KR102056704B1 - A method and apparatus for classifying class using multi-layer neural network with weighted fuzzy membership function - Google Patents

Abstract

The present invention relates to a method and an apparatus for class classification using a deep neural network with a multi-layer weighted fuzzy membership function. According to an embodiment of the present invention, the method for class classification using a deep neural network with a multi-layer weighted fuzzy membership function comprises: (a) a step of inputting input features selected from input data into a first deep neural network with a weighted fuzzy membership function (NEWFM); (b) a step of classifying output features outputted by using the first NEWFM into a first class and a second class in accordance with a class of records of the input features; (c) a step of inputting output features included in the first class into a second NEWFM, and inputting output features included in the second class into a third NEWFM; (d) a step of using the second NEWFM to calculate first output data, and using the third NEWFM to calculate second output data; and (e) a step of using the first and second output data to calculate the accuracy of the class of the records of the input features.

Description

Method and apparatus for classifying class using multi-layer neural network with weighted fuzzy membership function}

The present invention relates to a class classification method and apparatus, and more particularly, to a class classification method and apparatus using a multi-layer weighted fuzzy membership function based deep neural network composed of a plurality of NEWFM.

Neural network is a structure adopted by computers to solve problems in a way similar to how humans handle problems through the brain. When a neuron as a mathematical model is interconnected to form a network, a neural network or artificial It is called Neural Network.

Neural networks have excellent parallelism because each neuron acts as an independent processor, and because information is distributed over many connections, faults do not affect the whole system even if a few neurons fail. tolerance), and the ability to learn about a given environment.

Because of these characteristics, it is used for problem solving in artificial intelligence, and is used in various fields such as disease recognition, text recognition, speech recognition, classification, diagnosis, and prediction.

In the related art, a disease recognition model generation method using artificial intelligence technology such as deep learning has been disclosed. In the conventional disease recognition model generation method, a feature is extracted from input data, and the infection is classified based on the learned result. However, the development of a plan for improving the accuracy of disease infection is insufficient.

[Patent Document 1] Korean Patent Registration No. 10-1950395

The present invention has been made to solve the above-described problem, and an object thereof is to provide a class classification method and apparatus using a multilayer neural network based on a multilayer weighted fuzzy membership function.

In addition, the present invention calculates the first output data and the second output data by separately inputting the output features included in the first class and the output features included in the second class divided into two parts to two NEWFM It is an object of the present invention to provide a class classification method and apparatus.

The objects of the present invention are not limited to the above-mentioned objects, and other objects that are not mentioned will be clearly understood from the following description.

In order to achieve the above objects, a class classification method using a multilayer-weighted fuzzy membership function based deep neural network according to an embodiment of the present invention (a) a first weighted fuzzy selected input features selected from the input data; Inputting to a membership function based neural network with weighted fuzzy membership function (NEWFM); (b) classifying output features output using the first NEWFM into a first class and a second class according to a class of records of the input features; (c) inputting output features included in the first class into a second NEWFM, and inputting output features included in the second class into a third NEWFM; (d) calculating first output data using the second NEWFM and calculating second output data using the third NEWFM; And (e) calculating an accuracy of a class of records of the input features using the first output data and the second output data.

In an embodiment, the class classification method using the multilayer weighted fuzzy membership function based deep neural network may include: before the step (a), dividing the input data into a plurality of division groups; Extracting voxels for input data included in each of the plurality of division groups; Calculating an average voxel from the voxels; Extracting a plurality of features from the average voxel; And selecting the input features among the plurality of input groups and the plurality of features according to the accuracy using a fourth NEWFM.

In an embodiment, the step (b) compares the classes of output features output by the first NEWFM based on the class of records of the input features, classifying the output features into a first class and a second class. It may comprise;

In an embodiment, the first NEWFM, the second NEWFM, and the third NEWFM are included in one of the learning units of the first learning layer, wherein the first output data and the second output data include: It may be input to one of the learning units of the two learning layers.

In an embodiment, the number of learning units of the first learning layer may be greater than the number of learning units of the second learning layer.

In an embodiment, the calculated first output data includes Takagi-Sugeno defuzzification values corresponding to output features included in the first class, wherein the calculated first output data includes: The second output data may include Takagi-Sugeno inverse purge values corresponding to output features included in the second class.

In an embodiment, the step (e) may include inputting the first output data and the second output data to a fifth NEWFM to calculate an accuracy of a class of records of the input characteristics.

In an embodiment, a classifier using a multilayer weighted fuzzy membership function based deep neural network may include input features selected from input data based on a first weighted fuzzy membership function deep neural network (NEFM). ), Classify the first and second classes according to the class of the records of the input features output using the first NEWFM, and input the output features included in the first class to the second NEWFM. Inputting output features included in the second class into a third NEWFM, calculating first output data using the second NEWFM, calculating second output data using the third NEWFM, And a controller configured to calculate an accuracy of a class of records of the input features by using first output data and the second output data.

In an embodiment, the control unit divides the input data into a plurality of division groups, extracts voxels for input data included in each of the plurality of division groups, and averages voxels from the voxels. ), Extract a plurality of features from the average voxel, and select the input features among the plurality of division groups and the plurality of features according to accuracy using a fourth NEWFM.

In an embodiment, the controller may classify the output features into a first class and a second class by comparing classes of output features output by the first NEWFM based on the class of records of the input features. .

In an embodiment, the first NEWFM, the second NEWFM, and the third NEWFM are included in one of the learning units of the first learning layer, wherein the first output data and the second output data include: It may be input to one of the learning units of the two learning layers.

In an embodiment, the number of learning units of the first learning layer may be greater than the number of learning units of the second learning layer.

In an embodiment, the calculated first output data includes Takagi-Sugeno defuzzification values corresponding to output features included in the first class, wherein the calculated first output data includes: The second output data may include Takagi-Sugeno inverse purge values corresponding to output features included in the second class.

In an embodiment, the controller may input the first output data and the second output data to a fifth NEWFM to calculate an accuracy of a class of records of the input characteristics.

Specific details for achieving the above objects will be apparent with reference to the embodiments to be described later in detail with the accompanying drawings.

However, the present invention is not limited to the embodiments disclosed below, but may be configured in a variety of different forms, to make the disclosure of the present invention complete and those of ordinary skill in the art to which the present invention belongs ( The following description is provided to fully inform the scope of the invention to the ordinary technician.

According to an embodiment of the present invention, the first output data and the second output data by separately inputting the output features included in the first class and the output features included in the second class into two NEWFMs divided into two parts. By calculating the respective values, the accuracy of class classification can be improved.

The effects of the present invention are not limited to the above-described effects, and the tentative effects expected by the technical features of the present invention will be clearly understood from the following description.

1 is a diagram illustrating a method of operating a classifier using a multilayer weighted fuzzy membership function based deep neural network according to an embodiment of the present invention.
2 is a diagram illustrating a process of selecting input features from input data according to an embodiment of the present invention.
3A is a diagram illustrating an example of a voxel of input data according to an embodiment of the present invention.
3B illustrates an example of selection of input features in accordance with one embodiment of the present invention.
4 is a diagram illustrating a class classification structure using a weighted fuzzy membership function based deep neural network according to an embodiment of the present invention.
5 is a diagram illustrating a structure of a learning unit according to an embodiment of the present invention.
FIG. 6 is a diagram illustrating a functional configuration of a classifier using a multilayer weighted fuzzy membership function based deep neural network according to an embodiment of the present invention.

The present invention may be modified in various ways and may have various embodiments. Specific embodiments are illustrated in the drawings and described in detail.

Various features of the invention disclosed in the claims may be better understood in view of the drawings and detailed description. The apparatus, methods, recipes, and various embodiments disclosed herein are provided for purposes of illustration. The disclosed structural and functional features are intended to enable those skilled in the art to specifically practice the various embodiments, and are not intended to limit the scope of the invention. The terms and phrases disclosed are intended to explain various features of the disclosed invention in an easy-to-understand manner, and are not intended to limit the scope of the invention.

In the following description of the present invention, if it is determined that the detailed description of the related known technology may unnecessarily obscure the subject matter of the present invention, the detailed description thereof will be omitted.

Hereinafter, a class classification method and apparatus using a multilayer weighted fuzzy membership function (NEWFM) based on a multilayer weighted fuzzy membership function will be described.

1 is a diagram illustrating a method of operating a classifier using a multilayer weighted fuzzy membership function based deep neural network according to an embodiment of the present invention.

Referring to FIG. 1, in step S101, input characteristics selected from input data are input to a first NEWFM.

In an embodiment, before step S101, the input data is divided into a plurality of division groups, the voxels for the input data included in each of the plurality of division groups are extracted, and the average voxels are extracted from the voxels. , A plurality of features are extracted from the average voxel, and the fourth NEWFM may be used to select the input features among the plurality of division groups and the plurality of features according to the accuracy.

In step S103, the output features output using the first NEWFM are classified into a first class and a second class according to the class of records of the input features. Specifically, the output features may be classified into a first class and a second class by comparing classes of output features output by the first NEWFM based on the class of records of the input features. For example, the first class may mean a patient with Alzheimer's disease, and the second class may mean a general person without Alzheimer's disease.

In step S105, the output features included in the first class are input to the second NEWFM, and the output features included in the second class are input to the third NEWFM. That is, output features included in the first class divided into two parts and output features included in the second class may be learned using two NEWFMs separately.

In one embodiment, the first NEWFM, the second NEWFM, and the third NEWFM may be included in one of the learning units of the first learning layer. Here, the learning unit 422 may be referred to as a class pattern extraction unit (CPEU) or a term having an equivalent technical meaning.

In step S107, the first output data is calculated using the second NEWFM, and the second output data is calculated using the third NEWFM. In one embodiment, the calculated first output data includes Takagi-Sugeno defuzzification values corresponding to output features included in the first class, and the second output data is It may include Takagi-Sugeno inverse purge values corresponding to output features included in the second class.

In an embodiment, the first output data and the second output data may be input to one of the learning units of the second learning layer. In this case, the number of learning units of the first learning layer may be greater than the number of learning units of the second learning layer.

In step S109, using the first output data and the second output data, calculating the accuracy of the class of records of the input features.

2 is a diagram illustrating a process of selecting input features from input data according to an embodiment of the present invention.

Referring to FIG. 2, the input data division step 210 may divide the input data into a plurality of division groups according to input characteristics. For example, when the input data is a brain image, the brain image may be divided into a plurality of brain regions according to a function.

The voxel extraction step 220 may calculate voxels for input data included in each division group, and calculate one average voxel from the extracted voxels. 3A, the basic unit of the input data may be a voxel 310 which is a small cube. In this case, each voxel 310 may have a plurality of values according to the time series of the input data. Here, T may represent the number of brain images included in one record of the input data. Therefore, one average voxel can be obtained from a plurality of voxels along the time series.

Feature extraction step 230 may extract a plurality of features from the average voxel. For example, the Haar wavelet feature can be extracted from the average voxel using the d3, d4, a4 coefficients. However, the method of extracting a plurality of features is not limited thereto.

The feature selection step 240 may select the input features among the plurality of division groups and the input features among the plurality of division groups according to the accuracy using the fourth NEWFM. In one embodiment, a portion of the plurality of division groups with high accuracy may be selected as the input group and among the plurality of features as the input features. Here, each input group having the selected input feature may be used as an input to a multilayer weighted fuzzy membership function based deep neural network to be described below.

For example, referring to FIG. 3B, among the 32 features for each of the 90 brain regions, the 32 NEW brain regions and the 20 input features having the highest accuracy may be selected using the fourth NEWFM 320.

In one embodiment, the selected input groups may be ordered and grouped according to accuracy. For example, 32 brain regions can be grouped into eight groups, each group consisting of four brain regions. In this case, group 1 may be grouped into two input groups Cingulum_mid_r and Caudate_r having the highest accuracy and two input groups Amygdala_l and Frontal_sup_orb_l having the lowest accuracy.

4 is a diagram illustrating a class classification structure using a weighted fuzzy membership function-based deep neural network 400 according to an embodiment of the present invention.

Referring to FIG. 4, the weighted fuzzy membership function based deep neural network 400 may include an input layer 410, a learning layer 412, and an output layer 414.

The input layer 410 may receive input features 420 for each input group selected from the input data.

The learning layer 412 may be composed of a plurality of learning units 422. In one embodiment, referring to FIG. 5, each learning unit 422 may include a first NEWFM 510, a second NEWFM 520, and a third NEWFM 530.

In this case, input features 420 for each input group selected from the input data may be input to the first NEWFM 510. In addition, according to the class of records of the input features 420, the output features output from the first NEWFM 510 may be classified into a first class and a second class.

Thereafter, the output features included in the first class may be input to the second NEWFM 520, and the output features included in the second class may be input to the third NEWFM 530. In addition, the first output data s1 may be calculated using the second NEWFM 520, and the second output data s2 may be calculated using the third NEWFM 530. Here, the calculated first output data and the second output data may be input to the fifth NEWFM 424 of the output layer 414.

The learning layer 412 may be composed of at least one layer. In FIG. 4, the learning layer 412 is composed of three layers, but the number of layers is not limited. In this case, for example, the first learning layer may include 32 learning units, and outputs of four learning units of the first learning layer may be input as inputs to one learning unit of the second learning layer.

In addition, the second learning layer includes eight learning units, and outputs of two learning units of the second learning layer may be input as inputs to one learning unit of the third learning layer.

In addition, the third learning layer may include four learning units, and an output of the learning unit of the third learning layer may be input to the fifth NEWFM 424 of the output layer 414. The output layer 414 may output the accuracy of the class of records of the input features 420 using the fifth NEWFM 424 to which the first output data and the second output data are input.

For example, the accuracy of each layer can be expressed as shown in Table 1 below. Referring to <Table 1>, it can be seen that more accurate results can be obtained in an output layer having a multilayer NEWFM structure than an input layer.

Input group Input layer
accuracy First learning layer
accuracy Second learning layer
accuracy Output layer
accuracy Group 1 73.3% 89.6% 93.7% 97.3% Group 2 76.9% 88.5% Group 3 72.8% 89.1% 95.6% Group 4 71.7% 88.4% Group 5 77.1% 90.2% 94.3% Group 6 76.1% 90.7% Group 7 73.9% 88.1% 96.1% Group 8 75.2% 89.6%

In addition, the results using the multilayer NEWFM according to the present invention and the results using the conventional CNN technology can be represented as shown in Table 2 below. Referring to Table 2, it can be seen that the result of higher accuracy is obtained when using the multilayer NEWFM according to the present invention than the conventional CNN.

division accuracy CNN 96.8% Multilayer NEWFM 97.3%

FIG. 6 is a diagram illustrating a functional configuration of a classifier 600 using a multilayer weighted fuzzy membership function based deep neural network according to an embodiment of the present invention.

Referring to FIG. 6, the classifier 600 using the multilayer weighted fuzzy membership function-based deep neural network may include a controller 610, a communication unit 620, and a storage unit 630.

The controller 610 may select input features by preprocessing the input data. In addition, the controller 610 inputs the selected input features to the first NEWFM, classifies the output features output by using the first NEWFM into first and second classes according to the class of records of the input features. Inputting the output features included in the first class to the second NEWFM to produce first output data, inputting the output features included in the second class into the third NEWFM to calculate the second output data, and calculating the calculated first output The data and the second output data may be used to calculate the accuracy of the class of records of the input features.

In one embodiment, the controller 610 may include at least one processor or a micro processor, or may be part of a processor. In addition, the controller 610 may be referred to as a communication processor (CP). The controller 610 may control each operation of the class classification apparatus 600 using the multilayer weighted fuzzy membership function based deep neural network.

The communication unit 620 may receive input data from the outside through a wired communication network or a wireless communication network.

In one embodiment, the communication unit 620 may include at least one of a wired communication module and a wireless communication module. All or part of the communication unit 620 may be referred to as a 'transmitter', 'receiver' or 'transceiver'.

The storage unit 630 may store input data. In one embodiment, the storage unit 630 may store input data transmitted via offline means or a wireless communication network.

In an embodiment, the storage unit 630 may be configured as a volatile memory, a nonvolatile memory, or a combination of the volatile memory and the nonvolatile memory. In addition, the storage unit 630 may provide the stored data at the request of the controller 610.

The above description is merely illustrative of the technical spirit of the present invention, and various changes and modifications may be made by those skilled in the art without departing from the essential characteristics of the present invention.

Therefore, the embodiments disclosed herein are not intended to limit the technical spirit of the present invention, but to explain, and the scope of the present invention is not limited by these embodiments.

The scope of protection of the present invention should be interpreted by the claims, and all technical ideas within the scope equivalent thereto should be understood as being included in the scope of the present invention.

210: input data partitioning step
220: voxel extraction step
230: feature extraction step
240: feature selection step
310: voxel
320: fourth NEWFM
410: input layer
412: learning layer
414: output layer
420: input features
422: learning unit
424: Fifth NEWFM
510: First NEWFM
520: second NEWFM
530: Third NEWFM
600: Classifier using a multilayer weighted fuzzy membership function based deep neural network
610: control unit
620: communication unit
630: storage unit

Claims (14)

Dividing the input data into a plurality of division groups;
Extracting voxels for input data included in each of the plurality of division groups;
Calculating an average voxel from the voxels;
Extracting a plurality of features from the average voxel; And
Selecting input features among the plurality of division groups and a plurality of features according to accuracy using a fourth NEWFM;
Inputting the input features into a first neural network with weighted fuzzy membership function (NEWFM);
Classifying output features output using the first NEWFM into a first class and a second class according to a class of records of the input features;
Inputting output features included in the first class into a second NEWFM and inputting output features included in the second class into a third NEWFM;
Calculating first output data using the second NEWFM and calculating second output data using the third NEWFM; And
Calculating an accuracy of a class of records of the input features using the first output data and the second output data;
Containing,
An operating method of a processor for class classification using a multilayer weighted fuzzy membership function based deep neural network.
delete The method of claim 1,
The classifying step,
Comparing the classes of output features output by the first NEWFM based on the class of records of the input features, classifying the output features into a first class and a second class;
Containing,
An operating method of a processor for class classification using a multilayer weighted fuzzy membership function based deep neural network.
The method of claim 1,
The first NEWFM, the second NEWFM, and the third NEWFM are included in one of the learning units of the first learning layer,
The first output data and the second output data are input to one of the learning units of the second learning layer,
An operating method of a processor for class classification using a multilayer weighted fuzzy membership function based deep neural network.
The method of claim 4, wherein
The number of learning units of the first learning layer is greater than the number of learning units of the second learning layer,
An operating method of a processor for class classification using a multilayer weighted fuzzy membership function based deep neural network.
The method of claim 1,
The calculated first output data includes Takagi-Sugeno defuzzification values corresponding to output features included in the first class,
The calculated second output data includes Takagi-Sugeno inverse purge values corresponding to output features included in the second class.
An operating method of a processor for class classification using a multilayer weighted fuzzy membership function based deep neural network.
The method of claim 1,
The calculating step,
Inputting the first output data and the second output data into a fifth NEWFM to calculate an accuracy of the class of records of the input characteristics;
Containing,
An operating method of a processor for class classification using a multilayer weighted fuzzy membership function based deep neural network.
Divide the input data into a plurality of division groups,
Extracting voxels for input data included in each of the plurality of division groups,
Calculating an average voxel from the voxels,
Extract a plurality of features from the average voxel,
Select input features among the plurality of division groups and the input features among the plurality of division groups according to accuracy using a fourth NEWFM;
Input the input features to a first neural network with weighted fuzzy membership function (NEWFM),
Categorize into a first class and a second class according to a class of records of the input features output using the first NEWFM,
Input output features included in the first class to a second NEWFM, input output features included in the second class to a third NEWFM,
Calculating first output data using the second NEWFM, calculating second output data using the third NEWFM,
A processor that calculates the accuracy of the class of records of the input features using the first output data and the second output data;
Containing,
Classifier using multi-layer weighted fuzzy membership function-based deep neural network.
delete The method of claim 8,
The processor,
Compare the classes of output features output by the first NEWFM based on the class of records of the input features, classifying the output features into a first class and a second class,
Classifier using multi-layer weighted fuzzy membership function-based deep neural network.
The method of claim 8,
The first NEWFM, the second NEWFM, and the third NEWFM are included in one of the learning units of the first learning layer,
The first output data and the second output data are input to one of the learning units of the second learning layer,
Classifier using multi-layer weighted fuzzy membership function-based deep neural network.
The method of claim 11,
The number of learning units of the first learning layer is greater than the number of learning units of the second learning layer,
Classifier using multi-layer weighted fuzzy membership function-based deep neural network.
The method of claim 8,
The calculated first output data includes Takagi-Sugeno defuzzification values corresponding to output features included in the first class,
The calculated second output data includes Takagi-Sugeno inverse purge values corresponding to output features included in the second class.
Classifier using multi-layer weighted fuzzy membership function-based deep neural network.
The method of claim 8,
The processor,
Inputting the first output data and the second output data into a fifth NEWFM to calculate an accuracy of the class of records of the input characteristics,
Classifier using multi-layer weighted fuzzy membership function-based deep neural network.

Images