature vectors and the category dictionaries;
calculating a prediction error of the at least one signal segment based on the probability distribution and the actual category of the at least one signal segment;
obtaining, based on the prediction error, the prediction categories and category classifiers that minimize the errors of the categories; and
generating the initialized category classifiers as the obtained category classifiers;
containing,
Learning method for ECG-based authentication.
13. The method of claim 12,
generating a dictionary by merging the updated category dictionaries; and
generating a classifier by merging the updated category classifiers
further comprising,
Learning method for ECG-based authentication.
Claims 1 to 8, claims 11 to 13, and claims 15 to 23 of any one of claims for executing a program for executing a computer-readable recording medium recorded therein.
A feature vector for the electrocardiogram is generated based on a pre-learned dictionary, a classification result is generated for the electrocardiogram based on the feature vector and a pre-learned classifier, a classification result is obtained for a reference electrocardiogram, and the A processor for authenticating the ECG based on a voting mechanism for the classification result for the reference ECG and the classification result for the reference ECG
including,
the reference electrocardiogram comprises at least one reference signal segment,
The processor is
When performing the authentication for the electrocardiogram based on the voting mechanism,
generating a first distribution including the number of signal segments belonging to each of a plurality of categories based on the at least one label for the electrocardiogram, and a second distribution including the number of reference signal segments belonging to each of the plurality of categories obtaining two distributions, calculating a degree of similarity between the first distribution and the second distribution, and comparing the degree of similarity and a predefined threshold to perform the authentication;
An authentication device based on an electrocardiogram.
generating a plurality of feature vectors for the training ECGs using category dictionaries corresponding to the plurality of categories, determining prediction categories of the training ECGs based on the feature vectors, and determining the categories and the prediction category a processor that alternately updates the category dictionaries and category classifiers based on
including,
The processor is
When updating the category dictionaries and category classifiers alternately,
fixing the category classifiers, optimizing the category dictionaries to minimize the prediction categories and the error of the categories, fixing the optimized category dictionaries, and configuring the category classifiers to minimize the prediction categories and the error of the categories to optimize,
Learning device for ECG-based authentication.

Images