JP2022011411A - Inference device - Google Patents

Abstract

To increase the generalization when updating a map image for detecting a change point of a house shape or reading a transfer of a house.SOLUTION: In an inference device 1, an input unit 2 receives an input of a map image MP and an aircraft picture Ap corresponding to the region of the map image MP for updating the map image MP, and a generation unit 3 generates an overlapped image SI in which the map image MP and the aircraft picture AP corresponding to the region of the map image MP are overlapped with each other. An inference unit 4 infers a change area in the region by applying a learned model 4a obtained by a mechanical learning using as input data a learning overlapped image formed by overlapping a learning map image and a learning aircraft image corresponding to the region of the learning map image and using a change area in the region of the learning map image as teacher data, to an overlapped image SI.SELECTED DRAWING: Figure 2

Description

The present invention relates to an inference device that makes inferences based on a trained model obtained by machine learning.

Until now, image processing has been used to detect changes in the shape of a house on a map and to automate the interpretation of house changes. For example, in the house change interpretation support device described in Patent Document 1, the change candidate detection unit performs a house change in a 2 hour period based on spatial information measurement data consisting of DSM data and aerial photographed images in the new / old two periods. The location where the transfer is detected is extracted as a candidate region, and the stereoscopic image generation unit generates a stereoscopic image of the target region based on the spatial information measurement data. Then, the display screen generation unit displays on the screen of the display unit an image showing the detection result by the transfer candidate detection unit and the stereoscopic image by the stereoscopic image generation unit on the stereoscopic image by the stereoscopic image acquisition unit. By doing so, the interpreter can easily grasp the height information in the determination target area from the stereoscopic image by observing the physical image displayed on the screen, and new construction, loss, etc. It is possible to efficiently determine the change of a house that accompanies a change in height.

The image processing method as described in Patent Document 1 identifies the shape of a building from features such as shape and color, and detects a change point by taking a difference from map data or the like. However, there is a problem that the generalization and robustness are low only by image processing, and the detection accuracy is lowered depending on an unknown building or shooting conditions.

Japanese Unexamined Patent Publication No. 2012-146050

With the recent improvement in the detection performance of machine learning and deep learning technology, attempts are being made to automate map update work such as house shape change point detection and house change interpretation using deep learning. It is known that in deep learning, the shape of a building and the surrounding situation can be learned from a huge amount of teacher data, and high generalization can be obtained. The detection of the change point is realized by taking the difference between the inferred shape of the building and the map data before the update.

However, the method of inferring the shape of the building by deep learning and taking the difference from the map data before the update has the following problems. First, the shape of the building inferred by deep learning becomes unstable, such as being large, small, or distorted. Therefore, even if the difference between the inferred shape and the map data is taken, the standard for judging whether or not the difference is correct becomes a rule base and generalization is reduced. Therefore, the accuracy of updating the map image is lowered. In particular, when changing a building (two buildings are one building, an L-shaped building is a square, etc.), it is necessary to infer an accurate building shape, which is difficult to detect by the conventional method.

As an example of the problem to be solved by the present invention, it is possible to update the map image more accurately when the map image such as the house shape change point detection and the house change interpretation is updated.

In order to solve the above problems, the invention according to claim 1 has an input unit for inputting a map image, an aerial captured image for updating the map image corresponding to the area of the map image, and the map. A learning superimposed image in which a generation unit that generates a superimposed image in which an image and the aerial captured image are superimposed, a learning map image, and a learning aerial captured image corresponding to a region of the learning map image are superimposed. Is used as input data, and a trained model obtained by machine learning using the change area in the region represented by the superimposed image for training as teacher data is applied to the superimposed image to infer the change region in the superimposed image region. It is characterized by having a part and.

The invention according to claim 6 is an inference method executed by an inference device that makes inferences based on a trained model obtained by machine learning, and corresponds to a map image and a region of the map image. An input step for inputting an aerial captured image for updating a map image, a generation step for generating a superimposed image in which the map image and the aerial captured image are superimposed, a learning map image, and the learning map image. A trained model obtained by machine learning using an aerial image for learning corresponding to the area of the above and a superimposed image for learning superimposed as input data, and a change area in the area represented by the superimposed image for learning as teacher data. Is applied to the superimposed image to infer a change area in the region represented by the superimposed image, and the like.

The invention according to claim 7 is characterized in that the inference method according to claim 6 is executed by a computer as an inference program.

The invention according to claim 8 is characterized in that the inference program according to claim 7 is stored in a storage medium readable by a computer.

It is a schematic block diagram of the computer which functions as the inference device which concerns on one Embodiment of this invention. FIG. 3 is a functional configuration diagram of an inference device in which a computer functions as shown in FIG. 1. It is explanatory drawing about the generation of a superimposition image. It is a functional block diagram of the learning apparatus which generates the trained model which concerns on this Example. This is an example of a neural network formed in a trained model. It is a flowchart of the operation of the learning apparatus shown in FIG. It is explanatory drawing about the generation of teacher data. It is a flowchart of the operation of the inference device shown in FIG. It is a figure which showed the input superimposition image, the teacher data corresponding to the superimposed image, and the inference result performed using the trained model after learning. It is a table which compared the inference device shown in FIG. 2 with the prior art.

Hereinafter, the inference device according to the embodiment of the present invention will be described. In the inference device according to the embodiment of the present invention, a map image and an aerial captured image for updating the map image corresponding to the area of the map image are input to the input unit, and the generation unit is the map image and the air. Generates a superimposed image that is superimposed on the captured image. Then, the inference unit uses the learning superimposed image in which the learning map image and the learning aerial image captured corresponding to the area of the learning map image are superimposed as input data, and changes in the area represented by the learning superimposed image. The trained model obtained by machine learning using the region as teacher data is applied to the superimposed image, and the change region in the region represented by the superimposed image is inferred. By doing so, it is possible to infer the change area from the superimposed image in which the map image and the aerial image are superimposed by the trained model in which the change area of the target area is learned. It is not necessary to infer the shape of the inferred shape, and the decrease in generalization due to the instability of the inferred shape is suppressed. Therefore, it is possible to update the map image more accurately when the map image such as the house shape change point detection and the house change interpretation is updated.

In addition, the trained model is machine-learned using the building outline corresponding to each of the new installation, dismantling, and change in the superposed image for learning as teacher data, and the inference unit is used for each of the new installation, dismantling, and change in the aerial captured image. The corresponding range of change may be inferred. By doing so, it becomes possible to infer new construction, demolition, and changes of buildings on the map.

Further, the inference unit may infer the change area based on the probability of the change point for each pixel of the superimposed image output from the trained model. By doing so, by setting a predetermined threshold value or the like for the probability output for each pixel, it becomes possible to infer a change area such as new construction, demolition, or change of a building.

The superimposed image is configured by setting the data of the aerial image captured in the R channel and the G channel in the RGB format and the data of the map image in the B channel, respectively, and the superimposed image for learning is the R channel in the RGB format. It may be configured by setting the data of the aerial image for learning in the G channel and the data of the map image for learning in the B channel, respectively. By doing so, it is possible to construct a superimposed image or the like by using the RGB format. Further, by using the B channel as a map image or a learning map image, it becomes easy to distinguish the superimposed image from the aerial captured image and the learning aerial captured image even when the superimposed image is observed by a human.

Further, the teacher data may be an image in which label information corresponding to the characteristics of the changing region is set for each pixel. By doing so, by using the teacher data as an image, the input data and the format can be matched, and the learning efficiency and the processing efficiency can be improved.

Further, in the inference method according to the embodiment of the present invention, the map image and the aerial captured image for updating the map image corresponding to the area of the map image are input in the input step, and the map image is generated in the generation step. And an aerial captured image are superimposed to generate a superimposed image. Then, in the inference process, the learning superimposed image obtained by superimposing the learning map image and the learning aerial image captured corresponding to the area of the learning map image is used as input data, and the change in the area represented by the learning superimposed image is used. The trained model obtained by machine learning using the region as teacher data is applied to the superimposed image, and the change region in the region represented by the superimposed image is inferred. By doing so, it is possible to infer the change area from the superimposed image in which the map image and the aerial image are superimposed by the trained model in which the change area of the target area is learned. It is not necessary to infer the shape of the inferred shape, and the decrease in generalization due to the instability of the inferred shape is suppressed. Therefore, it is possible to update the map image more accurately when the map image such as the house shape change point detection and the house change interpretation is updated.

Further, the above-mentioned inference method may be executed by a computer as an inference program. By doing so, it is possible to infer the change area by the trained model in which the change area of the target area is learned from the superimposed image in which the map image and the aerial image are superimposed by using a computer. As in the case, it is not necessary to infer the shape of the building, and the decrease in generalization due to the instability of the inferred shape can be suppressed. Therefore, it is possible to update the map image more accurately when the map image such as the house shape change point detection and the house change interpretation is updated.

Further, the above-mentioned inference program may be stored in a computer-readable storage medium. By doing so, the program can be distributed as a single unit in addition to being incorporated in the device, and version upgrades and the like can be easily performed.

An inference device according to an embodiment of the present invention will be described with reference to FIGS. 1 to 10. FIG. 1 is a schematic configuration diagram of a computer 10 that functions as an inference device according to this embodiment.

As shown in FIG. 1, the computer 10 includes a calculation unit 11, a main storage unit 12, a communication unit 13, and an auxiliary storage unit 14. The arithmetic unit 11, the main storage unit 12, the communication unit 13, and the auxiliary storage unit 14 are connected to each other by the bus B.

The arithmetic unit 11 has an arithmetic processing unit such as a CPU (Central Processing Unit), an MPU (Micro-Processing Unit), and a GPU (Graphics Processing Unit), and reads and executes a program or the like stored in the auxiliary storage unit 14. As a result, various information processing, control processing, and the like related to the computer 10 are performed.

The main storage unit 12 is a temporary storage area such as a SRAM (Static Random Access Memory) and a DRAM (Dynamic Random Access Memory), and temporarily stores data necessary for the calculation unit 11 to execute a calculation process.

The communication unit 13 is a communication module for performing processing related to communication, and transmits / receives information to / from the outside.

The auxiliary storage unit 14 is a large-capacity memory, a hard disk, or the like, and stores programs and other data necessary for the arithmetic unit 11 to execute processing. Further, the auxiliary storage unit 14 stores the trained model 4a. The trained model 4a is a model for inferring the changed area on the map, and is a trained model generated by machine learning.

In addition to the above-mentioned components, the computer 10 may be provided with an external interface such as USB (Universal Serial Bus), a keyboard, a mouse, a display, and the like, if necessary.

FIG. 2 shows a functional configuration diagram of the inference device 1 in which the computer 10 described above functions. As shown in FIG. 2, the inference device 1 includes an input unit 2, a generation unit 3, and an inference unit 4.

For example, the communication unit 13 functions as the input unit 2, and the map image MP and the aerial photograph AP are input. The map image in this embodiment is an image of a map represented in a state where a road, a building, or the like is viewed from the sky in a plan view. Further, as is well known, the aerial photograph in this embodiment refers to a photograph of the ground surface taken by a camera from an in-flight flying object. Further, for the aerial photograph, it is preferable to use an ortho image in which the distortion caused by the terrain is corrected, but the distortion correction may not be performed. Further, in this embodiment, an aerial photograph will be described as an aerial image, but a satellite photograph may be used. Alternatively, it may be an image obtained by an aeronautical rider that radiates a laser beam from an aircraft toward the ground as an aerial image and measures the terrain or the like by the time difference of the laser beam reflected from the ground.

The generation unit 3 superimposes the map image MP and the aerial photograph AP input to the input unit 2 to generate the superposed image SI. Here, the generation of the superimposed image SI will be described with reference to FIG.

In FIG. 3, the map image MP is a map image before updating, and is vector data (polygon image) in this embodiment. The aerial photograph AP is a color image in RGB format in this embodiment. Further, the map image MP and the aerial photograph AP indicate the same area (region). That is, the aerial photograph AP is an aerial captured image for updating the map image MP corresponding to the area of the map image MP.

The aerial photograph AP is decomposed into each component of RGB. The R (red) component of the aerial photograph AP is designated by the reference numeral APr, the G (green) component of the aerial photograph AP is designated by the reference numeral APg, and the B (blue) component of the aerial photograph AP is designated by the reference numeral APb. Then, the R component image APr and the G component image APg are combined to generate the RG component image APrg. Then, an RGB format image in which the RG component image APrg is set as the R component, the B component image APb is set as the G component, and the map image MP is set as the B component is generated as the superimposed image SI.

In the superimposed image SI generated in this way, the house shape of the building or the like by the map image MP looks bluish to the human eye, and the aerial photograph AP part looks like a yellowish color. This is due to the difference in wavelength characteristics that the blue cones, green cones, and red cones of the photoreceptor cells that make up the human retina are sensitive to. Specifically, the blue cone is most sensitive to a wavelength of about 430 nm, or blue, the green cone is most sensitive to a wavelength of about 530 nm, that is, green, and the red cone is most sensitive to a wavelength of about 560 nm, that is. It is known to be the most sensitive to red. Therefore, by inserting the map image MP into the B component, the map image MP can be visually recognized as a bluish color.

The inference unit 4 applies the trained model 4a to the superimposed image SI to infer the change region in the region represented by the superimposed image SI. Here, inference is, as is well known in the field of AI (Artificial Intelligence), to apply to an inference (learned) model generated by learning and derive the result.

The trained model 4a uses a learning superimposed image in which a learning map image and a learning aerial image captured corresponding to the area of the learning map image are superimposed as input data, and is used in a region represented by the learning superimposed image. It was obtained by machine learning using the change area as teacher data. That is, the trained model 4a can be inferred by learning based on the teacher data.

Here, the trained model 4a will be described. The trained model 4a according to this embodiment is generated by a learning device 20 having a functional configuration as shown in FIG. 4, for example. The learning device 20 includes an input unit 22, a generation unit 23, and a learning unit 24. Further, the learning device 20 can be realized by operating the computer 10 shown in FIG.

For example, the communication unit 13 functions as the input unit 22, and the learning map image MPL and the learning aerial photograph APL are input. The generation unit 23 superimposes the learning map image MPL input to the input unit 22 and the learning aerial photograph APL, and generates a learning superposed image SIL. The input unit 22 and the generation unit 23 basically operate in the same manner as the input unit 2 and the generation unit 3. The learning map image MPL is different from the map image MP only in its usage, and there is no difference in the data format and the like. Similarly, the aerial photograph APL for learning has only a different purpose from the aerial image AP, and there is no difference in the data format and the like.

The learning unit 24 uses the learning superimposed image SIL generated by the generation unit 23 as input data, and machine-learns the change area in the region indicated by the learning map image MPL and the learning aerial photograph APL as teacher data, and is a trained model. Obtain 4a. In this embodiment, the change area indicates an area where the outer shape of the building has been updated, such as new construction, demolition, or change of the building, as will be described later.

The learning device 20 performs deep learning as a trained model 4a to learn a change area (update area such as new installation, disassembly, and change) from a learning map image in a learning aerial photograph, thereby performing a superposed image. A neural network is constructed (generated) that takes SI as an input and outputs a change area. For example, a neural network is a CNN (Convolution Neural Network), and has an input layer that accepts inputs of superimposed image SI, an output layer that outputs information indicating a change region, and an intermediate layer that extracts features (FIG. 5). reference).

In this embodiment, the trained model 4a is described as being a CNN, but the trained model 4a is not limited to the CNN and is constructed by other learning algorithms such as a neural network other than the CNN, semantic segmentation, and instance segmentation. It may be a trained model that has been trained.

The operation (learning method) in the learning device 20 will be described with reference to the flowchart of FIG. First, teacher data is generated (step S11). The generation of teacher data will be described with reference to FIG. 7.

First, the prediction is performed, and the attributes of new construction, demolition, and change of the building are given to the map data BE before editing and the map data AE after editing. The prediction is to extract the target to which the attribute is given by the worker by comparing the map data BE before editing and the map data AE after editing.

New data A, dismantling data D, and change data C are generated from the map data to which the attributes are given by performing the prediction. In the new construction data A, the dismantling data D, and the change data C shown in FIG. 7, the areas such as the newly constructed, demolished, and changed buildings in the target area of the map data are shown in white. That is, in the same range, the newly constructed building or the like is shown in the new construction data A, the dismantled building or the like is shown in the dismantling data D, and the changed building or the like is shown in the change data C. Here, the change means that the shape of the building or the like has been changed, for example, two buildings are one building and an L-shaped building is a square.

Then, the label image L is generated by synthesizing the newly installed data A, the disassembled data D, and the changed data C. The label image L is an image in which predetermined pixel values are set as label information for the pixels corresponding to the areas of new installation, disassembly, and change. In the example of FIG. 7, it is an 8-bit grayscale image in which the pixel value “0” is set for no change, the pixel value “1” is set for disassembly, the pixel value “2” is set for change, and the pixel value “3” is set for new installation. Therefore, even if the label image L is visually recognized by a human, it is difficult to identify each label, but there is no problem because it can be identified as data. The label image L generated in this way becomes the teacher data. That is, the teacher data is an image in which label information corresponding to the characteristics of the changing region is set for each pixel. Then, the trained model 4a is machine-learned using the building outline corresponding to each of new construction, dismantling, and change in the superposed image SIL for learning as teacher data.

The teacher data may be generated by the computer 10 or another computer or the like. When the computer 10 is used, the predicted map data before editing and the map data after editing may be input to the input unit 22 shown in FIG. 4, and the label image L may be generated by the generation unit 23.

Returning to the flowchart of FIG. 6, the learning map image MPL and the learning aerial photograph APL are input to the input unit 22 (step S12). The learning map image MPL and the learning aerial photograph APL input in this step indicate a region corresponding to the teacher data (label image L) generated in step S11.

Next, the generation unit 23 generates a learning superimposed image SIL (step S13). Then, learning is performed using the superposed image SIL for learning and the teacher data (step S14). The above operation is repeated by changing the learning map image MPL, the learning aerial photograph APL, and the teacher data, and the trained model 4a is generated.

That is, the learning device 20 inputs a learning map image MPL and a learning aerial photograph APL (learning aerial captured image) for updating the learning map image MPL corresponding to the area of the learning map image MPL. The input unit 22, the generation unit 23 that generates the learning superimposed image SIL by superimposing the learning map image MPL and the learning aerial photograph APL (learning aerial image captured image) corresponding to the area of the learning map image MPL, and the generation unit 23. It includes a learning unit 24 that obtains a trained model 4a by machine learning using the learning superimposed image SIL as input data and the change area in the area of the learning map image MPL as teacher data.

With the learning device 20 described above, the change area of the target area can be directly learned by the learning superimposed image SIL in which the learning map image MPL and the learning aerial image captured image APL are superimposed. It is not necessary to infer the shape of the inferred shape, and the decrease in generalization due to the instability of the inferred shape is suppressed.

Next, the operation (inference method) of the inference device 1 shown in FIG. 2 will be described with reference to the flowchart of FIG. The flowchart of FIG. 8 can be configured as a program (inference program) executed by the computer 10. Further, the inference program may be stored in a recording medium such as a memory card.

First, the map image MP and the aerial photograph AP are input to the input unit 2 (step S21). Next, the generation unit 3 generates the superimposed image SI (step S22). Then, the inference unit 4 infers a region corresponding to any of new establishment, disassembly, and change by the trained model 4a (step S23). From the trained model 4a, for example, the probability of whether to correspond to new installation, disassembly, or change for each pixel of the superimposed image SI is output. That is, the probability that it is the change point for each pixel of the superimposed image is output. That is, the probability that the target image is a point (image) included in the change area is output.

That is, step S21 functions as an input process, step S22 functions as a generation process, and step S23 functions as an inference process.

Then, the inference unit 4 determines whether the area is newly established, disassembled, or changed according to the output probability. This determination may be made by setting a threshold value for the probability in advance and determining that the threshold value is equal to or higher than the threshold value. For example, when the threshold value for determining new construction is 50%, it is determined that the area where the probability of new installation output from the trained model 4a is 50% or more is the area of the newly installed building. That is, the inference device 1 according to the present embodiment does not infer the shape of the building, but directly infers the changed region.

Finally, the experimental results obtained by inference with the inference device 1 having the above-mentioned configuration are shown. FIG. 9 shows the input superposed image SI, the teacher data (Ground Truth) corresponding to the superposed image SI, and the inference result performed using the trained model 4a after training.

As the training data, six 1 km square meshes (384tiles) composed of 8tiles × 8tiles were used. At the time of learning, Iterations = 250,000, Batch size = 8, Start Loss = 0.43, End Loss = 0.06, Training Time = 26h.

As shown in FIG. 9, it was clarified that the inference device 1 according to the present embodiment could substantially output the region corresponding to the region indicated by the teacher data as the inference result.

FIG. 10 shows Precision (correct answer rate) and Recall (correct answer rate) in the case of using the inference device 1 having the above-described configuration and the method of taking the difference between the inferred shape of the building and the map data as in the prior art. It is a comparison with the precision rate). FIG. 10 (a) is an inference device 1 according to this embodiment, and FIG. 10 (b) is a conventional technique. The conditions for learning and the like of the inference device 1 are the same as those in FIG.

As shown in FIG. 10, in the case of the inference device 1, both the precision and the precision were higher in both the case of new installation and the case of dismantling. In particular, there was a significant improvement in Precision. The change can be inferred only by the inference device 1. This is because the shape of the building is inferred in the prior art, but it cannot be determined whether or not the shape has been changed because the shape is unstable.

According to this embodiment, in the inference device 1, the map image MP and the aerial photograph AP for updating the map image MP corresponding to the area of the map image MP are input to the input unit 2, and the generation unit 3 is used. , A superposed image SI in which the map image MP and the aerial photograph AP corresponding to the area of the map image MP are superimposed is generated. Then, the inference unit 4 uses the learning superimposed image in which the learning map image and the learning aerial photograph corresponding to the area of the learning map image are superimposed as input data, and the change area in the area of the learning map image. The trained model 4a obtained by machine learning as teacher data is applied to the superimposed image SI to infer the change region in the above-mentioned region. By doing so, the change area can be inferred from the superposed image SI in which the map image MP and the aerial photograph AP are superimposed by the trained model 4a in which the change area of the target area is learned. In addition, it is not necessary to infer the shape of the building, and the decrease in generalization due to the instability of the inferred shape can be suppressed. Therefore, it is possible to update the map image more accurately when the map image such as the house shape change point detection and the house change interpretation is updated.

Further, the trained model 4a is machine-learned using the building outline corresponding to each of new installation, dismantling, and change in the superposed image for learning as teacher data, and the inference unit 4 is used for new installation, dismantling, and change in the aerial photography AP. The range of change corresponding to each is inferred. By doing so, it becomes possible to infer new construction, demolition, and changes of buildings on the map.

Further, the inference unit 4 infers the change area based on the probability of the change point for each pixel of the superimposed image SI output from the trained model 4a. By doing so, by setting a predetermined threshold value or the like for the probability output for each pixel, it becomes possible to infer a change area such as new construction, demolition, or change of a building.

Further, the superimposed image SI is configured by setting the aerial photograph AP data in the R channel and the G channel and the map image MP data in the B channel among the RGB formats, and the superimposed image SIL for learning is in the RGB format. Of these, the R channel and the G channel may be configured by setting the data of the aerial photograph SPL for learning, and the B channel may be configured by setting the data of the map image MPL for learning. By doing so, the superimposed image SI or the like can be formed by using the RGB format. Further, by using the map image MP or the learning map image MPL as the B channel, it becomes easy to distinguish the superimposed image SI from the aerial captured image AP and the learning aerial captured image APL even when the superimposed image SI is observed by a human. ..

Further, the teacher data is a label image L in which label information corresponding to the characteristics of the changing region (new installation, disassembly, change) is set for each pixel. By doing so, by using the teacher data as an image, the input data and the format can be matched, and the learning efficiency and the processing efficiency can be improved.

In the above-described embodiment, the superimposed image uses the RGB format, but it may be a 4-channel image in which a map image is superimposed on a color photograph (RGB).

Further, the present invention is not limited to the above examples. That is, those skilled in the art can carry out various modifications according to conventionally known knowledge within a range that does not deviate from the gist of the present invention. Even with such a modification, as long as the inference device of the present invention is provided, it is, of course, included in the category of the present invention.

1 Inference device 2 Input unit 3 Generation unit 4 Inference unit 4a Trained model MP Map image AP Aerial photograph (aerial image)
SI superimposed image MPL learning map image APL learning aerial photograph (aerial image)
Superimposed image for SIL learning

Claims (8)

An input unit for inputting a map image and an aerial captured image for updating the map image corresponding to the area of the map image.
A generation unit that generates a superposed image in which the map image and the aerial captured image are superimposed,
The learning superimposed image in which the learning map image and the learning aerial image captured corresponding to the area of the learning map image are superimposed is used as input data, and the change area in the area represented by the learning superimposed image is the teacher data. The inference unit that infers the change area in the region represented by the superimposed image by applying the trained model obtained by machine learning as
An inference device characterized by comprising. In the trained model, the building outline corresponding to each of new construction, dismantling, and change in the superposed image for learning is machine-learned as the teacher data.
The inference unit infers the change area corresponding to each of new installation, disassembly, and change in the aerial image.
The inference device according to claim 1. The inference device according to claim 1 or 2, wherein the inference unit infers the change area based on the probability of being a change point for each pixel of the superimposed image output from the trained model. The superimposed image is configured by setting the data of the aerial captured image in the R channel and the G channel of the RGB format and the data of the map image in the B channel, respectively.
The superposed image for learning is characterized in that it is configured by setting the data of the aerial image for learning in the R channel and the G channel of the RGB format and the data of the map image for learning in the B channel, respectively. The inference device according to any one of claims 1 to 3. The inference device according to any one of claims 1 to 4, wherein the teacher data is an image in which label information corresponding to the characteristics of the change region is set for each pixel. It is an inference method executed by an inference device that makes inferences based on a trained model obtained by machine learning.
An input process for inputting a map image and an aerial captured image for updating the map image corresponding to the area of the map image, and
A generation step of generating a superimposed image in which the map image and the aerial captured image are superimposed,
The learning superimposed image in which the learning map image and the learning aerial image captured corresponding to the area of the learning map image are superimposed is used as input data, and the change area in the area represented by the learning superimposed image is the teacher data. The inference step of applying the trained model obtained by machine learning as a method to the superimposed image to infer the change region in the region represented by the superimposed image, and
An inference method characterized by including. An inference program comprising executing the inference method according to claim 6 by a computer. A computer-readable storage medium comprising storing the inference program according to claim 7.

Images