JP7413049B2 - Underground radar data processing method, data processing program, and underground radar device - Google Patents

Description

The present invention relates to an underground radar data processing method, a data processing program, and an underground radar apparatus for investigating underground conditions.

For example, it is generally considered difficult to prepare learning data for underground radar image recognition in underground radar that searches for underground pipes, etc. There is a known method that generates a sufficient amount of learning data at high speed (see Patent Document 1).

However, in the image recognition of the underground radar disclosed in Patent Document 1, learning data is generated using an analytical model, that is, it is based on a simulation using the analytical model. For this reason, there is a possibility that the characteristics of the actual state, such as soil heterogeneity, etc., are not necessarily reflected. In other words, if radar-specific information is not fully utilized, the accuracy of image recognition may not be improved or the learning process may become burdensome.

Furthermore, as a method of generating new learning data from existing learning data, it is generally known to expand data by enlarging/reducing or changing gradation, etc. If the extension is applied to image recognition of underground radar, it may not necessarily be possible to obtain good learning data for underground analysis, and the accuracy may deteriorate.

JP2018-41178A

The present invention has been made in view of the above points, and provides a data processing method and data processing program for underground radar that can generate a sufficient amount of learning data in good condition for image recognition of underground radar. Furthermore, it is an object of the present invention to provide an underground radar device that has undergone machine learning using these methods and programs.

In order to achieve the above object, the underground radar data processing method expands underground data obtained through actual measurements by changing measurement parameters according to physical phenomena that occur during exploration with the underground radar.

In the data processing method for underground radar described above, learning data is generated by expanding the underground data obtained through actual measurements by changing measurement parameters. This corresponds to the physical phenomena that occur during medium radar exploration. In this case, by changing the measurement parameters by taking into account the physical phenomena specific to underground radar, while leaving the influence of the actual state of soil unevenness etc. in the actual measurement, it is possible to avoid situations that cannot occur in underground exploration. It is possible to avoid mixing of data. Therefore, a sufficient amount of learning data can be generated in good condition for image recognition of underground radar.

The data processing program for ground penetrating radar to achieve the above purpose prepares a reference image in which measurement parameters are extracted from underground data obtained through actual measurements, and analyzes the measurement parameters based on the physical phenomena that occur during exploration with ground penetrating radar. Data is expanded to generate a new composite image by changing the data accordingly, and machine learning is performed to make it possible to judge the status of buried objects underground based on the new composite image generated by data expansion.

The data processing program for the underground radar described above prepares a reference image in which measurement parameters are extracted from underground data obtained through actual measurements, and expands the prepared reference image by changing the measurement parameters to create learning data. A new composite image is generated, and at this time, the measurement parameters are changed in accordance with the physical phenomena that occur during exploration with ground penetrating radar. In this case, by changing the measurement parameters by taking into account the physical phenomena specific to underground radar while retaining the influence of the actual state of the soil, such as the unevenness of the soil in the actual measurement, the newly generated composite image , it is possible to avoid mixing up data from impossible situations during underground exploration. Therefore, it is possible to generate a composite image in good condition as a sufficient amount of learning data for image recognition using underground radar. In addition, in the above cases, by performing machine learning that makes it possible to judge the situation of buried objects underground based on the composite image, it is possible to create a trained model that performs image recognition of underground radar with high accuracy. become.

In a specific aspect of the present invention, a background image and a buried object reflection signal are extracted from a reference image, measurement parameters related to the buried object reflection signal are changed, and the changed buried object reflection signal is combined with the background image, Generate a new composite image. In this case, a new composite image can be generated by changing information on various parameters such as the position, shape, and number of buried objects as measurement parameters, and the composite image can be used as learning data.

In another aspect of the present invention, as a measurement parameter change, the arc spread of the buried object reflection signal is changed in accordance with a change in the dielectric constant of the soil. In this case, it is possible to generate a composite image that reflects changes in buried object reflection signals that occur during exploration with underground radar when soil properties change.

In yet another aspect of the present invention, the measurement parameter is changed by changing the signal intensity of the buried object reflected signal in accordance with the attenuation rate in the soil in the direction of depth underground. In this case, a composite image can be generated that takes into account attenuation of signal strength in the direction of depth underground.

In yet another aspect of the invention, the background image is generated by erasing the embedded object reflection signal from the reference image and filling the erased portion with adjacent background data. In this case, the background image can be extracted while retaining the characteristics of the actual state, such as the unevenness of the soil.

In still another aspect of the present invention, a buried object reflection signal extracted from one of the plurality of prepared reference images is combined with a background image extracted from another reference image to generate a new composite image. do. In this case, a new composite image can be generated by combining a plurality of prepared reference images.

In yet another aspect of the present invention, the position information of the buried object is included as label information in the measurement parameters extracted when preparing the reference image, and is newly set in the new composite image generated in data expansion. The location information of the buried object is stored as new label information. In this case, the new composite image can be provided with measurement parameters and can be applied as learning data in machine learning.

An underground radar device for achieving the above object includes a trained model obtained by machine learning using the underground radar data processing program described in any of the above, and a transmitting unit that transmits radio waves for exploration. It is equipped with a receiver that receives radio waves for exploration, extracts underground data based on the received signal received by the receiver, and uses the extracted underground data to determine the status of buried objects underground based on a trained model. to decide.

In the above ground penetrating radar device, by employing a trained model obtained by machine learning in the data processing program of the above ground penetrating radar, it is possible to perform image recognition with high accuracy.

FIG. 1 is a conceptual diagram illustrating an example of an underground radar system including a data processing device that performs data processing based on a data processing program according to an embodiment. FIG. 2 is a block diagram for explaining the configuration of an underground radar system. (A) and (B) are conceptual diagrams showing acquisition of underground data and various data processing. FIG. 2 is a conceptual diagram showing one aspect of data processing. FIG. 7 is a conceptual diagram illustrating another aspect of data processing. (A) and (B) are conceptual diagrams showing one aspect of data processing in a comparative example. FIG. 2 is a block diagram illustrating an example of a configuration for data processing in a data processing device. (A) to (C) are flowcharts for explaining an example of a series of operations in data processing. FIG. 2 is a conceptual diagram for explaining an example of image recognition processing using machine learning. FIG. 2 is a block diagram illustrating an example of an underground radar device equipped with a trained model.

Hereinafter, with reference to FIG. 1 etc., an example of a data processing device that performs data processing based on a data processing program or data processing method according to an embodiment of the present invention, and an underground radar system including the data processing device will be described. . FIG. 1 is a conceptual diagram illustrating an example of an underground radar system SR including a data processing device 100 that performs data processing based on a data processing program of the present embodiment, and FIG. 2 is a conceptual diagram illustrating the configuration of the underground radar system SR. FIG. 2 is a block diagram for explanation.

First, as shown in FIG. 1, the ground penetrating radar system SR includes a ground penetrating radar device 10, which is the main body of the ground penetrating radar that performs underground exploration and obtains raw data, and a ground penetrating radar device 10 that performs underground exploration to obtain raw data. It is composed of a data processing device 100 that performs processing. Here, the underground radar device 10 of the underground radar system SR acquires underground data through actual measurements. In addition, the data processing device 100 performs various processing on the underground data acquired by the underground radar device 10 to generate data that enables image recognition regarding underground conditions by machine learning. This makes it possible to explore underground conditions based on underground data.

In this embodiment, the data processing device 100 expands the limited amount of actual measurement data obtained by the underground radar device 10 etc. by changing measurement parameters, etc., which is sufficient for machine learning. The system generates a large amount of learning data. When generating this data, in particular, the measurement parameters are changed in accordance with the physical phenomena that occur during underground exploration, making it suitable as learning data for determining the status of buried objects underground. becomes ready. Note that in the following, machine learning has a broad meaning that includes deep learning and the like.

Of the underground radar system SR, the underground radar device 10 serves as the source for creating reference data (reference image) that is the target of data expansion in the data processing device 100, as described above. This is a device for acquiring actual measurement data. Therefore, the underground radar device 10 transmits radio waves for exploration as transmission waves (transmission signals), and also transmits a part of the reception waves (reception signals), which are components of the transmission waves reflected underground. It includes a data acquisition unit 20 that acquires underground data by receiving it as data, and a position measurement unit 30 that detects movement of the underground radar device 10 and measures the exploration position. As shown in the figure, the underground radar device 10 is configured by mounting the above-mentioned parts on various moving bodies MB such as automobiles, trains, and trolleys. The mobile body MB is a transport unit for changing the exploration position, and includes wheels WL and the like. The position measuring unit 30 detects the moving distance by detecting the rotation of the wheels WL, for example.

Further, as shown in the figure, here, the direction in which the mobile body MB or the underground radar device 10 moves (moving direction) is the x direction, and the direction of the depth underground (depth direction) is the y direction. The underground radar device 10 converts the underground data extracted by the data acquisition unit 20 into various data at the time of measurement, such as various numerical information about output signals for underground exploration and position information detected by the position measurement unit 30. It is also transmitted to the data processing device 100.

In the underground radar system SR, the data processing device 100 is a device composed of, for example, a PC, and includes a CPU, various storage devices, various input/output devices (GUI), and the like. The data processing device 100 is capable of data communication with the ground penetrating radar device 10 through a wired or wireless connection, and receives input of various data such as measurement results actually measured by the ground penetrating radar device 10. Process various data. Here, at least the data processing device 100 performs various processes in order to be able to create a learning model for making judgments regarding the underground situation based on data from the underground radar device 10. .

Hereinafter, one configuration example of the underground radar device 10 and the data processing device 100 will be described in more detail with reference to the block diagram of FIG. 2.

First, one specific aspect of the underground radar device 10 will be described. As described above, the underground radar device 10 includes the data acquisition unit 20 and the position measurement unit 30, and acquires underground data and position information of the underground radar device 10 at the time of acquisition.

Of the underground radar device 10, the data acquisition section 20 includes an antenna control circuit 21, a transmission section 22a, a reception section 22b, and an underground data creation section 23, and controls transmission and reception of radio waves for exploration. , to acquire underground data.

The antenna control circuit 21 controls the operation of the transmitting section 22a having the transmitting antenna ATx and the receiving section 22b having the receiving antenna ARx according to various command signals transmitted from the underground data generating section 23 which is the main control section. , it is possible to output the transmission wave S1 toward the underground soil SO and receive the response wave S2 from the underground object OB at appropriate timing.

The transmitting section 22a and the receiving section 22b perform various processes to perform such driving operations by the antenna control circuit 21. For this reason, the transmitter 22a is composed of an oscillator that generates a transmission wave (radio wave) S1 using pulses or chirps, a D/A converter that performs signal conversion, an amplifier that amplifies the transmission wave S1, etc. . In addition, the receiving unit 22b performs processing for noise processing, etc., in order to receive the response wave (radio wave) S2 reflected back from the buried object, cavity, or other object OB that is the object of exploration existing underground. It is composed of various filters, an amplifier for amplifying the response wave S2, an interphase processor for sampling processing, an A/D converter for signal conversion, and the like.

The underground data creation unit 23 serves as a main control unit that controls the acquisition operation of underground data, and is composed of, for example, a CPU and various types of storage. In addition to acquiring information, it is connected to the position measuring unit 30 to acquire information on the exploration position. The underground data creation unit 23 creates underground data based on this information.

Further, the underground data creation unit 23 is connected to the data processing device 100 and outputs the created underground data to the data processing device 100.

Here, the underground data output from the underground data creation section 23 includes information corresponding to various measurement parameters. For example, the underground data includes various information regarding the response wave S2 necessary for creating image data called a B scope image (see FIG. 3, etc.) showing the state of the underground, that is, various detection results at the receiving unit 22b. and reception timing information. Furthermore, in addition to this, information regarding the transmission wave S1, that is, various information regarding the radio signal used to measure the signal strength, frequency, phase, etc. at the time of transmission from the transmitter 22a, and the transmission timing of the transmission wave S1. The information is also output to the data processing device 100 as part of the underground data.

Next, one specific aspect of the data processing device 100 will be described. As described above, the data processing device 100 is composed of a PC having, for example, a CPU and various types of storage, but in the example shown in FIG. 60 and an input/output device (GUI) 70.

First, in the data processing device 100, the main control unit 50 reads out various programs and data stored in the storage unit 60, as necessary, and performs various arithmetic operations for acquiring or calculating various target data. I do.

Next, in the data processing device 100, the storage unit 60 is configured with a storage device, etc., and stores underground data such as various measurement parameters output from the underground radar device 10, as well as arithmetic processing in the main control unit 50. Various necessary data and programs are stored in various areas. It also stores results obtained by arithmetic processing in the main control unit 50, and further stores learned models such as programs or data structures as results of machine learning.

In order to achieve the above aspect, the storage unit 60 includes, for example, a reference image creation program SP for analyzing underground data and creating reference data (reference image), as well as data processing with the reference image as a processing target. A program storage section 61 is provided to store various programs such as a data expansion program EP for performing expansion, and a machine learning program LP for performing machine learning based on a new image (synthesized image) obtained by data expansion.

The storage unit 60 also includes, for example, a reference image data storage unit SD that stores various data obtained by analyzing underground data, and determines image conversion corresponding to changes in physical conditions and measurement parameters during data expansion. There is also a conversion data storage unit CD that stores table data and conversion functions, etc. for data expansion, a composite image data storage unit GD that stores composite images newly generated by data expansion, and a composite image data storage unit GD that stores composite images newly generated by data expansion. A data storage section 62 is provided for storing various data such as a trained model storage section MD for storing results, that is, trained models.

Finally, the input/output device 70 of the data processing device 100 receives command signals from the user of the data processing device 100, and displays various output results such as calculation results from the main control unit 50. It is the interface section of the computer, and consists of various things such as a touch panel, keyboard, and mouse.

In particular, in this embodiment, the main control unit 50 of the data processing device 100 extracts various measurement parameters and image data for checking the underground situation based on underground data from the underground radar device 10. , Furthermore, data for image data etc. is expanded by changing measurement parameters to generate learning data.

Hereinafter, with reference to FIGS. 3(A) and 3(B), a series of steps of acquisition of underground data by the underground radar device 10 and various data processing by the data processing device 100 will be briefly described. First, in FIG. 3A, as shown on the left side, underground data is acquired as a measurement result by the underground radar device 10, and the underground data is transferred to the data processing side, that is, the data processing device 100 side shown on the right side of the figure. Data, that is, various data such as measurement results from the underground radar device 10 and data regarding output signals are output.

In the data processing device 100, first, a reference image SG to be used as a target for data expansion is created from the received underground data and the like. Here, as the reference image SG, image data called a so-called B scope image is created as illustrated in FIG. 3(B).

Here, the B scope image is a signal in the x direction, which is the moving direction of the mobile body MB or the underground radar device 10 illustrated in FIG. 1, and in the y direction, which is the direction of the depth underground (depth direction). This is data that shows the intensity distribution, and serves as a cross-sectional image that shows the condition underground. Furthermore, in creating various data including B-scope images, various measurement parameters are created based on output information from the underground radar device 10. The measurement parameters created here include various parameters, and include, for example, position information about the buried object reflection signal RS shown as an arc-shaped partial image in the B scope image as the reference image SG. . This information corresponds to the position information of the buried object.

Furthermore, hereinafter, the B scope image and various measurement parameters corresponding thereto are collectively referred to as a reference image SG. In other words, the reference image SG includes information for estimating the soil condition shown in the B scope image and the intensity distribution of the reflected signal, as well as information on the reflected position, that is, the position of buried objects estimated from calculations regarding the reflected signal. , various measurement information including signal strength at the time of transmission, etc. In other words, the reference image SG is composed of various information such as measurement parameters extracted from underground data obtained through actual measurements.

Returning to FIG. 3A, in this embodiment, the data processing device 100 further performs data expansion (inflating) processing on the reference image SG created as including the various information described above. In order to make it possible to judge the situation of buried objects underground, we created multiple new composite images (inflated data) CG, and then used machine learning that used the multiple created composite images CG as learning data. creating a model.

As for the plurality of new composite images CG created from one reference image SG, for example, the one illustrated in FIG. 3(B) can be considered. For example, in the illustrated example, a new composite image CG is created by data expansion by changing various measurement parameters for the buried object reflection signal RS, which is a portion that appears to be arc-shaped in the reference image SG.

Here, in this embodiment, when creating a new composite image CG by data expansion as described above, the new composite image CG is created by changing measurement parameters according to physical phenomena that occur during exploration with the underground radar device 10. The composite image CG is prevented from becoming unrealistic in radar exploration. As a result, the new composite image CG resulting from data expansion is maintained in a state suitable for use as learning data.

For example, regarding data expansion by enlarging or reducing the buried object reflected signal RS, a composite image CG is created so as to reproduce the expansion of an arc according to the dielectric constant of the soil. Specifically, regarding the spread of the circular arc of the buried object reflection signal RS, the spread becomes larger as the depth becomes deeper, and the spread becomes smaller as the depth becomes shallower. It corresponds to the relative dielectric constant and its change. That is, the data processing device 100 changes the spread of the circular arc of the buried object reflection signal RS in accordance with the change in the dielectric constant of the soil as a change in the measurement parameter.

Furthermore, for example, regarding data expansion based on the signal strength at the time of reception, a composite image CG is created so as to reproduce the signal strength according to the attenuation rate of the soil. Specifically, the magnitude of the signal strength at the time of reception is made smaller as it gets deeper and increases as it gets shallower, and the degree of change at this time is determined according to the attenuation rate of the soil as one of the measurement parameters. I take it as a thing. In other words, the data processing device 100 changes the signal strength of the buried object reflection signal RS in accordance with the attenuation rate in the soil in the direction of depth underground, as a change in the measurement parameter.

In addition to the above, for example, as a physical phenomenon that occurs during exploration, when the output signal, that is, the transmitted wave S1 has a specific signal strength, frequency, or phase state, a unique reflection situation may be caused depending on the characteristics of the soil etc. In such a case, data expansion can be performed in accordance with the characteristic, based on, for example, information on the output signal at the time of transmission.

Hereinafter, with reference to FIG. 4, a specific aspect of data processing for data expansion as described above will be conceptually explained. Here, as an example, a case will be described in which a synthesized image is generated by changing the position of a buried object in the depth direction (y direction) with respect to one reference image SG. That is, in the reference image SG, among various measurement parameters, data expansion is performed by changing the position of the buried object reflection signal RS in the y direction.

First, as shown in state A1 in FIG. 4, one reference image SG in which measurement parameters are extracted from underground data obtained through actual measurements is prepared as a processing target for new image generation by data expansion. In the illustrated example, one buried object reflection signal RS is located at a slightly deep position in the y direction in the image, and here, as shown in state A7, it is finally reflected in the original reference image SG shown in state A1. A composite image CG is created that shows what happens when the buried object reflection signal RS is detected at a shallower position than the position shown. That is, the measurement parameters related to the position information regarding the buried object reflected signal RS are changed. Here, it is assumed that the physical conditions are not changed except for the position of the buried object reflected signal RS in the depth direction (y direction). That is, regarding the measurement results in the state shown in the standard image SG, if, among various conditions such as the measurement method and soil condition, only the position of the buried object in the depth direction differs from the original standard image SG and the situation. A composite image CG that is considered to be obtained if other conditions are the same is created.

In order to obtain the composite image CG as described above, first, as shown in state A2, one buried object reflection signal RS is cut out from the other background image BG as a rectangular region RD shown by a broken line, and as shown in state A3. Next, the rectangular area RD is removed from the reference image SG. Furthermore, as shown in state A4, the background image BG shown in state A5 is created by filling in the removed blank space using the information of the surrounding background part. That is, the background image BG is generated by erasing the buried object reflection signal RS from the reference image SG and filling the erased portion with adjacent background data.

On the other hand, regarding the rectangular region RD removed when changing from state A2 to state A3, it is assumed that the buried object reflection signal RS has a shape as shown in state B1. Further, among the measurement parameters at that time, the value in the depth direction (y direction) is defined as depth DP1. Here, as shown in state B2, the value in the depth direction (y direction) is changed to depth DP2 (<DP1). In this case, the shape of the buried object reflection signal RS is changed to the buried object reflection signal RSa as the measurement parameter regarding the depth is changed. Specifically, the expansion of the circular arc, etc. is changed.

In underground exploration, regarding the relationship between the position of a buried object in the depth direction and the shape of the corresponding reflected signal, for example, the shallower the position of the buried object is, that is, the closer it is to the ground surface, the narrower the shape of the reflected signal. It is known that it appears as a steep curve shape. However, the degree of deformation varies depending on the signal strength and frequency of the radio signal used for inspection, the relative dielectric constant (propagation speed) of the soil, etc. Here, the data processing device 100 takes this information into account and calculates the values according to various information such as signal strength acquired for the reference image SG as measurement parameters, the relative dielectric constant of the soil obtained by calculation, etc. A buried object reflection signal RSa is determined. Note that for calculation of the degree of deformation, etc., table data, conversion functions, etc. are prepared in advance in the data storage section 62 of the storage section 60 (see FIG. 2), and based on these, the degree of deformation is calculated. processing or image conversion processing is performed.

The rectangular area RDa including the buried object reflection signal RSa obtained as described above is pasted to a position corresponding to the depth DP2 in the background image BG, as shown in state A6. Thereafter, as shown in state A7, the target composite image CG is created by performing a process of blurring the boundary between the peripheral part of the buried object reflection signal RSa and the other background image BG.

In addition, during the above processing, the position information of the buried object, which is one of the measurement parameters extracted when preparing the reference image, is included as label information, and the new composite image generated in the data expansion as described above is In the CG, the position information of the newly set buried object may be held as new label information. Specifically, in the above embodiment, the information on the depth DP1 is configured such that the measurement parameters in the original reference image SG are incorporated as label information. Correspondingly, in the finally obtained new composite image CG, information on the depth DP2 is incorporated as new updated label information. In addition, although the manner in which the location information of a buried object, which is one of the measurement parameters, is handled as label information has been explained above, it is not limited to the location information of a buried object, but various measurement parameters can also be handled as label information. . By doing the above, the new composite image CG can have the same measurement parameters as the settings in the original reference image SG, that is, it can have information about the same room, so the composite image CG When applying it as learning data in machine learning, it becomes possible to treat it in the same way as the reference image SG.

Hereinafter, with reference to FIG. 5, another specific aspect of data processing for data expansion as described above will be conceptually explained. Here, as another example, a buried object reflection signal extracted from one of multiple prepared reference images is combined with a background image extracted from another reference image to generate a new composite image. Let me explain the case. In the illustrated example, data expansion is performed in which buried object reflection signals RS1 and RS2 extracted from one reference image SG1 are combined with a background image BG2 extracted from another reference image SG2 to generate a new composite image CG. ing.

That is, as shown in states C1 and D1 in FIG. 5, one reference image SG1 and another reference image SG2, in which measurement parameters are extracted from underground data obtained through actual measurements, are used as processing targets. prepare. In the illustrated example, in state C1, two buried object reflection signals RS1 and RS2 exist, while in state D1, one buried object reflection signal RS exists. Here, as finally shown in state D7, instead of one buried object reflection signal RS that existed in the original reference image SG2 shown in state D1, the signal shown in state C1 is shown in the soil of reference image SG2. A composite image CG is created that shows the situation when the two buried object reflection signals RS1 and RS2 that were present in the original reference image SG1 are detected.

Here, it is assumed that there is no change in the physical conditions other than the presence of buried objects, and the soil condition etc. are in the state shown in the reference image SG2. That is, regarding the measurement results in the state shown in the reference image SG2, if among various conditions such as the measurement method and soil condition, only the number and position of buried objects differ from the situation in the original reference image SG2, A composite image CG that would be obtained if other conditions were the same is created.

In order to obtain the composite image CG as described above, first, as shown in states C1 and C2, a rectangular region RD1 shown by a broken line is cut out from the reference image SG1 so as to include buried object reflection signals RS1 and RS2, and further, In consideration of pasting to the reference image SG1, a rectangular region RD1a is created in which the shapes of the buried object reflection signals RS1 and RS2 are changed to the buried object reflection signals RS1a and RS2a, as shown in state C3.

On the other hand, in parallel with the above, for the reference image SG2, as shown in states D1 to D5, one buried object reflection signal RS is cut out and removed as a rectangular area RD2, and the blank area where the rectangular area RD2 has been removed is , the background image BG2 is created by filling in the information using the information of the surrounding background part.

Thereafter, as shown in state D6, a rectangular area RD1a including the buried object reflection signals RS1a and RS2a is pasted on the background image BG2, and as shown in state D7, the surrounding areas of the buried object reflection signals RS1a and RS2a and other By performing processing to blur the boundary with the background image BG2, the desired composite image CG is created.

In the generation of a composite image by data expansion as described above, for example, in the data included as a B scope image, information regarding the influence of the actual state of soil unevenness in actual measurements is left in the background image BG, etc., and , the buried object reflected signal RS, etc., the measurement parameters are changed in consideration of the characteristics specific to the underground radar in order to correspond to the physical phenomena that occur during exploration with the underground radar. Thereby, it is possible to avoid mixing data of an impossible situation in underground exploration with the newly generated composite image CG.

On the other hand, for example, as shown in FIG. 6(A) and FIG. 6(B), which shows one aspect of data processing in the comparative example, data expansion using conventional simple enlargement/reduction, etc. If applied to images, there is a possibility that accuracy cannot be improved or the load on the learning process may increase. Note that the composite image CGx in FIG. 6(A) shows a case where simple scaling is performed in the x and y directions, and the image CGx in FIG. 6(B) shows a case in which simple gradation changes are performed. This shows the case where it was done. The presence of these composite images CGx may actually increase noise, making it difficult to perform appropriate learning for image recognition. In contrast, in the present embodiment, a composite image can be generated in good condition as a sufficient amount of learning data for image recognition using underground radar.

Hereinafter, with reference to the block diagram of FIG. 7 and the flowchart of FIG. 8, a configuration example of a process for data expansion in the data processing apparatus 100 and its operation will be described. That is, in the following, on the premise that the creation of the reference image has been completed through the processing in the data processing apparatus 100, the subsequent data expansion processing will be described. Here, it is assumed that a plurality of reference images (for example, about 50 images) are prepared, and a method of extracting background images and buried object reflection signals from these images and creating a new composite image by combining them. An example will be explained.

FIG. 7 is a block diagram conceptually showing the configuration of each part shown in FIG. 2 from a functional aspect. That is, each section shown in FIG. 7 is configured by the main control section (CPU) 50, the storage section 60, and the input/output device (GUI) 70, and their functions in cooperation.

Each part shown in FIG. 7 will be explained below along with its functions. As shown in the figure, the data processing device 100 includes an input data reading section 51, a buried object reflection signal extraction section 52, a background image generation section 53a, a reflection signal processing section 53b, and a physical condition calculation section 54. , and an image synthesis section 55.

The input data reading unit 51 reads reference image information, ie, image data, measurement parameters, etc., stored in a predetermined area of the storage unit 60. That is, the reference image SG shown in state A1 in FIG. 4 and the reference images SG1 and SG2 shown in states C1 and D1 in FIG. 5 are prepared.

The buried object reflection signal extraction section 52 extracts information on various signals including the buried object reflection signal and various data from the image data read by the input data reading section 51. In other words, the buried object reflection signals RS shown in states A2, A3, and B1 in FIG. 4 and states C2, D2, and D3 in FIG. 5 are extracted from the image data. At this time, various data including data about the position of the buried object in the depth direction, that is, the position of the buried object reflection signal (hereinafter referred to as buried object reflection position), is extracted as information regarding the buried object reflection signal RS. Additionally, it receives command signals from the input/output device (GUI) 70 and performs processing according to the command signals. For example, when a buried object reflection position is specified by a user as a command signal from the input/output device (GUI) 70 during processing, the specified buried object reflection position is extracted from the reference image. By changing the position to the reflection position, it becomes possible to generate a new composite image.

The background image generation section 53a generates a background image by performing image processing on the image data from which the buried object reflection signal RS and the like have been extracted as a result of the extraction process in the buried object reflection signal extraction section 52. That is, the background images BG and the like shown in states A3 to A5 in FIG. 4 and states D3 to D5 in FIG. 5 are created.

The reflected signal processing section 53b processes the buried object reflected signal RS, etc. extracted as a result of the extraction process by the buried object reflected signal extraction section 52, taking physical conditions into consideration. That is, the buried object reflected signal RS and the like shown in state B2 in FIG. 4 and state C3 in FIG. 5 are modified.

The physical condition calculation unit 54 identifies an appropriate processing method for processing in the reflected signal processing unit 53b. That is, the amount of change in the buried object reflection signal RS, etc. in the deformation process from state B1 to state B2 in FIG. 4 or from state C1 to state C2 in FIG. 5 is calculated.

The image synthesis section 55 pastes the buried object reflected signal RS etc. processed by the reflected signal processing section 53b onto the background image BG etc. prepared by the background image generation section 53a, and further performs necessary processing (blur processing etc.). ) to generate a composite image CG. That is, various processes shown in states A6 and A7 in FIG. 4 and states D6 and D7 in FIG. 5 are performed.

Hereinafter, with reference to the flowchart of FIG. 8, a series of operations regarding data expansion by each section such as the input data reading section 51 described with reference to FIG. 7 will be described. Note that FIG. 8(A) is a flowchart showing the entire operation regarding data expansion, FIG. 8(B) is a flowchart regarding background image generation processing in FIG. 8(A), and FIG. 8(A) is a flowchart relating to the processing of the buried object reflected signal.

As shown in FIG. 8A, first, the main control unit 50 reads the prepared reference image with the input data reading unit 51 (step S101), and in the buried object reflection signal extraction unit 52, the input data reading unit 51 reads the prepared reference image. A buried object reflection signal is extracted based on the image data and measurement parameters read in, or a command signal from the input/output device 70 (step S102).

Next, the main control unit 50 causes the background image generation unit 53a to generate a background image based on the remaining image data extracted as a result of the buried object reflection signal being extracted in step S102 (step S103a). The reflected signal processing section 53b and the physical condition calculation section 54 transform and process the image of the extracted buried object reflection signal (step S103b).

Next, the main control unit 50 pastes the background image generated in step S103a and the buried object reflection signal processed in step S103b (step S104), and In order to determine position data and remove edge noise occurring at the border of pasting, low-pass filter processing (blur processing) is performed (step S105) to obtain image data, and the image data obtained in S105 is and the new buried object position data determined in step S104 are output as data regarding a new composite image (step S106). As for the data on the buried object position, as described above, it is conceivable that the data is incorporated as label information.

Here, among the above, regarding the background image generation process in step S103a, image processing corresponding to the content shown with reference to states A2 to A5 in FIG. 4 and states D2 to D5 in FIG. 5 is performed, for example. . Specifically, as shown in FIG. 8(B), the background image generation unit 53a first performs a process of zero-filling the part of the buried object reflection signal in the image data, that is, creates an image for the part of the buried object reflection signal. By performing processing to set the data value of the region to zero, the relevant portion is treated as an extracted portion (step S201), and further processing is performed to backfill the extracted portion with surrounding data (step S202). In order to remove edge noise occurring at the boundary portion of the portion, processing (blur processing) is performed using a low-pass filter (step S203). Through the above steps, a background image is formed that retains the influence of the actual condition such as the unevenness of the soil.

Furthermore, among the above, regarding the buried object reflected signal processing processing in step S103b, for example, image processing corresponding to the contents shown with reference to states B1 and B2 in FIG. 4 and states C2 and C3 in FIG. 5 is performed. Ru. Specifically, as shown in FIG. 8C, the physical conditions to be changed are selected in the reflected signal processing section 53b based on the command signal from the input/output device 70 (step S301). As an example of the specific processing in step S301, for example, when the command signal from the input/output device 70 is assumed to be that the relative dielectric constant of the soil is lower than the specified value, that is, the propagation speed of radio waves is increased. If there is a signal specifying the physical condition to be changed, and this is accepted by the buried object reflected signal extracting section 52, the reflected signal processing section 53b uses the physical condition calculating section 54 to match the specification. Then, various measurement parameters are adjusted, and the amount of deformation of the circular arc shape of the buried object reflection signal to be deformed is calculated (step S302). For example, when changing the dielectric constant of the soil as a physical condition, a constant for converting the slope of the arc shape of the buried object reflection signal is calculated accordingly. Thereafter, the reflected signal processing section 53b processes the image according to the calculation result by the physical condition calculation section 54 (step S303). That is, the image is converted according to a conversion constant that corresponds to the change in the selected conditions.

In addition, in the above example, as a change in measurement parameters, a change in the dielectric constant of the soil, that is, a change according to the dielectric constant of the soil, was explained, but there are various other aspects of the target and content of the change. Conceivable. Specifically, as illustrated in FIG. 4, a change in the reflection position in the depth direction can be considered as a change in the physical conditions. In this case, a modification process is performed that takes into account, for example, a change in reception intensity depending on the attenuation rate of the soil. Furthermore, as illustrated in FIG. 5, it is also possible to combine one background image and another buried object reflection signal by pasting them together. Furthermore, it is also possible to change the frequency of the transmitted radio waves.

Through the data expansion process described above, a new composite image is generated from one or more reference images. For example, with respect to approximately 50 prepared reference images, ₅₀ C ₂ =1225 patterns of composite images can be created by combining one background image and other embedded object reflection signals. In addition, we created 10 patterns for data expansion by expansion or reduction, that is, data expansion that reproduces the spread of an arc according to the dielectric constant of the soil, and we created 10 patterns for data expansion based on the signal strength at the time of reception, that is, in the depth direction. We will create 10 patterns of data expansion that reproduces the signal intensity according to the attenuation rate of the soil due to changes in becomes possible. In this case, it is thought that a sufficient amount of data can be obtained as image data for performing machine learning that enables the determination of the situation of buried objects underground.

Machine learning based on the learning data acquired as described above will be described below. Various methods can be applied to machine learning, such as deep learning that constructs a multilayered neural network shown in FIG. 9, for example. In addition, the illustrated example has an input layer IL, a hidden layer (middle layer) ML, and an output layer OL, and in the output layer, the classification class is an image of a single buried pipe or an image of a plurality of buried pipes. It is classified into five categories: hollow, hollow, culvert, and other, and the probabilities are displayed.

In the learning stage, machine learning is performed based on the reference image and the composite image generated by the data expansion described above. That is, these image data are processed from the input layer IL, and learning is performed to improve the accuracy of the determination results. That is, construction is performed from the input layer IL to the hidden layer (middle layer) ML and the output layer OL. In the learned state, unknown underground data (image data) is processed starting from the input layer IL, and it is determined which of the previously learned classification classes it corresponds to.

Figure 10 is equipped with a trained model obtained through machine learning as described above, that is, machine learning using an underground radar data processing program, and determines the status of buried objects underground based on the learned model. FIG. 2 is a conceptual diagram for explaining an example of an underground radar device. Note that FIG. 10 is a block diagram corresponding to FIG. 2.

As illustrated, the underground radar device 500 according to the present embodiment includes a data acquisition unit 20, a position measurement unit 30, and a data processing device 300 equipped with a trained model. Note that the data acquisition unit 20 and the position measurement unit 30 have a structure similar to that illustrated in FIG. 2, and the data acquisition unit 20 includes an antenna control circuit 21 and a transmission unit that transmits exploration radio waves 22a, a receiving section 22b that receives radio waves for exploration, and an underground data creation section 23, and outputs a transmission wave S1 toward the underground soil SO at an appropriate timing via a transmission antenna ATx, A response wave S2 from the underground object OB is received via the receiving antenna ARx, and underground data is acquired. Furthermore, the position measuring unit 30 detects the moving distance of the underground radar device 500. The underground data creation unit 23 is connected to the data processing device 300 and outputs underground data including acquired data and various data to the data processing device 300.

On the other hand, of the underground radar device 500, the data processing device 300 is a device configured with, for example, a PC, and is configured with a CPU, various storage devices, various input/output devices (GUI), and the like. The data processing device 300 is capable of data communication with the data acquisition unit 20 through a wired connection, and receives input of various data such as measurement results actually measured by the data acquisition unit 20, and processes the received various data. . Here, the data processing device 300 includes a main control unit (CPU) 350, an input/output device (GUI) 370, a program storage unit 310 stored in various storage devices (not shown), and a data storage unit 320. The stored programs and data are read as appropriate by a main control unit (CPU) 350 and various processing operations are performed.

The program storage unit 310 stores, for example, an underground data analysis program AP and a judgment program TM that is a learned model.

The underground data analysis program AP is a program for analyzing the underground data acquired from the data acquisition unit 20 to create image data such as a B-scope image, and to extract various measurement parameters. That is, the main control unit 350 uses the underground data analysis program AP to make it possible to determine the status of underground objects based on the learned model TM with respect to the underground data extracted by the data acquisition unit 20. Prepare various data for

As mentioned above, the judgment program (learned model) TM is a program, data structure, etc. obtained by machine learning in the above-mentioned underground radar data processing program, and is prepared by the underground data analysis program AP. This is a program for performing analysis processing on a B-scope image, etc., and determining how the image is classified. The main control unit 350 uses the underground data analysis program AP to determine the status of underground objects. Specifically, as shown with reference to FIG. 9, for example, whether the image is an image of a single buried pipe, a plurality of buried pipes, a cavity, or a culvert. , others, and displays the probability of the classified results using the input/output device (GUI) 370, for example.

The data storage unit 320 stores numerical data necessary for various calculation processes as described above, and also stores calculation results based on the underground data analysis program AP and the judgment program TM, which is a learned model. Therefore, an analyzed data storage section AD and a judgment result storage section JD are provided. That is, the analyzed data storage unit AD stores the results of analyzing underground data by the underground data analysis program AP. The stored data is subject to judgment by the judgment program TM. Further, the determination result storage section JD stores the determination results of the determination program TM. The stored data indicates the display content on the input/output device (GUI) 370.

As described above, the underground radar data processing method according to the present embodiment expands underground data acquired through actual measurements by changing measurement parameters according to physical phenomena that occur during exploration with the underground radar. . In this case, when data is expanded to generate training data, the measurement parameters are changed in accordance with the physical phenomena that occur during exploration with underground radar, thereby avoiding situations that cannot occur during underground exploration. You can avoid mixing up data. Furthermore, by subjecting the data expansion to underground data obtained through actual measurements, the data can be created while retaining the influence of the actual state of the soil, such as unevenness in the actual measurements. Therefore, a sufficient amount of learning data can be generated in good condition for image recognition of underground radar.

In addition, in the data processing program for the underground radar, that is, in the data processing device 100 equipped with this program, a reference image SG in which measurement parameters are extracted from underground data obtained through actual measurements is prepared, and for the prepared reference image SG, By expanding the data by changing the measurement parameters, a new composite image CG is generated as learning data. At this time, the measurement parameters are changed in accordance with the physical phenomena that occur during exploration with underground radar. In this case, by performing machine learning that makes it possible to determine the situation of underground objects based on the composite image CG, it becomes possible to create a trained model that performs image recognition of underground radar with high accuracy.

Further, the underground radar device 500 includes a trained model TM acquired by machine learning in the data processing program of the underground radar, a transmitter 22a that transmits radio waves for exploration, and a transmitter 22a that receives radio waves for exploration. a receiving section 22b for extracting underground data based on the received signal received by the receiving section 22b, and determining the situation of buried objects underground based on the learned model TM with respect to the extracted underground data. There is. In this case, by employing the trained model TM obtained by the machine learning described above, it becomes possible to perform image recognition with high accuracy.

〔others〕
This invention is not limited to the embodiments described above, and can be implemented in various ways without departing from the spirit thereof.

First, among the above, in the underground radar system SR shown in FIGS. 1 and 2, the explanation has been made assuming that the underground radar device 10 and the data processing device 100 are separate bodies, but even if they are integrated. For example, the data processing device 100 may be installed in the ground penetrating radar device 10. In this case, for example, machine learning can be performed while acquiring underground data.

Furthermore, among the above, it is assumed that the data processing device 300 is installed in the underground radar device 500 shown in FIG. 500 data acquisition units 20 may also be adopted. In this case, for example, judgments can be made after the fact regarding the acquired underground data. Further, in this case, it is also possible to perform a judgment process on the underground data acquired by the plurality of underground radar devices 500 collectively in one data processing device 300 after the fact.

Further, the trained model explained above is an example, and various trained models can be adopted.

Furthermore, various configurations of the underground radar device can be adopted in addition to the examples shown in the conceptual diagram of FIG. 1 and the block diagrams of FIGS. 2 and 10.

Further, for example, in the above embodiment, the mobile body MB may have a drive mechanism and be able to move by itself, or may be manually moved by the user.

DESCRIPTION OF SYMBOLS 10... Underground radar device, 20... Data acquisition section, 21... Antenna control circuit, 22a... Transmission section, 22b... Receiving section, 23... Underground data creation section, 30... Position measurement section, 50... Main control section (CPU ), 51... Input data reading section, 52... Buried object reflection signal extraction section, 53a... Background image generation section, 53b... Reflection signal processing section, 54... Physical condition calculation section, 55... Image synthesis section, 60... Storage section, 61... Program storage section, 62... Data storage section, 70... Input/output device (GUI), 100... Data processing device, 300... Data processing device, 310... Program storage section, 320... Data storage section, 350... Main control section (CPU), 370...Input/output device (GUI), 500...Underground radar device, A1-A7, B1, B2, C1-C3, D1-D7...Status, AD...Analyzed data storage unit, AP...Underground Data analysis program, ARx...receiving antenna, ATx...transmitting antenna, BG, BG2...background image, CD...conversion data storage, CG, CGx...composite image, DP1, DP2...depth, EP...data expansion program, GD... Synthetic image data storage unit, IL...input layer, JD...judgment result storage unit, LP...machine learning program, MB...mobile object, MD...trained model storage unit, ML...hidden layer (middle layer), OB...object , OL... Output layer, RD, RD1, RD1a, RD2, RDa... Rectangular area, RS, RS1, RS2, RS1a, RS2a, RSa... Buried object reflected signal, S1... Transmitted wave, S2... Response wave, SD... Reference image Data storage unit, SG, SG1, SG2...Reference image, SO...Soil, SP...Reference image creation program, SR...Underground radar system, TM...Judgment program (learned model), WL...Wheel

Claims (9)

An underground radar data processing method for expanding underground data acquired through actual measurements by changing measurement parameters according to physical phenomena occurring during exploration with an underground radar, the method comprising:
In the data expansion, the buried object reflection signal of one of the plurality of reference images from which the measurement parameters have been extracted is combined with a background image in which the soil condition is as shown in the other reference image. A ground penetrating radar data processing method that generates a new composite image. Prepare one reference image and another reference image in which measurement parameters are extracted from underground data obtained through actual measurements,
When performing data expansion to generate a new composite image by changing the measurement parameters according to the physical phenomena that occur during exploration with underground radar, the buried object reflection signal of the first reference image is changed based on the soil condition. generate the new composite image by combining the background image with the background image in the state shown in the other reference image,
A data processing program for an underground radar that performs machine learning that makes it possible to determine the status of underground objects based on the new composite image generated by the data expansion. The measurement parameter regarding the buried object reflection signal of the one reference image is changed according to the attenuation rate in the soil in the background image in which the soil condition is as shown in the other reference image, and the The underground radar data processing program according to claim 2, which generates a new composite image. 4. The underground radar data processing program according to claim 3, wherein, as the change in the measurement parameter, the spread of the circular arc of the buried object reflection signal is changed in accordance with a change in the dielectric constant of the soil. According to any one of claims 3 and 4, the measurement parameter is changed by changing the signal intensity of the buried object reflected signal in accordance with the attenuation rate in the soil in the depth direction of the ground. ground penetrating radar data processing program. The background image of the other reference image is generated by erasing a buried object reflection signal included in the original image of the other reference image and filling the erased portion with adjacent background data. A data processing program for the underground radar according to any one of the items. 6. The buried object reflection signal included in the one reference image is different from the buried object reflection signal included in the original image of the other reference image in at least one of the number and position of buried objects. The ground penetrating radar data processing program described in . The measurement parameters extracted when preparing the reference image include position information of the buried object as label information,
The underground radar according to any one of claims 3 to 7, wherein position information of a newly set buried object is retained as new label information in the new composite image generated in the data expansion. data processing program. A trained model obtained by the machine learning in the underground radar data processing program according to any one of claims 2 to 8;
a transmitting unit that transmits radio waves for exploration;
and a receiving unit that receives the radio waves for exploration,
An underground radar device that extracts underground data based on a reception signal received by the receiver, and determines a state of a buried object underground based on the learned model based on the extracted underground data.

Images