Learning device, crack detection device, learning method, and crack detection method

Abstract

To provide a learning device which can accurately detect the occurrence of a crack from a camera image.SOLUTION: A crack detection device includes: a detection unit which accepts input of a face image being a determination object to a trained model which has learned a relation between a face image obtained by imaging a face sprayed with concrete and pixels corresponding to cracks included in the face image by using teacher data in which labels indicating cracks are given to pixels corresponding to cracks present in the sprayed concrete among pixels in the face image, and performs segmentation on the face image being the determination object to detect whether or not a region corresponding to a crack is included; and an output unit which outputs alert information if the detection unit detects the presence of a crack.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a learning device, a crack detection device, a learning method, and a crack detection method. [Background technology]

[0002] In mountain tunnel construction, "skin fall" can occur, where exposed ground peels off or falls off due to natural or man-made factors. If skin fall occurs during excavation work at the face (excavation surface), it can lead to serious disasters, so it is necessary to prevent this. In order to prevent skin fall, "mirror sprayed concrete" is applied to the face in a thickness of several centimeters. However, even if this method is applied, it is not possible to completely prevent collapse. Since rock mass collapses at the tunnel face occur instantly, it is difficult to evacuate after visually confirming the movement of the rock mass. Therefore, it is necessary to detect the signs of a tunnel face collapse at an early stage and ensure that workers have enough time to evacuate. It is known that in tunnel faces where mirror-sprayed concrete is applied, cracks of a few millimeters in width appear on the sprayed surface as a sign of impending collapse, and collapse occurs along the cracks a few minutes later. Therefore, if cracks appearing on the sprayed surface can be monitored in real time, it is expected that workers will be alerted before the collapse occurs, dramatically improving the possibility of early evacuation. Systems have also been disclosed that monitor the tunnel face by installing cameras around the tunnel face (for example, Patent Document 1). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2018-207194 A Summary of the Invention [Problem to be solved by the invention]

[0004] However, when installing surveillance cameras around the tunnel face during construction and detecting cracks from video or other images, there is a lot of noise in the images, such as changes in color due to the movement of workers and heavy machinery, and changes in lighting, making it difficult to accurately detect the occurrence of cracks from camera images using conventional technology.

[0005] In order to solve the above problems, the present invention aims to provide a learning device, a crack detection device, a learning method, and a crack detection method that can accurately detect the occurrence of cracks from camera images. [Means for solving the problem]

[0006] In order to solve the above-mentioned problems, one aspect of the present invention is to provide a method for detecting a face image of a face on which shotcrete has been applied. The shape and color of the robot are specified by the technician operating the control terminal. The image forming apparatus includes a synthesis unit that synthesizes an image representing a crack; an assignment unit that assigns a first label representing a crack to a first pixel corresponding to the image representing the crack based on the synthesized image and assigns a second label representing something other than a crack to a second pixel other than the first pixel; a teacher data generation unit that obtains teacher data by assigning the first label and the second label to the synthesized image; and a learning unit that learns the relationship between a face image and pixels corresponding to cracks contained in the face image using the teacher data in which the first label representing a crack is assigned to the first pixel corresponding to a crack in the sprayed concrete among the pixels in the synthesized image and the second label is assigned to the second pixel corresponding to something other than the crack.

[0007] In addition, one aspect of the present invention is to provide a method for detecting a face image of a face on which shotcrete has been applied. The shape and color of the robot are specified by the technician operating the control terminal.The system has a detection unit that inputs a face image to be judged and detects whether or not there is a region corresponding to a crack by segmenting the face image to be judged, and an output unit that outputs alert information when the detection unit detects the presence of a crack.The system has a trained model that has learned the relationship between a face image and pixels corresponding to cracks contained in the face image using training data in which, among the pixels in an image synthesized from an image representing a crack, pixels that correspond to cracks in the sprayed concrete are given a label representing a crack, and pixels that correspond to other than the cracks are given a label representing other than a crack.

[0008] Further, one aspect of the present invention is a learning method in which a learning device, which is a computer, learns, and a synthesis unit synthesizes a face image in which a face to which shotcrete has been applied is captured, The shape and color of the robot are specified by the technician operating the control terminal. A learning method in which an image representing a crack is synthesized, and an assignment unit assigns a first label representing a crack to a first pixel corresponding to the image representing the crack based on the synthesized image, and assigns a second label representing something other than a crack to a second pixel other than the first pixel, and a learning unit of the computer learns the relationship between a face image and pixels corresponding to cracks contained in the face image using training data in which the first label representing a crack is assigned to the first pixel in the synthesized image that corresponds to a crack in the sprayed concrete, and the second label is assigned to the second pixel corresponding to something other than the crack.

[0009] In one aspect of the present invention, a detection unit of the crack detection device detects a face image of a face on which shotcrete has been applied. A composite image of the cracks with shapes and colors specified by the engineer using the control terminal. Among the pixels in the image, a label representing a crack is assigned to a pixel corresponding to a crack present in the shotcrete. , and a label indicating a non-crack is assigned to the pixels corresponding to the non-cracks.This is a crack detection method in which a face image to be judged is input to a trained model that has learned the relationship between a face image and pixels corresponding to cracks contained in the face image using training data, and the face image to be judged is segmented to detect whether or not there is an area corresponding to a crack, and the output unit of the crack detection device outputs alert information if the detection unit detects the presence of a crack. Effect of the Invention

[0010] As described above, according to the present invention, the occurrence of cracks can be detected with high accuracy from camera images. [Brief description of the drawings]

[0011] [Figure 1] 1 is a schematic block diagram showing the configuration of a face monitoring system 1 according to an embodiment of the present invention. [Diagram 2] FIG. 13 is a diagram showing an example of a face image captured of a face where a crack has actually occurred in the mirror-sprayed concrete. [Diagram 3] 13 is a diagram showing an example of a case in which a crack image is synthesized with a face image captured of a face where mirror-sprayed concrete has been applied. FIG. [Figure 4] FIG. 13 is a diagram showing an example of an image in which a region to be a crack image is specified for an actually occurring crack in a face image. [Diagram 5] 4 is a flowchart illustrating the operation of the crack detection device 40 in the learning phase. [Figure 6] 4 is a flowchart illustrating the operation of the crack detection device 40 in the execution phase. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0012] Hereinafter, a face monitoring system using a crack detection device according to one embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a schematic block diagram showing the configuration of a face monitoring system 1 according to an embodiment of the present invention. The tunnel face monitoring system 1 includes a camera 10, terminal devices 20 (terminal device 20a, terminal device 20b), an operation terminal 30, and a crack detection device 40.

[0013] Camera 10 is installed on the side wall of the tunnel at a position a predetermined distance (e.g., about 10 meters) rearward of the tunnel face, and captures an image of the tunnel face F. Camera 10 outputs the captured image as a face image to crack detection device 40. Camera 10 captures images at a rate of several to a dozen frames per second, and sequentially outputs the face images to crack detection device 40. Here, at the tunnel face F, "mirror sprayed concrete" is applied, in which concrete is sprayed onto the ground exposed by the tunnel excavation to a thickness of about several centimeters. The camera 10 captures an image of the tunnel face on which the mirror-sprayed concrete has been applied, and outputs the tunnel face image.

[0014] A plurality of terminal devices 20 (e.g., terminal device 20a, terminal device 20b, etc.; simply referred to as terminal device 20 when no particular distinction is made) are used at the site where excavation work is carried out on the face. For example, a smartphone or a tablet can be used as the terminal device 20. The terminal device 20 is carried by a worker performing work such as excavation. The terminal device 20 is connected to the crack detection device 40 wirelessly or via a wire, and has a function of communicating with the crack detection device 40. The terminal device 20 also includes a speaker, a touch panel, etc. Although the case where there are two terminal devices 20 (terminal device 20a and terminal device 20b) will be described here, there may be only one terminal device, or three or more terminal devices.

[0015] The operation terminal 30 has input devices such as a mouse and a keyboard, and a display device such as a display. It can accept various operations from users such as engineers. A computer may be used as this operation terminal 30.

[0016] The crack detection device 40 includes a data processing unit 401 , an assignment unit 402 , a teacher data generation unit 403 , a learning unit 404 , a storage unit 405 , a detection unit 406 , and an output unit 407 .

[0017] Here, in the crack detection device 40, the teacher data is roughly divided into two types. The first teacher data is teacher data using an image taken of a state where a crack has occurred at a face where mirror-sprayed concrete has been applied, and the second teacher data is teacher data using a composite image in which a crack image, which is an image showing a crack, is combined with a portion where no crack has occurred in an image taken of a face where mirror-sprayed concrete has been applied. The shape of such a crack is generally linear.

[0018] The first training data can be an image of a face where cracks have occurred in the sprayed concrete at an actual construction site. FIG. 2 is a diagram showing an example of a face image obtained by capturing an image of a face where a crack has actually occurred in the mirror-coated concrete. The second training data can be a composite image that is drawn on the face image by an engineer using the operation terminal 30 to display the crack image based on his experience and knowledge of what the condition will be when a crack occurs in an image of the face where mirror sprayed concrete has been applied. 3 is a diagram showing an example of a case where a crack image is synthesized with a face image obtained by capturing an image of a face where mirror-coated concrete has been applied. In this example, a crack image such as that shown by reference numeral 300 is synthesized.

[0019] When using the first training data, there is an advantage that the accuracy of crack determination by the trained model obtained by learning is high because images of cracks that actually occurred in the mirror shotcrete are used. On the other hand, it may not be easy to prepare the number of images required for learning. When the second teaching data is used, although the image is not an image of an actual crack, various crack patterns can be drawn, and there is an advantage that the number of images required for learning can be easily prepared. Here, the crack image in the composite image is drawn by reflecting the experience and knowledge of the engineer, so it is possible to draw it so as to reproduce the crack that actually occurred with high accuracy. In addition, the image showing the crack is drawn by the engineer inputting the operation from the operation terminal 30 to the image of the face where the mirror sprayed concrete has been applied. The crack image is expressed by the shape, color, etc., taking into consideration the condition of the face and the environment of the construction site.

[0020] The data processing unit 401 processes data necessary for the learning unit 404 to perform learning. For example, the data processing unit 401 synthesizes an image representing a crack with a face image in response to the operation contents from the keyboard, mouse, etc. of the operation terminal 30. Here, when synthesizing a crack image with a face image, the composite image may be generated by rewriting the pixel values ​​(each RGB value in the case of an RGB image) of pixels included in an area corresponding to the crack in the face image, or the composite image may be generated by drawing the crack image on a layer other than the layer of the face image and superimposing these layers. Here, when the crack image is generated as a layer other than the face image, the coordinate data indicating the area that is a crack can also be associated based on the layer corresponding to the crack image.

[0021] The assigning unit 402 assigns a label (annotation) representing a crack to a pixel corresponding to the crack image based on the combined image. For example, the assigning unit 402 assigns a label representing "crack" to each of the pixels in the composite image that correspond to the crack image. Here, for the composite image, the crack image composited with the face image is the subject, and a label representing "crack" is assigned to each pixel corresponding to this crack image. For example, in the composite image, a label representing "crack" is assigned to pixels within the area drawn by the engineer as the crack image. On the other hand, for a face image of a face where a crack has actually occurred in the mirror-sprayed concrete, it is necessary to identify which part is the crack image. Here, for example, an engineer operates the operation terminal 30 to specify that the part of the face image that corresponds to the crack is a crack, thereby identifying the crack image. 4 is a diagram showing an example of an image in which a region is specified as a crack image for a crack that has actually occurred in a face image. In this diagram, a linear image portion indicated by reference numeral 400 is shown as being specified as the region of the crack image. When specifying pixels corresponding to the crack, the engineer may specify the region corresponding to the crack, or the engineer may specify pixels corresponding to the crack.

[0022] When the area is defined as the crack image in this manner, the labeling unit 402 labels each of the pixels included in the area of ​​this crack image with a label representing "crack." Furthermore, the assigning unit 402 assigns a label other than "crack" (for example, concrete, heavy machinery, worker, lighting, etc.) to pixels in the face image or the composite image other than the region of the crack image. In this way, by the assignment unit 402 assigning a label indicating "crack" to each pixel, it becomes possible to identify areas of an image in which adjacent pixels having a label indicating "crack" are assigned in the face image or the composite image as areas in which cracks exist.

[0023] The teacher data generating unit 403 obtains, as teacher data, data in which the pixels of the composite image are labeled by the labeling unit 402. For example, the teacher data generating unit 403 generates teacher data by obtaining, as teacher data, a face image or a composite image to which a label has been added.

[0024] The learning unit 404 uses training data in which pixels in a face image of a face where mirror sprayed concrete has been applied are given labels indicating cracks, and learns the relationship between the face image and pixels corresponding to cracks contained in the face image. This learning uses AI (artificial intelligence) technology. For example, semantic segmentation is used for this learning. Semantic segmentation is an algorithm that uses deep learning to assign a label to each pixel in an image. This semantic segmentation can detect objects on a pixel-by-pixel basis, so it has high recognition resolution even for linear cracks that occur in large numbers, and also has high discrimination performance against noise, which is image changes caused by the movement of workers and heavy machinery during construction. Therefore, it is possible to efficiently identify cracks even under conditions where the face during construction is being filmed by the camera 10. In semantic segmentation, for example, the relationship between information about a pixel corresponding to a crack and information about pixels surrounding that pixel is learned. The learning unit 404 performs learning to generate a learned model that has learned the relationship between a face image and pixels corresponding to cracks contained in the face image.

[0025] The storage unit 405 stores the trained model obtained by the training unit 404. The storage unit 405 is configured by a storage medium, for example, a hard disk drive (HDD), a flash memory, an electrically erasable programmable read only memory (EEPROM), a random access read / write memory (RAM), a read only memory (ROM), or any combination of these storage media. This storage unit 405 can be, for example, a non-volatile memory.

[0026] The detection unit 406 uses training data in which pixels in a face image of a face where sprayed concrete has been applied are given labels indicating cracks for pixels that correspond to cracks present in the sprayed concrete, and inputs the face image to be judged into a trained model that has learned the relationship between a face image and pixels corresponding to cracks contained in the face image, and detects whether or not there is an area corresponding to a crack by segmenting the face image to be judged. When the detection unit 406 detects that there is an area corresponding to a crack in the face image, the detection unit 406 outputs, to the output unit 407, information indicating that a crack has been detected.

[0027] When the detection unit 406 detects the presence of a crack, the output unit 407 outputs alert information to the terminal device 20. By transmitting the alert information to the terminal device 20 when the presence of a crack is detected, the output unit 407 can urge workers to evacuate before a face collapse occurs.

[0028] In this crack detection device 40, the data processing unit 401, the assignment unit 402, the teacher data generation unit 403, the learning unit 404, the detection unit 406, and the output unit 407 may be composed of a processing unit such as a CPU (central processing unit) or a dedicated electronic circuit.

[0029] Next, the operation of the crack detection device 40 will be described. Learning Phase FIG. 5 is a flowchart illustrating the operation of the crack detection device 40 in the learning phase. In the learning phase, the data processing unit 401 inputs a face image (step S101). This input may be obtained from the camera 10, or, if an already captured face image is stored in an external or internal storage device, the face image may be read out from the storage device and input. Here, a face image of a face where a crack has actually occurred in the mirror-sprayed concrete may be input, or a composite image in which a crack image is combined with the face image may be input as the face image, or both of these may be input.

[0030] Next, the data processing unit 401 identifies crack images from the input face image (step S102). When this crack image is a face image taken of a face where a crack has actually occurred in the mirror-sprayed concrete, the pixels corresponding to the crack specified by the engineer are designated on the operation terminal 30, and the crack image is identified based on the designated pixels. Furthermore, if the face image is a composite image, an area in the face image that has been processed as a crack image can be identified as the crack image. Also, a face image of a face where no cracks have occurred in the mirror-sprayed concrete may be input in step S101, and in step S102, a composite image may be generated by inputting the crack image by an engineer from the operation terminal 30. Then, the crack image may be identified from the generated composite image.

[0031] Next, the assigning unit 402 assigns a label of "crack" to pixels corresponding to the crack image identified in the face image by the data processing unit 401 (step S103). In addition, for areas of the face image that are not crack images, the assigning unit 402 assigns labels other than "crack" (concrete, heavy machinery, worker, lighting, etc.) according to the information of the pixels in those areas. Once the labels have been assigned, the training data generation unit 403 generates the labeled face images as training data and inputs them to the learning unit 404.

[0032] The learning unit 404 uses the input teacher data to learn the relationship between the face image and the pixels corresponding to the cracks included in the face image (step S104). After learning, the learning unit 404 writes the learned model into the storage unit 405.

[0033] Execution Phase Next, the execution phase will be described. FIG. 6 is a flowchart illustrating the operation of the crack detection device 40 in the execution phase. In the execution phase, the camera 10 captures an image of the tunnel face to be monitored (step S201). The tunnel face captured here is the tunnel face after the construction of the mirror-sprayed concrete, and is the tunnel face to be monitored for the occurrence of cracks. The camera 10 outputs the captured image of the tunnel face to the crack detection device 40 as a tunnel face image. Here, the camera 10 continuously captures the tunnel face. For example, the camera 10 captures images at intervals that provide a predetermined number of frames per second, and sequentially outputs the tunnel face images to the crack detection device 40.

[0034] The detection unit 406 of the crack detection device 40 reads out the learned model stored in the memory unit 405, and when a face image is obtained from the camera 10 (step S202), the detection unit 406 inputs the obtained face image into the learned model, and determines whether or not a region corresponding to a crack has been detected in the input face image, thereby determining whether or not a crack has occurred in the mirror sprayed concrete (step S203). When no area corresponding to a crack is detected in the face image (step S204-NO), the detection unit 406 proceeds to step S201. This makes it possible to continuously monitor the face.

[0035] On the other hand, when a region corresponding to a crack is detected in the face image (step S204-YES), the detection unit 406 outputs to the output unit 407 that a crack has been detected. When the detection unit 406 detects the presence of a crack, the output unit 407 outputs alert information indicating that a crack has been detected to the terminal device 20 (terminal device 20a, terminal device 20b) (step S205). As a result, a message such as "A crack has been detected at the face. Please evacuate" is displayed on the display screen of each terminal device 20, and an alert sound is emitted from the speaker of each terminal device 20. By outputting such an alert, workers can recognize the occurrence of a crack at an early stage and take necessary measures such as evacuation. In this way, when a crack is detected, a warning can be issued from the terminal device, so that signs of a skin fall at the face can be detected and workers can be urged to take necessary measures such as evacuation. In addition, since the detection can be performed in real time, it is immediate.

[0036] According to the embodiment described above, segmentation is used to assign meaning to each pixel of the face image obtained from the camera, such as whether or not it corresponds to a crack, and the occurrence status of a crack can be detected based on whether or not an area corresponding to a crack has been detected. In addition, according to this embodiment, crack images in the tunnel face image are assigned a label indicating "crack" on a pixel-by-pixel basis, and then deep learning is used to learn using semantic segmentation, an algorithm. This makes it possible to obtain a trained model that can detect linear cracks on a pixel-by-pixel basis. Furthermore, the obtained trained model is used to detect areas corresponding to cracks in the tunnel face image by segmentation, thereby improving the accuracy of crack detection.

[0037] In addition, in the conventional background difference method, the difference between frames is calculated for the entire image, and whether or not an object has been detected is detected based on the difference, so for example, if a worker or heavy machinery is captured in the camera image, or if there is a change in the way the lighting hits the image, the change will cause a change in the color of the image, etc., resulting in a difference between frames. In this case, it is difficult to distinguish whether a difference has been detected due to the image of a crack or the image of a worker or heavy machinery, so there is a high possibility of false detection. In contrast, according to this embodiment, as described above, since a meaning is given to each pixel by segmentation, even if a worker or heavy machinery is captured in the camera image, it is possible to detect an area based on each pixel, and it is possible to identify whether the detected area is a crack or not. Therefore, since the difference between frames is not used for judgment as in the background subtraction method, it is resistant to noise and can reduce erroneous judgment. In addition, since it can be detected in real time, it is responsive.

[0038] Furthermore, according to the above-mentioned embodiment, even if a face image containing noise data is input in the face image captured of the monitored face, it is possible to use a monitoring camera (wearable camera) or monitoring system equipped with a trained AI capable of recognizing cracks to monitor the occurrence of cracks, which are a precursor phenomenon of the face falling, in real time and promptly inform workers of the risk of the face falling. This makes it possible to detect cracks, which are a precursor phenomenon of the collapse of the mirror-sprayed concrete, and quickly share the information among face workers, allowing them to evacuate to a safe place before the collapse occurs.

[0039] In the above-mentioned embodiment, the data processing unit 401 generates a composite image by adding a crack image to a face image, but it is also possible to generate a composite image by adding an image that may be noise (an image of a heavy machine, an image of a worker, an image of a change in the lighting, etc.) to the face image as a noise pattern image, and generate teacher data by labeling this noise pattern image as "noise". This enables the learning unit 404 to learn the relationship between the face image and pixels that correspond to the noise pattern contained in the face image, and when determining whether or not there is a crack, it is also possible to detect whether or not there is an area that corresponds to a noise pattern, thereby further improving resistance to noise.

[0040] In addition, in this embodiment, not only a face image of a tunnel face where a crack has occurred in the mirror-sprayed concrete, but also a composite image in which a crack image has been composited with the face image is used. In an actual tunnel face, cracks rarely occur in the mirror-sprayed concrete, and the crack patterns obtained by experiments are limited. Therefore, it is not easy to prepare a large number of images of a face where cracks have actually occurred, but according to this embodiment, a composite image in which cracks are artificially expressed by photo composition or the like is used, so that a large number of crack patterns and noise patterns can be created as learning data, which makes it possible to improve the learning efficiency and judgment accuracy of AI.

[0041] In the above-described embodiment, the crack detection device 40 has the data processing unit 401, the assignment unit 402, the teacher data generation unit 403, and the learning unit 404. However, the functions of the data processing unit 401, the assignment unit 402, the teacher data generation unit 403, and the learning unit 404 may be provided in a device other than the crack detection device 40, and the other device may be configured as a learning device. Then, the learned model obtained by learning in the learning device may be written to the storage unit 405 of the crack detection device 40.

[0042] The trained model may also be stored in a computer installed at the site of face construction, and a program for implementing the functions of the detection unit 406 and the output unit 407 may be executed on the computer. This allows for the face image captured by the camera 10 to be sent to the computer via the Internet, and the computer detects the presence or absence of cracks and sends the results to the terminal device 20, making it possible to detect and warn of cracks without going through the Internet, thereby providing immediacy without being affected by delays caused by going through the Internet.

[0043] The functions of the data processing unit 401, the assignment unit 402, the teacher data generation unit 403, and the learning unit 404 in the above-mentioned embodiment may be realized by a computer as a learning device. The functions of the detection unit 406 and the output unit 407 in the above-mentioned embodiment may be realized by a computer as a crack detection device. In that case, each program for realizing these functions may be recorded in a computer-readable recording medium, and the program recorded in the recording medium may be read into a computer system and executed to realize the functions. Note that the "computer system" here includes hardware such as an OS and peripheral devices. In addition, the "computer-readable recording medium" refers to a portable medium such as a flexible disk, an optical magnetic disk, a ROM, a CD-ROM, and a storage device such as a hard disk built into a computer system. Furthermore, the "computer-readable recording medium" may include a medium that dynamically holds a program for a short period of time, such as a communication line when transmitting a program via a network such as the Internet or a communication line such as a telephone line, and a medium that holds a program for a certain period of time, such as a volatile memory inside a computer system that is a server or client in that case. Furthermore, the above program may be for realizing part of the functions described above, or may be capable of realizing the functions described above in combination with a program already recorded in a computer system, or may be realized using a programmable logic device such as an FPGA (Field Programmable Gate Array).

[0044] Although an embodiment of the present invention has been described in detail above with reference to the drawings, the specific configuration is not limited to this embodiment, and designs that do not deviate from the gist of the present invention are also included. [Explanation of symbols]

[0045] 1... face monitoring system, 10... camera, 20, 20a, 20b... terminal device, 30... operation terminal, 40... crack detection device, 401... data processing unit, 402... assignment unit, 403... teacher data generation unit, 404... learning unit, 405... storage unit, 406... detection unit, 407... output unit

Claims

1. a synthesis unit that synthesizes an image representing a crack having a shape and color designated by an engineer operating an operation terminal onto an image of a face where a shotcrete has been applied; an assigning unit that assigns a first label representing a crack to a first pixel corresponding to the image representing the crack based on the combined image, and assigns a second label representing something other than a crack to a second pixel other than the first pixel; a training data generation unit that obtains training data by assigning the first label and the second label to the synthesized image; a learning unit that uses the teacher data in which the first pixel corresponding to a crack present in the sprayed concrete among the pixels in the synthesized image is assigned the first label representing a crack and the second pixel corresponding to other than the crack is assigned the second label to learn a relationship between a face image and pixels corresponding to a crack included in the face image; A learning device having the above configuration.

2. a detection unit which inputs the face image to be judged into a trained model which has learned the relationship between a face image and pixels corresponding to cracks contained in the face image using training data in which, among the pixels in an image obtained by synthesizing an image representing a crack having a shape and color specified by an engineer by operating an operation terminal with a face image where a face image where sprayed concrete has been applied is imaged, pixels corresponding to cracks present in the sprayed concrete are assigned a label representing a crack, and pixels corresponding to other than the cracks are assigned a label representing something other than a crack; and which detects whether or not there is a region corresponding to a crack by inputting the face image to be judged and performing segmentation on the face image to be judged into a trained model which has learned the relationship between a face image and pixels corresponding to cracks contained in the face image using training data in which the pixels corresponding to cracks present in the sprayed concrete are assigned a label representing a crack, and pixels corresponding to other than the cracks are assigned a label representing something other than a crack; an output unit that outputs alert information when the detection unit detects that there is a crack; A crack detection device having the above structure.

3. A learning method in which a learning device that is a computer learns, comprising the steps of: The synthesis unit synthesizes an image representing a crack having a shape and a color designated by an engineer operating the operation terminal with respect to the face image captured of the face on which the shotcrete has been applied; an assigning unit assigns a first label representing a crack to a first pixel corresponding to the image representing the crack based on the combined image, and assigns a second label representing something other than a crack to a second pixel other than the first pixel; A learning unit of the computer learns a relationship between a face image and pixels corresponding to cracks included in the face image by using training data in which the first label representing a crack is assigned to the first pixel corresponding to a crack present in the sprayed concrete among the pixels in the synthesized image and the second label is assigned to the second pixel corresponding to other than the crack. How to learn.

4. A detection unit of the crack detection device A face image to be judged is input to a trained model that has learned the relationship between a face image and pixels corresponding to cracks contained in the face image, using training data in which a label representing a crack is assigned to pixels corresponding to cracks present in the shotcrete and a label representing a non-crack is assigned to pixels corresponding to other than the cracks, and the face image to be judged is input to the trained model that has learned the relationship between a face image and pixels corresponding to cracks contained in the face image, and the face image to be judged is segmented to detect whether or not there is a region corresponding to a crack, The output unit of the crack detection device When the detection unit detects that there is a crack, alert information is output. Crack detection methods.

Images