Anomaly image generation device and smoke image generation device

Abstract

To provide an anomaly image generation device and a smoke image generation device, capable of generating anomaly occurrence images or learning smoke images with less time and effort by using an image generation model.SOLUTION: An anomaly image generation device is provided, for generating anomaly occurrence images of an object when anomaly occurs using an image generation model obtained through machine learning using multiple different images including images without the object and anomaly. The anomaly image generation device comprises a generation unit for causing the image generation model to generate multiple anomaly occurrence images based on externally entered input information.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present disclosure relates to an abnormality image generating device and a smoke image generating device. [Background technology]

[0002] In waste storage pits where waste (garbage) is collected, the waste may ignite. Because smoke is generated before the waste ignites, attempts have been made to automatically detect the smoke before it ignites. In relation to this, Patent Document 1 discloses the use of a learning model to detect combustion smoke emitted from combustible materials mixed with waste in waste storage pits. Patent Document 2 discloses the generation of a fire learning image by synthesizing a fire smoke image with a normal surveillance image, which is an image of the normal state of the surveillance area captured by a surveillance camera. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2021-103345 A [Patent Document 2] JP 2018-088630 A Summary of the Invention [Problem to be solved by the invention]

[0004] However, in the technology disclosed in Patent Document 1, it is necessary to obtain a learning model by machine learning using teacher images including images containing combustion smoke as input data, but the frequency of combustion smoke generation in waste storage pits is relatively low, and collecting such teacher images is cumbersome. Also, in the technology disclosed in Patent Document 2, there is room for improvement in the complication of the synthesis process for obtaining fire learning images. Therefore, an object of the present disclosure is to provide an abnormality image generation device and a smoke image generation device that can generate images of abnormalities or learning smoke images with less effort by using an image generation model. [Means for solving the problem]

[0005] One aspect of the present disclosure is an abnormality image generation device that generates an image of an abnormality occurring, which is an image of a subject when an abnormality occurs, by an image generation model obtained by machine learning using a plurality of different images including the subject and images other than the abnormality, and the abnormality image generation device includes a generation unit that causes the image generation model to generate a plurality of images of the abnormality occurring based on input information input from outside. Effect of the Invention

[0006] According to the above aspect, by using an image generation model, a large number of training smoke images can be generated with less effort. [Brief description of the drawings]

[0007] [Figure 1] 1 is an overall diagram showing an example of the configuration of a smoke image generation device 100, an image generation server 200, a learning device 300, and a smoke detection system 400. FIG. [Diagram 2] FIG. 13 is a diagram showing an example of designating a smoke region. [Diagram 3] This is an example of an image showing dust, dirt, sand, and other particles that resemble smoke, as well as water spray from a water cannon. [Figure 4] This is an example of an image showing flames instead of smoke. [Diagram 5] An example image of a waste storage pit with nothing happening. [Figure 6] This is an example of an image in which something other than smoke is generated as a smoke image. [Figure 7] 1 is an example of a training smoke image that includes a crane in the background. [Figure 8] FIG. 4 is a diagram illustrating another example of the configuration of the first embodiment. [Figure 9] FIG. 13 is a diagram illustrating an example of machine learning of a learning model 322. [Figure 10] FIG. 11 is a configuration diagram of a second embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0008] First Embodiment Hereinafter, a first embodiment of a smoke image generating device, a smoke detection system, a smoke image generating method, a smoke detection method, an abnormality image generating device, and a program of the present disclosure will be described with reference to the drawings. The smoke image generating device, the smoke detection system, the smoke image generating method, the smoke detection method, and the program of the present disclosure are for detecting whether smoke has been generated in a waste storage pit of a waste treatment facility and detecting the position where the smoke has been generated. A waste storage pit is a facility for temporarily storing waste collected by a garbage truck or the like. There are more than 1500 waste storage pits in Japan, but there are only about several dozen waste storage pits from which a single company can freely obtain images. In order to stably perform learning for generating a smoke image, it is desirable to learn based on several thousand teacher images. However, since there are only several dozen background patterns of waste storage pits, it is necessary to use a large number (several hundred levels) of images of waste storage pits with the same background pattern. In that case, the accuracy of learning may be insufficient. There is a premise that a large number of training smoke images that serve as teacher images are necessary to sufficiently learn a learning model used in a smoke detection system.

[0009] 1 is an overall diagram showing an example of the configuration of a smoke image generation device 100, an image generation server 200, a learning device 300, and a smoke detection system 400. Below, an overview of each device, system, and server will be given, and each component will be described later.

[0010] The smoke image generation device 100, the image generation server 200, the learning device 300, and the smoke detection system 400 communicate with each other via a network NW. The network NW includes, for example, the Internet, a LAN (Local Area Network), a wireless base station, a provider device, and the like.

[0011] The smoke image generating device 100 causes the image generating server 200 to generate a learning smoke image. The learning smoke image is an image of the waste storage pit captured from above the waste storage pit, and is an image in which smoke is generated. The learning smoke image is an image for the learning device 300 to learn. The image generating server 200 generates a plurality of learning smoke images using the image generating model 210. The image generating server 200 accesses an information source 220 such as an Internet site, and uses words such as "common knowledge" to cause the learning model 210 to learn in advance and acquire general knowledge. When the content is specified by text or the like, the learning model 210 can automatically generate an image that matches the content using AI (artificial intelligence). The learning device 300 learns the learning smoke image and generates a learning model 322. The learning model 322 further performs additional learning using input information described later and smoke images of different patterns (types) that are not generated by the image generating model. The smoke detection system 400 is a system that captures an image of a waste storage pit from above the waste storage pit at a waste treatment facility and detects smoke using a learning model 322.

[0012] Each unit will be described below. The smoke image generating device 100 includes, for example, an acquisition unit 110, an annotation unit 120, a generation unit 130, a selection unit 140, and a storage unit 150. The acquisition unit 110 acquires input information input from the outside. The input information is, for example, the position (coordinates) of the smoke, the area generating the smoke, the type of smoke, and images of multiple waste storage pits. The input from the outside is performed, for example, using an input / output device 160 connected to the smoke image generating device 100.

[0013] The annotation unit 120 annotates the training smoke image using input information such as the position of the smoke, the area of ​​the smoke, and the type of smoke. Annotation refers to, for example, adding information or annotations to an image, or classifying it by adding attributes. The area of ​​the smoke is information that specifies the area that indicates the part of the image of the waste storage pit that is emitting smoke. This is called area information. The annotation unit 120 may perform annotation by accepting a human operation.

[0014] FIG. 2 is a diagram showing an example of specifying a smoke area. FIG. 2 shows an image of a waste storage pit from which smoke is being generated. The area is represented by a rectangle. Point A and point B indicate the position of the smoke. The coordinates of point A (X1.Y1) and point B (X2.Y2) are input, and a rectangle is drawn with points A and B as vertices and the intersection of lines extending from each point, and the enclosed area is designated as the smoke area. The area information may be only point A or point B, or may be the coordinates of four points. The annotation unit 120 indicates the area where it is determined that smoke is being generated by drawing a rectangle as shown in FIG. 2.

[0015] The generation unit 130 provides the input information to the image generation model 210 to obtain a training smoke image. The generation unit 130 performs tuning by providing the image generation model 210 with an actual smoke image and input information before and after obtaining the training smoke image. Furthermore, the training smoke image generated by the image generation model 210 may be used for additional learning of the image generation model 210. Additional learning is learning to generate training smoke images including fermentation gas, dust, soil, sand, and other particles other than smoke at the time of a fire, and water spray from a water cannon, in addition to the above-mentioned tuning, using training smoke images including dust, dust, soil, sand, and other particles other than smoke at the time of a fire, and water spray from a water cannon. The image generation model 210 becomes able to detect smoke caused by a fire by learning using these training smoke images. FIG. 3 is an example of an image in which dust, soil, sand, and other particles similar to smoke, and water spray from a water cannon are flying up.

[0016] The selection unit 140 causes the selection model 152 to select the learning smoke images generated by the image generation model 210. The learning smoke images are images of the waste storage pit captured from above the waste storage pit with smoke, but images other than images with smoke are also generated. Since it is desired that the learning model 322 learn only images with smoke, the selection unit 140 needs to select and exclude learning smoke images with things other than smoke. The selection model 152 is a learning model that has learned a large number of images related to smoke and waste storage pits. The selection model 152 can determine whether an image includes smoke or a waste storage pit, and excludes images of smoke and images that do not include a waste storage pit (for example, an image of a garbage dump with garbage piled up on flat ground). The images to be excluded are, for example, the images in FIG. 4 to FIG. 6. Examples of the images include images of a state in which smoke is generated in a place other than the waste storage pit, and images of a waste storage pit with things other than smoke (including flames) generated. FIG. 4 is an example of an image showing flames but not smoke. FIG. 5 is an example of an image of a waste storage pit where nothing is occurring. FIG. 6 is an example of an image where something other than smoke is generated as a smoke image. These images are to be rejected. The selection of images may be achieved, for example, by an image processing algorithm.

[0017] The storage unit 150 stores the selection model 152. The selection model 152 is as described above. The storage unit 150 may be realized by a hard disk drive (HDD), an electrically erasable programmable read only memory (EEPROM), a read only memory (ROM), a random access memory (RAM), or the like.

[0018] The components of the smoke image generating device 100 other than the storage unit 150 are realized by, for example, a hardware processor such as a CPU (Central Processing Unit) executing a program (software). Some or all of these components may be realized by hardware (including circuitry) such as an LSI (Large Scale Integration), an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), a GPU (Graphics Processing Unit), or a TPU (Tensor Processing Unit), or may be realized by cooperation between software and hardware. The program may be stored in advance in a storage device such as an HDD or flash memory (a storage device having a non-transient storage medium), or may be stored in a removable storage medium such as a DVD or CD-ROM (a non-transient storage medium), and installed by mounting the storage medium in a drive device.

[0019] The image generation server 200 includes an image generation model 210. As described above, the image generation model 210 is a model obtained by learning in advance based on information acquired from the information source 220. For example, when "Please generate an image of smoke" is input as the first input information, the image generation model 210 generates an image of smoke. The generated training smoke images are images with different backgrounds. The background includes the wall of the waste storage pit. The image generation model 210 also generates a training smoke image including a crane in the background of the training smoke image. FIG. 7 is an example of a training smoke image including a crane as the background. Since the waste storage pit also performs work of stirring the waste, the image generation model 210 also generates a training smoke image including the stirred waste. The image generation server 200 may be an image generation device, a device equipped with an image generation function, or a device that holds an image generation function.

[0020] The learning device 300 includes, for example, a model generation unit 310 and a storage unit 320. The model generation unit 310 generates a learning model 322. The storage unit 320 stores the learning model 322. The components other than the storage unit 320 are realized by, for example, a hardware processor such as a CPU executing a program (software). Some or all of these components may be realized by hardware (including circuitry) such as an LSI, ASIC, FPGA, or GPU, or may be realized by cooperation between software and hardware. The program may be stored in advance in a storage device such as an HDD or flash memory (a storage device having a non-transient storage medium), or may be stored in a removable storage medium such as a DVD or CD-ROM (a non-transient storage medium), and may be installed by mounting the storage medium in a drive device.

[0021] The smoke detection system 400 includes, for example, an imaging unit 410, a determination unit 420, a notification unit 430, and a storage unit 440. The imaging unit 410 captures an image of the waste storage pit from above the waste storage pit. The imaging unit 410 is connected to a camera (not shown) wirelessly or via a wire, and constantly shares captured images. The camera is, for example, a seamless camera. A seamless camera is a camera that can continue capturing images without frame loss by automatically switching recording media even if the capacity of the recording media runs out during capturing.

[0022] The determination unit 420 determines whether or not smoke is occurring in the image linked from the imaging unit 410 using the learning model 322. If the determination unit 420 determines that smoke is occurring, it further determines what type of smoke it is. The determination unit 420 links the determination result (determination result) to the notification unit 430.

[0023] The storage unit 440 includes the learning model 322. The learning model 322 is a learning model 322 linked from the learning device 300.

[0024] The components of the smoke detection system 400 other than the storage unit 440 are realized, for example, by a hardware processor such as a CPU executing a program (software). Some or all of these components may be realized by hardware (including circuitry) such as an LSI, ASIC, FPGA, or GPU, or may be realized by a combination of software and hardware. The program may be stored in advance in a storage device such as an HDD or flash memory (a storage device having a non-transient storage medium), or may be stored in a removable storage medium (non-transient storage medium) such as a DVD or CD-ROM, and installed by inserting the storage medium into a drive device.

[0025] Fig. 8 is a configuration diagram of another example of the configuration of the first embodiment. In addition to the above-mentioned configuration, for example, as shown in Fig. 8, the configuration of the smoke image generation device 100 and the image generation server 200 may be implemented as one smoke image generation device A, and the learning device 300 and the judgment unit 420 and the memory unit 440 of the smoke detection system 400 may be implemented as one smoke detection device B. In that case, the imaging unit 410 and the notification unit 430 operate as a smoke detection device C. Furthermore, the smoke image generation device 100, the image generation server 200, the learning device 300, and the judgment unit 420 and the memory unit 440 of the smoke detection system 400 may be implemented as one smoke detection device.

[0026] An example of the operation of the present disclosure will be described below. The acquisition unit 110 acquires input information and an image of smoke. The input information includes first input information and second input information. The first input information is, for example, text such as "Please generate an image of smoke." The first input information may include type information (pattern) indicating the type of smoke. The second input information is, for example, text d such as "Please generate smoke of pattern 2." For example, the patterns are pattern 1 for smoke generated when a fire occurs, pattern 2 for smoke generated when fermentation gas occurs, pattern 3 for dust particles such as dirt, soil, and sand, and pattern 4 for water spray from a water gun. The acquisition unit 110 acquires the image generation model 210 by additional learning using the first input information and the image of smoke. This is called an acquisition step.

[0027] Annotation unit 120 annotates the acquired training smoke image. It annotates which parts of the training smoke image are areas where smoke is occurring. Annotation unit 120 uses input information for the annotation. Annotation unit 120 transmits the annotated training smoke image to learning device 300 (dashed arrow (1) in the figure).

[0028] The generation unit 130 inputs the second input information and causes the image generation model 210 to generate training smoke images (dashed arrow (2) in the figure). The image generation model 210 generates a plurality of training smoke images. This is referred to as the generation step.

[0029] The image generation model 210 transmits the generated training smoke image to the generation unit 130 (dashed arrow (3) in the figure).

[0030] The generated training smoke images may contain images other than smoke, so it is necessary to select only smoke images. The selection unit 140 uses the selection model 152 to select only images in which smoke is generated in the waste storage pit from the training smoke images. This is called the selection step.

[0031] 9 is a diagram showing an example of machine learning of the learning model 322. The model generation unit 310 learns the parameters of the machine learning model by a method such as backpropagation so that the output of the machine learning model when a teacher image including a plurality of learning smoke images is used as input data (323A) approaches the output parameters (323B) of the judgment result of the learning smoke image. For example, the machine learning model at the point in time when the above process is executed a prescribed number of times is determined as the learning model 322.

[0032] The learning device 300 transmits the generated learning model 322 to the smoke detection system 400 (dashed arrow (4) in the figure).

[0033] The imaging unit 410 captures an image of the waste storage pit from above the waste storage pit. The imaging unit 410 transmits the captured image to the determination unit 420. The determination unit 420 uses the learning model 322 to determine whether or not smoke is occurring in the waste storage pit. The determination unit 420 transmits the determination result to the notification unit 430. The determination unit 420 can detect even when multiple smokes are occurring in one waste storage pit.

[0034] The notification unit 430 changes the notification method depending on the determination result. For example, when the notification unit 430 is linked with the determination result that smoke is being generated, the notification unit 430 notifies the fact that smoke is being generated and the location. The notification unit 430 may be connected to a terminal device and output the determination result. The terminal device is, for example, a tablet terminal.

[0035] For example, when the determination result that smoke is not being generated is linked, the notification unit 430 may not need to perform any particular notification. Also, the notification unit 430 may display on the tablet terminal that smoke is not being generated.

[0036] According to the first embodiment described above, the annotation unit 120 performs annotation, which enables efficient learning when learning smoke images in the learning model 322. The selection model 152 can select the necessary learning smoke images, which reduces the amount of work required. By having the learning model 322 learn the learning smoke images that have been subjected to these processes, smoke generation in the waste storage pit can be detected more accurately.

[0037] The smoke image generating device is an image generating device that generates training smoke images, which are images of a waste storage pit when smoke is generated, using an image generation model obtained by machine learning using multiple different images to train a learning model that outputs information regarding the presence of image portions of smoke when an image is input, and is equipped with an acquisition unit that acquires input information indicating the amount and type of smoke input from the outside, and a generation unit that causes the image generation model to generate multiple training smoke images by inputting the input information into the image generation model. By having the smoke image generation device generate training smoke images using an image generation model, a large number of training smoke images can be prepared with little effort.

[0038] The image generation model is incrementally trained using input information and smoke images that are not generated by the image generation model. Through additional learning, the smoke image generation device can obtain an image generation model that can more reliably generate images of the waste storage pit when smoke is generated.

[0039] The input information is an image used as a background image by the image generation model, and includes an image of the waste storage pit when no smoke is being generated, and area information indicating an area in the image of the waste storage pit when no smoke is being generated that generates smoke. By inputting the above-mentioned input information, the burden of having the image generation model learn the waste storage pit when no smoke is generated is reduced, and the smoke image generation device can generate learning smoke images of the waste storage pit when smoke is generated, with smoke occurring in various positions.

[0040] The image generation model is additionally trained using input information, smoke images not generated by the image generation model, and images of the waste storage pit when no smoke is being generated. Through additional training, the image generation model can generate training smoke images of waste storage pits during smoke generation, including a limited number of waste storage pits, as well as a variety of smoke images.

[0041] The input information includes area information indicating an area in a training smoke image generated by the image generation model that generates smoke. By inputting the above-mentioned input information, the image generation model can generate training smoke images of a waste storage pit where smoke is occurring at various positions.

[0042] A smoke image generating device that can handle a plurality of types of waste storage pits, and input information includes pit type information that indicates the type of waste storage pit when no smoke is being generated. By inputting the above-mentioned input information, the image generation model is able to generate training smoke images according to the type of waste storage pit.

[0043] The smoke image generating device is capable of handling a plurality of types of smoke, and the input information includes smoke type information indicating the type of smoke. By inputting the above-mentioned input information, the image generation model becomes able to generate training smoke images according to the type of smoke generated in the waste storage pit.

[0044] A plurality of pieces of region information are input to an image generation model as input information. By inputting multiple region information, the image generation model can generate images of the waste pit with smoke occurring in multiple locations.

[0045] A smoke image generating device capable of handling multiple types of smoke, wherein the input information includes smoke type information indicating the type of smoke, and further comprises an annotation unit that annotates areas where smoke is generated in multiple learning smoke images using the input information and area information. By providing an annotation section, the effort required for manually performing annotation work can be reduced.

[0046] The generator generates a different background for each of the plurality of training smoke images. By the generation unit generating training smoke images with a variety of different backgrounds, it is possible to prevent a learning model that learns from training smoke images with the same background from determining that the background is a feature.

[0047] The generator generates a training smoke image that includes a wall of a waste storage pit in the background. By the generation unit generating training smoke images in which the walls of various different waste storage pits are in the background, a learning model that learns from training smoke images in which the walls of the waste storage pits are the same can be prevented from determining that the background including similar waste storage pit walls is a feature.

[0048] The generator generates a training smoke image including a crane in the background. By the generation unit generating training smoke images that include various cranes in the background that dump waste into the combustion furnace, a learning model that learns from training smoke images that include the same cranes in the background can be prevented from determining that the cranes are a feature.

[0049] The generator generates a plurality of training smoke images including an image of the agitated waste. By the generation unit generating multiple training smoke images including images of agitated waste, it becomes easier to generate training smoke images of agitated waste, and it is possible to prevent a learning model that learns training smoke images including agitated waste from determining that agitated waste is a characteristic feature.

[0050] The smoke detection system detects smoke generated in a waste storage pit in which combustible waste is stored, and includes an imaging unit that captures an image of the top surface of waste stored in the waste storage pit, a determination unit that determines whether or not an image portion of smoke is present in the image captured by the imaging unit, and a notification unit that issues a notification in accordance with the determination result by the determination unit. The determination unit inputs the image captured by the imaging unit to a learning model, and makes a determination based on a probability parameter that indicates the probability of the presence of an image portion of smoke output from the learning model. The learning model receives a plurality of teacher images, which are images of the top surface of waste stored in the waste storage pit and include images in which an image portion of smoke is present, as input data, and a teacher image The learning model is generated in advance by machine learning, using as training data an input / output dataset whose output parameters are the judgment results of the image portions of smoke contained in each of the training data, the input data is generated by an image generation model obtained by machine learning using first input information from the outside and images of smoke, the judgment results are information input to the image generation model, each of the training images includes at least two or more of an image portion of a first type of smoke, an image portion of a second type of smoke, and an image portion of a third type of smoke, and the judgment results are a result of judging whether the smoke corresponding to the image portion of the smoke contained in each of the training data is the first type of smoke, the second type of smoke, or the third type of smoke. The above-mentioned smoke detection system can accurately determine the type of smoke generated in the waste storage pit using a learning model generated by machine learning using multiple training smoke images generated by an image generation model.

[0051] The smoke image generating method is a method for generating a learning model for smoke detection in which a smoke image generating device generates a learning model capable of detecting smoke generation in a waste storage pit based on an image generation model obtained by machine learning using a plurality of mutually different images, and includes an acquisition step of acquiring an image generation model by additional learning using first input information from the outside and smoke images, a generation step of generating multiple learning smoke images, which are images of a waste storage pit from which smoke is generated, based on second input information from the outside using the image generation model acquired in the acquisition step, a selection step of selecting the multiple learning smoke images generated in the generation step, and a deep learning step of obtaining a learning model by deep learning using the learning smoke images selected in the selection step. By executing the above steps, a large number of smoke images can be prepared with little effort because training smoke images are generated using an image generation model, and a training model can be obtained by training the prepared images using deep learning. By selecting training smoke images and training the training model using the correct training smoke images, detection accuracy can be improved.

[0052] There are a plurality of types of smoke, and each of the first input information and the second input information includes type information indicating the type of smoke. By inputting the above-mentioned smoke type information, the detection accuracy of the learning model obtained through deep learning is improved.

[0053] The selection step determines whether or not the multiple training smoke images generated in the generation step contain flame image portions, and selects training smoke images determined not to contain flame image portions from among the multiple training smoke images generated in the generation step. By selecting training smoke images that include the flame image portions described above, the detection accuracy of the learning model obtained by deep learning is improved.

[0054] The acquisition step acquires an image generation model by additional learning using an image of the waste storage pit and an image of smoke based on first input information and second input information from outside. By having the image generation model generate training smoke images that include not only images of smoke but also images of waste storage pits, the number of which is limited, the generated training smoke images can be used for learning, improving the detection accuracy of the learning model obtained through deep learning.

[0055] <Second embodiment> Fig. 10 is a configuration diagram of the second embodiment. The difference from the first embodiment is that the smoke image generation device 100 holds an image generation model 210. The storage unit 150 stores a selection model 152 and the image generation model 210. In the first embodiment, the smoke image generation device 100 and the image generation server 200 communicate with each other to generate the image generation model 210, but this can be completed within the smoke image generation device 100. According to the second embodiment, it is possible to achieve the same effects as the first embodiment.

[0056] The above describes the form for implementing the present disclosure using embodiments, but the present disclosure is in no way limited to these embodiments, and various modifications and substitutions can be made within the scope that does not deviate from the gist of the present disclosure. [Explanation of symbols]

[0057] 100 Smoke image generating device 110 Acquisition Department 120 Annotation Section 130 Generation part 140 Sorting Department 150 Storage section 152 Selection Model 160 I / O Devices 200 Image Generation Server 210 Image Generation Model 220 Sources of information 300 Learning Device 310 Model Generation Unit 320 Storage section 322 Learning Model 400 Smoke Detection System 410 Imaging unit 420 Judgment section 430 Information Department 440 Storage section NW Network

Claims

1. An abnormality image generating device that generates an abnormality occurrence image, which is an image of a subject when an abnormality occurs, by an image generation model obtained by machine learning using a plurality of images different from each other, the images including the subject and images other than the abnormality, A generation unit that causes the image generation model to generate a plurality of images at the time of the abnormality occurrence based on input information input from an external source. Anomaly image generating device.

2. The image generation model is additionally trained using the input information, the image when an abnormality occurs, and the image of the subject when no abnormality occurs. The abnormality image generating device according to claim 1 .

3. The input information includes area information indicating an area in the image when an abnormality occurs, the area being generated by the image generation model, where the abnormality occurs.

3. The abnormality image generating device according to claim 1 or 2.

4. A smoke image generating device that generates training smoke images, which are images of a waste storage pit when smoke is generated, using an image generation model obtained by machine learning using a plurality of mutually different images including images other than the waste storage pit and smoke, for use in training a learning model that outputs information regarding the presence of smoke when an image is input, the training smoke images being images of the waste storage pit when smoke is generated, A generation unit that generates a plurality of the learning smoke images in the image generation model by inputting input information into the image generation model. Smoke image generator.

5. The image generation model is additionally trained using the input information and smoke images not generated by the image generation model. The smoke image generating device according to claim 4 .

6. The input information is information indicating the position, type, and amount of smoke in the learning smoke image.

6. The smoke image generating device according to claim 4 or 5.

Images