JP7328144B2 - Face property estimation device and estimation method - Google Patents

Description

The present invention relates to a face property estimation device, an estimation method, and the like.

In tunnel excavation work, the geology of the natural ground is surveyed in advance, the construction plan is formulated, and the actual construction is carried out. However, there are many cases where the geology of the natural ground is different from the result of the preliminary investigation when the construction is actually carried out. For this reason, excavation contractors generally check the geology of the natural ground around the face by determining the properties of the face. This determination is made periodically during construction. The construction company considers the results of face property determination, and formulates safety measures and construction plans for excavation work.

Since the judgment of the face properties is made by a human based on knowledge and experience, the accuracy of the judgment depends on the judge. For example, conventionally, face properties were determined by geologists by comprehensively referring to face photographs or observation data. On the other hand, in recent years, technology has been developed that allows artificial intelligence (AI) to perform geological evaluation. For example, Patent Literature 1 discloses a technique for increasing the accuracy of rock strength determination performed by AI.

JP 2018-17570 A

As described above, the face property determination results are also used to develop safety measures for construction work. Therefore, it is required that the face properties be determined with high accuracy. However, when the AI of the prior art such as the AI described in Patent Document 1 is used to determine the face properties, the accuracy rate is only about 90% at the highest. Therefore, unconditionally using these AI judgment results for consideration of safety measures and construction plans has a high risk, and AI judgment of face properties has low practicality.

One aspect of the present invention is made in view of the above problems. An object of one aspect of the present invention is to construct a trained model suitable for estimating face properties in a tunnel under construction.

In order to solve the above-described problems, a face property estimation apparatus according to the present invention acquires the judgment result of the property of a certain face determined by an expert based on the measurement data of the face having a tunnel under construction. The determination result acquisition unit acquires from teacher data composed of a combination of measurement data on various faces including the certain face and the determination results corresponding to each of the various faces. using an extraction unit for extracting one or more teacher data containing judgment results that match or are similar to the judgment result as relearning teacher data , and one or more relearning teacher data extracted by the extraction unit a re-learning unit for re-learning a trained model obtained by machine-learning the correlation between the measurement data of the various faces and the determination result corresponding to each of the various faces , wherein the measurement data is , at least one of a stereo image and a three-dimensional image, and a spectral image.

According to the above configuration or method, the trained model is re-learned using re-learning teacher data that matches or is similar to the result of judgment by an expert regarding the nature of the face with a tunnel under construction. to run. As a result, it is possible to construct a trained model that is more suitable for estimating the face properties of a tunnel under construction.

The face property estimating device includes a measurement data acquisition unit that acquires the measurement data, and inputs the measurement data of various faces to the learned model after executing the re-learning, so that the measurement data is and an estimating unit for estimating the property of each face shown.

According to the above configuration, the properties of the face can be estimated from the measurement data using the trained model after re-learning. The trained model after re-learning is a model that is more suitable for estimating the characteristics of the face of the tunnel than the trained model before re-learning. Therefore, according to the above configuration, the face properties can be estimated with high accuracy using the learned model.

In the face property estimating device, the estimating unit pre-estimates the property of the certain face by inputting the measurement data of the certain face into the learned model before executing the re-learning, The determination result of the face may be created by correcting the result of preestimating the property of the certain face by the estimation unit.

According to the above configuration, a more correct determination result can be obtained by having the expert correct the pre-estimation result of the trained model. Further, by re-learning the learned model based on the determination result, it is possible to improve the determination accuracy of the learned model.

In the face property estimating device, the extracting unit includes at least part of a result of preestimating the property of the certain face by the estimating unit and a determination result of the property of the certain face acquired by the determination result acquiring unit. is different, the relearning unit extracts the relearning teacher data, and the relearning unit causes the trained model to perform the relearning when the extracting unit extracts the relearning teacher data. good.

Further, according to the above configuration, the result of pre-estimation of the trained model and the judgment result of the expert are collated, and re-learning is executed when at least a part of them is different. The learned model after re-learning is a model that is more suitable for face determination of the tunnel than the learned model before re-learning. Therefore, according to the above configuration, it is possible to construct a trained model capable of highly accurately determining the face properties by re-learning.

In the apparatus for estimating face properties, the extraction unit converts teacher data created by combining the determination result acquired by the determination result acquisition unit and the measurement data corresponding to the determination result into relearning teacher data. You may extract one or more as

According to the above configuration, the learned model can be re-learned using teacher data created using the result of judgment by an expert. This allows the trained model to perform appropriate re-learning.

In a storage device storing a plurality of the training data, the plurality of training data are images for generating at least one of a stereo image and a three-dimensional image included in the training data and a spectral image. A plurality of data sets may be classified and stored according to the photographing area, and the extraction unit may extract the data set including the relearning teacher data.

It is desirable to comprehensively estimate the properties of the face based on the geology of the ground around the face rather than determining it under uniform conditions. According to the above configuration, it is possible to construct a learned model that is more suitable for estimating the face properties of a tunnel under construction.

In the face property estimation device, the measurement data may include a near-infrared photograph.

From the near-infrared photograph, the amount of water contained in the face (water content) can be specified. Therefore, according to the above configuration, it is possible to construct a learned model that can more precisely specify the water content among the face properties. Moreover, when the face property is estimated using the learned model, a more precise estimation result can be obtained for the water content of the face.

In order to solve the above-mentioned problems, an estimation method according to the present invention is a determination result of acquiring a determination result of a certain face property determined by an expert based on measurement data of a certain face of a tunnel under construction. an acquiring step, and from among teacher data composed of a combination of measurement data for various faces including the certain face and the determination results corresponding to each of the various faces, the determination result acquired in the determination result acquisition step An extraction step of extracting one or more pieces of teacher data containing determination results that match or are similar to the determination result as relearning teacher data , and using the one or more relearning teacher data extracted in the extraction step. , a re-learning step of re-learning a learned model obtained by machine-learning the correlation between the measurement data of the various faces and the determination results corresponding to the various faces, wherein the measurement data is At least one of a stereo image and a three-dimensional image, and a spectral image. According to the estimation method, the same effects as those of the face property estimation device can be obtained.

In the estimation method, the measurement data may include near-infrared photographs.

From the near-infrared photograph, the amount of water contained in the face (water content) can be specified. Therefore, according to the above estimation, it is possible to construct a trained model capable of more precisely specifying the water content among the face properties. Moreover, when the face property is estimated using the learned model, a more precise estimation result can be obtained for the water content of the face.

According to one aspect of the present invention, it is possible to construct a trained model suitable for estimating face properties in a tunnel under construction.

1 is a diagram showing an outline of a face estimation system according to Embodiment 1; FIG. FIG. 2 is a block diagram showing the main configuration of various devices included in the face estimation system; FIG. FIG. 4 is a diagram schematically showing an example of a trained model according to Embodiment 1; FIG. 6 is a flow chart showing the flow of processing related to re-learning of a trained model in the estimation device according to Embodiment 1. FIG. FIG. 11 is a diagram showing an outline of a face estimation system according to Embodiment 2; FIG. 2 is a block diagram showing a configuration of a main part of a device included in the face estimation system; FIG. 9 is a flow chart showing an example of the processing flow of the estimation device according to the second embodiment; FIG. 12 is a diagram schematically showing an example of a trained model according to Embodiment 4;

[Embodiment 1]
≪System Overview≫
FIG. 1 is a diagram showing an outline of a face estimation system 100 according to this embodiment. The face estimation system 100 is a system for re-learning a trained model for estimating face properties so as to become a more appropriate model.

Here, "face property" means information on the geological properties or conditions of the face and its surroundings, such as the lithology of the face, the degree of weathering, the degree of unevenness, the presence or absence of cracks, the presence or absence of water, the amount of water contained (water content), etc. means The face properties can be determined by an expert with reference to various measurement data. In the present embodiment, the term "expert" refers to a person skilled in determining face properties, such as a geological expert or geotechnical engineer. The terminal device 3 transmits to the estimation device 1 information indicating the input determination result of the face property. Hereinafter, the information indicating the determination result of the face properties will be referred to as determination result information.

The face estimation system 100 includes, for example, a measurement device 2, a terminal device 3, an estimation device 1, and a storage device 4. Note that the terminal device 3 is not an essential component. The estimation device 1 and the terminal device 3, the estimation device 1 and the storage device 4, and the terminal device 3 and the storage device 4 are connected so as to be able to communicate with each other.

The measuring device 2 is a device for measuring the face of a tunnel under construction. In the present embodiment, the measuring device 2 photographs the following (1) and (2) with the face as the subject.
(1) Multiple Monocular Images from Different Viewpoints (2) Multiple Images Recording Electromagnetic Waves in Different Wavelength Bands The measuring device 2 generates at least one of a stereo image and a three-dimensional image. A “stereo image” is an image that can be viewed stereoscopically, generated from the plurality of monocular images themselves of (1) or the image group of (1). A "three-dimensional image" is a three-dimensional stereoscopic image generated from a plurality of monocular images of (1). Also, the measuring device 2 generates a spectrum image from (2). A “spectrum image” is an image in which electromagnetic waves in a plurality of wavelength bands are recorded and generated from at least one of the images of (2). Electromagnetic waves in at least one wavelength band such as visible light, ultraviolet light, infrared light, and far infrared light are recorded in the spectrum image. Hereinafter, the data generated by the measurement device 2 will be referred to as "measurement data". That is, in this embodiment, the measurement data includes at least one of stereo images and three-dimensional images, and spectral images. Note that the measurement data may include the image groups (1) and (2) that are the sources of these images.

Measurement data from the measurement device 2 is supplied to the terminal device 3 . For example, the measurement device 2 transmits measurement data to the terminal device 3 connected by wire or wirelessly. Alternatively, the measurement device 2 may record the measurement data in an external recording medium such as a USB flash memory and an SD card. Then, the external recording medium may be connected to the terminal device 3 and read by the terminal device 3 . Thereby, measurement data can be transferred from the measuring device 2 to the terminal device 3 . In this way, when supplying measurement data via an external recording medium, communication connection between the measuring device 2 and the terminal device 3 is not essential.

The terminal device 3 is a device for viewing measurement data by an expert. Further, the terminal device 3 is a device for the expert to input the determination result of the face properties of the face indicated by the measurement data browsed by him/herself. The terminal device 3 transmits to the estimating device 1 information indicating the input determination result of the face properties. Henceforth, the determination result of a face property is also simply called a "determination result." Information indicating the determination result is referred to as "determination result information".

The terminal device 3 also associates the received measurement data with the determination result corresponding to the measurement data and transmits them to the storage device 4 . In the storage device 4, a set of the measurement data and the judgment result is stored as one teacher data. In other words, the teacher data is a combination of measurement data for various faces and determination results corresponding to each of the various faces. Note that the terminal device 3 may transmit teacher data to the estimation device 1 instead of the determination result information.

The estimation device 1 is a device that re-learns a trained model. The estimating device 1 extracts teacher data used for re-learning the trained model from the storage device 4 based on the determination result information or the teacher data received from the terminal device 3 . The estimation device 1 re-learns the trained model using the teacher data used for the re-learning. Here, the trained model is a model obtained by machine-learning the correlation between the measurement data and the judgment results for various faces.

The storage device 4 is a storage device that stores a plurality of teacher data. In the storage device 4, the training data may be classified and stored in a plurality of data sets according to the measurement area of the measurement data included in the training data. The term "area" used here refers to an area classified according to the geological properties of the ground, and does not necessarily correspond to a geographical area. The storage device 4 reads out the stored training data in response to a request from the estimation device 1 and transmits it to the estimation device 1 .

When constructing a tunnel, it is common to determine the face properties multiple times according to the progress of the construction. In addition, it is desirable to comprehensively determine the properties of the face based on the geology of the ground around the face, rather than determining it under uniform conditions.

According to the face estimation system 100 shown in FIG. 1, the learned model is re-learned using the judgment result of a certain face property and the teacher data in which the result matches or is similar. Here, the "certain face" is a face in the natural ground for which face property estimation is to be performed using a trained model after re-learning. More specifically, "some face" is, for example, the face of a tunnel under construction. As a result, it is possible to construct a trained model suitable for estimating the face properties of a tunnel under construction.

In addition, the tendency of rock quality or soil quality of the natural ground varies depending on the region. For example, even if the lithology is the same, the appearance of the face may differ depending on the region. In addition, even the face of the same rock quality may look different depending on the degree of weathering. Therefore, when building a trained model for estimating the face properties, the measured data in the area where the face to be estimated exists, that is, the tunnel under construction, and the judgment results of the face properties are used as teacher data. Building a model is desirable.

However, the number of measurement data obtained in the same area is limited, and the content of the measurement data is often similar. Therefore, the number of training data that can be obtained from the same area is limited, and moreover, they often become similar training data. For example, when a trained model is constructed using the face measurement and property judgment of a tunnel under construction as teacher data, there is a risk that the number of similar teacher data will increase. When machine learning is performed using such similar training data, the versatility of the trained model is lost, and the range in which the face properties can be estimated (for example, the area in which the face properties can be estimated) is limited. There is a risk that it will be lost.

In contrast, the face estimation system 100 extracts from the storage device 4 teacher data whose results match or are similar to the determination result of a certain face property, and re-learns. The training data extracted here is not necessarily training data derived from measurement data of the same region. Therefore, even if the number of measurement data obtained in the same area is limited, it is possible to ensure the diversity of training data.

As described above, in the face estimation system 100, the terminal device 3 is not an essential component. When the face estimation system 100 does not include the terminal device 3 , the estimation device 1 may also function as the terminal device 3 . That is, in the face estimation system 100, the terminal device 3 and the estimation device 1 may be integrated. In this case, the estimation device 1 includes a measurement data acquisition unit for acquiring measurement data from the measurement device 2 . Also, the measuring device 2 transmits measurement data to the estimating device 1 . The expert may then browse the measurement data on the estimation device 1 . Then, the expert may input the judgment result to the estimation device 1 .

≪Main part composition≫
FIG. 2 is a block diagram showing the essential configuration of the devices included in the face estimation system 100. As shown in FIG. Although the terminal device 3 is also illustrated in FIG. 2 , the terminal device 3 is not an essential component in the face estimation system 100 .

(Measuring device 2)
The measurement device 2 includes at least a stereo camera 22 and a spectral camera 23 . Furthermore, the measuring device 2 may include one or more of the control section 20 , the communication section 21 , the light source 24 and the storage section 25 . In addition, the measuring device 2 may include an input section such as a button, a mouse, and a touch panel, and a display section such as a display.

The stereo camera 22 is a camera for capturing stereo images of the face or images for generating three-dimensional images. The detailed structure and performance of the stereo camera 22 are not particularly limited as long as the stereo image of the entire face can be captured.

The spectral camera 23 is a camera for capturing an image for generating a spectral image of the face. The detailed structure and performance of the spectral camera 23 are not particularly limited as long as the spectral image of at least a portion of the face can be captured.

In the measurement device 2, the stereo camera 22 and the spectrum camera 23 are preferably arranged and fixed so that they can photograph the same subject at the same direction and angle. Moreover, it is desirable that the measuring device 2 be mounted on a movable body capable of carrying the measuring device 2 to the vicinity of the face of the tunnel. As a result, it is possible to shorten the time during which the construction is interrupted for photographing the face. For example, the measuring device 2 is fixed to the bed of a truck and transported to the vicinity of the face.

Light source 24 is a standard light source when at least one of stereo camera 22 and spectral camera 23 captures images. The light source 24 can be implemented by, for example, a light using light emitting diodes. By using the light source 24, it is possible to reduce the influence of light in the tunnel (for example, light from lights for construction work), ie, background light, on the captured image.

The light from light source 24 is preferably similar in color and intensity to sunlight. For example, it is desirable that the light from the light source 24 has a saturation within a predetermined range. Here, the predetermined range indicates a range of chroma values such that when the light from the light source 24 is viewed by human eyes, it looks like white light. Further, for example, the light from the light source 24 may be light with a hue that looks like white light. Also, the light from the light source 24 may be white light with a color temperature of about 3500 Kelvin to 5000 Kelvin. As a result, the influence of the color of the light from the light source 24 on the captured image can be minimized. Moreover, it is desirable that the light source 24 is a flat light source capable of irradiating diffused light from a plane. As a result, light can be emitted evenly during photographing.

The communication unit 21 is a communication interface that performs communication between the measuring device 2 and other devices. For example, the communication section 21 communicates with the terminal device 3 under the control of the control section 20 . For example, the communication unit 21 transmits measurement data input from the control unit 20 to the terminal device 3 . The communication unit 21 may also receive an instruction to start imaging with the measuring device 2 from another device (not shown). When receiving an instruction to start shooting, the communication unit 21 outputs the instruction to the control unit 20 .

The control unit 20 controls the measuring device 2 in an integrated manner. The control unit 20 controls the stereo camera 22 and the spectral camera 23 to photograph the face. For example, control unit 20 may cause stereo camera 22 and spectral camera 23 to perform imaging in response to an instruction to start imaging received from another device via communication unit 21 . Further, for example, the control unit 20 may receive a user's input operation for instructing the start of imaging via an input device or the like provided in or connected to the measuring device 2 . Then, when the input operation is accepted, the control unit 20 may cause the stereo camera 22 and the spectrum camera 23 to perform photographing.

The imaging by the stereo camera 22 and the imaging by the spectral camera 23 are performed at the same timing, more preferably substantially at the same time. The control unit 20 acquires at least one of the stereo image and the three-dimensional image, and the spectrum image.

The control unit 20 may associate measurement data obtained by one measurement. For example, the control unit 20 may associate at least one of the stereo image and the three-dimensional image generated from the captured images at the same timing with the spectral image. The control unit 20 causes the storage unit 25 to store measurement data for each measurement. Note that when the images included in the measurement data are associated with each other, the control unit 20 may align the faces shown in the images.

For example, the control unit 20 may perform photographing after previously aligning the photographing regions of the stereo camera 22 and the spectrum camera 23 with the photographing region of the optical camera as a reference. In this case, the measuring device 2 comprises an optical camera (not shown). Then, the control unit 20 positions the optical camera and the spectral camera 23 so that one point at the upper left corner, the lower left corner, the upper right corner, or the lower right corner of the photographing area of each camera or the central point is at the same position. Align. Next, the control unit 20 sets the photographing areas of the optical camera and the stereo camera 22 so that one point at the four corners or one point in the center is the same position as in the alignment of the optical camera and the spectral camera 23. Align as follows. As a result, the measurement device 2 can obtain an image aligned at any one of the points described above when photographing is performed. The measurement device 2 is thus capable of producing registered stereo, three-dimensional and spectral images.

Alternatively, face image editing software may be executed on the measurement device 2 or the terminal device 3, and each image included in the measurement data may be aligned using the software. For example, in a general face editor software, the faces may be aligned by superimposing the shape drawing of the face on the outline of the face of each image included in the measurement data.

Further, the control unit 20 may read the measurement data stored in the storage unit 25 in response to a request from the terminal device 3 and transmit the measurement data to the terminal device 3 via the communication unit 21 . Also, the control unit 20 may autonomously transmit measurement data to the terminal device 3 for each measurement.

Moreover, when the measurement device 2 includes the light source 24 , the control section 20 may control at least one of the position and orientation of the light source 24 . Also, if the measurement device 2 includes a light source 24 , the controller 20 may control at least one of the intensity and color of the light from the light source 24 .

The storage unit 25 is a storage device that stores various data necessary for processing of the measuring device 2 . The storage unit 25 also stores measurement data. In this embodiment, the storage unit 25 stores at least one of a stereo image and a three-dimensional image based on the monocular image captured by the stereo camera 22 and a spectral image based on the image captured by the spectral camera 23 as measurement data.

(Terminal device 3)
The terminal device 3 is, for example, a personal computer (PC) or the like. The terminal device 3 includes a display section for displaying measurement data, and an input section such as a mouse, buttons, and keyboard for the expert to input determination results. Also, the terminal device 3 has a communication unit for communicating with the estimation device 1 and the measurement device 2 . The terminal device 3 also has an interface for connecting an external recording medium.

The terminal device 3 transmits determination result information to the estimation device 1 . In addition, the terminal device 3 associates the measurement data acquired from the measurement device 2 with the determination result for the measurement data input by the expert, and transmits the result to the storage device 4 . That is, the terminal device 3 creates teacher data and transmits it to the storage device 4 . Note that the terminal device 3 may transmit teacher data instead of determination result information to the estimating device 1 .

Also, the terminal device 3 may record the created teacher data in an external recording medium connected to the terminal device 3 itself. When the external storage medium is connected, the storage device 4 may read and store the teacher data from the external storage medium.

(Estimation device 1)
The estimation device 1 includes a communication unit 11 , a storage unit 12 , and a control unit (face property estimation device) 10 . The estimating device 1 may have an interface to which an external recording medium can be connected. For example, the estimating device 1 may have an interface to which a USB flash memory, SD card, or the like can be connected. Also, the estimation device 1 may include an input unit such as a button, mouse, and touch panel, and a display unit such as a display. When the face estimation system 100 does not include the terminal device 3 , the estimation device 1 has a configuration corresponding to the various members of the terminal device 3 described above in order to realize the function of the terminal device 3 .

The communication unit 11 realizes communication between the estimation device 1 and other devices. The communication unit 11 communicates with the terminal device 3 and the storage device 4 . For example, the communication unit 11 receives determination result information or teacher data from the terminal device 3 . For example, the communication unit 11 receives the relearning data set from the storage device 4 . Note that the communication unit 11 does not need to communicate with the terminal device 3 when acquiring the determination result information via an external recording medium. Further, for example, the communication unit 11 receives the relearning data set from the storage device 4 . When the face estimation system 100 does not include the terminal device 3 , the communication section 11 receives measurement data from the measurement device 2 and outputs the measurement data to the control section 10 .

The storage unit 12 is a storage device that stores data necessary for processing of the estimation device 1 . Storage unit 12 stores learned model 121 . The trained model 121 is a trained model obtained by machine-learning the correlation between at least one of stereo images and three-dimensional images of various faces, spectral images, and determination results. The structure of the trained model 121 is not particularly limited. For example, the trained model 121 can be realized by a model having a CNN (Convolutional Neural Network) structure. As an example, the face estimation system 100 according to the present embodiment uses a trained model 121 having an AlexNet model structure.

Also, the learning method for initially constructing the trained model 121 and the content of the data set for learning are not particularly limited. For example, a combination of measurement data of tunnel faces constructed in various regions in the past and determination results of face properties for each face may be used as teaching data for learning. The trained model 121 is relearned by the relearning unit 104, which will be described later.

Note that the storage unit 12 may be an external device of the estimation device 1 . For example, the storage unit 12 may be a storage device such as a server communicably connected to the estimation device 1 .

The control unit 10 comprehensively controls the estimation device 1 . Control unit 10 includes determination result acquisition unit 101 , data set determination unit 102 , extraction unit 103 , and relearning unit 104 .

The determination result acquisition unit 101 acquires determination result information or teacher data. When the estimation device 1 and the terminal device 3 are connected for communication, the determination result acquisition unit 101 acquires determination result information or teacher data from the terminal device 3 via the communication unit 11 . Also, when the determination result information or the training data is transferred via an external recording medium, the determination result acquisition unit 101 reads the determination result information or the training data from the external recording medium connected to the interface of the estimating device 1 . The determination result acquisition unit 101 outputs determination result information or teacher data to the dataset determination unit 102 .

The data set determination unit 102 determines one or more data sets to be used for re-learning the trained model 121 from the re-learning data sets 41 in the storage device 4 according to the determination result information or the teacher data. For example, the dataset determination unit 102 accesses the storage device 4 and refers to the relearning dataset 41 to determine the dataset to be used for relearning. Data set determination section 102 outputs information indicating the determined data set to extraction section 103 .

More specifically, the dataset determining unit 102 specifies at least teacher data that includes a determination result that matches or is similar to the determination result indicated by the determination result information, and uses the dataset that includes the teacher data as data to be used for re-learning. Determine as a set.

When the data set determination unit 102 acquires the teacher data from the determination result acquisition unit 101, the data set determination unit 102 specifies the determination results that are “correct data” among the teacher data and the teacher data that includes the determination results that match or are similar to each other. Then, the data set containing the teacher data may be determined as the data set to be used for re-learning.

When the teacher data is acquired from the determination result acquisition unit 101, the dataset determination unit 102 determines the measurement data (i.e., , measurement data corresponding to the determination result), the teacher data matching or similar in at least one of the parameters may be specified from the group of relearning data sets 41 .

Then, the dataset determining unit 102 may determine the dataset containing the identified teacher data as the dataset to be used for re-learning. For example, the measurement data may include the name of the measurement area, the name of the mountain at the measurement point, the name of the mountain range, information indicating geographically the measurement point, information indicating the date and time of measurement, and the like. In this case, the learned model 121 can be made more suitable for the region and the date and time for which the face properties are estimated by performing re-learning in consideration of the measurement data as described above.

The extraction unit 103 extracts one or more datasets determined by the dataset determination unit 102 , ie, one or more teacher data, from the relearning dataset 41 of the storage device 4 . The extraction unit 103 outputs one or more extracted data sets to the relearning unit 104 .

The relearning unit 104 relearns the trained model 121 using one or more teacher data included in the data set acquired by the extraction unit 103 as relearning teacher data . The re-learning method is not particularly limited.

The storage device 4 is a storage device that stores a group of relearning data sets. Hereinafter, a group of relearning data sets will be collectively referred to as a relearning data set 41 . The relearning data set 41 is a group of data sets including one or more pieces of relearning teacher data. Each data set of the relearning data set 41 is obtained by classifying the teacher data according to the measurement area of the measurement data, that is, the shooting area of at least one of the stereo image and the three-dimensional image and the spectral image. . The storage device 4 reads the relearning data set from the relearning data set 41 in response to a request from the estimating device 1 and transmits it to the estimating device 1 . Note that the storage unit 12 and the storage device 4 described above may be configured integrally. That is, the trained model 121 and the relearning data set 41 may be stored in the same storage device.

It is preferable to comprehensively estimate the properties of the face based on the geology of the ground around the face, rather than determining it under uniform conditions. According to the above configuration, it is possible to construct a learned model that is more suitable for estimating the face properties of a tunnel under construction.

<<Details of trained model>>
FIG. 3 is a diagram schematically showing an example of the trained model 121 according to this embodiment. Measured data is input to the trained model 121 as shown. In the example of FIG. 3, stereo images and spectral images are input as measurement data. It is desirable that the position of the face in the stereo image and the position of the face in the spectral image be aligned.

The trained model 121 consists of, for example, a convolution layer, a pooling layer, and a connection layer. In the convolutional layer, input data is convolved with information by filtering. Data that has undergone convolution is subjected to pooling processing in the pooling layer. This can improve the model's ability to recognize changes in the position of features in the data. The data that has undergone the pooling process is processed in the coupling layer to be converted into the output data of the learned model 121, that is, the face property estimation result format and output.

That is, by passing the measurement data input to the trained model 121 through each layer shown in FIG. etc., the result of estimating the face properties is output. Note that the output format of the estimation result is not particularly limited. For example, various properties may be indicated numerically as index values, or may be indicated by binary values such as "crack present" and "no crack". Alternatively, the estimation results of various properties may be indicated by text data.

<<Process flow>>
FIG. 4 is a flow chart showing the flow of processing related to re-learning of a trained model in the estimation device 1 . In addition, the figure shows the flow of processing when the estimation device 1 acquires the determination result information from the terminal device 3 .

The determination result acquisition unit 101 of the estimation device 1 acquires determination result information from the terminal device 3 (S10). The determination result acquisition unit 101 outputs determination result information to the data set determination unit 102 . The data set determination unit 102 determines a data set to be used for re-learning the trained model 121 based on the determination result indicated by the input determination result information (S11). Data set determination section 102 outputs information indicating the determined data set to extraction section 103 .

The extraction unit 103 extracts and acquires the relearning data set determined by the data set determination unit 102 from the relearning data set 41 of the storage device 4 (S12). The extraction unit 103 outputs the acquired data set to the relearning unit 104 . The relearning unit 104 causes the trained model 121 to perform relearning using the data set input from the extracting unit 103 (S13).

[Embodiment 2]
The estimation device according to an aspect of the present invention inputs the measurement data acquisition unit that acquires the measurement data, and the measurement data of various faces into the trained model 121 after executing re-learning, so that the measurement data is and an estimating unit for estimating the property of each face shown.

Further, the above-described estimation unit may pre-estimate the properties of a certain face by inputting the measurement data of a certain face into the trained model 121 before executing re-learning. Then, the determination result of a certain face may be created by correcting the result of preestimating the properties of the certain face by the estimating unit.

In this way, when preestimating or estimating the properties of the face, the estimating device 5 may display the estimation result of the estimating unit on a display unit or the like (not shown). Alternatively, the estimation device 5 may be connected to an output device such as a printer, and output the estimation result via the output device.

A second embodiment of the present invention will be described below. For convenience of description, members having the same functions as those of the members described in the first embodiment are denoted by the same reference numerals, and description thereof will not be repeated. This also applies to subsequent embodiments.

≪System Overview≫
FIG. 5 is a diagram showing an overview of the face estimation system 200 according to this embodiment. Face estimation system 200 has a configuration in which estimation device 1 is replaced with estimation device 5 . The face estimation system 200 is similar to the first embodiment in that the measuring device 2 transmits measurement data to the terminal device 3 and the estimating device 5, and that the estimating device 5 transmits the estimation result of the face properties to the terminal device 3. It is different from the face estimation system 100 according to.

The estimating device 5 uses the learned model 121 to pre-estimate the properties of the face indicated by the measured data. Here, pre-estimation refers to estimation performed before re-learning the trained model 121 . The estimation device 5 transmits the result of preliminary estimation (preliminary estimation result) to the terminal device 3 . An expert who operates the terminal device 3 browses the measurement data and the preliminary estimation result of the estimation device 5, and if there is an error in the preliminary estimation result, corrects the preliminary estimation result to a correct result. A determination result of face properties may be created. The terminal device 3 transmits determination result information or training data to the estimation device 5 . Note that the determination results included in the training data are also the determination results obtained by correcting the pre-estimation results as described above.

Also, the estimation device 5 may use the learned model 121 after re-learning to estimate the properties of the face indicated by the measurement data from the measurement data. Then, the estimation result may be output via an output device such as a printer.

According to the above configuration, the properties of the face can be estimated from the measurement data using the learned model 121 after re-learning. The learned model 121 after re-learning is a model more suitable for estimating the properties of the tunnel face than the learned model 121 before re-learning. Therefore, according to the above configuration, it is possible to highly accurately estimate the face properties using the learned model 121 after re-learning .

Moreover, according to the above configuration, an error-free determination result can be obtained by having an expert correct the pre-estimation result of the trained model 121 before re-learning. Therefore, a re-learning data set can be extracted based on the face property determination result according to the pre-estimation result of the trained model 121 . Therefore, a data set that is more suitable for re-learning can be extracted.

Note that FIG. 5 describes an example in which measurement data is transmitted from the measuring device 2 to both the terminal device 3 and the estimating device 5 . However, in the face estimation system 200 , the measurement device 2 may transmit measurement data to the terminal device 3 via the estimation device 5 . Specifically, when the estimation device 5 transmits the estimation result to the terminal device 3 , the estimation device 5 may also transmit measurement data corresponding to the estimation result to the terminal device 3 .

≪Main part composition≫
FIG. 6 is a block diagram showing the main configuration of various devices included in the face estimation system 200. As shown in FIG. The face estimation system 200 differs from the face estimation system 100 in that an estimation device 5 is included instead of the estimation device 1 .

In the same manner as the face estimation system 100, the face estimation system 200 also realizes delivery of various data among the measurement device 2, the terminal device 3, the storage device 4, and the estimation device 5 via an external recording medium. good. Estimating device 5 includes a measurement data acquiring unit 105 and an estimating unit 106 in addition to various components included in estimating device 1 .

The measurement data acquisition unit 105 acquires measurement data. When the estimation device 5 and the measurement device 2 are connected for communication, the measurement data acquisition unit 105 acquires measurement data from the measurement device 2 via the communication unit 11 . Also, when the measurement data is transferred via an external recording medium, the measurement data acquisition unit 105 reads the measurement data from the external recording medium connected to the interface of the estimating device 5 . Measured data acquiring section 105 outputs the measured data to estimating section 106 .

The estimating unit 106 pre-estimates the face properties by inputting measurement data to the trained model 121 after re-learning. The estimating unit 106 may also estimate the face properties by inputting measurement data into the trained model 121 before re-learning. The estimation unit 106 transmits the pre-estimation result to the terminal device 3 via the communication unit 11 .

The terminal device 3 displays the measurement data and the preliminary estimation result of the estimation device 5 to the expert. Moreover, the terminal device 3 receives an input operation for correcting the pre-estimation result. The terminal device 3 creates a corrected preliminary estimation result, that is, a face property determination result, and transmits determination result information indicating the determination result to the estimation device 5 . In addition, the terminal device 3 creates teacher data including the judgment result and transmits it to the storage device 4 . Note that the terminal device 3 may transmit teacher data to the estimation device 5 instead of the determination result information.

<<Process flow>>
FIG. 7 is a flowchart showing an example of the flow of processing executed by the estimation device 5. As shown in FIG. In addition, the figure shows the flow of processing when the estimation device 5 acquires the determination result information from the terminal device 3 . The measurement data acquisition unit 105 acquires measurement data from the measurement device 2 (S20). Measured data acquiring section 105 outputs the measured data to estimating section 106 . The estimation unit 106 inputs the measured data to the learned model 121 (S21), and acquires the pre-estimated result of the face properties output from the learned model 121 (S22). The estimation unit 106 transmits the acquired pre-estimation result to the terminal device 3 via the communication unit 11 (S23).

After obtaining the preliminary estimation result, the terminal device 3 displays the preliminary estimation result and the corresponding measurement data. The terminal device 3 receives an input operation by an expert to instruct correction of the pre-estimation result. The terminal device 3 corrects the preliminary estimation result in accordance with the input operation, thereby creating the determination result of the face properties. The terminal device 3 transmits determination result information to the estimation device 5 .

The determination result acquisition unit 101 of the estimation device 5 acquires determination result information from the terminal device 3 via the communication unit 11 (S24). Subsequent steps S25 to S27 are the same as steps S11 to S13 in FIG.

(Modification 1)
Note that the estimation device 5 may be configured not to transmit the preliminary estimation result to the terminal device 3 . Then, similarly to the terminal device 3 according to the first embodiment, the terminal device 3 may create an expert judgment result from the measurement data obtained from the measurement device 2 and transmit the judgment result to the estimation device 5. good.

In this case, the determination result acquisition unit 101 of the estimation device 5 receives the determination result and outputs it to the data set determination unit 102 . Also, the estimating unit 106 of the estimating device 5 pre-estimates the face properties using the learned model 121 and outputs the pre-estimated result to the data set determining unit 102 . Subsequent processing is the same as that of the above-described second embodiment.

(Modification 2)
In addition, the data set determination unit 102 compares the result of pre-estimating the property of a certain face by the estimating unit 106 with the determination result of the property of the certain face acquired by the determination result acquisition unit 101, and at least one of these A decision may be made to extract the data set if the parts are different. Then, the relearning unit 104 may cause the trained model 121 to perform relearning when the extracting unit 103 has extracted the relearning teacher data.

According to the above configuration, the pre-estimation result of the trained model 121 and the judgment result of the expert are collated, and re-learning is executed when at least a part of them is different. The learned model 121 after re-learning is a model that is more suitable for face determination of the tunnel than the learned model 121 before re-learning. Therefore, according to the above configuration, it is possible to construct the learned model 121 capable of highly accurately determining the face properties by re-learning.
(Modification 3)
In addition, the data set determination unit 102 compares the result of pre-estimating the property of a certain face by the estimating unit 106 with the determination result of the property of the certain face acquired by the determination result acquisition unit 101, and at least one of these A decision may be made to extract the data set if the parts are different. Then, the relearning unit 104 may cause the trained model 121 to perform relearning when the extracting unit 103 has extracted the relearning teacher data.

In addition, the data set determination unit 102 provides teacher data (or a data set of teacher data) including measured data that matches or resembles the determination result information and that matches or resembles the measurement data corresponding to the determination result of the determination result information. ) may be determined as the re-learning teacher data (or data set) to be extracted. Then, the extraction unit 103 may extract the teacher data (or data set) determined by the data set determination unit 102 from the storage device 4 as re-learning teacher data . For example, the measurement data may include the name of the measurement area, the name of the mountain at the measurement point, the name of the mountain range, information indicating geographically the measurement point, information indicating the date and time of measurement, and the like. In this case, the trained model 121 can be made more suitable for the region and the date and time for estimating the face properties by extracting the re-learning teacher data in consideration of the measurement data as described above.

(Modification 4)
The data set determination unit 102 may acquire measurement data of a certain face from the measurement data acquisition unit 105 . Further, the data set determination unit 102 may acquire the determination result information of the certain face from the determination result acquisition unit 101 . Moreover, the data set determination unit 102 may create teacher data by combining the determination result indicated by the determination result information and the measurement data corresponding to the determination result. The dataset determination unit 102 may then output the created teacher data to the extraction unit 103 .

In this case, the extraction unit 103 handles the created teacher data in the same manner as the extracted teacher data for re-learning (for example, a data set for re-learning). That is, the extraction unit 103 outputs the created teacher data to the relearning unit 104 . Then, the relearning unit 104 causes the trained model to relearn using the teacher data. Note that the dataset determining unit 102 may determine the relearning dataset described in the second embodiment in addition to creating the teacher data described above. The extraction unit 103 may also extract the relearning data set from the storage device 4 . In this case, the extraction unit 103 outputs both the created teacher data and the relearning data set to the relearning unit 104 . Then, the relearning unit 104 causes the trained model 121 to perform relearning using both the created teacher data and the relearning data set. According to the above configuration, the trained model 121 can be re-learned using the teacher data created using the judgment result of the expert. This allows the trained model to perform appropriate re-learning.

(Modification 5)
Also, the data set determination unit 102 may analyze the relationship between the measurement data and the pre-estimation result when using the measurement data. Then, the dataset determination unit 102 may determine the relearning dataset in consideration of the result of the analysis. The “analysis” referred to here means, for example, when performing pre-estimation using the trained model 121, the trained model 121 determines which part of the measured data (for example, which parameter or which region) Focusing on this, it is an analysis as to whether the pre-estimation results have been output. Henceforth, the above-mentioned point of interest of the trained model 121 is also simply referred to as a "point of interest."

More specifically, for example, the dataset determining unit 102 may identify points of interest in at least one of the spectral image, the three-dimensional image, and the stereo image. In addition, the said specific method is not specifically limited. For example, the dataset determining unit 102 can identify the point of interest by using Grad-CAM or Guided Grad-CAM technology.

Then, the dataset determination unit 102 selects a plurality of teacher data (or a dataset including the teacher data) containing measurement data similar to the identified point of interest (that is, the region of interest in the image) as a relearning dataset. can be determined. This allows the trained model 121 to perform more effective relearning.

(Modification 6)
Moreover, when the trained model 121 includes configurations other than the CNN, the dataset determination unit 102 may specify how the trained model 121 is affected by the CNN. Here, the "configuration other than CNN" is, for example, a configuration for a preprocessing step before inputting data to the CNN, or a configuration for a postprocessing step that converts the output result of the CNN into the format of the estimation result. etc. Then, the data set determination unit 102 may correct incorrect influences among the influences that the trained model 121 receives from the CNN, and use the corrected data as re-learning data.

Specifically, the data set determination unit 102 collates the pre-estimation result for a face with the determination result for the face acquired by the determination result acquisition unit 101, as in the third modification described above. Next, the data set determination unit 102 identifies a location where the pre-estimation result differs from the determination result in the collation result as "in the pre-estimation, the learned model 121 was erroneously affected by the CNN." . Subsequently, the data set determination unit 102 corrects the identified affected location according to the determination result, and generates data combining the measured data and the corrected pre-estimation result as relearning teacher data. The data set determination unit 102 outputs the relearning teacher data to the relearning unit 104, and the relearning unit 104 causes the trained model 121 to perform relearning using the relearning teacher data. This allows the trained model 121 to perform more effective relearning.

[Embodiment 3]
In addition to the stereo camera 22 and the spectral camera 23, the measuring device 2 may include a camera capable of taking near-infrared photographs. The measurement data may include a near-infrared photograph of the face taken by the camera.

In this case, the result of determining the properties of the face is the result of determination by an expert based on at least one of the stereo image and the three-dimensional image of the face, the spectral image, and the near-infrared photograph. The measurement data included in the training data also includes near-infrared photographs of each face. In addition, the trained model 121 is a trained model obtained by machine-learning the correlation between at least one of stereo images and three-dimensional images of various faces, spectral images, and near-infrared photographs, and the determination result. .

From the near-infrared photograph, the amount of water contained in the face (water content) can be specified with high accuracy. Therefore, according to the above configuration, it is possible to construct the learned model 121 capable of more precisely specifying the water content among the face properties. In addition, when the face property is estimated (or pre-estimated) using the learned model 121, a more precise estimation result can be obtained for the water content of the face.

[Embodiment 4]
Moreover, the stereo camera 22 and the spectral camera 23 according to the first to third embodiments may be realized by a plurality of camera groups 26 to which bandpass filters 27 are attached. The bandpass filter 27 of each camera in the camera group 26 transmits light in different wavelength bands. Thereby, a spectrum image can be generated based on a plurality of monocular images from the camera group 26 . Also, it can be said that a plurality of monocular images from the camera group 26 are compound eye images. This compound eye image can be used as a stereo image, or a three-dimensional image can be generated based on this compound eye image.

By configuring the camera group 26 in this way, even if a camera for capturing the original image of the spectral image and a camera for capturing the original image for the stereo image or the three-dimensional image are not separately prepared, The same camera can be used to capture the original images required for both image generation. That is, it is possible to generate at least one of a stereo image or a three-dimensional image and a spectral image based on a plurality of monocular images obtained by one shot. Note that the number of monocular cameras included in the camera group 26 may be appropriately set according to the shooting target range.

Also, the trained model 121 according to Embodiments 1 to 3 may be composed of a plurality of sub-models. FIG. 8 is a diagram schematically showing an example of the trained model 121 according to this embodiment. As shown, in this embodiment, the trained model 121 consists of two sub-models, a first trained model 121-1 and a second trained model 121-2. The structure of each submodel is not particularly limited. For example, the first trained model 121-1 and the second trained model 121-2 may each have the model structure of AlexNet, like the trained model 121 according to the first embodiment.

Note that FIG. 8 shows an example in which the stereo camera 22 and the spectral camera 23 are implemented by a plurality of camera groups 26 to which bandpass filters 27 are attached. However, the configuration of trained model 121 shown in FIG. 8 can also be realized when stereo camera 22 and spectrum camera 23 are provided separately. Also, in the example of FIG. 8, the measurement data includes a three-dimensional image and a spectrum image.

The first trained model 121-1 inputs spectral images generated from images of a plurality of cameras, and outputs the degree of weathering, the degree of unevenness, and the presence or absence of cracks among the face properties as intermediate results. is. The second trained model 121-2 uses, as input data, intermediate results, which are the output data of the first trained model 121-1, and three-dimensional images generated from images of a plurality of cameras, and the final face properties are It is a model that outputs accurate estimation results. For example, in the example of FIG. 8, the second trained model 121-2 outputs estimation results such as rock quality, degree of weathering, degree of unevenness, presence or absence of cracks, and presence or absence of water.

The estimating unit 106 according to the present embodiment first obtains an intermediate result by inputting the spectral image of the measurement data to the first trained model 121-1. Next, the estimation unit 106 obtains the final estimation result by inputting the intermediate result and the three-dimensional image of the measurement data to the second trained model 121-2. The estimating unit 106 can also perform pre-estimation of the face properties using the first trained model 121-1 and the second trained model 121-2, similarly to the estimation of the face properties.

Further, when causing the trained model 121 to perform relearning, the relearning unit 104 according to the present embodiment decomposes the measurement data included in the relearning teacher data so as to match the input data of each sub-model. may be used for re-learning.

By configuring the trained model 121 with a plurality of sub-models in this way, the face properties can be estimated more precisely. For example, it is generally possible to estimate the degree of weathering of the face, the degree of unevenness, the presence or absence of cracks, etc. from the spectrum image. Therefore, the spectrum image is input to the first trained model 121-1 for estimation, and then the estimation result and the three-dimensional image are combined to brush up the estimation result with the second trained model 121-2. , it is possible to obtain a more precise estimation result of the face properties. It should be noted that it is possible to obtain a more precise preestimation result for preestimation as well.

[Embodiment 5]
In addition, the estimation unit 106 described in each of the above embodiments may perform rule-based estimation of the face properties and estimation of the face properties by the learned model 121 in combination. In this case, the storage unit 12 stores a database or the like defining the rules.

For example, the storage unit 12 may store a database (DB) in which predetermined features of spectral images are associated with index values of lithology and weathering degree corresponding to the features. Then, the trained model 121 is a trained model obtained by machine-learning the correlation between the lithology and weathering index values in the DB, the stereo image or the three-dimensional image, and the estimation result of the face properties. good.

In this case, when the measurement data is input from the measurement data acquisition unit 105, the estimation unit 106 first identifies features from the spectral image in the measurement data, and calculates index values of lithology and weathering degree corresponding to the features. is specified by referring to the DB of the storage unit 12 . Then, the estimating unit 106 acquires the final face property estimation result by inputting the identified lithology and weathering degree index values and the stereo image or the three-dimensional image into the learned model 121 . Note that the estimating unit 106 can also perform pre-estimation of the face properties using the DB of the storage unit 12 and the learned model 121 in the same manner as the estimation of the face properties.

According to the above configuration, measurement data, particularly data with a large amount of information, such as image data, can be first converted into information such as numerical values on a rule basis, and then estimated by the trained model 121 . As a result, the amount of processing required for estimating the face properties can be reduced. It should be noted that it is possible to similarly reduce the amount of processing required for prior estimation.

Also, the DB described above may be stored in the storage device 4 . Then, the estimation unit 106 may access the DB of the storage device 4 via the communication unit 11 when estimating the face properties.

[Example of realization by software]
Control blocks of estimation devices 1 and 5 (at least one of determination result acquisition unit 101, data set determination unit 102, extraction unit 103, re-learning unit 104, measurement data acquisition unit 105, and estimation unit 106) are integrated circuits. It may be implemented by a logic circuit (hardware) formed in an (IC chip) or the like, or may be implemented by software.

In the latter case, the estimating devices 1 and 5 are equipped with computers that execute program instructions, which are software that implements each function. This computer includes, for example, one or more processors, and a computer-readable recording medium storing the program. In the computer, the processor reads the program from the recording medium and executes it, thereby achieving the object of the present invention. As the processor, for example, a CPU (Central Processing Unit) can be used. As the recording medium, a "non-temporary tangible medium" such as a ROM (Read Only Memory), a tape, a disk, a card, a semiconductor memory, a programmable logic circuit, or the like can be used. In addition, a RAM (Random Access Memory) for developing the above program may be further provided. Also, the program may be supplied to the computer via any transmission medium (communication network, broadcast wave, etc.) capable of transmitting the program. Note that one aspect of the present invention can also be implemented in the form of a data signal embedded in a carrier wave in which the program is embodied by electronic transmission.

The present invention is not limited to the above-described embodiments, but can be modified in various ways within the scope of the claims, and can be obtained by appropriately combining technical means disclosed in different embodiments. is also included in the technical scope of the present invention.

Reference Signs List 1, 5 estimation device 2 measurement device 3 terminal device 4 storage device 10, 20 control unit 11, 21 communication unit 12, 25 storage unit 22 stereo camera 23 spectral camera 24 light source 41 re-learning data set 100, 200 face estimation system 101 Determination result acquisition unit 102 Data set determination unit 103 Extraction unit 104 Re-learning unit 105 Measurement data acquisition unit 106 Estimation unit 121 Trained model

Claims (9)

a determination result acquisition unit that acquires the determination result of the property of a tunnel face determined by an expert based on the measurement data of the tunnel face during construction;
The determination result obtained by the determination result obtaining unit matches or is selected from teacher data consisting of a combination of measurement data for various faces including the certain face and the determination result corresponding to each of the various faces. an extraction unit that extracts one or more teacher data containing similar determination results as relearning teacher data ;
Using one or more re-learning teaching data extracted by the extraction unit, machine learning of the correlation between the measurement data of the various faces and the determination results corresponding to the various faces. and a re-learning unit for re-learning the model,
A face property estimation apparatus, wherein the measurement data includes at least one of a stereo image and a three-dimensional image, and a spectrum image. a measurement data acquisition unit that acquires the measurement data;
an estimating unit for estimating properties of each face indicated by the measurement data by inputting the measurement data of various faces into the trained model after the re-learning. , The face property estimating device according to claim 1. The estimating unit pre-estimates the properties of the certain face by inputting the measurement data of the certain face into the learned model before executing the re-learning,
3. The face property estimating apparatus according to claim 2, wherein the judgment result of the certain face is created by correcting the result of preestimating the property of the certain face by the estimating unit. When at least a part of the result of preestimation of the property of the certain face by the estimation unit and the determination result of the property of the certain face acquired by the determination result acquisition unit is different, the extraction unit performs the re-evaluation . Extract training training data,
The face property estimation apparatus according to claim 3, wherein the re-learning unit causes the trained model to perform the re-learning when the extraction unit has extracted the re -learning teacher data. The extraction unit extracts one or more teacher data created by combining the determination result acquired by the determination result acquisition unit and the measurement data corresponding to the determination result as relearning teacher data. The face property estimation device according to any one of claims 2 to 4, characterized by: In a storage device storing a plurality of the training data, the plurality of training data are images for generating at least one of a stereo image and a three-dimensional image included in the training data and a spectral image. It is classified and stored in multiple data sets according to the shooting area of
The face property estimation apparatus according to any one of claims 1 to 5, wherein the extraction unit extracts the data set including the re-learning teacher data. The face property estimation apparatus according to any one of claims 1 to 6, wherein said measurement data includes near-infrared photographs. a determination result acquisition step of acquiring a determination result of the property of a tunnel face determined by an expert based on the measurement data of the tunnel face under construction;
Matches the determination result obtained in the determination result obtaining step from teacher data consisting of a combination of measurement data for various faces including the certain face and the determination result corresponding to each of the various faces. Alternatively, an extraction step of extracting one or more teacher data containing similar determination results as relearning teacher data ;
Using one or more re-learning teaching data extracted in the extraction step, machine learning of the correlation between the measurement data of the various faces and the determination results corresponding to the various faces. a retraining step of retraining the trained model;
An estimation method, wherein the measurement data includes at least one of a stereo image and a three-dimensional image, and a spectral image. 9. The estimation method according to claim 8, wherein said measurement data includes near-infrared photographs.

Images