KR102193952B1 - Method of learning and inferring based on convolution neural network for minimizing clutter in passive sonar signal - Google Patents

Abstract

In accordance with the present invention, a convolutional neural network-based inference method is provided for minimizing sound gram clutter. The method includes a convolutional neural network configuration step of configuring a convolutional neural network for removing clutter from the sound tomgram, performed by a control unit; A learning data generation step of generating learning data by using sound gram data; And a sonicgram inference step of inferring a sound gram by removing sound gram clutter and detecting a tonal component through the learned convolutional neural network structure and learning weight, and automatically removing the clutter to detect the tonal component. It is possible to provide a method for detecting and inferencing soundgrams through improved accuracy and automation.

Description

{Method of learning and inferring based on convolution neural network for minimizing clutter in passive sonar signal}

The present invention relates to a method for detecting and inferring a tomgram by removing clutter from a tomgram. In more detail, it relates to a convolutional neural network-based learning and inference method for minimizing the tomgram clutter.

In a marine environment, underwater and surface ships can be detected/identified by using a sonar utilizing the transmission characteristics of acoustic signals.

In this regard, time-frequency analysis such as LOFAR (LOw Frequency Analysis and Recording) or DEMON (Detection of Envelope Modulation On Noise) after multiple signal processing processes after recording acoustic signals of the ocean using an underwater earpiece/ Using the expression method, a sound probe is generated that can visually check the sound signal. Thereafter, the tonal component is detected by analyzing the sound track with the human eye, and the identification process can be performed by comparing the frequency position with known underwater and suspicious vessels.

However, such a manual analysis/identification process has a problem in that the difficulty in the identification process is very high due to the clutter mixed in the sound tomgram.

Therefore, there is a need for a method of detecting and inferencing a tonal gram through automation and improving accuracy of tonal component detection by automatically removing clutter.

An object of the present invention is to provide a method for detecting and inferencing a tonal gram through automatic removal of clutter, improving accuracy of tonal component detection, and automating.

In addition, the object of the present invention is to It is to provide a convolutional neural network-based learning and inference method for minimizing clutter and a convolutional neural network-based learning and inference device using the same.

In order to solve the above problems, a convolutional neural network-based inference method is provided for minimizing sonar gram clutter according to the present invention. The method includes a convolutional neural network configuration step of configuring a convolutional neural network for removing clutter from the sound tomgram, performed by a control unit; A learning data generation step of generating learning data by using sound gram data; And a sonicgram inference step of inferring a sound gram by removing sound gram clutter and detecting a tonal component through the learned convolutional neural network structure and learning weight, and automatically removing the clutter to detect the tonal component. It is possible to provide a method for detecting and inferencing soundgrams through improved accuracy and automation.

In one embodiment, the structure of the convolutional neural network for identification of a tonogram includes a convolutional layer, an activation layer, and a batch normalization layer excluding a fully connected layer.

In one embodiment, the convolutional neural network structure is, when the convolutional layer has a kernel size of (rowsize, colsize), a kernel size that satisfies the condition of rowsize≥colsize in order to reflect time information of the tonegram well. It is characterized by having.

In one embodiment, the training data includes an original sound tomgram and a labeling image corresponding to the original sound tomgram. Meanwhile, in the generating of the training data, a size of the labeling image and a value of each pixel are set. In this case, the size of the labeling image may be set equal to the size of a result output by using the original sound tomgram as an input to the convolutional neural network.

In an embodiment, in the generating of the training data, a portion of the original sound tomgram determined to be a tonal component may be displayed as a value representing the tonal component at a corresponding position of the first labeling image. In addition, a portion of the original sound tomgram that is determined not to be a tonal component may be displayed as a value representing the clutter at a corresponding position in the second labeling image.

In an embodiment, in the generating of the training data, the training data may be generated by using the sound tomgram data in which the size of the labeling image and the value of each pixel are reflected. In this case, the learning data may be augmented and generated in consideration of the time and frequency characteristics of the sound tomgram.

In one embodiment, in the step of inferring the tomgram, values corresponding to the respective pixels for the input of the tomgram are between 0 and 1 based on the learned convolutional neural network structure and a learning model according to a learning weight. We can infer the soundgram to have a real value. In this case, each pixel may be subjected to binary classification by comparing with a specific threshold value, and a tonal component that continuously appears 2 pixels or more at a specific time position with respect to the output value of the convolutional neural network may be changed to a size of 1 pixel. In addition, by comparing the output value with a pixel value at the same location of the original sound tomgram, a portion representing a value having the highest difference may be replaced with a preset value representing the tonal component. An output value of a pixel of the at least two pixels excluding a pixel having the largest difference may be displayed as a preset value representing the clutter.

In accordance with another aspect of the present invention, a convolutional neural network-based inference apparatus for minimizing sonar gram clutter is provided. The device may include an interface unit configured to receive image information for a sound tomgram; And constructing a convolutional neural network for removing the clutter of the sonic gram, generating training data using the sonic gram data, removing the sonic gram clutter through the learned convolutional neural network structure and learning weight, and And a control unit configured to detect the component and infer a sound tomgram.

In an embodiment, the structure of the convolutional neural network for identification of a tonogram may include a convolutional layer, an activation layer, and a batch normalization layer excluding a fully connected layer. In this case, the control unit may configure a kernel size to apply the convolutional neural network structure. On the other hand, the convolutional neural network structure, when the convolutional layer has a kernel size of (rowsize, colsize), has a kernel size that satisfies the condition of rowsize≥colsize in order to reflect the temporal information of the tonegram well. It is characterized.

In one embodiment, the training data includes an original sound tomgram and a labeling image corresponding to the original sound tomgram. Meanwhile, the controller may set a size of the labeling image and a value of each pixel. In this case, the size of the labeling image may be set equal to the size of a result output by using the original sound tomgram as an input to the convolutional neural network.

In an embodiment, the control unit may display a portion of the original sound tomgram that is determined to be a tonal component as a value representing the tonal component at a corresponding position of the first labeling image. In addition, a portion of the original sound tomgram that is determined not to be a tonal component may be displayed as a value representing the clutter at a corresponding position in the second labeling image.

In an embodiment, the controller may generate the training data by using the sound gram data in which the size of the labeling image and the value of each pixel are reflected. In addition, the learning data may be augmented and generated in consideration of the time and frequency characteristics of the sound tomgram.

In an embodiment, the control unit calculates a real value between 0 and 1 in which values corresponding to the respective pixels are based on the learned convolutional neural network structure and a learning model according to a learning weight. You can infer the sound tomgram to have. In this case, each pixel may be subjected to binary classification by comparing with a specific threshold value, and a tonal component that continuously appears 2 pixels or more at a specific time position with respect to the output value of the convolutional neural network may be changed to a size of 1 pixel. A portion indicating a value having the highest difference compared to the pixel value at the same position of the original sound tomgram may be replaced with a preset value representing the tonal component. An output value of a pixel of the at least two pixels excluding a pixel having the largest difference may be displayed as a preset value representing the clutter.

According to the present invention, there is an advantage in that it is possible to support a process of automatically/semi-automatically identifying a tonal frequency by minimizing clutter from a sound tomgram to emphasize tonal frequency information.

In addition, through the present invention, there is an advantage that it is possible to ultimately expect improvement in identification/detection accuracy using a sonic gram.

1 is a block diagram of a convolutional neural network-based inference apparatus for minimizing sound gram clutter according to the present invention.
2 shows the configuration of a procedure for detecting (detecting) a tonal component for inferring a tongram according to the present invention.
3 shows an example of a convolutional neural network structure according to the present invention.
4 shows a method of generating learning data according to the present invention.
5 shows an example of a data augmentation technique in consideration of the sonic gram characteristic according to the present invention.
6 shows an example of a convolutional neural network-based inference method according to the present invention.
7 is a flowchart of a convolutional neural network-based learning and inference method for minimizing sonar gram clutter according to the present invention.

The features and effects of the present invention described above will become more apparent through the following detailed description in connection with the accompanying drawings, and accordingly, those of ordinary skill in the technical field to which the present invention pertains can easily implement the technical idea of the present invention. I will be able to.

In the present invention, various modifications may be made and various embodiments may be provided, and specific embodiments will be illustrated in the drawings and described in detail in the detailed description. However, this is not intended to limit the present invention to a specific embodiment, it is to be understood to include all changes, equivalents, or substitutes included in the spirit and scope of the present invention.

In describing each drawing, similar reference numerals are used for similar elements.

Terms such as first and second may be used to describe various elements, but the elements should not be limited by the terms. These terms are used only for the purpose of distinguishing one component from another component.

For example, without departing from the scope of the present invention, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element. The term and/or includes a combination of a plurality of related listed items or any of a plurality of related listed items.

Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs.

Terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with the meaning in the context of the related technology, and should not be interpreted as an ideal or excessively formal meaning unless explicitly defined in this application. Shouldn't.

The suffix modules, blocks, and parts for the components used in the following description are given or mixed in consideration of only the ease of writing the specification, and do not have meanings or roles that are distinguished from each other by themselves.

Hereinafter, preferred embodiments of the present invention will be described with reference to the accompanying drawings so that those of ordinary skill in the art can easily implement them. In the following description of the embodiments of the present invention, when it is determined that a detailed description of a related known function or a known configuration may unnecessarily obscure the subject matter of the present invention, a detailed description thereof will be omitted.

In the following, the sound gram according to the present invention A convolutional neural network-based learning and inference method for minimizing clutter and a convolutional neural network-based learning and inference device using the same will be described.

First, with respect to the configuration of the present invention, the present invention is a convolutional neural network-based learning and inference method for minimizing the tomgram clutter, and includes the following configurations and techniques.

-Convolutional neural network structure for clutter removal: Convolutional neural network structure reflecting the characteristics of tonal components

-Learning data generation method using sound gram data:

1) How to create training data

2) Data augmentation technique considering the characteristics of sound gram

-Convolutional neural network-based reasoning method: result analysis method

Meanwhile, FIG. 1 shows a configuration diagram of a convolutional neural network-based inference apparatus for minimizing sound gram clutter according to the present invention. Referring to FIG. 1, a convolutional neural network-based inference device 100 includes an interface unit 110, a control unit 120, and a memory 130. Here, the meaning of “sound tomgram” is used in the same or similar meaning as a passive sonar signal. In more detail, the sound tomgram may be an image signal in which a passive sonar signal is mapped on a time axis and a frequency axis.

The interface unit 110 is configured to receive image information for a sound tomgram. Meanwhile, the control unit 120 is configured to configure a convolutional neural network for removing the clutter of the tomgram, and to generate training data by using the tomgram data. In addition, the memory 130 is configured to store information on sound tomgrams, information on learning data, and/or information on inferred sound tomgrams.

On the other hand, the controller 120 may be configured to infer the sound tomgram by removing the sound tomgram clutter and detecting the tonal component through the learned convolutional neural network structure and the learning weight.

Meanwhile, FIG. 2 shows the configuration of a procedure for detecting (detecting) a tonal component for inferring a tonal gram according to the present invention. In this regard, the procedure for detecting (detecting) the tonal component according to FIG. 2 may be performed by the controller 120. Alternatively, some of the procedures for tonal component detection (detection), such as generating a time-series sonar wave, may be performed by a separate external device. At this time, the interface unit 110 receives the sonar waveform, and accordingly, the control unit 120 may configure a convolutional neural network, generate training data, and perform sound tomgram inference (analysis).

In relation to the sonar gram data, the controller 120 may generate a time-series sonar wave through a simulator based on a simulation parameter. Meanwhile, the controller 120 may generate a ground truth table related to data learning according to the simulation parameter. In addition, the control unit 120 may generate training data by using the tomgram data generated from the time-continuous sonar waveform and the Ground Truth Table. In this case, the tonegram data may be referred to as LOFAR (Low Frequency Analysis and Recording), but is not limited thereto and may be variously modified according to applications.

On the other hand, the process of acquiring sound tomgram data such as LOFARgram, that is, Spectrogram, from the time-continuous sonar waveform is as follows. In this regard, a Short-Time Fourier Transform may be performed to transform a time-continuous sonar waveform into a frequency domain. Thereafter, the Fourier-transformed sonar signal may be configured as sound tomgram data such as LOFARgram, that is, a spectrogram. In this case, the tonegram data may include DEMONgram related to DEMON (DEModulation Of Noise) in addition to LOFARgram, but is not limited thereto and may be changed according to an application.

On the other hand, tonal line detection is performed by using the tongram data. In addition, source detection and classification may be performed so as to be distinguishable from clutter from the tonal line detection through the tonal line detection.

On the other hand, the convolutional neural network structure for identifying the tomgram may include a convolutional layer, an activation layer, and a batch normalization layer excluding a fully connected layer. At this time, the controller 120 may configure a kernel size to apply the convolutional neural network structure. In this regard, the convolutional neural network structure, when the convolutional layer has a kernel size of (rowsize, colsize), a kernel size that satisfies the condition of rowsize≥colsize in order to reflect the temporal information of the tonegram well It can be characterized by having.

Meanwhile, the training data may include an original sound tomgram and a labeling image corresponding to the original sound tomgram. In this regard, the controller 120 may set the size of the labeling image and a value of each pixel. In this case, the controller 120 may set the size of the labeling image to be the same as the size of a result output by using the original sound tomgram as an input to the convolutional neural network.

Meanwhile, the control unit 120 may display a portion of the original sound tomgram that is determined to be a tonal component as a value representing the tonal component at a corresponding position of the first labeling image. In addition, the controller 120 may display a portion of the original sound tomgram that is determined not to be a tonal component as a value representing the clutter at a corresponding position of the second labeling image.

Meanwhile, the control unit 120 may generate the learning data by using the tonegram data in which the size of the labeling image and the value of each pixel are reflected. In this case, the learning data may be generated by augmentation in consideration of the time and frequency characteristics of the tonegram.

Accordingly, the controller 120 may perform data augmentation on the training data generated by using the sound gram data and the ground truth table. In addition, the control unit 120 may perform training and evaluation for detecting a tonal component of a sound gram using the increased learning data. Through such a series of procedures, tonal line detection in the tonal gram is possible.

On the other hand, the control unit 120 is sound probed so that values corresponding to each pixel have a real value between 0 and 1 for the input of the sound tomgram based on the learned convolutional neural network structure and a learning model according to a learning weight. Gram can be deduced. In this case, each pixel may be subjected to binary classification by comparing with a specific threshold value, and a tonal component that continuously appears 2 pixels or more at a specific time position with respect to the output value of the convolutional neural network may be changed to a size of 1 pixel. In addition, by comparing the output value with a pixel value at the same location of the original sound tomgram, a portion representing a value having the highest difference may be replaced with a preset value representing the tonal component. An output value of a pixel of the at least two pixels excluding a pixel having the largest difference may be displayed as a preset value representing the clutter.

On the other hand, a detailed operation of the convolutional neural network-based learning and reasoning apparatus for minimizing the sonic gram clutter according to the present invention and its principle will be described as follows.

In this regard, FIG. 3 shows an example of a convolutional neural network structure according to the present invention. Referring to FIG. 3, a kernel size, a channel size (# Channel), and an activation function are shown for each layer. In this regard, specific matters are as follows.

-The neural network structure of FIG. 3 is an example of a convolutional neural network for removing tomgram clutter.

-The vertical axis of the sound tomgram is the time axis, and the horizontal axis is the frequency axis.

-When the kernel size of the convolution layer is defined as (rowsize, colsize), it can be defined to satisfy the condition of rowsize≥colsize in order to reflect the time axis information of the tonegram well.

-Convolutional layer using ReLU Activation can be used as multiple layers, and activation can be set so that only the last convolutional layer can express the presence or absence of clutter/tonal components with a value between 0 and 1.

Meanwhile, FIG. 4 shows a method of generating learning data according to the present invention. Referring to FIG. 4, technical characteristics of a method of generating learning data including a labeling image from original sound gram data are as follows.

1) It can be defined as learning data that includes the original sound tomgram and the corresponding labeling image.

2) The size of the output result may be set equal to the size of the labeling image by using the original sound tomgram as an input to the convolutional neural network for identification.

3) In the original sound tomgram, the part that is considered to be a tonal component can be expressed as a value representing the tonal component (eg, 1) at a corresponding position in the labeling image.

4) In the original sound gram, the part that is not considered to be a tonal component can be marked as a value representing clutter (eg, 0) at a corresponding position in the labeling image.

Meanwhile, FIG. 5 shows an example of a data augmentation technique in consideration of the sound gram characteristics according to the present invention. Referring to FIG. 5, the technical characteristics of the data enhancement technique in consideration of the sound gram characteristics according to the present invention are as follows.

1) A function f(x) that is continuous in two dimensions and has one function value (eg, Gaussian probability density function) can be selected.

2) The following function g(x) that utilizes the function f(x) is converted into g(x) = d * m * f(x) + s, where d∈{-1, 1}, m∈{real}, It can be defined as s∈{real}.

3) The g(x) value can be continuously extracted as much as the number of bins on the time axis of the original sound gram.

4) The frequency axis can be moved by the extracted g(x) at each time position of the original sound tomgram.

Meanwhile, FIG. 6 shows an example of an inference method based on a convolutional neural network according to the present invention. Referring to FIG. 6, the technical features of the convolutional neural network-based reasoning method according to the present invention are as follows.

1) Based on the model learned by the above method, values corresponding to each pixel with respect to the inference result for the input of the tomgram have a real value between 0 and 1.

2) For binary classification, a threshold value of 0 or 1 is applied based on a specific threshold value α.

3) With respect to the output value of the convolutional neural network, a portion that appears as a tonal component signal at a specific time position may appear continuously for 2 pixels or more. When it is necessary to change this to the size of 1 pixel, the output value and the pixel value at the same position of the original tomgram are compared, and the part representing the highest difference value is replaced with a preset value representing the tonal component, and the pixel is not Is expressed as a preset value representing clutter.

Based on the above description, a method of learning and inference based on a convolutional neural network for minimizing sound gram clutter according to the present invention will be described as follows. In this regard, FIG. 7 shows a flow chart of a convolutional neural network-based learning and inference method for minimizing sonar gram clutter according to the present invention. Meanwhile, the convolutional neural network-based learning and inference method may be performed by the controller 120 of the convolutional neural network-based inference apparatus 100 of FIG. 1.

Referring to FIG. 7, the convolutional neural network-based learning and inference method includes a convolutional neural network construction step (S110), a training data generation step (S120), and a tonegram inference step (S130).

In the convolutional neural network configuration step (S110), a convolutional neural network is configured to remove clutter from the tomgram. On the other hand, in the learning data generation step (S120), learning data is generated by using the sound tomgram data.

In addition, in the sound tomgram inference step (S130), the sound tomgram may be inferred (classified or classified) by removing the tonal component and removing the tonal component through the learned convolutional neural network structure and the learning weight.

In this case, the structure of the convolutional neural network for identifying the tomgram may include a convolutional layer, an activation layer, and a batch normalization layer excluding a fully connected layer.

On the other hand, the convolutional neural network structure, when the convolutional layer has a kernel size of (rowsize, colsize), has a kernel size that satisfies the condition of rowsize≥colsize in order to reflect the temporal information of the tonegram well. It can be characterized.

Meanwhile, the training data may include an original sound tomgram and a labeling image corresponding to the original sound tomgram. In this regard, in the learning data generation step S120, the size of the labeling image and the value of each pixel may be set. In this case, the size of the labeling image may be set equal to the size of a result output by using the original sound tomgram as an input to the convolutional neural network.

Meanwhile, in the training data generation step S120, a portion determined as a tonal component in the original sound tomgram may be displayed as a value representing the tonal component at a corresponding position of the first labeling image. In addition, a portion of the original sound tomgram that is determined not to be a tonal component may be displayed as a value representing the clutter at a corresponding position in the second labeling image.

Meanwhile, in the training data generation step S120, the training data may be generated by using the sound tomgram data in which the size of the labeling image and the value of each pixel are reflected. In this case, the learning data may be augmented and generated in consideration of the time and frequency characteristics of the tonegram.

On the other hand, in the tomgram inference step (S130), based on the learned convolutional neural network structure and a learning model according to a learning weight, values corresponding to the respective pixels for the input of the tomgram are real numbers between 0 and 1 You can infer the sound tomgram to have a value.

Specifically, each pixel may be subjected to binary classification by comparing it with a specific threshold value, and a tonal component that continuously appears 2 pixels or more at a specific time position with respect to the output value of the convolutional neural network may be changed to a size of 1 pixel. In addition, by comparing the output value with a pixel value at the same location of the original sound tomgram, a portion representing a value having the highest difference may be replaced with a preset value representing the tonal component. An output value of a pixel of the at least two pixels excluding a pixel having the largest difference may be displayed as a preset value representing the clutter.

On the other hand, with respect to the above-described steps, it is possible for some steps to change their order or to repeatedly perform some steps. For example, while the training data generation step S120 is performed, the convolutional neural network configuration step S110 may be performed again to partially change the neural network structure. Accordingly, the training data generation step S120 and the sound gram inference step S130 may be performed based on the partially changed volute neural network.

In the above, a learning and inference method based on a convolutional neural network for minimizing sonar gram clutter and a convolutional neural network-based learning and inference apparatus using the same have been described. The technical effects of the convolutional neural network-based learning and inference method and the convolutional neural network-based learning and inference apparatus using the same for minimizing the sonic gram clutter according to the present invention are as follows.

According to at least one embodiment of the present invention, there is an advantage that it is possible to support a process of automatically/semi-automatically identifying a tonal frequency by highlighting tonal frequency information by minimizing clutter from a sound tomgram.

In addition, according to at least one embodiment of the present invention, there is an advantage that it is possible to ultimately expect an improvement in identification/detection accuracy using a sonic gram.

According to the software implementation, not only the procedures and functions described in the present specification, but also design and parameter optimization for each component may be implemented as a separate software module. The software code can be implemented as a software application written in an appropriate programming language. The software code may be stored in a memory and executed by a controller or a processor.

Claims (13)

In the convolutional neural network-based inference method for minimizing sound gram clutter, the method is performed by a control unit,
A convolutional neural network configuration step of configuring a convolutional neural network for removing the clutter of the sound tamgram;
A learning data generation step of generating training data including an original sound tomgram and a labeling image corresponding to the original sound tomgram using the sound tomgram data; And
A sound gram inference step of inferring a sound gram by removing sound gram clutter and detecting a tonal component through a learned convolutional neural network structure and a learning weight generated based on the learning data,
The generating of the training data includes a process of setting a size of the labeling image and a value of each pixel,
The step of inferring tomgrams,
Based on the learned convolutional neural network structure and a learning model according to a learning weight, a sound tomgram is inferred so that values corresponding to each pixel have a real value between 0 and 1 for the input of the sound tomgram, The process of binary classification by comparing the pixels with a specific threshold,
With respect to the output value of the convolutional neural network, when a tonal component signal continuously appears in at least two pixels at a specific time position, the difference between the output value and the pixel value at the same position of the original sound tomgram is the largest pixel. Substituting an output value with a preset value representing the tonal component, and displaying an output value of a pixel excluding a pixel having the largest difference among the at least two pixels as a preset value representing the clutter Including, convolutional neural network-based reasoning method.
The method of claim 1,
The convolutional neural network structure for identifying a tomgram, characterized in that it includes a convolutional layer, an activation layer, and a batch normalization layer excluding a fully connected layer.
The method of claim 2,
The convolutional neural network structure,
When the convolutional layer has a kernel size of (rowsize, colsize), the convolutional neural network-based inference, characterized in that it has a kernel size that satisfies the condition of rowsize≥colsize in order to reflect the temporal information of the tonegram well Way. The method of claim 1,
The size of the labeling image is set equal to the size of a result output by using the original sound tomgram as an input to the convolutional neural network.
The method of claim 4,
In the learning data generation step,
For a portion determined to be a tonal component in the original sound tomgram, a value representing the tonal component is displayed at a corresponding position in the first labeling image,
A convolutional neural network-based reasoning method, characterized in that a portion of the original tonegram that is determined not to be a tonal component is displayed as a value representing a clutter at a corresponding position of the second labeling image.
The method of claim 5,
In the learning data generation step,
Generating the training data by using the sound tomgram data reflecting the size of the labeling image and the value of each pixel,
In consideration of the time and frequency characteristics of the tonegram, characterized in that generating by augmenting the learning data, convolutional neural network-based inference method. In the convolutional neural network-based inference device for minimizing the tomgram clutter,
An interface unit configured to receive image information for the sound tomgram;
Construct a convolutional neural network for removing the clutter of the tonegram,
Using the tomgram data, training data including an original tomgram and a labeling image corresponding to the original tomgram is generated,
A control unit configured to infer the sound tomgram by removing the tonal component and removing the tonal component through the learned convolutional neural network structure and training weight generated based on the learning data,
The control unit sets the size of the labeling image and the value of each pixel,
Based on the learned convolutional neural network structure and a learning model according to a learning weight, a sound tomgram is inferred so that values corresponding to each pixel have a real value between 0 and 1 for the input of the sound tomgram, Binary classification by comparing the pixels to a specific threshold,
When a tonal component signal continuously appears in at least two pixels at a specific time position in the output value of the convolutional neural network, the output value of the pixel having the largest difference between the output value and the pixel value at the same position of the original sound tomgram A convolutional neural network for substituting a preset value representing the tonal component and displaying an output value of a pixel excluding a pixel having the largest difference among the at least two pixels as a preset value representing the clutter Based reasoning device.
The method of claim 7,
The convolutional neural network structure for identifying a tomgram, characterized in that it includes a convolutional layer, an activation layer, and a batch normalization layer excluding a fully connected layer.
The method of claim 8,
The convolutional neural network structure,
When the convolutional layer has a kernel size of (rowsize, colsize), the convolutional neural network-based inference, characterized in that it has a kernel size that satisfies the condition of rowsize≥colsize in order to reflect the temporal information of the tonegram well Device.
The method of claim 7,
The size of the labeling image is set equal to the size of a result output by using the original sound tomgram as an input to the convolutional neural network.
The method of claim 10,
The control unit,
For a portion determined to be a tonal component in the original sound tomgram, a value representing the tonal component is displayed at a corresponding position in the first labeling image,
A convolutional neural network-based reasoning apparatus, characterized in that a portion of the original tonegram that is determined not to be a tonal component is displayed as a value representing a clutter at a corresponding position of the second labeling image.
The method of claim 10,
The control unit,
Generating the training data by using the sound tomgram data reflecting the size of the labeling image and the value of each pixel,
A convolutional neural network-based inference device, characterized in that for generating by augmenting the learning data in consideration of the time and frequency characteristics of the tonegram. delete

Images