KR20060082677A - A biometics system and method using electrocardiogram - Google Patents

Abstract

The present invention provides a biometric system and method using an electrocardiogram for performing user recognition using a lead III electrocardiogram waveform. The biometric method includes: a first input of a lead III electrocardiogram waveform of a corresponding user to remove baseline noise; step; A second step of determining a basic shape of each waveform according to a user after determining the basic shape of each waveform; Checking a vertex of each waveform to determine a change in the basic shape of the waveform based on the number of vertices to classify the waveform; Extracting a feature of the corresponding waveform for user recognition when the waveform is classified; A fifth step of generating recognition information for recognizing a user by a predetermined weight based on the extracted feature; A sixth step of performing a user recognition based on the recognition information.

Biometrics, electrocardiogram, lead III waveform, feature extraction, neural network

Description

Biometric system and method using ECG {A BIOMETICS SYSTEM AND METHOD USING ELECTROCARDIOGRAM}             

1 is a diagram showing an example of a general electrocardiogram measurement method.

2 is a block diagram of a biometric system using an electrocardiogram according to the present invention.

3 is a detailed configuration diagram of the feature shape extraction unit of FIG. 2;

4 is an operational flowchart of the present invention.

5 is a diagram illustrating a baseline noise cancellation signal using a conventional IIR band filter.

6 is a diagram for explaining distortion of an ECG waveform by a conventional IIR band filter.

Figure 7 illustrates the basic form of a waveform defined in the present invention.

8 is a view showing a change in the UTL and LTL in accordance with the increase TC in the present invention.

9 shows an example of vertex analysis in the present invention.

10 is an exemplary diagram of interval extraction of waveforms for Type 1 according to the present invention;

11 is a view showing the details of the waveform defined in the present invention.

12 is a view showing a feature extraction result in the present invention.

Figure 13 is a schematic structural diagram of a neural network in the present invention.

14 illustrates a test pattern environment of the present invention.                 

Figure 15 shows a representative waveform for the experiment of the present invention.

Fig. 16 is a diagram showing a test waveform form in the present invention.

17 is a table which shows the output result of the test pattern in this invention.

<Description of the symbols for the main parts of the drawings>

100: input unit 200: feature shape extraction unit

210: baseline noise canceller 220: waveform classifier

230: feature extraction unit 300: neural network

      400: memory 500: control unit

600: output unit

The present invention relates to a biometric system and method using an electrocardiogram. More particularly, the present invention relates to a biometric system and method using an electrocardiogram for performing user recognition using a Limb Lead III waveform among electrocardiogram (ECG) waveforms. .

In general, biomechanics refers to a technology that utilizes unique characteristics of a person as a unit of identity based on human physiological or behavioral characteristics. In other words, it is a technical field that uses human biological characteristics for identification through an automated device.                         

Biometric technology is safer and more convenient than conventional password or password identification methods because the recognized person must actually exist at the time of recognition and eliminate the need to remember the password or carry a passkey. The bio-parts used for biometrics are diverse such as fingerprint, retina, iris, face, hand, vein, voice, signature, body smell, DNA, etc. However, the most popular parts are bio-prints using fingerprint, voice, iris, and face. It is awareness.

In particular, fingerprint recognition is one of the oldest technologies that has been used successfully in many applications. Fingerprints are raised in the sweat glands to form a constant flow. Retina and Iris are more reliable in terms of reliability and stability due to their unique characteristics that are different from person to person and that do not change for life. It is estimated to be higher than the vein, facial, and the like, and is used as an efficient personal authentication method.

In general, conventional fingerprint recognition methods include thinning, Fourier transform or wavelet transform in frequency space, or neural network or fuzzy logic. Most of them require preprocessing and postprocessing to reduce noise. Will be done. In addition, a syntactic method, such as syntactic, statistical, rule-based, neural network configuration is commonly used between them, chain code or ridge tracking is used as a characteristic method. In the conventional fingerprint recognition algorithm based on the feature point, if the resolution of the fingerprint image, the pixel value distribution characteristic, and the sensing method of the fingerprint input sensor are different, even the fingerprint image of the same person generates geometrical deformations on the extracted feature points to provide proper recognition. There is a problem that cannot be extended, and the ridge number information extraction and use method is vulnerable to fingerprint rotation and local deformation, and has a problem of poor reproducibility.

Retinal scanning techniques evaluate vascular patterns in the human retina. A disadvantage of this technique is that the porch pattern may change over time when alcohol is included in the blood or when the glasses or contact lenses are used irregularly. In addition, the user may feel uncomfortable with the eye being illuminated while retinal scanning is performed, and the eye may be contaminated by contact with the scanning device.

The use of speech patterns as identification features has the disadvantage of widespread changes in human speech over time, the presence of background noise during the evaluation, and the possibility of forgers to deceive the system by recording the voice of an authorized individual. .

In addition, face recognition technology occupies an important position in the field of image recognition and machine vision. Recently, more active research is being conducted as a next generation biometric algorithm for human interface, but face recognition is also modifiable. There is this.

On the other hand, today, the so-called "ubiquitous environment", which accesses the network and enjoys the Internet and multimedia regardless of time and place anytime and anywhere, has become an issue. At the same time, the movement to check the disease and health status and cope with the danger situation in real time by integrating it into medical equipment is becoming an essential element in the aging society. To this end, development of personalized medical devices should be a priority.                         

In general, electrocardiogram is a curve showing and recording the active current according to the contraction of the heart, and amplifies the minute current of the heart muscle, so that each person has different repeatable characteristics, and it is impossible to modulate, non-irritating, and even a single channel is biometric. Because of this possibility, it is possible to develop into the next generation biometric technology. The electrocardiogram is the most basic diagnostic tool used for diagnosing heart disease, and the lead II waveform is mainly used for diagnosing heart disease.

As shown in FIG. 1, the electrocardiogram is based on the measurement site or the number of electrodes. In addition to the standard induction of both hands (first induction), right hand and left foot (second induction), left hand and left foot (third induction), chest induction, unipolar induction, etc. There is this.

ECG waveforms vary slightly from person to person, and even in the same person, there are significant differences in body condition, measurement environment, and in particular, noise. Therefore, it is not easy to define standardized rules for accurate diagnosis, and it is difficult for the individual to make accurate diagnosis with the current ECG waveform test.

For automated personalized diagnosis, biometric technology is essential. The study for the automatic diagnosis of heart disease has been steadily progressed up to now, and the diagnosis method considering the age and gender of the measurer has been studied in part, but the diagnosis method considering the individual characteristics has not been developed.

Accordingly, the present invention has been made in view of the above, and the present invention cannot extract the characteristics of the lead III (third induction) waveform having relatively superior reproducibility among the electrocardiogram waveforms and perform user recognition using neural networks. In addition, the purpose of the present invention is to provide a biometric system and method using an electrocardiogram considering individual characteristics.

A biometric system using an electrocardiogram according to the present invention for achieving the above object comprises: feature shape extraction means for removing noise from an electrocardiogram waveform input through an input means and then extracting a feature for user recognition; A neural network for determining a relationship between each feature and a user as a weight value by learning the feature values extracted by the feature shape extracting means, and generating recognition information for recognizing the corresponding user using the weight value; Memory means for storing a weight value obtained from the neural network; And control means for recognizing the corresponding user according to the recognition information from the neural network.

The feature shape extracting means may include: a baseline noise canceling unit configured to remove baseline noise of an electrocardiogram waveform of a user input through the input means; A waveform classifier classifying the ECG waveform from which the baseline noise is removed through the baseline noise canceller into three basic types according to the shape thereof, and determining a basic shape of the entire waveform according to a user; And a feature extractor for extracting a predetermined number of features to be used as information for user recognition with respect to the corresponding waveform whose noise is removed from the baseline noise remover and the waveform classifier and whose basic shape is determined. .

The baseline noise removing unit separates and extracts waveforms based on the peak point to remove the baseline noise by using the average value of each waveform as a new baseline value, and the waveform classifier determines the basic shape of the entire waveform according to the user. The basic shape of the entire waveform is compared with the basic shape of each waveform, and the waveform having a different basic shape is judged to be removed as a signal containing noise, and after determining the basic shape of the entire waveform, the vertex of each waveform is determined. If the basic shape is different by determining the change in the basic shape of the waveform based on the number of vertices, the corresponding waveform is determined to be removed by determining that the signal includes noise.

A biometric method using an electrocardiogram according to the present invention for achieving the above object includes a first step of receiving an electrocardiogram waveform of a corresponding user and removing baseline noise thereof; A second step of determining a basic shape of each waveform according to a user after determining the basic shape of each waveform; Checking a vertex of each waveform to determine a change in the basic shape of the waveform based on the number of vertices to classify the waveform; Extracting a feature of the corresponding waveform for user recognition when the waveform is classified; A fifth step of generating recognition information for recognizing a user by a predetermined weight based on the extracted feature; And a sixth step of performing user recognition on the basis of the recognition information.

Hereinafter, the present invention will be described in more detail with reference to the accompanying drawings. However, the following examples are merely to illustrate the present invention is not limited to the contents of the present invention.

Figure 2 shows a block diagram of a biometric system using an electrocardiogram according to the present invention.

As shown, after extracting the noise from the input unit 100 receiving the lead III electrocardiogram waveform of the user, the lead III electrocardiogram waveform input through the input unit 100, feature extraction for feature extraction for user recognition By the learning of the feature values extracted by the feature shape extracting unit 200 and 200, the relationship between each feature and the user is determined as a weight value, and the user can be recognized using the weight. The neural network 300 generating recognition information, the memory 400 in which the weight value obtained from the neural network 300 is stored, and the control unit 500 performing recognition of the user according to the recognition information from the neural network 300. ), An output for displaying the overall state of the biometric system using the electrocardiogram under the control of the controller 500 and a message according to the user recognition result from the controller 500 The unit 600 is configured.

As shown in FIG. 3, the feature shape extractor 200 separately extracts waveforms based on the peak point of the ECG waveform of the lead III of the user input through the input unit 100 to extract the average value of each waveform. The baseline noise canceller 210 removes baseline noise using a new baseline value, and the lead III electrocardiogram waveform from which the baseline noise is removed through the baseline noise canceller 210 according to its shape. After classifying into basic form, it removes the signal containing noise while determining the basic form of the entire waveform according to the user.After determining the basic form of the waveform, it checks the vertices of each waveform and then applies them based on the number of vertices. The waveform classifier 220 removes the noise of the waveform, and the baseline noise remover 210 and the waveform classifier 220 remove the noise, and the basic shape of the waveform is determined. The feature extractor 230 extracts seventeen features to be used as information for user recognition by using an interval and a detail shape of the waveform.

The operation of the biometric system using the electrocardiogram according to the present invention configured as described above will be described with the flowchart of FIG. 4.

First, in the present invention, in order to extract a feature for user recognition from an electrocardiogram waveform, the removal of noise components must be prioritized, and the noise components are the baseline noise removal of the waveform, the noise removal through the type determination of the waveform, and the vertex analysis. It is removed through three processes.

That is, when the user's lead III electrocardiogram waveform is input through the input unit 100 (S101), the feature shape extraction unit 200 is input to extract the feature shape.

To extract the feature shape, first, baseline noise removal is performed through the baseline noise canceller 210 (S103).

Conventionally, the 1 ~ 50 [Hz] IIR band filter has been frequently used to remove the baseline noise. Although the band filter is effective in removing noise when there is a baseline variation due to noise as shown in FIG. 5B, as shown in FIG. 6, a vertex and a signal interval, which are important features of a waveform, may be distorted.

Therefore, the present invention proposes a dynamic baseline method. This algorithm separates and extracts waveforms based on the peak points of consecutive ECG waveforms and uses the average value of each waveform as a new baseline value. In this case, since the baseline adjusted for each waveform is used for the variation of the baseline due to noise, the shape of the original signal can be maintained without distortion through noise removal.

After removing the baseline noise through the baseline noise canceller 210, the waveform classifier 220 determines the basic shape of the waveform to classify the waveform.

That is, the signal from which the baseline noise has been removed is classified into three base types as shown in FIG. 7. The reason for classifying the waveform into three basic forms is that there are differences in feature extraction methods for each waveform. For example, since type 3 does not have a signal falling down like type 1 and type 2, the method of determining the frequency of signal spacing of the waveform is different.

In addition, since the waveform shape of the measurement person does not change except noise or body change, the signal whose basic shape of the waveform changes during the measurement may be determined as a signal including a noise component and may be removed from the information for recognition.

In order to classify the basic shape of the waveform, the upper threshold line (UTL) and the lower threshold line (LTL) were defined as Equations (1) and (2) below (S105). These values serve as reference points for extracting waveform features such as vertices and intervals.

         UTL = Baseline Value + TC × (Maximum Peak Value + Baseline Value) (1)

  LTL = Baseline Value-┃ UTL-Baseline Value ┃ (2)

Here, the TC (Threshold Coefficient) has a value of 0.3 to 0.9. The initial value of TC is 0.3. Then, the TC value is increased by 0.1, and the waveform is examined to determine whether a change in the waveform type occurs. As the TC increases, the distance between the UTL and the LTL becomes wider as shown in FIG. 8.

UTL does not exceed the maximum peak point and uses the TC value at the point when the basic shape of the waveform is finally changed.

When the basic shape of each waveform is determined, the basic shape of the entire waveform according to the user is determined (S107 and S109). The basic shape of the determined total waveform is compared with the basic shape of each waveform, and the waveform having a different basic shape is determined as a signal containing noise and is removed.

Once the basic shape of the waveform is determined, the vertices of each waveform are examined. Vertex extraction checks for sign changes by taking derivatives from two points sampled from the moment the waveform crosses the threshold line (TL).

As shown in FIG. 9, the differential change rate sign of a1 is '+', and the differential change rate sign of a2 is '-'. The point (b) where the signs of the two differential rates change is recognized as the vertex. The number of vertices obtained here is compared with the average number of vertices of the immediately preceding waveform. If the difference between the vertices is 4 or more, the TC is increased by 0.1, and then the waveform classifier 220 classifies the waveform. When the basic shape is different from the basic shape change of the waveform, the waveform is judged as a signal including noise and removed (S111, S113).

The signal from which the noise has been removed is extracted by the feature extractor 230 using seventeen features using the detailed shape and the interval of the waveform (S115).                     

The spacing of the waveforms varies depending on the basic shape of the waveform. 10 is an example of interval extraction for a waveform of type1. FIG. 10A is based on the UTL, FIG. 10B is based on the LTL, and FIG. 10C is an interval of the waveform extracted from the base line. However, in the case of Type 3, the interval as shown in FIG. 10B cannot be obtained, and only the intervals having the shapes of FIGS. 10A and 10C are extracted as characteristics of the waveform.

Then, the detailed shape of the waveform is extracted.

The detailed shape of the waveform is used as a shape characteristic of the signal by inspecting each of the signals above and below the baseline with respect to the shape of the waveform as a whole. The waveform shape was defined as eight in total as shown in FIG.

(A) to (c) of FIG. 11 define the upper shape of the waveform with respect to the baseline, and FIG. 11 (d) to (f) define the lower shape of the waveform with respect to the baseline. (G) and (h) of FIG. 11 are shapes for determining whether the position of the minimum point is in front of or behind the maximum peak point, and extracting features from the ECG signals of the measurers having the type 3 shape. useful.

For example, to classify the waveform shown in FIG. 6 for feature extraction, the basic shape corresponds to type 3 of FIG. 7, and the detailed shape corresponds to FIG. 11C. And since the minimum point is to the left of the maximum peak point corresponds to FIG. 11G.

In the present invention, a total of 17 features were extracted as a method of classifying waveform intervals and detailed shapes. FIG. 12 is a result of extracting 17 features of the waveform shown in FIG. 6. Item a in FIG. 12 is an average value of the basic shape of the waveform. If there is a record in which some signals are noisy and the waveform is determined differently, the value is displayed as a decimal point. Item b expresses the type value from item a as an integer. Item c is the number of vertices, and item d and item e represent the magnitude from baseline to maximum peak point and the value from baseline to minimum peak point, respectively. Item f represents a ratio value of item d and item e, items g through i are interval values of the waveform defined in FIG. 10, and items j through q include information on eight detailed shapes defined in FIG. .

As such, when the feature is extracted, it is input to the neural network 300 through the control unit 500, and the neural network 300 may recognize the user by a weight already assigned by learning and stored in the memory 400. The recognition information may be generated (S117). Based on the recognition information, the control unit 500 performs recognition of the user (S119), and a message according to the user recognition result is output to the output unit 600 under the control of the control unit 500.

Here, the neural network 300 will be briefly described.

When the 17 features are extracted by the feature shape extractor 200, the features are input into the neural network 300, and when the feature is input, the feature is called a target value. When a person named B, this feature is named C, etc., the target value is set and the neural network 300 stores the relationship between each feature and the person as a weight value in the memory 400.

The weight is a value indicating the degree of "what characteristics are important when comparing people", and the trained neural network 300 inputs 17 newly measured ECG characteristics of the unknown person D as a test (test). When the instruction to insert and test is applied to the input of the weights that have already been learned to the input of the currently input features to inform the user of the characteristics closest to the user recognition information, the structure of the neural network 300 is shown in FIG. As shown, it consists of an input layer, a hidden layer, and an output layer. In the present invention, since there are 17 features, there are 17 inputs and the output is the number of people.

Next, look at the experimental results of the present invention.

We experimented with 7 users for user recognition using ECG. For user recognition, as shown in FIG. 12, the waveform of 17 items was used as the input of the neural network 300, and the output layer was configured by allocating one neuron representing each user. And in order to optimize the transfer function of BP neural networks, we tested three types of functions (log-sigmoid, tan-sigmoid, linear-transfer function) and found that as shown in Figure 13, the hidden layer log-sigmoid and output layer log-sigmoid The highest recognition rate was found in the combination of functions.

Meanwhile, in FIG. 12, the meaning of the end portion 00100010 of FIG. 11 is the upper part of the waveform from a to c in FIG. 11, the lower part of the waveform to d to f from FIG. 11, and the g to h are the positions of the minimum values in Type 3. j to l is the upper part of the waveform, n to o is the lower part of the waveform, p to q is the position of the minimum value in Type 3, and the upper part of the waveform is up to (j to l). It is 001 because there is a vertex on the right side of the peak (Peak point), and it is 000 from n to o because there is no bottom part of the waveform, and because it is type 3, the minimum value of the waveform is 10 on the left side of the vertex. Therefore, the feature is extracted as 00100010.

 Next, in the experiment of the present invention, 58 data were extracted from seven electrocardiogram waveforms, 33 patterns were used for learning the neural network 300, and the remaining 25 patterns were used as test patterns. There were two major tests.

First, part of the measured ECG signal for 5 minutes was used for training, and the remainder was used for testing. In this case, there was very little variation in the measurement environment and the physical changes of the measurer, resulting in 100% recognition rate.

In the second method, only signals measured in the steady state were used for learning, and the test waveform was measured by using an ECG after artificially changing the body. In order to generate an artificial user change, the ECG measurement result after physical stress through exercise such as ingestion of coffee, tobacco and alcohol, and up and down stairs were used. 14 is a graph of 25 test environments.

FIG. 15 is a representative waveform shape of seven people, and FIG. 16 is an example of six test pattern waveform shapes. And the table of FIG. 17 is an output result of the test pattern of FIG. Referring to the table of FIG. 17, seven users were correctly recognized from the test patterns 1 to 4, but the test pattern 5 and the test pattern 6 were not recognized. Test pattern 5 was user 2 but was recognized as user 3. The reason for this is that the waveform type, the number of vertices and the interval of the waveform are very similar to the user 3. The test pattern 6 is a waveform measured by applying artificial stress to the heart by severe movement (step up and down). In this case, feature extraction failed due to the noise and rhythm change of the waveform. Among all 25 test patterns, all test patterns except test patterns 5 and 6 of FIG. 17 were recognized.

As described above, although described with reference to a preferred embodiment of the present invention, those skilled in the art various modifications of the present invention without departing from the spirit and scope of the invention described in the claims below Or it can be changed.

As described above, the biometric system and method using the electrocardiogram according to the present invention extracts the characteristics of the lead III waveform having a relatively excellent reproducibility of the electrocardiogram waveform and performs user recognition using a neural network. It can be used as a next-generation biometric technology because it is measured without intention, and it can be used as a technology that can automatically recognize the measurer and make a customized diagnosis in a situation where the measurer does not recognize in the ubiquitous environment. In addition, it can be applied to personalized medical services such as personal data management and metabolism management, and can be developed for military use and can be used to identify the exhibition location through soldier biometrics and to check whether gunshot wounds or life and death are without a separate interface.

Claims (15)

In the biometric system using the electrocardiogram, Feature shape extraction means for removing noise from an electrocardiogram waveform input through the input means, and then performing feature extraction for user recognition; A neural network for determining a relationship between each feature and a user as a weight value by learning the feature values extracted by the feature shape extracting means, and generating recognition information for recognizing the corresponding user using the weight value; Memory means for storing a weight value obtained from the neural network; And a control means for recognizing the user according to the recognition information from the neural network. The method of claim 1, wherein the feature shape extraction means A baseline noise removing unit for removing baseline noise of an electrocardiogram waveform of a user input through the input means; A waveform classifier classifying the ECG waveform from which the baseline noise is removed through the baseline noise canceller into three basic types according to the shape thereof, and determining a basic shape of the entire waveform according to a user; And a feature extractor for extracting a predetermined number of features to be used as information for user recognition with respect to the corresponding waveform in which the basic shape of the waveform is determined while the noise is removed through the baseline noise canceller and the waveform classifier. Biometric system using electrocardiogram. The EKG waveform of claim 1, wherein the ECG waveform is Biometric system using an electrocardiogram, characterized in that the lead III electrocardiogram waveform. The method of claim 2, wherein the baseline noise canceller A biometric system using an electrocardiogram, which extracts waveforms based on peak points and removes baseline noise by using a mean value of each waveform as a new baseline value. 3. The waveform classifier of claim 2, wherein the waveform classifier After determining the basic shape of the entire waveform according to the user, the basic shape of the determined total waveform is compared again with the basic shape of each waveform, and the waveforms having different basic shapes are judged to be removed as a signal containing noise. Live chain system using electrocardiogram. 6. The waveform classifier of claim 5, wherein the waveform classifier After determining the basic shape of the entire waveform, by examining the vertex of each waveform to determine the change in the basic shape of the waveform based on the number of vertices, if the basic shape is different, the corresponding waveform is determined to remove the signal containing noise Biometric system using an electrocardiogram, characterized in that. The method of claim 2, wherein the feature extraction unit A biometric system using an electrocardiogram, characterized by extracting a feature to be used as information for user recognition using a spacing and a detailed shape of a waveform. A first step of receiving an electrocardiogram waveform of the user and removing baseline noise thereof; A second step of determining a basic shape of each waveform according to a user after determining the basic shape of each waveform; Checking a vertex of each waveform to determine a change in the basic shape of the waveform based on the number of vertices to classify the waveform; Extracting a feature of the corresponding waveform for user recognition when the waveform is classified; A fifth step of generating recognition information for recognizing a user by a predetermined weight based on the extracted feature; And a sixth step of performing user recognition on the basis of the recognition information. The EKG waveform of claim 8, wherein the ECG waveform is Biometric method using an electrocardiogram, characterized in that the lead III ECG waveform. 9. The method of claim 8, wherein the noise cancellation in the first step is A method of biometric recognition using an electrocardiogram, characterized by removing the baseline noise by separating and extracting waveforms based on peak points of an electrocardiogram waveform and using the average value of each waveform as a new baseline value. The method of claim 8, wherein the second step And comparing the basic shape of the determined whole waveform with the basic shape of each waveform again while determining that the waveform having a different basic shape is determined as noise and removing the basic shape of the entire waveform. The method of claim 8, wherein the third step The vertex takes a derivative at two points sampled from the moment when the waveform crosses the TL (Threshold Line), checks the variation of the sign, and recognizes the point where the sign of the differential change rate changes as a vertex. Recognition method. 13. The method of claim 12, wherein the number of vertices of the corresponding waveform is compared with the average number of vertices of the immediately preceding waveform. The biometric method using an electrocardiogram, characterized in that the waveform is determined to be eliminated when it is different from each other. 9. The method of claim 8, wherein the feature extraction in the fourth step is Biometric method using an electrocardiogram, characterized in that the extraction using the interval and detail of the waveform. The method of claim 8, wherein the weight is Biometric method using an electrocardiogram characterized in that the relationship between each feature and the user by learning the extracted feature values.

Images