JP2019046416A - Premises change estimation apparatus, premises change learning apparatus, premises change estimation method, parameter generation method for discriminator, and program - Google Patents

Abstract

To provide a premises change estimation apparatus capable of improving estimation accuracy of premises change on the basis of images of a land, premises change learning apparatus, premises change estimation method, parameter generation method of a discriminator, and program.SOLUTION: An estimation processing execution unit 76 inputs estimation input data in a learned change discriminator 34, the estimation input data representing a difference in a pair of estimation images of a land captured at different time points and representing whether or not a building is shown at a prescribed position in an at least one estimation image. The estimation processing execution unit 76 estimates premises change between the prescribed position in the at least one estimation image, and a position associated with the prescribed position in the other estimation image on the basis of the output when the estimation input data are input in the learned change discriminator 34.SELECTED DRAWING: Figure 8

Description

  The present invention relates to a house change estimation device, a house change learning device, a house change estimation method, a classifier parameter generation method, and a program.

  Patent Document 1 describes a method of detecting a change in the shape of a structure used for an appearance inspection or the like of the structure. In the method described in Patent Document 1, a shape change of a structure is detected by inputting a pair of images of a structure such as a tunnel captured at different times by a camera on the ground into a convolutional neural network (CNN). ing.

JP 2017-62776 A

  Based on the output when a pair of land images such as aircraft images and satellite images captured at different times are input to a machine learning model such as CNN, for example, new construction, loss, rebuilding, etc. , We are examining a method to estimate the house change shown in the image. For example, high accuracy is required when estimating house changes for fixed asset tax services etc. using the method under consideration.

  In this respect, in the technique described in Patent Document 1, it is premised that the same part of the structure appears in both of the pair of images input to the CNN. Therefore, the technique described in Patent Document 1 can not exhibit sufficient accuracy in estimation of new construction and loss such as a house being shown only in one of a pair of images.

  The present invention has been made in view of the above problems, and one of its purposes is a house change estimation device, a house change learning device, and a house change estimation method capable of improving the estimation accuracy of house changes based on an image of land. It is providing the parameter generation method and program of a discriminator.

  (1) In the house change estimation device according to the present invention, a building is displayed at a predetermined position in the learning image of at least one of the learning images of a pair of lands photographed at different time points. Teaching data indicating house change between the learning input data indicating whether or not the predetermined position in one learning image and the position corresponding to the predetermined position in the other learning image And a difference between estimated data of a pair of lands captured at different points in time, and a building being indicated at a predetermined position in at least one of the estimated images. Based on a change discriminator input unit that inputs estimated input data indicating whether or not to the change discriminator, and an output when the estimated input data is input to the change discriminator , Including a house change estimating means for estimating a house change between the position associated with the said predetermined position in the predetermined position and the other of said estimated image in one of the estimated image.

  (2) In the house change estimation apparatus according to (1), presence or absence feature extraction is performed to output feature information indicating whether a building is shown at a predetermined position in the image according to the input of the image of the land. And change feature input means for specifying feature information which is output when the pair of estimated images is input to the presence feature extractor, the change discriminator input means including the presence feature information The estimated input data including the feature information identified by the identifying means may be input to the change discriminator.

  (3) In the house movement estimation apparatus according to (1) or (2), an image of a pair of lands photographed at different times and a building at a predetermined position in the image of at least one of the images A change feature extractor that outputs feature information indicating a difference between the pair of land images according to an input of an image indicating whether or not the symbol is indicated, the pair of estimated images, and at least one of the estimates Change feature information specifying means for specifying feature information which is an output when an image indicating whether a building is shown at a predetermined position in the estimated image for an image is input to the change feature extractor; The change discriminator input means may be configured to input the estimated input data including the feature information identified by the change feature information identifying means to the change discriminator.

  (4) In the house movement estimation apparatus according to (3), the change feature extraction unit is configured to generate an image of a pair of lands captured at different points in time and a predetermined one of the images of at least one of the images. The characteristic information indicating the difference between the pair of land images is output according to the input of the image indicating the shape of the building indicated in the position, and the change characteristic information identification unit determines the pair of estimated images; A feature information is specified, which is an output when an image indicating a shape of a building shown at the predetermined position in the estimated image of at least one of the estimated images is input to the change feature extractor. be able to.

  (5) In the house movement estimation apparatus according to (1), presence or absence feature extraction is performed to output feature information indicating whether a building is shown at a predetermined position in the image according to the input of the image of the land. , Presence / absence feature specifying means for specifying feature information which is an output when the pair of estimated images is inputted to the presence / absence feature extractor, an image of a pair of lands photographed at different points in time, and A change feature extractor that outputs feature information indicating a difference between the pair of land images according to an input of an image indicating whether a building is shown at a predetermined position in the image for the image of An output when the pair of estimated images and an image indicating whether a building is shown at a predetermined position in at least one of the estimated images are input to the change feature extractor The apparatus further includes: change feature information specifying means for specifying certain feature information; and combining means for combining the feature information specified by the presence / absence feature information specifying means with the feature information specified by the change feature information specifying means The discriminator input means may be configured to input the estimated input data including the feature information combined by the combining means to the change discriminator.

  (6) In the house change estimation device according to (1) to (5), a screen generation means for generating a screen on which information representing a house change estimated by the house change estimation means is arranged, and the screen generation means And a display control unit configured to control a screen generated by the control unit to be displayed.

  (7) In the house change learning apparatus according to the present invention, a building is displayed at a predetermined position in the learning image of at least one of the learning images of a pair of lands photographed at different time points. Input means for inputting, to the change discriminator, learning input data indicating whether or not it has been input, an output when the learning input data is inputted to the change discriminator, the predetermined position in one of the learning images A comparison means for comparing teacher data indicating house change between a position corresponding to the predetermined position in the other learning image, and a parameter of the change discriminator based on a result of the comparison And parameter changing means for changing.

  (8) In the house change estimation method according to the present invention, the difference between the learning images of a pair of lands taken at different points in time, and the building is displayed at a predetermined position in the learning image for at least one of the learning images. Teaching data indicating house change between the learning input data indicating whether or not the predetermined position in one learning image and the position corresponding to the predetermined position in the other learning image And a change discriminator obtained by learning a plurality of learning data, the difference between estimated images of a pair of lands photographed at different time points, and a building being indicated at a predetermined position in at least one of the estimated images One of the steps of inputting estimated input data indicating whether or not to the change discriminator, and an output when the estimated input data is input to the change discriminator; Comprising the steps of: estimating a house change between the position associated with the said predetermined position in the predetermined position and the other of said estimated image in serial estimation image.

  (9) The parameter generation method for a classifier according to the present invention is a method for generating a difference between learning images of a pair of lands photographed at different points in time and at a predetermined position in the learning image for at least one of the learning images. Inputting, to the change discriminator, learning input data indicating whether or not is indicated, an output when the learning input data is input to the change discriminator, and the predetermined position in one of the learning images Comparing the parameter of the change discriminator with the step of comparing teacher data indicating house change between the other and the position corresponding to the predetermined position in the other learning image, and the result of the comparison. And changing.

  (10) In the program according to the present invention, is the difference between the learning images of a pair of lands taken at different times, and whether a building is shown at a predetermined position in the learning image for at least one of the learning images? Learning input data indicating whether or not there is teaching data indicating house change between the predetermined position in one of the learning images and a position corresponding to the predetermined position in the other learning image; The change discriminator, which has learned a plurality of learning data, includes a difference between estimated images of a pair of lands captured at different time points, and whether a building is shown at a predetermined position in at least one of the estimated images A step of inputting estimated input data indicating the input to the change discriminator, and based on an output when the estimated input data is input to the change discriminator, Procedure for estimating the house change between the position associated with the said predetermined position in the predetermined position and the other of said estimated image of the inner, causing the computer to execute.

  (11) In another program according to the present invention, a building is shown at a predetermined position in the learning image of at least one of the learning images of a pair of lands photographed at different time points. Procedure for inputting learning input data indicating whether or not to the change discriminator, an output when the learning input data is input to the change discriminator, the predetermined position in one of the learning images, and the other A procedure of comparing with teacher data indicating house change between a position corresponding to the predetermined position in the learning image, and a procedure of changing a parameter of the change discriminator based on a result of the comparison Make it run on a computer.

  According to the present invention, estimation accuracy of house change based on an image of a land is improved.

It is a block diagram of the house change estimation apparatus which concerns on one Embodiment of this invention. It is a figure which shows an example of a presumed target image. It is a figure which shows an example of a presumed target image. It is a figure which shows an example of a presumed result screen. It is a figure which shows an example of a presumed building detection image. It is a figure which shows an example of a presumed building detection image. It is a figure which shows the example of the chip | tip image about the building estimated that there existed a change in two time points. It is a figure which shows an example of the learned machine learning model used for estimation. It is a figure which shows an example of the machine learning model at the time of learning using learning data being performed. It is a functional block diagram showing an example of a function of a house change estimating device concerning one embodiment of the present invention. It is a flow figure showing an example of a flow of study processing of a machine learning model concerning one embodiment of the present invention. It is a flow figure showing an example of a flow of study processing of a machine learning model concerning one embodiment of the present invention. It is a flow figure showing an example of the flow of presumption processing performed with the house change estimation device concerning one embodiment of the present invention. It is a flow figure showing an example of the flow of presumption processing performed with the house change estimation device concerning one embodiment of the present invention. It is a figure which shows an example of a presumed chip | tip image. It is a figure which shows an example of a machine learning model. It is a figure which shows an example of a machine learning model. It is a figure which shows an example of a machine learning model. It is a figure which shows an example of a machine learning model.

  Hereinafter, an embodiment of the present invention will be described in detail based on the drawings.

  FIG. 1 is a block diagram of a house change estimation apparatus 10 according to an embodiment of the present invention. The house change estimation device 10 according to the present embodiment is a computer such as a personal computer. As shown in FIG. 1, the house change estimation device 10 includes, for example, a processor 12, a storage unit 14, a display unit 16, and an operation unit 18.

  The processor 12 is a program control device such as a CPU that operates according to a program installed in the house change estimation device 10, for example.

  The storage unit 14 is a storage element such as a ROM or a RAM, a hard disk drive, or the like. The storage unit 14 stores programs executed by the processor 12 and the like.

  The display unit 16 is a display device such as a liquid crystal display, and displays various images according to an instruction of the processor 12.

  The operation unit 18 is a user interface such as a keyboard and a mouse, and receives a user's operation input and outputs a signal indicating the content to the processor 12.

  The house change estimation device 10 may include a communication interface such as a network board, an optical disk drive for reading an optical disk such as a DVD-ROM or Blu-ray (registered trademark) disk, a USB (Universal Serial Bus) port, etc. .

  In the house change estimation device 10 according to the present embodiment, a house change such as new construction, loss, rebuilding, etc. is estimated based on the images of a pair of lands photographed at different points in time. FIG. 2A is a view schematically showing an example of an estimated target image 20a which is an image captured at a first time point. FIG. 2B is a view schematically showing an example of an estimated target image 20b which is an image captured at a second time point after the first time point. In the present embodiment, for example, based on the estimated target image 20a and the estimated target image 20b, house change is estimated between the time when the estimated target image 20a is photographed and the time when the estimated target image 20b is photographed. . The estimated target image 20a and the estimated target image 20b do not have to be an aircraft image captured by an aircraft, and may be satellite images captured by artificial satellites.

  FIG. 3 is a diagram showing an example of an estimation result screen 22 expressing house changes estimated based on a pair of images captured at different points in time. In the present embodiment, for example, the estimation result screen 22 illustrated in FIG. 3 is displayed on the display unit 16 according to an operation of selecting the estimated target image 20 a and the estimated target image 20 b.

  On the estimation result screen 22 shown in FIG. 3, an estimated target image 20b on which a house change corresponding image 24 associated with a house change, such as new construction, loss, rebuilding, etc., is superimposed is arranged. In the example of FIG. 3, an image of a solid line surrounding a building estimated to be a new building is shown as a house change corresponding image 24 a. Moreover, in the example of FIG. 3, the image of the dotted line surrounding the building (not shown in FIG. 3) estimated to be lost is shown as the house change corresponding | compatible image 24b. Moreover, in the example of FIG. 3, the image of the dashed-dotted line surrounding the building estimated to be rebuilding is shown as a house change corresponding | compatible image 24c.

  Thus, in the present embodiment, the user can view the estimation result of house change.

  In the estimation of a house change, in the present embodiment, for example, first, the position and the shape of a building shown in the estimated target image 20a and the estimated target image 20b are detected by a known method. FIG. 4A is a view showing an estimated building detection image 26a which is an example of the detection result of the position and shape of the building shown in the estimated target image 20a shown in FIG. 2A. FIG. 4B is a view showing an estimated building detection image 26b which is an example of the detection result of the position and shape of the building shown in the estimated target image 20b shown in FIG. 2B. In the estimated building detection image 26a shown in FIG. 4A and the estimated building detection image 26b shown in FIG. 4B, the position and shape of the detected building are indicated by polygons expressed in white.

  In the present embodiment, it is assumed that the same place is indicated by the estimated target image 20a and the estimated target image 20b. For example, a pixel of a coordinate value in the estimated target image 20a and a pixel of the coordinate value in the estimated target image 20b are associated with each other as pixels indicating the same place. Further, in the present embodiment, the pixel of a certain coordinate value in the estimated target image 20a or in the estimated target image 20b indicates the same place in the pixel of the coordinate value in the estimated building detection image 26a or in the estimated building detection image 26b. It is assumed that they correspond to each other as pixels. The estimated target image 20a and the estimated target image 20b do not have to indicate exactly the same place. The pixel of a certain coordinate value in the estimated target image 20a and the pixel of the corresponding coordinate value in the estimated target image 20b may not indicate exactly the same place.

  Then, for each of the plurality of polygons shown in the estimated building detection image 26a shown in FIG. 4A and the estimated building detection image 26b shown in FIG. 4B, a predetermined position in the polygon is specified. Here, for example, for all polygons shown in the estimated building detection image 26a and all polygons shown in the estimated building detection image 26b, the position of the center of the polygon is specified.

  Then, an estimated chip image 28 of predetermined vertical and horizontal lengths centering on the specified position is generated. Here, for example, as shown in FIG. 5, for each of a plurality of specified positions, an estimated chip image 28 occupying a rectangular area centered on the positions is an estimated target image 20 a, an estimated target image 20 b, and an estimated building It may be extracted from each of the detection images 26b. In this case, three estimated chip images 28 will be generated based on one position to be identified. For example, the estimated chip image 28a is extracted from the estimated target image 20a, the estimated chip image 28b is extracted from the estimated target image 20b, and the estimated chip image 28c is extracted from the estimated building detection image 26b.

  FIG. 5 shows an example of an estimated chip image 28 of a building estimated to have a house change in the present embodiment. The upper part of FIG. 5 shows three estimated chip images 28 associated with a building estimated to be a new building. These three estimated chip images 28 are generated based on the position of the center of the polygon shown in the estimated building detection image 26b. In the middle part of FIG. 5, three estimated chip images 28 associated with the buildings estimated to be lost are shown. These three estimated chip images 28 are generated based on the position of the center of the polygon shown in the estimated building detection image 26a. In the lower part of FIG. 5, three estimated chip images 28 associated with a building estimated to be a rebuild are shown. These three estimated chip images 28 are generated based on the position of the center of the polygon shown in the estimated building detection image 26b.

  And in this embodiment, the presumed result of house change is specified based on the result of having inputted the generated presumed chip picture 28 into the learned machine learning model 30 illustrated in Drawing 6, for example.

  The machine learning model 30 shown in FIG. 6 includes a processing unit 32 and a change discriminator 34. The original estimated input data corresponding to the estimated chip image 28 described above is input to the processing unit 32. Then, the processing unit 32 generates processed estimated input data by processing the original estimated input data. The change discriminator 34 outputs the estimation result of the house change according to the input of the processing estimation input data generated by the processing unit 32.

  The processing unit 32 according to the present embodiment is implemented as a machine learning model. The processing unit 32 according to the present embodiment includes two presence feature extractors 32a (32aa and 32ab) and one variation feature extractor 32b. In the machine learning model shown in FIG. 6, two presence / absence identification devices 32c (32ca and 32cb) which are used for learning of the machine learning model 30 described later but not used for estimation of house change are shown by broken lines. . In the present embodiment, the same values are set as values of parameters such as weights in the two presence / absence feature extractors 32a. Further, in the present embodiment, the same value is set in the two presence / absence identification devices 32c as the value of the parameter such as weight.

  In the present embodiment, for example, the presence feature extractor 32a and the presence identifier 32c are implemented as a convolutional neural network (CNN) as a whole. The presence / absence feature extractor 32a corresponds to, for example, a convolution layer and a pooling layer in CNN, and the presence / absence identifier 32c corresponds to, for example, a total combined layer in CNN.

  Further, in the present embodiment, for example, the change feature extractor 32 b and the change discriminator 34 are implemented as a CNN as a whole. The variation feature extractor 32b corresponds to, for example, the convolution layer and the pooling layer in CNN, and the variation discriminator 34 corresponds to, for example, the total combination layer in CNN.

  In the present embodiment, for example, three estimated chip images 28 generated based on the position of one polygon shown in the estimated building detection image 26 a or the estimated building detection image 26 b are machine learning models as original estimated input data. It is input to 30.

  For example, the estimated chip image 28a is input to the presence / absence feature extractor 32aa as the input 40. Then, in response to the input, the presence / absence feature extractor 32aa outputs, as an output 42, feature information indicating whether a building is shown at the center of the estimated chip image 28a.

  Further, the estimated chip image 28 b is input as the input 44 to the presence / absence feature extractor 32 ab. Then, according to the input, the presence / absence feature extractor 32ab outputs, as an output 46, feature information indicating whether a building is shown at the center of the estimated chip image 28b.

  Further, data (multi-channel image) is generated by combining the estimated chip image 28 a input as the input 40, the estimated chip image 28 b input as the input 44, and the estimated chip image 28 c input as the input 48. In FIG. 6, the combination of estimated chip images 28a is shown as arrow A1, the combination of estimated chip images 28b is shown as arrow A2, and the combination of estimated chip images 28c is shown as arrow A3. The combined data (multi-channel image) is then input as the input 50 to the variation feature extractor 32b. Then, the variation feature extractor 32b outputs, as an output 52, feature information indicating the difference between the estimated chip image 28a and the estimated chip image 28b according to the input.

  For example, a feature map composed of 512 images (maps) in which the number of vertical and horizontal pixels is L, which is the output of the CNN intermediate layer, corresponds to the above-described feature information. The number of images (the number of maps) included in the feature map is not limited to 512.

  Then, data combining the feature information output as the output 42, the feature information output as the output 46, and the feature information output as the output 52 is generated. In FIG. 6, the combination of feature information output as output 42 is shown as arrow A4, the combination of feature information output as output 46 is shown as arrow A5, and the combination of feature information output as output 52 is shown as arrow A6. It is done. For example, this combined data corresponds to the above-mentioned processing estimation input data. For example, if the feature information output as the output 42, the feature information output as the output 46, and the feature information output as the output 52 are feature maps each having 512 maps, the processing estimation input data is the number of maps The feature map is 1536.

  Then, the processing estimation input data generated in this manner is input to the change discriminator 34 as the input 54. Then, the change discriminator 34 outputs, as an output 56, estimation result data indicating a house change between the center of the estimated chip image 28a and the center of the estimated chip image 28b.

  The said estimation result data may be data which show the probability about each of the kind of house change, such as new construction, loss, rebuilding, no change, etc., for example. In this case, the type of house change associated with the probability of taking the maximum value may be specified as an estimation result of a house change between the center of the estimated chip image 28a and the center of the estimated chip image 28b. Further, the estimation result data may be, for example, data taking any value among multiple values associated with the type of house change, such as new construction, loss, rebuilding, no change, and the like. In this case, the value indicated by the estimation result data may be specified as an estimation result of house movement between the center of the estimated chip image 28a and the center of the estimated chip image 28b.

  In this embodiment, for example, for each polygon shown in the estimated building detection image 26a and the estimated building detection image 26b, three estimated chip images 28 generated based on the position of the center of the polygon are used. Estimates of house changes were made. Then, based on the result of the estimation, the estimation result screen 22 is generated and displayed on the display unit 16.

  Hereinafter, learning of the machine learning model 30 will be described with reference to FIG.

  In the present embodiment, for example, learning of the machine learning model 30 is performed by supervised learning using original learning data including original learning input data and original teacher data. Further, in the learning of the machine learning model 30 in the present embodiment, the processing unit 32 generates processing learning input data by processing the original learning input data. Then, the learning of the change discriminator 34 is executed by supervised learning using the process learning data including the process learning input data generated by the process unit 32 and the process teacher data which is at least a part of the original teacher data. . Although the change discriminator 34 is a part of the machine learning model 30, since the change discriminator 34 itself is also a machine learning model, learning of the machine learning model 30 is performed by supervised learning using processing learning data as described above. The learning of the change discriminator 34 is performed as a part of.

  The original learning input data includes, for example, a learning image of a pair of lands photographed at different points in time, and a learning building detection image indicating the position and shape of a building shown in any of the learning images. . Hereinafter, each of the pair of learning images is referred to as a first learning image and a second learning image. Here, the learning building detection image may be, for example, an image generated based on the detection result of the position and shape of the building shown in the learning image, similar to the estimated building detection image 26.

  In the original teacher data, for example, presence / absence data indicating whether a building is shown at the center of the learning image, and a change indicating a house change between the center of the first learning image and the center of the second learning image Data is included. Hereinafter, presence / absence data indicating whether a building is indicated at the center of the first learning image is referred to as first presence / absence data, and presence / absence data indicating whether a building is indicated at the center of the second learning image is We call it 2 presence data.

  The presence / absence data is, for example, binary data which takes 1 as a value when a building is shown at the center, and takes 0 as a value when a building is not shown at the center. The change data is, for example, data taking any value of multiple values associated with the type of house change, such as new construction, loss, rebuilding, no change, and the like.

  In learning of the machine learning model 30, for example, the first learning image included in the original learning input data is input as the input 40 to the presence / absence feature extractor 32aa. Further, the second learning image included in the original learning input data is input as the input 44 to the presence / absence feature extractor 32ab.

  Also, data obtained by combining the learning image included in the original learning input data input as the input 40 and the input 44 and the learning building detection image included in the original learning input data input as the input 48 (multichannel Image) is generated. In FIG. 7, the combination of the first learning images is shown as arrow A7, the combination of the second learning images is shown as arrow A8, and the combination of the learning building detection images is shown as arrow A9. The combined data (multi-channel image) is then input as the input 50 to the variation feature extractor 32b.

  The output 42 of the presence feature extractor 32aa is input to the presence identifier 32ca, and the output 46 of the presence feature extractor 32ab is input to the presence identifier 32cb.

  Also, data is generated by combining the output 42 of the presence / absence feature extractor 32 aa, the output 46 of the presence / absence feature extractor 32 ab, and the output 52 of the variation feature extractor 32 b. In FIG. 7, the combination of the output 42 of the presence / absence feature extractor 32aa is shown as arrow A10, the combination of the output 46 of the presence / absence feature extractor 32ab is shown as arrow A11, and the combination of the output 52 of the variation feature extractor 32b is shown as arrow A12. It is done. For example, this combined data corresponds to the above-mentioned processing learning input data.

  Then, the process learning input data generated in this manner is input to the change discriminator 34 as the input 54.

  Then, in the present embodiment, for example, the output 58 of the presence / absence identifier 32 ca is compared with the value of the first presence / absence data included in the teacher data associated with the learning input data input to the machine learning model 30. Then, values of parameters such as weights of the presence / absence feature extractor 32aa and the presence / absence identification device 32ca are changed based on a comparison result (for example, an error) by an error back propagation method (back propagation).

  Further, in the present embodiment, for example, the output 60 of the presence / absence identifier 32cb is compared with the value of the second presence / absence data included in the teacher data associated with the learning input data input to the machine learning model 30. Then, values of parameters such as weights of the presence / absence feature extractor 32ab and the presence / absence identification device 32cb are changed based on a comparison result (for example, an error) by an error back propagation method (back propagation).

  Further, in the present embodiment, the value of the other parameter is also updated in conjunction with the change of the value of one of the parameters so that the value of the parameter of the presence / absence feature extractor 32aa and the value of the parameter of the presence / absence feature extractor 32ab become the same. . Further, in the present embodiment, the value of the other parameter is also updated in conjunction with the change of the value of one of the parameters so that the value of the parameter of the presence / absence identifier 32ca and the value of the parameter of the presence / absence identifier 32cb become the same.

  Further, in the present embodiment, for example, the value (for example, the value of change data) of the processed teacher data included in the original teacher data associated with the original learning input data input to the machine learning model 30 by the output 56 of the change discriminator 34 Compared with. Then, values of parameters such as the weight of the change feature extractor 32 b and the change discriminator 34 are changed based on a comparison result (for example, an error) by an error back propagation method (back propagation) or the like. Further, in the present embodiment, the values of the parameters of the presence / absence feature extractor 32aa and the values of the parameters of the presence / absence feature extractor 32ab are also changed based on the comparison result (error) by an error back propagation method (back propagation). It has become so.

  In the present embodiment, for example, a learning process of causing the machine learning model 30 to learn a plurality of learning data is executed. Then, estimation of a house change is performed using the learned machine learning model 30 in which such a learning process is performed.

  In the present embodiment, the processing estimation input data input to the change discriminator 34 includes not only feature information associated with the difference between the pair of land images captured at different points in time, but also predetermined ones in at least one of the images. It is also shown whether the building is shown at a position (e.g. center). For example, as indicated by arrows A4 to A6 in FIG. 6, whether or not a building is shown at the center of the estimated chip image 28a or the estimated chip image 28b in the feature information indicating the difference between the estimated chip image 28a and the estimated chip image 28b Process estimation input data is generated in which feature information indicating. Then, estimation of house movement is performed based on the processing estimation input data of many dimensions (number of maps), which is reinforced in this way as to whether or not a building is shown. Therefore, according to this embodiment, the estimation accuracy of the house change based on the image of the land is improved.

  Hereinafter, the function of the house change estimation device 10 according to the present embodiment and the process performed by the house change estimation device 10 according to the present embodiment will be further described.

  FIG. 8 is a functional block diagram showing an example of functions implemented by the house change estimation device 10 according to the present embodiment. In addition, in the house change estimation apparatus 10 which concerns on this embodiment, all the functions shown in FIG. 8 do not need to be mounted, and functions other than the function shown in FIG. 8 may be mounted.

  As shown in FIG. 8, the house change estimation apparatus 10 according to the present embodiment functionally includes, for example, a machine learning model 30, a learning data acquisition unit 70, a learning process execution unit 72, and an estimation target data acquisition unit 74. An estimation process execution unit 76, an estimation result identification unit 78, a screen generation unit 80, and a display control unit 82 are included. The machine learning model 30 includes a processing unit 32 and a change discriminator 34. The processing unit 32 further includes a presence feature extractor 32a, a change feature extractor 32b, and a presence identifier 32c. The machine learning model 30 is implemented mainly by the processor 12 and the storage unit 14. The learning data acquisition unit 70, the learning process execution unit 72, the estimation target data acquisition unit 74, the estimation process execution unit 76, the estimation result identification unit 78, and the screen generation unit 80 are mainly implemented with the processor 12. The display control unit 82 mainly includes the processor 12 and the display unit 16.

  The machine learning model 30, the learning data acquisition unit 70, and the learning process execution unit 72 play a role as, for example, a house change learning device that executes a learning process of the machine learning model 30. The machine learning model 30, the estimation target data acquisition unit 74, the estimation process execution unit 76, the estimation result specification unit 78, the screen generation unit 80, and the display control unit 82 have, for example, a role as a house movement estimation device for estimating house movement. I am responsible.

  The above functions may be implemented by causing the processor 12 to execute a program including a command corresponding to the above functions installed in the house change estimation device 10 that is a computer. For example, this program may be supplied to the house movement estimation apparatus 10 via a computer readable information storage medium such as an optical disk, a magnetic disk, a magnetic tape, a magneto-optical disk, a flash memory, or the Internet. Good.

  In the embodiment, the processing unit 32 generates, for example, processing learning input data based on the above-described original learning input data in learning, and processing estimation input based on the above-described original estimation input data in estimation Generate data.

  In the present embodiment, for example, the presence / absence feature extractor 32a included in the processing unit 32 outputs feature information indicating whether a building is indicated at a predetermined position in the image according to the input of the image of the land. . Here, for example, the presence / absence feature extractor 32a may output feature information indicating whether a building is shown at the center of the image according to the input of the image. Also, as described above, the presence / absence feature extractor 32a may include the presence / absence feature extractor 32aa and the presence / absence feature extractor 32ab.

  In the variation feature extractor 32 b included in the processing unit 32, in the present embodiment, for example, an image of a pair of lands captured at different time points, and a building at a predetermined position in the image of at least one of the images An image is input indicating whether or not is indicated. Then, in the present embodiment, for example, the change feature extractor 32b outputs feature information indicating the difference between the pair of land images according to the input.

  In the present embodiment, for example, the presence / absence identifier 32c included in the processing unit 32 is a machine learning model that learns whether or not a building is shown at a predetermined position (for example, the center) in an image. As described above, the presence or absence identifier 32c may include the presence or absence identifier 32ca and the presence or absence identifier 32cb.

  For example, feature information indicating whether or not a building is shown at a predetermined position in the learning image is input to the presence / absence identifier 32c. In the above-described example, the feature information output from the presence / absence feature extractor 32a is input to the presence / absence identifier 32c.

  The teacher data used for the learning of the presence / absence discriminator 32c is, for example, data indicating whether or not a building is indicated at a predetermined position in the learning image. For example, the presence / absence data described above corresponds to teacher data used for learning of the presence / absence identifier 32c.

  The change discriminator 34 is, for example, a machine learning model that learns the type of house change in the present embodiment. The aforementioned change learning input data is input to the change discriminator 34. The process learning input data indicates, for example, the difference between a pair of learning images captured at different points in time, and data indicating whether a building is indicated at a predetermined position in the learning image for at least one of the learning images. It is. In the above-described example, data obtained by combining the feature information output from the presence / absence feature extractor 32a and the feature information output from the variation feature extractor 32b corresponds to processing learning input data.

  In addition, for example, processing teacher data which is teacher data used for learning of the change discriminator 34 is between a predetermined position in one learning image and a position associated with the predetermined position in the other learning image. It is data showing house change. Here, the processing teacher data may be data indicating a house change between the center in one learning image and the center in the other learning image. In the above-mentioned example, for example, change data taking any value of multiple values associated with the type of house change, such as new construction, loss, rebuilding, no change, etc. corresponds to processing teacher data.

  The learning data acquisition unit 70 acquires, for example, original learning data including a combination of original learning input data and original teacher data input to the machine learning model 30 in the present embodiment.

  The original learning input data is the difference between a pair of learning images taken at different points in time, and whether or not a building is indicated at a predetermined position (for example, the center) in the learning image for at least one of the learning images. It may include the data shown. In the above-mentioned example, the combination of the first learning image, the second learning image, and the learning building detection image corresponds to the original learning input data.

  The original teacher data may also include data indicating house change between a predetermined position in one learning image and a position associated with the predetermined position in the other learning image. Here, the original teacher data may be data indicating a house change between the center in one learning image and the center in the other learning image. In the above-described example, the combination of the first presence data, the second presence data, and the change data corresponds to the original teacher data.

  In the present embodiment, for example, the learning process execution unit 72 has a plurality of original learning data acquired by the learning data acquisition unit 70 as the presence / absence feature extractor 32a, the change feature extractor 32b, the change identifier 34, and the presence / absence identifier 32c. The learning is performed by the machine learning model 30 including. Here, as described above, the presence / absence feature extractor 32a changes, based on the comparison result of the output when the learning process execution unit 72 inputs the processing learning input data to the change discriminator 34, and the value of the processing teacher data. The learning of the feature extractor 32 b and the change discriminator 34 may be performed.

  In the present embodiment, the estimation target data acquisition unit 74 acquires, for example, data to be estimated for house change. Here, for example, the estimated target image 20a and the estimated target image 20b may be acquired.

  The estimation process execution unit 76 executes, for example, a process of estimating a house change in the present embodiment. In this estimation process, the estimation process execution unit 76 indicates whether the difference between a pair of estimated images captured at different points in time and whether a building is shown at a predetermined position in at least one of the estimated images. The estimated input data may be input to the change discriminator 34. Here, for example, the estimated chip image 28 in the above-described example corresponds to the estimated image. Then, based on the output when the processing estimation input data is input to the change discriminator 34, the estimation processing execution unit 76 associates the predetermined position in one estimated image with the predetermined position in the other estimated image. It is possible to estimate the house change between the location where the

  The estimation process execution unit 76 may also specify feature information that is an output when a pair of estimated images is input to the presence / absence feature extractor 32a. Then, the estimation process execution unit 76 may input, to the change discriminator 34, the processing estimation input data including the specified feature information.

  Further, the estimation process execution unit 76 inputs, to the change feature extractor 32b, a pair of estimated images and an image indicating whether or not a building is shown at a predetermined position in the estimated image of at least one of the estimated images. Feature information that is an output of the event may be identified. Here, an image showing whether a building is shown at a predetermined position in the estimated image is, for example, a shape of the building shown at a predetermined position in the estimated image, such as the above-described estimated chip image 28c. It may be an image shown. Then, the estimation process execution unit 76 may input, to the change discriminator 34, the processing estimation input data including the specified feature information.

  Further, as described above, the estimation process execution unit 76 may combine the feature information that is the output of the presence / absence feature extractor 32a with the feature information that is the output of the variation feature extractor 32b. Then, the estimation processing execution unit 76 may input the processing estimation input data including the combined feature information to the change discriminator 34.

  Further, for example, the estimation process execution unit 76 may generate the estimated building detection image 26 or generate the estimated chip image 28. Then, based on the output when the estimated chip image 28 is input to the machine learning model 30, the house change shown in the estimated chip image 28 may be estimated.

  In the present embodiment, for example, the estimation result specifying unit 78 specifies the execution result of the estimation process by the estimation process execution unit 76. Here, for example, the type of house change (for example, new construction, loss, rebuilding, or no change) for each of the buildings shown in the estimated target image 20a or the estimated target image 20b may be identified.

  In the present embodiment, the screen generation unit 80 is, for example, information indicating an estimated house change such as the estimation result screen 22 illustrated in FIG. 3 based on the execution result of the estimation process specified by the estimation result identification unit 78. To create a screen where is placed.

  In the present embodiment, the display control unit 82 causes the display unit 16 to display a screen such as the estimation result screen 22 generated by the screen generation unit 80, for example.

  Hereinafter, an example of the flow of the learning process of the machine learning model 30 performed in the house change estimation apparatus 10 according to the present embodiment will be described with reference to the flow diagrams illustrated in FIGS. 9A and 9B.

  First, the learning data acquisition unit 70 acquires a plurality of original learning data (S101). The original learning data includes original learning input data and original teacher data.

  Then, the learning process execution unit 72 selects one of the original learning data acquired in the process shown in S101 for which the following process has not been performed yet (S102).

  Then, the learning process execution unit 72 inputs the first learning image included in the original learning input data of the original learning data selected in the process shown in S102 to the presence / absence feature extractor 32aa (S103).

  Then, the learning process execution unit 72 specifies the feature information output by the presence / absence feature extractor 32aa according to the input in the process shown in S103 (S104).

  Then, the learning process execution unit 72 inputs the second learning image included in the original learning input data of the original learning data selected in the process shown in S102 to the presence / absence feature extractor 32ab (S105).

  Then, the learning process execution unit 72 specifies the feature information output by the presence / absence feature extractor 32ab according to the input in the process shown in S105 (S106).

  The learning process execution unit 72 is data obtained by combining the first learning image and the second learning image included in the original learning input data of the original learning data selected in the process shown in S102 and the learning building detection image. A multi-channel image is generated (S107).

  Then, the learning process execution unit 72 inputs the multi-channel image generated in the process shown in S107 to the variation feature extractor 32b (S108).

  Then, the learning process execution unit 72 specifies the feature information output from the change feature extractor 32b according to the input in the process shown in S108 (S109).

  Then, the learning process execution unit 72 inputs the feature information identified in the process shown in S104 to the presence / absence identification device 32ca (S110).

  Then, the learning process execution unit 72 specifies the output of the presence / absence identification device 32ca according to the input in the process shown in S110 (S111).

  Then, the learning process execution unit 72 compares the output specified in the process shown in S111 with the value of the first presence / absence data included in the original teacher data of the original learning data (S112).

  Then, the learning process execution unit 72 determines the values of the parameters of the presence / absence feature extractor 32 aa, the presence / absence feature extractor 32 ab, the presence / absence identification device 32 ca, and the presence / absence identification device 32 cb based on the error that is the comparison result in the processing shown in S112. Change (S113).

  Then, the learning process execution unit 72 inputs the feature information identified in the process shown in S106 to the presence / absence identification device 32cb (S114).

  Then, the learning process execution unit 72 specifies the output of the presence / absence identifier 32cb according to the input in the process shown in S114 (S115).

  Then, the learning process execution unit 72 compares the output specified in the process shown in S115 with the value of the second presence / absence data included in the original teacher data of the original learning data (S116).

  Then, the learning process execution unit 72 determines the values of the parameters of the presence / absence feature extractor 32 aa, the presence / absence feature extractor 32 ab, the presence / absence identification device 32 ca, and the presence / absence identification device 32 cb based on the error that is the comparison result in the processing shown in S116. Change (S117).

  Then, the learning process execution unit 72 combines the feature information identified in the process shown in S104, the feature information identified in the process shown in S106, and the feature information identified in the process shown in S109. Are generated (S118).

  Then, the learning process execution unit 72 inputs the process learning input data generated in the process shown in S118 to the change discriminator 34 (S119).

  Then, the learning process execution unit 72 identifies the output of the change discriminator 34 according to the input in the process shown in S119 (S120).

  Then, the learning process execution unit 72 compares the output specified in the process shown in S120 with the value of the change data included in the original teacher data of the original learning data, which corresponds to the processing teacher data (S121).

  The learning process execution unit 72 then determines the values of the parameters of the presence / absence feature extractor 32 aa, the presence / absence feature extractor 32 ab, the change feature extractor 32 b, and the change discriminator 34 based on the error that is the comparison result in the process shown in S121. Is changed (S122).

  Then, the learning process execution unit 72 checks whether or not the original learning data that is not selected in the process shown in S102 exists (S123).

  If the existence is confirmed here (S123: Y), the process returns to the process shown in S102.

  If it is confirmed that there is not (S123: N), the processing shown in this processing example is ended.

  The execution order of the processes shown in S101 to S123 is not limited to the above-described one, and these processes may be performed in an execution order different from the order shown above.

  Hereinafter, an example of a flow of a house change estimation process performed in the house change estimation apparatus 10 according to the present embodiment will be described with reference to flowcharts illustrated in FIGS. 10A and 10B.

  First, the estimation target data acquisition unit 74 acquires the estimation target image 20a and the estimation target image 20b selected by the user (S201).

  Then, the estimation process execution unit 76 generates an estimated building detection image 26a based on the estimated target image 20a obtained in the process shown in S201, and detects an estimated building based on the estimated target image 20b obtained in the process shown in S201. An image 26b is generated (S202).

  The estimation process execution unit 76 then calculates the centers of the polygons for all polygons shown in the estimated building detection image 26a generated in the process shown in S202 and all polygons shown in the estimated building detection image 26b. The position is specified (S203).

  Then, the estimation process execution unit 76 selects one of the positions of the centers specified in the process shown in S203 that has not yet been subjected to the processes shown in S205 to S214 (S204).

  Then, the estimation process execution unit 76 generates an estimation chip image 28a, an estimation chip image 28b, and an estimation chip image 28c based on the position selected in the process shown in S204 (S205). Here, the estimated chip image 28a is generated, for example, by extracting a part of the estimated target image 20a. The estimated chip image 28b is generated by, for example, extracting a part of the estimated target image 20b. The estimated chip image 28c is generated, for example, by extracting a part of the estimated building detection image 26b.

  Then, the estimation process execution unit 76 inputs the estimated chip image 28a generated in the process shown in S205 to the presence / absence feature extractor 32aa (S206).

  Then, the estimation process execution unit 76 specifies the feature information output by the presence / absence feature extractor 32aa according to the input in the process shown in S206 (S207).

  Then, the estimation process execution unit 76 inputs the estimated chip image 28b generated in the process shown in S206 to the presence / absence feature extractor 32ab (S208).

  Then, the estimation process execution unit 76 specifies the feature information output by the presence / absence feature extractor 32ab according to the input in the process shown in S208 (S209).

  Then, the estimation process execution unit 76 generates a multi-channel image which is data obtained by combining the estimation chip image 28a generated in the process shown in S205, the estimation chip image 28b, and the estimation chip image 28 (S210).

  Then, the estimation process execution unit 76 inputs the multi-channel image generated in the process shown in S210 to the variation feature extractor 32b (S211).

  Then, the estimation process execution unit 76 specifies the feature information output from the variation feature extractor 32b according to the input in the process shown in S211 (S212).

  Then, the estimation process execution unit 76 combines the feature information identified in the process shown in S207, the feature information identified in the process shown in S209, and the feature information identified in the process shown in S212, and then the processing estimation input data Are generated (S213).

  Then, the estimation process execution unit 76 inputs the process estimation input data generated in the process shown in S213 to the change discriminator 34 (S214).

  Then, the estimation result specifying unit 78 specifies the output of the change discriminator 34 according to the input of the processing estimation input data in the process shown in S214 as estimation result data associated with the position specified in the process shown in S204 ( S215).

  Then, the estimation process execution unit 76 confirms whether or not there is a position that has not been selected in the process shown in S204 (S216).

  If the existence is confirmed here (S216: Y), the process returns to the process shown in S204.

  If it is confirmed that there is not (S216: N), the screen generation unit 80 generates the estimation result screen 22 based on the estimation result data specified in the process shown in S215 (S217). In the process shown in S217, the screen generation unit 80 may draw, for example, a house change-corresponding image 24 according to the value of the estimation result data, surrounding a polygon centered on the position associated with the estimation result data. . Then, the screen generation unit 80 may generate the estimation result screen 22 in which the estimated target image 20b on which the house change corresponding image 24 is superimposed is arranged.

  Then, the display control unit 82 causes the display unit 16 to display the estimation result screen 22 generated in the process shown in S217 (S218), and the process shown in this process example is ended.

  The execution order of the processes shown in S201 to S218 is not limited to the above, and these processes may be executed in an execution order different from the order shown above.

  The present invention is not limited to the above-described embodiment.

  For example, in the original learning input data, both a learning building detection image showing the position and shape of the building shown in the first learning image and a learning building detection image showing the position and shape of the building shown in the second learning image May be included. Then, both of these two learning building detection images may be input to the machine learning model 30 as the input 48. In this case, in the estimation, not only the estimated chip image 28 (estimated chip image 28 c) extracted from the estimated building detection image 26 b but also the estimated chip image 28 extracted from the estimated building detection image 26 a is input as the machine learning model 30. Will be input to

  Further, for example, as a learning building detection image used for learning, an image indicating the position and shape of only the building shown at the center may be used. Also, for example, in estimation, instead of the estimated chip image 28c, an estimated chip image 28d illustrated in FIG. 11 may be used which shows the position and shape of only the building shown at the center. The estimated chip image 28d shown in FIG. 11 is generated, for example, by deleting polygons of buildings other than the building shown at the center from the estimated chip image 28c.

  Also, for example, in learning of the machine learning model 30, a learning building detection image may be input to the presence / absence feature extractor 32a instead of the learning image. In this case, instead of the estimated chip image 28 (estimated chip image 28a) extracted from the estimated target image 20a in estimation, the estimated chip image 28 extracted from the estimated building detection image 26a is used as the input 40 as the presence / absence feature extractor 32aa Will be input to Also, instead of the estimated chip image 28 (estimated chip image 28b) extracted from the estimated target image 20b, the estimated chip image 28 (estimated chip image 28c) extracted from the estimated building detection image 26b is used as the input 44 as a presence / absence feature extractor It will be input to 32ab.

  Also, for example, in learning of the machine learning model 30, a learning building detection image generated based on a learning image may be used instead of presence / absence data which is binary data included in the entire teacher data.

  Also, for example, in learning, a difference image indicating the difference between a pair of learning images may be used instead of the pair of learning images. In this case, in the estimation, instead of the estimated chip image 28a and the estimated chip image 28b, a difference image indicating the difference between the estimated chip image 28a and the estimated chip image 28b may be used. Here, for example, from the pixel value of the pixel of the coordinate value included in the second learning image as the pixel value of the pixel of the coordinate value included in the difference image, the pixel value of the pixel of the coordinate value included in the first learning image A value obtained by subtracting may be set.

  Also, for example, in learning, instead of the learning building detection image included in the learning input data, a learning building detection image generated based on the first learning image and a learning building detection image generated based on the second learning image A difference image may be used that indicates the difference between In this case, in the estimation, instead of the estimated chip image 28c, a difference image indicating the difference between the estimated chip image 28 extracted from the estimated building detection image 26a and the estimated chip image 28c extracted from the estimated building detection image 26b is used It may be done.

  For example, as shown in FIG. 12, the machine learning model 30 may not include the presence / absence feature extractor 32a. In this case, in learning, the output 52 of the change feature extractor 32b may be input not only to the change identifier 34 but also to the presence / absence identifier 32c (32ca and 32cb). Then, based on a comparison result (error) between the output 58 of the presence / absence identifier 32ca and the first presence / absence data by an error back propagation method or the like, the weight of the change feature extractor 32b and the presence / absence identifier 32ca Parameter values may be changed. Further, based on a comparison result (error) between the output 60 of the presence / absence identifier 32cb and the second presence / absence data by an error back propagation method or the like, weights of the change feature extractor 32b and the presence / absence identifier 32cb, etc. Parameter values may be changed.

  For example, as shown in FIG. 13, the machine learning model 30 may not include the presence / absence feature extractor 32a and the presence / absence identifier 32c. In this case, in learning and estimation, an image indicating whether or not a building is shown at a predetermined position (for example, the center) in the first learning image may be input as the input 84. Further, an image indicating whether a building is shown at a predetermined position (for example, the center) in the second learning image may be input as the input 86. The image input as the input 84 or the input 86 may be, for example, a monochrome image of the same size (for example, the number of vertical and horizontal pixels is L) as the feature map which is the output 52 of the variation feature extractor 32b. . Here, for example, a white image may be used when a building is shown, and a black image may be used as an image showing whether a building is shown if a building is not shown. Then, instead of the feature information output from the presence / absence feature extractor 32a, the image input as the input 84 or the input 86 is combined with the output 52 of the variation feature extractor 32b, and then the change discriminator 34 is input as the input 54. May be input. Here, if the feature information output as the output 52 is, for example, a feature map having 512 maps, the data input to the change discriminator 34 as the input 54 is a feature map having 514 maps. .

  For example, as shown in FIG. 14, in learning and estimation, a multi-channel image combined with an image input as the input 84 and the input 86 may be input to the change feature extractor 32 b. In this case, the output 52 of the variation feature extractor 32b may be input to the variation discriminator 34 as it is.

  For example, in learning, a learning building detection image may not be input to the change feature extractor 32b, and only the first learning image and the second learning image may be input to the change feature extractor 32b. Further, in this case, in the estimation, the estimated chip image 28 extracted from the estimated building detection image 26 is not input to the change feature extractor 32b, and only the estimated chip image 28 extracted from the estimated target images 20a and 20b is a change feature extractor It may be input to 32b.

  For example, as shown in FIG. 15, the machine learning model 30 may have only one presence / absence feature extractor 32a. Moreover, the machine learning model 30 may have only one presence / absence identifier 32c. In this case, in learning, the input to the presence / absence feature extractor 32a of the first learning image, the specification of the output according to the input, the input to the presence / absence feature extractor 32a of the second learning image, and the output according to the input The identification of is performed sequentially. The data obtained by combining the output when the first learning image is input to the presence / absence feature extractor 32a, the output when the second learning image is input to the presence / absence feature extractor 32a, and the output of the change feature extractor 32b are It may be input to the change discriminator 34. In this case, in the estimation, the input to the presence / absence feature extractor 32a of the estimated chip image 28a and the specification of the output according to the input, and the input to the presence / absence feature extractor 32a of the estimated chip image 28b and the specification of the output according to the input May be sequentially executed. Then, data obtained by combining the output when the estimated chip image 28a is input to the presence / absence feature extractor 32a, the output when the estimated chip image 28b is input to the presence / absence feature extractor 32a, and the output of the variation feature extractor 32b is It may be input to the change discriminator 34.

  Also, for example, a Dilated Convolution layer may be used instead of the pooling layer included in the machine learning model 30. This enables feature extraction without compression of the amount of information generated when the pooling layer is used. Also, instead of all the joint layers included in the machine learning model 30, a convolutional layer may be used. In this case, it may be possible to realize a machine learning model 30 which is independent of the reduction of calculation cost and the input image size.

  Further, the estimated target image 20b may be an image captured at a second time point before the first time point. In this case, it is necessary to make it possible to estimate that a building that is presumed to be a new building in the above example is a loss and a building that is assumed to be a loss is a new building.

  Further, instead of the estimated building detection images 26a and 26b, an image in which the position and the outline of the building are described may be used.

  Further, the above-described specific character strings and numerical values, and the specific character strings in the drawings are merely examples, and the present invention is not limited to these character strings and numerical values.

  DESCRIPTION OF SYMBOLS 10 house change estimation apparatus, 12 processors, 14 memory | storage parts, 16 display parts, 18 operation parts, 20a, 20b estimated target image, 22 estimated result screen, 24, 24a, 24b, 24c house change corresponding | compatible image 26, 26, 26a, 26b Estimated building detection image 28, 28a, 28b, 28c, 28d Estimated chip image, 30 machine learning model, 32 processing units, 32a, 32aa, 32ab presence feature extractor, 32b change feature extractor, 32c, 32ca, 32cb presence distinction , 34 change discriminators, 40, 44, 48, 50, 54, 84, 86 inputs, 42, 46, 52, 56, 58, 60 outputs, 70 learning data acquisition unit, 72 learning process execution unit, 74 estimation target Data acquisition unit, 76 estimation processing execution unit, 78 estimation result identification unit, 80 screen generation unit, 82 display control .

Claims (11)

Learning input data indicating a difference between learning images of a pair of lands photographed at different points in time, and whether or not a building is indicated at a predetermined position in the learning image for at least one of the learning images; Change identification in which learning data including a plurality of teacher data indicating house change between the predetermined position in the learning image and the position corresponding to the predetermined position in the other learning image is learned And the
Estimated input data indicating differences between estimated images of a pair of lands taken at different times and whether a building is shown at a predetermined position in at least one of the estimated images is input to the change discriminator Change discriminator input means,
Between the predetermined position in one of the estimated images and the position associated with the predetermined position in the other estimated image, based on the output when the estimated input data is input to the change discriminator House change estimation means for estimating house changes in
House movement estimation device characterized by including. A presence / absence feature extractor that outputs feature information indicating whether a building is shown at a predetermined position in the image according to an input of the image of the land;
Further including presence / absence feature information specifying means for specifying feature information which is an output when the pair of estimated images is input to the presence / absence feature extractor;
The change discriminator input means inputs the estimated input data including the feature information specified by the presence / absence feature information specifying means to the change discriminator.
The house change estimation device according to claim 1, characterized in that: The pair of lands is input according to an image of a pair of lands taken at different points in time and an image indicating whether a building is shown at a predetermined position in the image for at least one of the images. A change feature extractor that outputs feature information indicating differences in images of
A feature that is an output when the pair of estimated images and an image indicating whether or not a building is shown at a predetermined position in the estimated image of at least one of the estimated images is input to the change feature extractor A change feature information identification means for identifying information;
The change discriminator input means inputs the estimated input data including the feature information specified by the change feature information specifying means to the change discriminator.
The house change estimation device according to claim 1 or 2, characterized in that The variation feature extractor is responsive to input of an image of a pair of lands taken at different points in time and an image showing a shape of a building shown at a predetermined position in the image of at least one of the images. Output feature information indicating the difference between the pair of land images,
The change feature information specifying means may be an image showing the shape of a building shown at the predetermined position in the estimated image of the pair of estimated images and the estimated image of the at least one of the estimated images. Identify the feature information that is the output when input to
The house change estimation device according to claim 3, wherein A presence / absence feature extractor that outputs feature information indicating whether a building is shown at a predetermined position in the image according to an input of the image of the land;
Presence / absence feature information specifying means for specifying feature information which is an output when the pair of estimated images is input to the presence / absence feature extractor;
The pair of lands is input according to an image of a pair of lands taken at different points in time and an image indicating whether a building is shown at a predetermined position in the image for at least one of the images. A change feature extractor that outputs feature information indicating differences in images of
It is an output at the time of inputting into the change feature extractor an image indicating whether the building is indicated at a predetermined position in the estimated image of the pair of estimated images and at least one of the estimated images. Variation feature information specifying means for specifying feature information;
The information processing apparatus further includes combining means for combining the feature information specified by the presence / absence feature information specifying means with the feature information specified by the change feature information specifying means.
The change discriminator input means inputs the estimated input data including the feature information combined by the combining means to the change discriminator.
The house change estimation device according to claim 1, characterized in that: Screen generation means for generating a screen on which information representing house changes estimated by the house change estimation means is arranged;
And a display control unit configured to control to display a screen generated by the screen generation unit.
The house movement estimation apparatus according to any one of claims 1 to 5, characterized in that: Change identification of learning input data indicating differences between learning images of a pair of lands taken at different times and whether or not a building is shown at a predetermined position in at least one of the learning images. Input means for input to the
A house between the output when the learning input data is input to the change discriminator and the position corresponding to the predetermined position in the other learning image in the one learning image A comparison means for comparing teacher data indicating a transfer;
Parameter changing means for changing the parameter of the change discriminator based on the result of the comparison;
House change learning device characterized by including. Learning input data indicating a difference between learning images of a pair of lands photographed at different points in time, and whether or not a building is indicated at a predetermined position in the learning image for at least one of the learning images; Change identification in which learning data including a plurality of teacher data indicating house change between the predetermined position in the learning image and the position corresponding to the predetermined position in the other learning image is learned Said change discriminator showing estimated input data indicating the difference between estimated images of a pair of lands taken at different times and whether a building is shown at a predetermined position in at least one of said estimated images Step to
Between the predetermined position in one of the estimated images and the position associated with the predetermined position in the other estimated image, based on the output when the estimated input data is input to the change discriminator Estimating house changes in
The house movement estimation method characterized by including. Change identification of learning input data indicating differences between learning images of a pair of lands taken at different times and whether or not a building is shown at a predetermined position in at least one of the learning images. Input to the
A house between the output when the learning input data is input to the change discriminator and the position corresponding to the predetermined position in the other learning image in the one learning image Comparing with teacher data indicating a transfer;
Changing parameters of the change discriminator based on a result of the comparison;
A method of generating a parameter of a discriminator, comprising: Learning input data indicating a difference between learning images of a pair of lands photographed at different points in time, and whether or not a building is indicated at a predetermined position in the learning image for at least one of the learning images; Change identification in which learning data including a plurality of teacher data indicating house change between the predetermined position in the learning image and the position corresponding to the predetermined position in the other learning image is learned Said change discriminator showing estimated input data indicating the difference between estimated images of a pair of lands taken at different times and whether a building is shown at a predetermined position in at least one of said estimated images Steps to enter in
Between the predetermined position in one of the estimated images and the position associated with the predetermined position in the other estimated image, based on the output when the estimated input data is input to the change discriminator Procedure to estimate house change in
A program characterized by causing a computer to execute. Change identification of learning input data indicating differences between learning images of a pair of lands taken at different times and whether or not a building is shown at a predetermined position in at least one of the learning images. Procedure to input to the
A house between the output when the learning input data is input to the change discriminator and the position corresponding to the predetermined position in the other learning image in the one learning image Procedure to compare with the teacher data showing the transfer,
Changing the parameter of the change discriminator based on the result of the comparison;
A program characterized by causing a computer to execute.

Images