JP2020030730A - House movement reading system, house movement reading method, house movement reading program, and house loss reading model - Google Patents

Abstract

To accurately read the loss of a house on the basis of aerial photographs taken in two periods of time.SOLUTION: A house movement reading device 1 comprises: an acquisition unit for acquiring a first aerial photograph taken in a first period, a second aerial photograph taken in a second period, and house diagram data in the first period; an extraction unit for extracting a first house area image and a second house area image from the first aerial photograph and the second aerial photograph respectively, based on a position and a planar shape of a house shown by the house diagram data; a learning data generation unit for generating learning data including the first house area image and the second house area image having shape information of the house and correct data showing whether the house has been lost or not; a model generation unit for generating a loss reading model that performs machine learning on the basis of the learning data and makes the first and second house area images used as inputs and house loss information, showing whether the house has been lost or not, used as outputs; and a model output unit for outputting the generated loss reading model.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a house change interpretation system, a house change interpretation method, a house change interpretation program, and a loss interpretation model.

  For example, for the purpose of investigating taxation, etc., it has been practiced to compare the aerial photographs of the two periods to determine a change in a house (lost, new construction, and renovation). For example, a technique for displaying image data based on the aerial photograph data of the previous year and image data based on the aerial photograph data of the current year on a screen in order to compare them is known (for example, see Patent Document 1). .

JP 2018-4389 A

  Conventionally, the interpretation of a change in a house based on comparison of aerial photographs in two periods has been performed by a person viewing the screen displaying the changes. However, visual interpretation was very troublesome. In addition, the image processing by a machine extracts difference information such as color, texture, and height information from images of two periods, and interprets a change in a house. Contains many noises due to differences in image capturing conditions and the like, resulting in low interpretation accuracy.

  Accordingly, an object of the present invention is to accurately and quickly read the loss of a change in a house based on two aerial photographs.

  A house transfer interpretation system according to one embodiment of the present invention provides a system for interpreting a house loss between a first time and a second time after the first time based on an aerial photograph. A house transfer interpretation system for generating an interpretation model, comprising: a first aerial photograph taken at a first time, a second aerial photograph taken at a second time, and a first aerial photograph at a first time. And an acquisition unit that acquires house diagram data corresponding to the area represented in the second aerial photograph, wherein the house diagram data includes one or more house data, and the house data includes at least a plane related to one house. Based on the acquisition unit including the shape information indicating the shape and the position information indicating the position, and the position and the planar shape of the house indicated by the house data, one of the first aerial photograph and the second aerial photograph is used. To the location and range of the house An extraction unit that extracts a first house area image and a second house area image that are images of corresponding areas, and a first house area image and a second house area image that include at least shape information; A learning data generation unit that generates learning data including correct data indicating whether or not a house has been lost in an area indicated by the second house area image; and a first learning apparatus that performs machine learning based on the learning data and includes shape information. And a model generation unit that generates a loss interpretation model that outputs the house loss information indicating whether or not the house shown in the first house area image has been lost as an input of the second house area image and the generated loss interpretation. A model output unit for outputting a model.

  The method for interpreting house transfer according to one aspect of the present invention includes a method for interpreting loss of a house between a first time and a second time after the first time based on an aerial photograph. A house change interpretation method in a house change interpretation system that generates a interpretation model, comprising: a first aerial photograph taken at a first time, a second aerial photograph taken at a second time, and a first aerial photograph. An acquisition step of acquiring house diagram data corresponding to an area represented in the first and second aerial photographs at a time, wherein the house diagram data includes one or more house data, and the house data includes one or more house data. An acquisition step including at least shape information indicating a planar shape and position information indicating a position of the house, and a first aerial photograph and a second aerial photograph based on the position and the planar shape of the house indicated by the house data; Each An extraction step of extracting a first house area image and a second house area image which are images of an area corresponding to the position and range of one house, and a first house area image and a second house area image including at least shape information. A learning data generating step of generating learning data including a second house area image, and learning data including correct data indicating presence or absence of a house in the areas indicated by the first and second house area images; and a machine based on the learning data. A model that performs learning, and generates a loss interpretation model that outputs first and second house area images accompanied by shape information and outputs loss information indicating the presence or absence of loss of the house indicated in the first house area image. A generating step; and an output step of outputting the generated loss interpretation model.

  A house change interpretation program according to one embodiment of the present invention interprets a computer based on an aerial photograph of loss of a house between a first time and a second time that is a time later than the first time. A house change interpretation program for functioning as a house change interpretation system for generating a loss interpretation model for a first aerial photograph taken at a first time and taken at a second time by a computer A second aerial photograph, and an acquisition function for acquiring house map data corresponding to an area represented in the first and second aerial photographs at the first time, wherein the house map data includes one or more houses. Including the data, the house data includes at least the shape information representing one house and the position information indicating the position indicating the position, an acquisition function, and the position and the plane shape of the house indicated by the house data. Then, extraction is performed to extract a first house area image and a second house area image, which are images of an area corresponding to the position and range of one house, from each of the first aerial photograph and the second aerial photograph. Learning including the function and the first house area image and the second house area image with at least the shape information, and the correct answer data indicating the presence or absence of the house in the areas shown in the first and second house area images. A learning data generation function of generating data, and performing machine learning based on the learning data, and inputting the first and second house area images accompanied by the shape information as inputs to determine the loss of the house shown in the first house area image. A model generation function for generating a loss interpretation model that outputs loss information indicating presence / absence and an output function for outputting the generated loss interpretation model are realized.

  A loss interpretation model according to one embodiment of the present invention is a machine for interpreting loss of a house between a first time and a second time later than the first time based on an aerial photograph. A loss-reading model, which is a learned model including a neural network generated by learning, wherein a first house area image and a second house area image accompanying shape information representing a planar shape of a house are input values, and The first and second house area images are configured by machine learning based on learning data having output values of correct data indicating the presence or absence of a house in the areas shown in the first and second house area images. Are based on the house data including the shape information indicating the planar shape and the position information indicating the position of one house, based on the first aerial photograph taken at the first time and the second time at the second time. Extracted from each of the second aerial photographs that are shaded, it is an image of the region corresponding to the position and extent of one house.

  According to the above aspect, based on the house map data at the first time, from the first and second aerial photographs, the first and second aerial photographs are images of the area corresponding to the position and range of one house. A second house area image is extracted. The first and second house area images extracted in this manner include image features indicating the presence or absence of one house at each of the first and second periods. In addition, the first and second house areas which are the houses shown in the first house area image and which include shape information indicating the planar shape of the house included in the house diagram data at the first time. Since the image is used for the learning data of the machine learning, the first and second house area images are appropriately associated with the planar shape of the house and learned. Then, the correctness data indicating the presence or absence of loss of the house is further associated and learned to generate the loss interpretation model. Therefore, based on the image of the area of the house at each of the first and second times, the house is determined based on the image of the house. It is possible to obtain a model capable of outputting information indicating the possibility of occurrence of the loss with high accuracy at high speed.

  In the house change interpretation system according to another aspect, the extraction unit generates house area data including a square area including graphic data representing a planar shape of the house, based on the shape information of the house indicated by the house data. , An area corresponding to the house area data in each of the first aerial photograph and the second aerial photograph is extracted as a first house area image and a second house area image, and the learning data generation unit The learning data including the house area image, the second house area image, the house area data, and the correct answer data may be generated.

  According to the above aspect, the house area data including the rectangular area including the graphic data representing the planar shape of the house is generated, and the first and second house area images including the area having the same shape as the house area are generated. Is generated. The thus generated house area data and the first and second house area images are used as learning data. Since the house area data generated in this manner surely includes the information on the plane shape of the house at the first time, the information on the plane shape of the house is appropriately included in the first and second house area images. Machine learning can be performed in association with it.

  In the house change interpretation system according to another aspect, the extraction unit is configured to extract the house indicated in the shape information from each of the first aerial photograph and the second aerial photograph based on the house shape information indicated by the house data. A region having a shape corresponding to the planar shape is extracted as a first house region image and a second house region image, and the learning data generation unit outputs a first house region image having a shape corresponding to the planar shape of the house and The learning data including the second house area image and the correct answer data may be generated.

  According to the above aspect, the region having the shape corresponding to the planar shape of the house at the first time is extracted from the first and second aerial photographs as the first and second house region images. The first and second house area images include information on the planar shape of the house at the first time. Therefore, the first and second house region images extracted in this manner are used as learning data for machine learning without adding data representing a separate house plane shape, thereby enabling information on the house plane shape to be obtained. Appropriate machine learning can be performed, and the amount of data provided for machine learning can be reduced.

  In the house change interpretation system according to another aspect, the correct answer data is image data having a predetermined shape, and all pixels of the correct answer data indicating that the house has been lost have the first color and the second color. , And all pixels of the correct answer data indicating that the house has not been lost may have the other of the first color and the second color.

  According to the above aspect, it is possible to obtain a loss interpretation model that outputs image data having a predetermined shape and composed of pixels of the first or second color as house loss information indicating whether a house has been lost. Become. In addition, in the house loss information, the classification result of whether or not the house has been lost is indicated as one of the binary colors of each pixel. Therefore, the ratio of the pixels of the first color and the pixels of the second color Since the possibility of the loss of the house is reflected on this, the presence or absence of the loss of the house can be easily read.

  In addition, the house change interpretation system according to one aspect of the present invention uses the above-described loss interpretation model to determine whether a house has been lost between a first time and a second time that is a time later than the first time. Is a house transfer interpretation system that interprets the loss of a house based on an aerial photograph, wherein the first aerial photograph taken at the first time, the second aerial photograph taken at the second time, And an acquisition unit for acquiring house diagram data corresponding to the areas represented in the first and second aerial photographs at the first time, wherein the house diagram data includes one or more house data, A first aerial photograph and a second aerial photograph based on an acquisition unit including at least shape information representing a planar shape and position information indicating a position with respect to one house, and a position and a planar shape of the house indicated by the house data. From each of the aerial photographs An extraction unit that extracts a first house area image and a second house area image that are images of an area corresponding to the position and range of one house, and a first house area image and a second house image that include at least shape information. The system includes a reading unit that inputs a house area image to a loss interpretation model and acquires an output value of the loss interpretation model as house loss information, and a loss information output unit that outputs house loss information acquired by the interpretation unit.

  In addition, the method for interpreting house transfer according to one aspect of the present invention uses the above-described model for judging loss and eliminates loss of a house between a first time and a second time after the first time. A house change interpretation method in a house change interpretation system that interprets the aerial photograph based on an aerial photograph, wherein a first aerial photograph taken at a first time, a second aerial photograph taken at a second time, and In an acquisition step of acquiring house map data corresponding to an area represented in the first and second aerial photographs at a first time, the house map data includes one or more house data, and the house data includes: The first aerial photograph and the second aerial photograph based on the acquisition step and the position and the planar shape of the house indicated by the house data, the information including at least the shape information indicating the planar shape and the position information indicating the position of the one house. Aerial photography An extraction step of extracting a first house area image and a second house area image that are images of an area corresponding to the position and range of one house, and a first house area including at least shape information. A reading step of inputting the image and the second house area image to the loss interpretation model and acquiring an output value of the loss interpretation model as house loss information; and a loss information output step of outputting the house loss information acquired in the interpretation step. And

  In addition, the house change interpretation program according to one embodiment of the present invention uses a computer to execute the loss interpretation model according to claim 7 in a first period and a second period that is a period later than the first period. A house change interpretation program for functioning as a house change interpretation system for interpreting the loss of a house during a period based on an aerial photograph, wherein the first aerial photograph taken at a first time, An acquisition function for acquiring the second aerial photograph taken at the time of and the house map data corresponding to the area represented by the first and second aerial photographs at the first time, Includes one or more house data, the house data includes at least shape information representing a planar shape and position information indicating a position regarding one house, an acquisition function, and a position of the house indicated by the house data. A first house area image and a second house area image, which are images of an area corresponding to the position and range of one house, from each of the first aerial photograph and the second aerial photograph based on the first and second aerial photographs. And a reading function of inputting at least the first house area image and the second house area image accompanying the shape information to the loss interpretation model, and acquiring the output value of the loss interpretation model as house loss information. And a loss information output function of outputting house loss information acquired by the reading function.

  According to the above aspect, based on the house map data at the first time, from the first and second aerial photographs, the first and second aerial photographs are images of the area corresponding to the position and range of one house. A second house area image is extracted. The first and second house area images extracted in this manner include image features indicating the presence or absence of one house at each of the first and second periods. Since the first and second house region images extracted in this way are input to the loss interpretation model together with the shape information representing the planar shape of the house, the house between the first time and the second time is The presence or absence of loss is accurately read.

  In the house change interpretation system according to another aspect, the house loss information is an image having a shape corresponding to the shapes of the first and second house region images, and each pixel of the house loss information has a first color. And the second color different from the first color, and the ratio of the pixel of the first color and the pixel of the second color in the house loss information is represented by the first and second house area images. The possibility of loss of a house in the indicated area may be indicated.

  According to the above aspect, in the house loss information, the ratio of the pixels of the first color and the pixels of the second color reflects the possibility of the loss of the house. Can be read.

  According to an aspect of the present invention, it is possible to accurately and quickly read the loss of a change in a house based on the aerial photographs of two periods.

It is a block diagram showing the functional composition of the house change interpretation system concerning this embodiment. It is a figure showing the hardware constitutions of a house change interpretation device. FIG. 3A is a diagram schematically illustrating an example of an aerial photograph. FIG. 3B is a diagram schematically illustrating an example of house diagram data. It is a figure showing generation of house area data. It is a figure explaining polygon integration processing which is pre-processing about house figure data. It is a figure which shows the example of extraction of a house area image. FIGS. 7A and 7B are diagrams each showing an example of learning data. FIGS. 7A and 7B are diagrams illustrating another example of extraction of a house area image and generation of learning data. FIGS. 7A and 7B are diagrams illustrating other examples of learning data. It is a figure which shows typically the data set input into a loss interpretation model, and the example of the house loss information output from the model. It is a figure showing an example of house loss information outputted from a loss interpretation model. It is a flowchart which shows the process content of the house change interpretation method implemented in the house change interpretation apparatus in the aspect of model generation by machine learning. It is a flowchart which shows the process content of the house change interpretation method implemented in a house change interpretation apparatus in the phase of house loss interpretation using the learned model. FIGS. 13A and 13B are diagrams each showing a configuration of a house change interpretation program.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. In the description of the drawings, the same or equivalent elements will be denoted by the same reference symbols, without redundant description.

  FIG. 1 is a diagram showing a device configuration of a house change reading system according to the present embodiment and a functional configuration of a house change reading device. As shown in FIG. 1, the house change interpretation device 1A constitutes a house change interpretation system in a model generation phase by machine learning. Specifically, the house change interpretation device 1A reads the loss of the house between the first time and the second time after the first time based on the aerial photograph. Generate a model.

  In addition, the house change interpretation device 1B constitutes a house change interpretation system in the phase of house loss interpretation. Specifically, the house change interpretation device 1B interprets the loss of the house between the first time and the second time based on the aerial photograph using the loss interpretation model.

  As illustrated in FIG. 1, the house change interpretation device 1A includes an acquisition unit 11, an extraction unit 12, a learning data generation unit 13, a model generation unit 14, and a model output unit 15. The house change interpretation device 1B includes an acquisition unit 11B, an extraction unit 12B, an interpretation unit 16, and a loss information output unit 17.

  Although the house change reading device 1A and the house change reading device 1B of the present embodiment are each configured as separate devices, they may be formed integrally. Further, each functional unit included in the house change reading device 1A and the house change reading device 1B may be configured to be distributed to a plurality of devices. Each functional unit included in the house change reading device 1A and the house change reading device 1B is realized by any one of hardware and software, or any combination.

  As shown in FIG. 1, the house change interpretation system can include storage means such as an aerial photograph storage unit 21, a house map data storage unit 22, and a model storage unit 23. These storage units 21 to 23 are configured to be accessible from the house change reading device 1A and the house change reading device 1B via a network or predetermined communication means. Further, the storage units 21 to 23 may be provided in the house change interpretation device 1.

  Prior to the description of the functional units of the house change reading device 1A and the house change reading device 1B, the storage units 21 to 23 will be described. The aerial photograph storage unit 21 is storage means for storing data of aerial photographs taken at a predetermined time. In the present embodiment, in order to interpret the loss of the house between the first time and the second time, which is a time after the first time, based on the aerial photograph, the aerial photograph storage unit 21 stores A first aerial photograph taken at a first time and a second aerial photograph taken at a second time are stored. Specifically, for the purpose of investigating a taxable house, for example, the first time is January 1 of a certain year, and the second time is January 1 of the next year. Day.

  The house diagram data storage unit 22 is a storage unit that stores house diagram data. The house diagram data includes house data of each house existing in a predetermined geographical area at one time. The house data includes shape information indicating a planar shape of one house and position information indicating a position. The house data may include information on the size (area) of the house. Further, the house data may be data in which each information is associated with identification information for uniquely identifying each house. When the house diagram data is represented on a two-dimensional map, the house data may be represented as a polygon representing the position, shape, and size of the house. The house diagram data and the house data can be, for example, vector data. The house map data is created by a local government or the like based on, for example, a survey in a certain period. In the present embodiment, the house diagram data storage unit 22 stores house diagram data at the first time.

  The model storage unit 23 is a storage unit that stores the loss interpretation model created and output by the house change interpretation device 1A. The house change interpretation device 1B can use the loss interpretation model stored in the model storage unit 23 to perform a process of interpreting the loss of the house during the two o'clock period based on the aerial photographs of the two periods.

  FIG. 2 is a hardware configuration diagram of the house change interpretation device 1. As shown in FIG. 2, the house change reading device 1 physically includes a CPU 101, a main storage device 102 including a memory such as a RAM and a ROM, an auxiliary storage device 103 including a hard disk and a memory, and a communication control device. It is configured as a computer system including the device 104 and the like. The house change interpretation device 1 may further include an input device 105 such as a keyboard, a touch panel, and a mouse, which are input devices, and an output device 106 such as a display. Note that the house change interpretation device 1 may be configured to include a GPU.

  Each function shown in FIG. 1 operates the communication control device 104 and the like under the control of the CPU 101 by reading predetermined computer software on hardware such as the CPU 101 and the main storage device 102 shown in FIG. At the same time, it is realized by reading and writing data in the main storage device 102 and the auxiliary storage device 103. Data and databases required for processing are stored in the main storage device 102 and the auxiliary storage device 103.

  With reference to FIG. 1 again, the functional unit of the house change reading device 1A will be described. The acquisition unit 11 includes a first aerial photograph taken at the first time, a second aerial photograph taken at a second time after the first time, and a house at the first time. Get figure data. In the present embodiment, the presence or absence of the house existing at the first time in the second time is read. Specifically, the acquiring unit 11 acquires, from the aerial photograph storage unit 21, the first and second aerial photographs in which the area to be interpreted as the loss of the house is indicated. In addition, the acquisition unit 11 acquires, from the house diagram data storage unit 22, house diagram data corresponding to the areas represented in the first and second aerial photographs.

  FIG. 3A is a diagram schematically illustrating an example of an aerial photograph. As shown in FIG. 3A, the aerial photograph A1 is a photograph taken from the sky of a target area where the loss of a house is to be interpreted, and shows a plurality of houses.

  FIG. 3B is a diagram schematically illustrating an example of house diagram data. The example shown in FIG. 3B is an example in which house diagram data is represented on a two-dimensional map. As shown in FIG. 3B, the house map data B1 corresponds to the area shown in the aerial photograph A1 shown in FIG. 3A and includes house data b11 to b18. Each of the house data b11 to b18 includes position information indicating the position of each house and shape information indicating a planar shape. The position information in the house data is represented by, for example, coordinate values. The shape information is represented by polygon data, which is graphic data, as shown in FIG.

  Note that the acquisition unit 11 may normalize the pixel values of the acquired aerial photograph by a well-known method in order to correct a difference in color tone caused by a difference in photographing conditions of the aerial photograph. Further, the acquisition unit 11 may remove the shadow in the acquired aerial photograph by using a known image processing technique for removing the shadow.

  The extraction unit 12 is an image of an area corresponding to the position and range of one house from each of the first aerial photograph and the second aerial photograph based on the house position and the planar shape indicated by the house data. A first house area image and a second house area image are extracted. The extraction process of the house area image will be described with reference to FIGS.

  The extraction unit 12 generates house area data based on house data of the house drawing data before extracting the house area image. FIG. 4 is a diagram illustrating generation of house area data. FIG. 4 shows partial house drawing data B2 which is a part of the house drawing data. The partial house diagram data B2 includes house data b21.

  The extraction unit 12 calculates the coordinates of the center point CP (for example, the center of gravity) of the polygon indicating the shape of the house indicated by the house data b21 and the area of the polygon. Subsequently, the extraction unit 12 sets a rectangular area r having a size capable of surrounding the outer shape of the polygon with the center point CP of the polygon as the center according to the area of the polygon in the house data b21. In the present embodiment, the square region r has a square shape, but is not limited to a square shape. The size of the rectangular area r may be, for example, a size obtained by adding a fixed margin to the size of the polygon, or one of a plurality of predetermined sizes set in advance may be applied. .

  Then, the extracting unit 12 extracts a square region r which is a square region including the polygon of the house data b21 from the partial house diagram data B2, and generates house region data c21.

  FIG. 5 is a diagram for explaining polygon integration processing, which is preprocessing for house diagram data. In the house diagram data, one house may be represented by being divided into a plurality of polygons due to a rule in creating the house diagram data. In such a case, the extraction unit 12 may execute a polygon integration process before generating the house area data.

  For example, the extraction unit 12 extracts a plurality of polygons (shape information) associated with one piece of position information from the house map data. The left side of the arrow in FIG. 5 illustrates an example in which one house is represented by a plurality of polygons, and the shape of one house is represented by two polygons, polygon b31 and polygon b32. The extraction unit 12 performs an integration process on the polygons b31 and b32, and generates a polygon b33 as one polygon representing one house.

  FIG. 6 is a diagram illustrating an example of extracting a house area image. FIG. 6 shows house area data c41. The house area data c41 includes house data b41 which is a polygon indicating the planar shape of one house. The extraction unit 12 converts the area corresponding to the house area data in each of the first aerial photograph taken at the first time and the second aerial photograph taken at the second time into a first house area image. And a second house area image.

  In the example illustrated in FIG. 6, the extraction unit 12 extracts a region corresponding to the house region data c41 in the first aerial photograph, and generates a first house region image a41. Further, the extraction unit 12 extracts a region corresponding to the house region data c41 in the second aerial photograph, and generates a second house region image a42.

  The learning data generation unit 13 corrects the presence or absence of a house in at least the first house area image and the second house area image accompanying the house shape information and the area shown in the first and second house area images. Generate learning data including data.

  In the present embodiment, as an example, learning data including a first house area image, a second house area image, house area data, and correct answer data is generated. FIG. 7 is a diagram illustrating an example of learning data.

  In the example illustrated in FIG. 7A, the learning data generation unit 13 acquires the first house area image a51 and the second house area image a52 extracted by the extraction unit 12. Further, the learning data generation unit 13 acquires the house area data c5 referred to when the first and second house area images a51 and a52 are extracted.

  Further, the learning data generation unit 13 converts the correct answer data e5 indicating the presence or absence of the house in the areas shown in the first and second house area images a51 and a52 into the first and second house area images a51 and a52. In addition, it is associated with the house area data c5.

  The correct answer data is image data having a predetermined shape, and all pixels of the correct answer data indicating that the house has been lost have one of the first color and the second color, and the house has been lost. All pixels of the correct answer data indicating that there is no other of the first color and the second color;

  Specifically, the correct answer data e5 is image data indicating that the house represented in the first house area image a51 has been lost in the second house area image a52. The correct answer data e5 is image data having a predetermined shape. For example, the correct answer data e5 may have a shape corresponding to the shapes of the first and second house area images a51 and a52, but is not limited thereto. Further, in the present embodiment, since the loss has been indicated by white, all the pixels of the correct answer data e5 have white of two colors of white and black. The association of the correct answer data e5 with the first and second house area images a51 and a52 and the house area data c5 is performed by the user who has visually determined whether or not the first and second house area images a51 and a52 have been lost. It may be that.

  Then, the learning data generation unit 13 generates learning data d5 including a set of the first house area image a51, the second house area image a52, the house area data c5, and the correct answer data e5. As described above, the learning data d5 includes the house area data c5 including the house data b5, and thus includes information on the planar shape (polygon shape) of the house indicated by the house data b5. Therefore, by such learning data d5, machine learning in which information on the planar shape of the house is appropriately associated with the first and second house area images a51 and a52 is realized.

  In the example illustrated in FIG. 7B, the learning data generation unit 13 acquires the first house area image a61 and the second house area image a62 extracted by the extraction unit 12. Further, the learning data generation unit 13 acquires the house area data c6 referred to when extracting the first and second house area images a61 and a62.

  Further, the learning data generation unit 13 converts the correct answer data e6 indicating the presence or absence of the house in the areas shown in the first and second house area images a61 and a62 into the first and second house area images a61 and a62. In addition, it is associated with the house area data c6.

  Specifically, the correct answer data e6 is image data indicating that the house represented in the first house area image a61 has not been lost in the second house area image a62. The correct answer data e6 is image data having a predetermined shape. For example, the correct answer data e6 may have a shape corresponding to the shapes of the first and second house area images a61 and a62, but is not limited thereto. In this embodiment, since no loss is indicated by black, all the pixels of the correct answer data e6 have black of two colors of white and black. The association of the correct answer data e6 with the first and second house area images a61 and a62 and the house area data c6 is performed by the user who has visually determined whether or not the first and second house area images a61 and a62 have been lost. It may be that.

  Then, the learning data generation unit 13 generates learning data d6 including a set of the first house area image a61, the second house area image a62, the house area data c6, and the correct answer data e6. As described above, the learning data d6 includes the house area data c6 including the house data b6, so that the learning data d6 includes information on the planar shape (polygon shape) of the house indicated by the house data b6. Accordingly, such learning data d6 implements machine learning in which the information on the planar shape of the house is appropriately associated with the first and second house area images a61 and a62.

  FIG. 8 is a diagram illustrating another example of extracting a house area image and generating learning data. In another example described here, the extraction unit 12 outputs the plane of the house indicated by the shape information from each of the first aerial photograph and the second aerial photograph based on the shape information of the house indicated by the house data. A region having a shape corresponding to the shape is extracted as a first house region image and a second house region image.

  In the example illustrated in FIG. 8A, house data b7 is shown as an example of information of one house included in the house diagram data. The house data b7 has polygon data as shape information indicating the planar shape of the house. The extraction unit 12 extracts a region having a shape corresponding to the house data b7 in the first aerial photograph, and generates a first house region image a71. Further, the extraction unit 12 extracts a region having a shape corresponding to the house data b7 in the second aerial photograph, and generates a second house region image a72. Note that the extraction unit 12 extracts, from the first and second aerial photographs, an area substantially similar to the polygon with a fixed margin area added outside the outline of the polygon shown in the house data b7, The first and second house area images may be generated.

  Since the house shown in the first house area image a71 is lost in the second house area image a72, the first and second house area images a71 and a72 indicate the loss of the house. The correct answer data e7 can be associated manually, for example.

  The learning data generation unit 13 generates learning data d7 including a first house area image a71, a second house area image a72, and correct answer data e7. As described above, the first and second house area images a71 and a72 have the same plane shape as the house by being extracted based on the plane shape of the house. As a result, the first and second house area images a71 and a72 include information on the planar shape of the house.

  In the example shown in FIG. 8B, house data b8 is shown as an example of information of one house included in the house diagram data. The house data b8 has polygon data as shape information indicating the planar shape of the house. The extraction unit 12 extracts a region having a shape corresponding to the house data b8 in the first aerial photograph, and generates a first house region image a81. Further, the extraction unit 12 extracts a region having a shape corresponding to the house data b8 in the second aerial photograph, and generates a second house region image a82. Note that the extraction unit 12 extracts, from the first and second aerial photographs, an area substantially similar in shape to the polygon with a fixed margin area added to the outside of the outline of the polygon shown in the house data b8, The first and second house area images may be generated.

  Since the house represented in the first house area image a81 has not been lost in the second house area image a82, no house has been lost in the first and second house area images a81 and a82. Correct data e8 indicating the fact can be associated manually, for example.

  The learning data generation unit 13 generates learning data d8 including the first house area image a81, the second house area image a82, and the correct answer data e8. As described above, the first and second house area images a81 and a82 have the same plane shape as the house by being extracted based on the plane shape of the house. As a result, the first and second house area images a81 and a82 include information on the planar shape of the house.

  By using the learning data d7 and d8 as shown in FIGS. 8A and 8B for machine learning, the information on the planar shape of the house can be appropriately adjusted without adding data representing the separate planar shape of the house. In addition to performing the machine learning associated with, the amount of data provided for the machine learning can be reduced.

  Referring again to FIG. 1, model generation unit 14 performs machine learning based on the learning data generated by learning data generation unit 13 and inputs first and second house area images accompanied by house shape information. A loss interpretation model that outputs house loss information indicating the presence or absence of loss of the house shown in the first house region image in the second house region image is generated.

  In the present embodiment, for example, a loss reading model composed of a CNN (Convolution Neutral Network) is generated. However, if the model configuration is applicable in the field of machine learning and deep learning, this example is used. Not limited. The model generation unit 14 generates a loss interpretation model by machine learning based on the learning data generated by the learning data generation unit 13 to determine whether a house has been lost as an image classification problem.

  The loss interpretation model of the present embodiment is configured to include a neural network. The loss interpretation model, which is a model including a learned neural network, is read or referred to by a computer, and can be regarded as a program that causes a computer to execute a predetermined process and realize a predetermined function.

  That is, the loss interpretation model of the present embodiment is used in a computer including a CPU (or / and a GPU) and a memory. Specifically, the CPU (or / and GPU) of the computer responds to input from the input layer of the neural network in accordance with instructions from the loss-and-decided model stored in the memory, by using the learned data corresponding to each layer. An operation is performed based on a weighting coefficient and a response function, and the result is output from the output layer.

  The input data to the neural network in the present embodiment is, for example, first and second house area images accompanied by house shape information. Specifically, the pixel values of the first and second house area images are appropriately converted into a data format as input data. The result of the output from the neural network is the house loss information output in a format corresponding to the correct answer data included in the learning data. As will be described later, the house loss information indicates a possibility that house loss has occurred.

  The model output unit 15 outputs the loss interpretation model generated by the model generation unit 14. Specifically, the model output unit 15 causes the model storage unit 23 to store the generated loss interpretation model, for example.

  As described above, according to the house change interpretation device 1A of the present embodiment, based on the house map data at the first time, the position and the range of one house correspond to each of the first and second aerial photographs. First and second house region images, which are images of the region to be moved, are extracted. The first and second house area images extracted in this manner include image features indicating the presence or absence of one house at each of the first and second periods. In addition, the first and second house area images represented by the first house area image and accompanied by the shape information indicating the planar shape of the house included in the house plan data at the first time are used as the machine images. Since it is used for the learning data of the learning, the first and second house area images are appropriately associated with the planar shape of the house and learned. Then, the correctness data indicating the presence or absence of loss of the house is further associated and learned to generate the loss interpretation model. Therefore, based on the image of the area of the house at each of the first and second times, the house is determined based on the image of the house. It is possible to obtain a model capable of outputting information indicating the possibility of occurrence of the loss with high accuracy.

  Next, with reference to FIG. 1 again, the function of the house change reading apparatus 1B constituting the house change reading system in the house loss judgment phase will be described. The acquisition unit 11B has the same function as the acquisition unit 11. That is, the obtaining unit 11B obtains the first aerial photograph, the second aerial photograph, and the house map data at the first time, which are to be interpreted as the loss of the house.

  The extraction unit 12B has the same function as the extraction unit 12. That is, the extraction unit 12B determines the position and the position of one house from each of the first aerial photograph and the second aerial photograph acquired by the acquiring unit 11B based on the position and the planar shape of the house indicated by the house data. A first house area image and a second house area image, which are images of an area corresponding to the range, are extracted.

  The interpretation unit 16 inputs the first house area image and the second house area image accompanied by the shape information on the planar shape of the house to the loss interpretation model, and acquires the output value of the loss interpretation model as house loss information. FIG. 9 is a diagram schematically illustrating an example of a data set input to the loss interpretation model and house loss information output from the model.

  The data set d9 to be input to the loss interpretation model includes a first house area image a91, a second house area image a92, and house area data c9. The house area data c9 is data extracted based on house data relating to a house to be interpreted as being lost, and includes polygon information representing the planar shape of the house. The first house area image a91 is an image extracted from the first aerial photograph taken at the first time, and represents a house to be interpreted as being lost. The second house area image a92 is an image extracted from the second aerial photograph taken at the second time, and represents the same area as the geographical area shown in the first house area image a91. ing.

  The interpretation unit 16 inputs the data set d9 into the loss interpretation model, and acquires the house loss information f9 as an output value. In the house loss information, the high possibility of house loss is reflected in the ratio of white pixels to black pixels. Specifically, in the learning data, the correct data indicating that the house has been lost is defined as image data composed of white pixels, and the correct data indicating that the house is not lost is defined as image data composed of black pixels. Therefore, the higher the ratio of white pixels in the house loss information, the higher the possibility that the house has been lost. The house loss information f9 shown in FIG. 9 indicates that there is a high possibility that the house has been lost because the ratio of white pixels is extremely high.

  FIG. 10 is a diagram illustrating an example of house loss information output from the loss interpretation model. In the house loss information f11 of the house loss information f11 to f16 shown in FIG. 10, the ratio of white pixels is extremely high. Therefore, the house loss information f11 indicates that there is a high possibility that the house has been lost. On the other hand, in the house loss information f16, the ratio of white pixels is extremely low. Therefore, the house loss information f16 indicates that the possibility that the house has been lost is low. As described above, the house loss information is composed of a plurality of pixels, and the classification result of the presence or absence of the house is indicated as one of the binary colors of each pixel. The possibility of the loss of the house is reflected in the ratio of the pixels of the second color, and the presence or absence of the loss of the house can be easily read.

  The loss information output unit 17 outputs the house loss information acquired by the reading unit 16. Specifically, the loss information output unit 17 outputs and stores the house loss information to a predetermined storage unit. Further, the loss information output unit 17 may display the house loss information on a predetermined display device.

  As described above, according to the house transfer interpretation device 1B of the present embodiment, based on the house map data at the first time, each of the first and second aerial photographs corresponds to the position and range of one house. First and second house region images, which are images of the region to be moved, are extracted. The first and second house area images extracted in this manner include image features indicating the presence or absence of one house at each of the first and second periods. Since the first and second house region images extracted in this way are input to the loss interpretation model together with the shape information representing the planar shape of the house, the house between the first time and the second time is The presence or absence of loss is accurately read.

  Next, with reference to FIG. 11 and FIG. 12, an operation of the house change interpretation system of the present embodiment will be described. FIG. 11 is a flowchart showing the processing contents of the house change reading method performed in the house change reading apparatus 1A in the phase of model generation by machine learning.

  In step S1, the acquisition unit 11 acquires a first aerial photograph taken at a first time, a second aerial photograph taken at a second time, and house map data at the first time.

  In step S2, the extraction unit 12 extracts the first house area image and the second house from each of the first aerial photograph and the second aerial photograph based on the position and the planar shape of the house indicated by the house data. Extract the region image.

  In step S3, the learning data generation unit 13 determines whether or not a house has been lost in the first and second house area images accompanying the house shape information and the areas shown in the first and second house area images. The learning data including the correct answer data is generated.

  In step S4, the model generation unit 14 performs machine learning based on the learning data generated by the learning data generation unit 13, receives the first and second house area images accompanying the house shape information, A loss interpretation model that outputs house loss information indicating the presence or absence of loss in the second house region image of the house shown in the house region image is generated.

  In step S5, the model output unit 15 outputs the loss interpretation model generated by the model generation unit 14 in step S4.

  FIG. 12 is a flowchart showing the processing contents of the house change interpretation method performed in the house change interpretation apparatus 1B in the phase of house loss judgment using the trained model.

  In step S11, the acquisition unit 11B acquires the first aerial photograph, the second aerial photograph, and the house map data at the first time, which are to be interpreted as the loss of the house.

  In step S12, the extraction unit 12B extracts a first aerial photograph and a second aerial photograph acquired by the acquisition unit 11B in step S11 based on the position and the planar shape of the house indicated by the house data. The first house area image and the second house area image are extracted.

  In step S13, the interpretation unit 16 inputs the first house area image and the second house area image accompanying the shape information on the planar shape of the house to the loss interpretation model.

  In step S14, the interpretation unit 16 acquires the output value from the loss interpretation model as house loss information.

  In step S16, the loss information output unit 17 outputs the house loss information acquired by the reading unit 16.

  Next, a house change interpretation program for causing a computer to function as a house change interpretation device will be described with reference to FIG. The house change interpretation program P1A shown in FIG. 13A is a program for causing a computer to function as the house change interpretation device 1A in the model generation aspect, and includes a main module m10, an acquisition module m11, an extraction module m12, and learning. It includes a data generation module m13, a model generation module m14, and a model output module m15.

  The house change reading program P1B shown in FIG. 13B is a program for causing a computer to function as the house change reading device 1B in the house loss reading phase using a trained model, and includes a main module m10B, An acquisition module m11B, an extraction module m12B, a reading module m16, and a loss information output module m17 are included.

  The main modules m10 and m10B are parts that comprehensively control the house transfer interpretation processing. The functions realized by executing the acquisition module m11, the extraction module m12, the learning data generation module m13, the model generation module m14, and the model output module m15 are respectively the acquisition unit 11 of the house transfer interpretation device 1A shown in FIG. The functions are the same as those of the extraction unit 12, the learning data generation unit 13, the model generation unit 14, and the model output unit 15.

  The functions realized by executing the acquisition module m11B, the extraction module m12B, the interpretation module m16, and the loss information output module m17 are respectively the acquisition unit 11B, the extraction unit 12B, and the interpretation unit of the house change interpretation device 1B shown in FIG. 16 and the function of the loss information output unit 17.

  The house transfer interpretation programs P1A and P1B are provided by storage media M1A and M1B such as a magnetic disk, an optical disk, or a semiconductor memory. Further, the house change interpretation programs P1A and P1B may be provided as a computer data signal superimposed on a carrier wave via a communication network.

  The present invention has been described in detail based on the embodiments. However, the present invention is not limited to the above embodiment. The present invention can be variously modified without departing from the gist thereof.

  1, 1A, 1B ... house transfer reading device, 11, 11B ... acquisition unit, 12, 12B ... extraction unit, 13 ... learning data generation unit, 14 ... model generation unit, 15 ... model output unit, 16 ... reading unit, 17 ... Loss information output unit, 21 ... Aerial photograph storage unit, 22 ... House map data storage unit, 23 ... Model storage unit, m10, m10B ... Main module, m11, m11B ... Acquisition module, m12, m12B ... Extraction module, m13 ... Learning data generation module, m14: model generation module, m15: model output module, m16: interpretation module, m17: loss information output module, M1A, M1B: storage medium, P1A, P1B: house transfer interpretation program.

Claims (11)

A house transfer interpretation system that generates a loss interpretation model for interpreting loss of a house between a first time and a second time after the first time based on an aerial photograph. ,
Represented in a first aerial photograph taken at the first time, a second aerial photograph taken at the second time, and the first and second aerial photographs at the first time; An acquisition unit for acquiring house map data corresponding to an area, wherein the house map data includes one or more house data, and the house data includes shape information and a position representing at least a planar shape of one house. An acquisition unit including location information indicating
A first image which is an image of an area corresponding to a position and a range of one house from each of the first aerial photograph and the second aerial photograph based on the position and planar shape of the house indicated by the house data. An extraction unit for extracting a house area image and a second house area image of
At least the first house area image and the second house area image accompanying the shape information, and correct data indicating whether or not a house has been lost in the areas shown in the first and second house area images. A learning data generation unit that generates learning data;
Machine learning is performed based on the learning data, and house loss information indicating whether or not the house shown in the first house area image has been lost is output using the first and second house area images accompanying the shape information as input. A model generation unit that generates the loss interpretation model,
A model output unit that outputs the generated loss interpretation model,
House transfer interpretation system equipped with. The extraction unit generates, based on the shape information of the house indicated by the house data, house area data including a rectangular area including graphic data representing a planar shape of the house, and generates the first aerial photograph and the house area data. In each of the second aerial photographs, an area corresponding to the house area data is extracted as the first house area image and the second house area image,
The learning data generation unit generates the learning data including the first house area image, the second house area image and the house area data, and the correct answer data.
The house transfer interpretation system according to claim 1. The extracting unit is configured to determine, based on the shape information of the house indicated by the house data, a shape corresponding to the planar shape of the house indicated by the shape information from each of the first aerial photograph and the second aerial photograph. Are extracted as the first house area image and the second house area image,
The learning data generation unit generates the learning data including the first house area image and the second house area image having a shape corresponding to the planar shape of the house and the correct answer data,
The house transfer interpretation system according to claim 1. The correct answer data is image data having a predetermined shape, and all pixels of the correct answer data indicating that the house has been lost have one of the first color and the second color, and the house has been lost. All the pixels of the correct data indicating that they do not have the other of the first color and the second color,
The house change interpretation system according to any one of claims 1 to 3. House change in a house change interpretation system that generates a loss interpretation model for interpreting loss of a house between a first time and a second time after the first time based on an aerial photograph. The interpretation method,
Represented in a first aerial photograph taken at the first time, a second aerial photograph taken at the second time, and the first and second aerial photographs at the first time; Acquiring house map data corresponding to the region, wherein the house map data includes one or more house data, and the house data includes shape information and a position representing at least a planar shape of one house. An acquisition step including location information indicating
A first image which is an image of an area corresponding to a position and a range of one house from each of the first aerial photograph and the second aerial photograph based on the position and planar shape of the house indicated by the house data. An extraction step of extracting a house area image and a second house area image of
At least the first house area image and the second house area image accompanying the shape information, and correct data indicating whether or not a house has been lost in the areas shown in the first and second house area images. A learning data generating step of generating learning data;
Machine learning is performed based on the learning data, and the first and second house area images accompanying the shape information are input and loss information indicating the presence or absence of the house shown in the first house area image is output. A model generation step of generating the loss interpretation model,
An output step of outputting the generated loss interpretation model;
House transfer interpretation method having a. A house transfer interpretation system that generates a loss interpretation model for interpreting a loss of a house between a first time and a second time after the first time based on an aerial photograph. A house transfer reading program to function as a
On the computer,
Represented in a first aerial photograph taken at the first time, a second aerial photograph taken at the second time, and the first and second aerial photographs at the first time; An acquisition function for acquiring house map data corresponding to an area, wherein the house map data includes one or more house data, and the house data includes at least a shape information and a position representing a planar shape of one house. An acquisition function including location information indicating
A first image which is an image of an area corresponding to a position and a range of one house from each of the first aerial photograph and the second aerial photograph based on the position and planar shape of the house indicated by the house data. An extraction function for extracting a house area image and a second house area image of
At least the first house area image and the second house area image accompanying the shape information, and correct data indicating whether or not a house has been lost in the areas shown in the first and second house area images. A learning data generation function for generating learning data;
Machine learning is performed based on the learning data, and the first and second house area images accompanying the shape information are input and loss information indicating the presence or absence of the house shown in the first house area image is output. A model generation function for generating the lost interpretation model,
An output function of outputting the generated loss interpretation model,
A house transfer reading program that realizes. A trained machine learning generated neural network for interpreting, based on aerial photographs, the loss of a house between a first time and a second time later than the first time. The model, which is a model of loss interpretation,
The first house area image and the second house area image accompanying the shape information representing the planar shape of the house are used as input values, and the presence or absence of houses in the areas shown in the first and second house area images is indicated. It is configured by machine learning based on learning data with correct answer data as an output value,
Each of the first house area image and the second house area image is based on house data including shape information indicating a planar shape and position information indicating a position with respect to one house, at the first time. It is an image of an area corresponding to the position and range of one house extracted from each of the first aerial photograph taken and the second aerial photograph taken at the second time.
Lost interpretation model. A house that reads the loss of a house between a first time and a second time that is a time after the first time based on the aerial photograph using the loss reading model according to claim 7. A house transfer interpretation system for interpreting loss,
Represented in a first aerial photograph taken at the first time, a second aerial photograph taken at the second time, and the first and second aerial photographs at the first time; An acquisition unit for acquiring house map data corresponding to an area, wherein the house map data includes one or more house data, and the house data includes shape information and a position representing at least a planar shape of one house. An acquisition unit including location information indicating
A first image which is an image of an area corresponding to a position and a range of one house from each of the first aerial photograph and the second aerial photograph based on the position and planar shape of the house indicated by the house data. An extraction unit for extracting a house area image and a second house area image of
A reading unit that inputs at least the first house area image and the second house area image with the shape information to the loss interpretation model, and acquires an output value of the loss interpretation model as house loss information;
A loss information output unit that outputs the house loss information acquired by the reading unit,
House transfer interpretation system equipped with. The house loss information is an image having a shape corresponding to the shape of the first and second house area images,
Each pixel of the house loss information has one of a first color and a second color different from the first color,
The ratio of the pixel of the first color and the pixel of the second color in the house loss information indicates a possibility of house loss in the area shown in the first and second house area images.
The house change interpretation system according to claim 8. A house transfer that uses the aerial photograph to interpret the loss of a house between a first time and a second time that is a time later than the first time, using the loss reading model according to claim 7. A method of interpreting house transfer in the interpretation system,
Represented in a first aerial photograph taken at the first time, a second aerial photograph taken at the second time, and the first and second aerial photographs at the first time; Acquiring house map data corresponding to the region, wherein the house map data includes one or more house data, and the house data includes shape information and a position representing at least a planar shape of one house. An acquisition step including location information indicating
A first image which is an image of an area corresponding to a position and a range of one house from each of the first aerial photograph and the second aerial photograph based on the position and planar shape of the house indicated by the house data. An extraction step of extracting a house area image and a second house area image of
A reading step of inputting at least the first house area image and the second house area image with the shape information to the loss interpretation model, and acquiring an output value of the loss interpretation model as house loss information;
A loss information output step of outputting the house loss information obtained in the reading step,
House transfer interpretation method having a. A computer, using the loss interpretation model according to claim 7, interprets the loss of the house between the first time and the second time after the first time based on the aerial photograph. A house transfer interpretation program for functioning as a house transfer interpretation system,
On the computer,
Represented in a first aerial photograph taken at the first time, a second aerial photograph taken at the second time, and the first and second aerial photographs at the first time; An acquisition function for acquiring house map data corresponding to an area, wherein the house map data includes one or more house data, and the house data includes at least a shape information and a position representing a planar shape of one house. An acquisition function including location information indicating
A first image which is an image of an area corresponding to a position and a range of one house from each of the first aerial photograph and the second aerial photograph based on the position and planar shape of the house indicated by the house data. An extraction function for extracting a house area image and a second house area image of
A reading function of inputting at least the first house area image and the second house area image with the shape information to the loss interpretation model, and acquiring an output value of the loss interpretation model as house loss information;
A loss information output function for outputting house loss information obtained by the reading function,
House transfer reading program that realizes.