Generator construction method, generator construction apparatus, generator construction program, generator, estimation method, estimation apparatus, and estimation program

Abstract

To estimate a two-dimensional distribution that is not limited to a horizontal direction when estimating a distribution of another physical amount different from one physical amount by means of image conversion from an image showing a distribution of the one physical amount about an earth structure.SOLUTION: A generator construction method comprises: the step (S22) of converting a first physical amount about an earth structure (S) into a discrete value according to a property, generating an image before conversion (I1) showing a distribution of the discrete values and inputting an image for input (I2) generated by geometrically converting the image before conversion to a generator (g, G) to output a generated image (I3) indicating a distribution of a second physical amount; the step (S23) of analyzing the image before conversion to obtain a correct answer image (I4); the step (S24) of inputting the generated image and the correct answer image to a discriminator to output a discrimination result of authenticity of the generated image; and the step (S26) of causing the generator to perform learning when the discrimination result is false.SELECTED DRAWING: Figure 8

Description

[Technical Field]

[0001] The present invention relates to a generator construction method, a generator construction apparatus, a generator construction program, a generator, an estimation method, an estimation apparatus, and an estimation program. [Background technology]

[0002] For example, as described in Non-Patent Document 1 below, a method has been proposed in which a generator is constructed using a generative adversarial network to generate an image showing the distribution of pressure head within a soil structure when an image showing the distribution of the permeability coefficient within the soil structure is input, and the distribution of pressure head within the soil structure is estimated using this generator. [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] Alexander Y. Sun,Discovering state-parameter mappings in subsurface models using generative adversarial networks,Geophysical Research Letters,October 2018 [Overview of the project] [Problems that the invention aims to solve]

[0004] However, conventional techniques, such as those described in Non-Patent Document 1 above, could only estimate the pressure head distribution in the horizontal direction. The present invention aims to enable the estimation of a two-dimensional distribution, not limited to the horizontal direction, when estimating the distribution of another physical quantity different from the first physical quantity related to an earth structure by image transformation from an image showing the distribution of that first physical quantity. [Means for solving the problem]

[0005] A method for constructing a generator according to one aspect of the present invention includes the steps of: converting a first physical quantity relating to a soil structure to be analyzed into discrete values ​​according to its characteristics; generating a pre-conversion image having the same shape as the cross-section of the soil structure showing the distribution of the discrete values; inputting a rectangular input image generated by geometrically transforming the pre-conversion image into the generator and outputting a generated image showing the distribution of a second physical quantity; analyzing the pre-conversion image to obtain a ground truth image showing the distribution of the second physical quantity; inputting the generated image and the ground truth image into a classifier and outputting a result of identifying whether the generated image is genuine or not; and training the generator if the identification result is found to be false.

[0006] A generator construction apparatus according to another aspect of the present invention comprises: a generated image output unit that converts a first physical quantity relating to a soil structure to be analyzed into discrete values ​​according to its characteristics, generates a pre-conversion image having the same shape as the cross-section of the soil structure showing the distribution of the discrete values, inputs a rectangular input image generated by geometrically transforming the pre-conversion image into the generator and outputs a generated image showing the distribution of a second physical quantity; an analysis unit that analyzes the pre-conversion image and obtains a correct image showing the distribution of the second physical quantity; an identification result output unit that inputs the generated image and the correct image into a classifier and outputs an identification result of whether the generated image is genuine or not; and a learning unit that trains at least the generator if the identification result is false.

[0007] Another aspect of the present invention is a generator constructed using a generative adversarial network, which converts a first physical quantity relating to a soil structure to be analyzed into discrete values ​​according to its characteristics, generates a pre-conversion image having the same shape as the cross-section of the soil structure showing the distribution of the discrete values, inputs a rectangular input image generated by geometrically transforming the pre-conversion image into the generator, outputs a generated image showing the distribution of a second physical quantity, inputs a ground truth image showing the distribution of the second physical quantity obtained by analyzing the generated image and the pre-conversion image into a classifier, outputs a genuine / fake identification result of the generated image, and trains the generator if the identification result is fake, thereby generating an estimated image showing the distribution of the second physical quantity when an input image showing the distribution of the first physical quantity is input.

[0008] Another aspect of the present invention relates to an estimation method which involves converting a first physical quantity relating to a soil structure to be analyzed into discrete values ​​according to its characteristics, and generating a rectangular input image by geometrically transforming a pre-transformation image of the cross-sectional shape of the soil structure showing the distribution of the discrete values. This input image is then input to a generator constructed using a generative adversarial network, which generates an estimation image showing the distribution of a second physical quantity when the input image showing the distribution of the first physical quantity is input, thereby generating the estimation image.

[0009] An estimation device according to another aspect of the present invention includes an estimation unit that generates an estimated image by converting a first physical quantity relating to a soil structure to be analyzed into discrete values ​​according to its characteristics, and then geometrically transforming a pre-transformation image of the cross-sectional shape of the soil structure showing the distribution of the discrete values ​​to generate a rectangular input image, which is constructed using a generative adversarial network and, upon receiving the input image showing the distribution of the first physical quantity, generates an estimated image showing the distribution of a second physical quantity.

[0010] Another embodiment of the present invention provides a generator construction device comprising: a first conversion unit that converts physical quantities relating to an earth structure to be analyzed into discrete values ​​corresponding to the characteristics of the physical quantities; a generation unit that generates a pre-conversion image having the same shape as the cross-section of the earth structure, showing the distribution of the discrete values ​​within the earth structure; and a second conversion unit that geometrically converts the pre-conversion image into a rectangular input image.

[0011] Each aspect of the present invention may be implemented by a computer, in which case a generator construction program or estimation program that enables the computer to implement the generator construction device or estimation device by operating the computer as each part (software element) of the generator construction device or estimation device, and a computer-readable recording medium on which they are recorded, also fall within the scope of the present invention. [Effects of the Invention]

[0012] According to one aspect of the present invention, when estimating the distribution of another physical quantity different from the one physical quantity from an image showing the distribution of the one physical quantity related to an earth structure by image conversion, a two-dimensional distribution not limited to the horizontal direction can be estimated.

Brief Description of Drawings

[0013] [Figure 1] It is a diagram showing an example of an image to be processed by an image processing method according to an embodiment of the first invention. [Figure 2] It is a flowchart showing the flow of the method. [Figure 3] It is a diagram showing an example of the content performed in the method. [Figure 4] It is a diagram showing an example of the content performed in the method. [Figure 5] It is a diagram showing the difference between the content performed in the method and the content performed in a method different from the method. [Figure 6] It is a diagram showing an example of the content performed in the method. [Figure 7] It is a block diagram showing a schematic configuration of an image processing apparatus according to an embodiment of the second invention. [Figure 8] It is a flowchart showing the flow of a generator construction method according to an embodiment of the third invention. [Figure 9] It is a diagram showing an example of the content performed in the method. [Figure 10] It is a diagram showing the operation of a generator constructed by the method. [Figure 11] It is a block diagram showing a schematic configuration of a generator construction apparatus according to an embodiment of the fourth invention. [Figure 12] It is a flowchart showing the flow of an estimation method according to an embodiment of the fifth invention. [Figure 13] It is a diagram showing an example of the content performed in the method. [Figure 14] It is a block diagram showing a schematic configuration of an estimation apparatus according to an embodiment of the sixth invention. [Figure 15]This is a histogram showing the distribution of the evaluation values ​​(SSIM) of each pixel in the output image obtained using each method or apparatus according to the embodiments of the first to sixth inventions. [Figure 16] This is a histogram showing the distribution of evaluation values ​​(errors) for each pixel of the output image obtained using each method or apparatus according to the embodiment. [Modes for carrying out the invention]

[0014] <Embodiment of the First Invention> First, an embodiment of the first invention (image processing method) will be described.

[0015] [Image to be processed] The image processing method according to this embodiment processes images showing the distribution of physical quantities related to the soil structure S being analyzed. The soil structure S being analyzed includes, for example, all structures constructed of soil, including embankments (dam bodies). The cross-sectional shape of the soil structure S is generally a shape other than a rectangle. If the soil structure S is a dam body, its cross-sectional shape is trapezoidal. Furthermore, the soil structure S has a unique distribution of permeability coefficients depending on the properties of the soil used. That is, the soil structure S may have a uniform permeability coefficient, or it may have two or more parts with different permeability coefficients. The following explanation will use an image showing the cross-section of a dam body S having a first part P1, a second part P2, and a third part P3, as shown in Figure 1, as an example. The second part P2 is the central part of the dam body and has the lowest permeability coefficient. The first part P1 is the upstream side of the dam body and has a higher permeability coefficient than the second part P2. The third part P3 is the downstream side of the dam body and has a higher permeability coefficient than the first part P1. The above permeability coefficient distribution is just one example, and the permeability coefficient distribution of the soil structure S processed by this image processing method is not limited to that shown above.

[0016] [Image Processing Flowchart] As shown in Figure 2, the image processing method according to this embodiment includes a first acquisition step S11, a first conversion step S12, a generation step S13, a second conversion step S14, and a first output step S15.

[0017] (First acquisition step) In the initial first acquisition step S11, physical quantities related to the soil structure S to be analyzed are acquired. These physical quantities can be acquired, for example, using the first acquisition unit 131 (see Figure 7) of the image processing device 1 (second invention) described later.

[0018] (First conversion step) After acquiring the physical quantity, the process moves to the first conversion step S12. In the first conversion step S12, the physical quantity acquired in the first acquisition step S11 is converted into a discrete value according to the characteristics of that physical quantity. The physical quantity in this embodiment is the permeability coefficient or the pressure head.

[0019] When the physical quantity is the permeability coefficient, in the first conversion step S12, the permeability coefficient is toned down such that the brightness or saturation changes by one step for every tenfold increase, as shown in the upper part of Figure 3. -1 ~10 -8 This is a physical quantity that can take values ​​on a wide order of magnitude, in the order of cm / s. Therefore, if the permeability coefficient is simply toned down, the difference in color tone becomes too small, making it difficult to distinguish differences in the values ​​of the permeability coefficient. However, by using the above configuration, even small differences in value can result in relatively large differences in color tone. As a result, it is possible to obtain an input image I2 (details described later) that can generate an estimated image I6 (details described later) that accurately reflects the distribution of the permeability coefficient.

[0020] On the other hand, when the physical quantity is pressure head, in the first conversion step S12, as shown in the lower part of Figure 3, for example, the saturation of one hue increases as the value of pressure head increases in the positive direction, and the saturation of other hues increases as the value of pressure head increases in the negative direction. In this way, the region corresponding to the area that is neither saturated nor unsaturated (where the pressure head is 0) (the line L indicating the infiltration surface) is shown in white, which has the lowest saturation. Therefore, an input image I2 can be obtained that can generate an estimated image I6 in which the position of the line L indicating the infiltration surface within the soil structure S is accurately reflected.

[0021] The above discrete value conversions can be performed, for example, using the first conversion unit 132 (see Figure 7) of the image processing device 1, which will be described later.

[0022] (Generation step) After converting the physical quantities into discrete values, the process moves to generation step S13, as shown in Figure 2. In generation step S13, a pre-conversion image I1 is generated, which has the same shape as the cross-section of the soil structure S and shows the distribution of discrete values ​​within the soil structure S. The pre-conversion image I1 is an image that has the same shape as the cross-section of the soil structure S and shows the distribution of discrete values ​​within the soil structure S.

[0023] The generation step S13 according to this embodiment includes a partitioning step and an extraction step. In the first partitioning step, the pre-conversion image I1 is partitioned into a plurality of minute regions. In the partitioning step according to this embodiment, the pre-conversion image I1 is partitioned such that the upper minute regions of the pre-conversion image I1 are finer than the lower minute regions. Specifically, the pre-conversion image I1 is partitioned by gradually changing the size of the minute regions so that they become finer as you move towards the top of the pre-conversion image I1. After the pre-conversion image I1 has been partitioned into a plurality of minute regions, the process moves to the extraction step. In the extraction step, one pixel p is extracted from each minute region. In this way, a group of pixels is obtained in which the density increases towards the top of the pre-conversion image I1, for example, as shown on the left side of Figure 4.

[0024] The generation of the pre-conversion image I1 described above can be performed, for example, using the generation unit 133 (see Figure 7) of the image processing device 1, which will be described later.

[0025] (Second transformation step) After generating the pre-conversion image I1, the process moves to the second conversion step S14, as shown in Figure 2. In the second conversion step S14, the pre-conversion image I1 is geometrically transformed into a rectangular input image I2. Specifically, multiple pixels p arranged within the pre-conversion image I1 (trapezoid), as shown on the left side of Figure 4, are moved so that they are distributed throughout the entire rectangle, as shown on the right side of Figure 4. The distribution may be perfectly uniform or may have some degree of unevenness. Then, in the second conversion step S14 according to this embodiment, after moving the pixels p, interpolation is performed between the pixels p. As described above, the arrangement of the pixels p extracted in the extraction step becomes denser towards the top of the pre-conversion image I1. Therefore, pixels p at the top of the pre-conversion image I1 are moved more significantly. This makes it possible to obtain an input image I2 in which the entire image is uniformly interpolated. In particular, if the spacing between pixels p becomes too wide, it becomes impossible to determine whether the line L representing the invasive surface between pixels p is convex upwards or downwards, making it impossible to accurately interpolate the curved portion of the line L representing the invasive surface (resulting in interpolation being linear, as shown on the left side of Figure 5). However, by doing as described above, the spacing between pixels p does not become too wide. Therefore, as shown on the right side of Figure 5, the curved portion of the line L representing the invasive surface can be accurately interpolated. Note that in the second transformation step S14, the pre-transformation image I1 (trapezoid) may be geometrically transformed into a rectangular input image I2 that is smaller than the pre-transformation image I1 (in order to aggregate the pixels).

[0026] Furthermore, if the soil structure S has two or more parts with different permeability coefficients, as shown in Figure 1, and the physical quantity shown in the pre-conversion image I1 is the permeability coefficient, then in the second conversion step S14, the pre-conversion image I1 may be geometrically transformed so that the area of ​​the region showing the part with a relatively high permeability coefficient in the input image I2 is larger than the area of ​​the region showing the part with a relatively high permeability coefficient in the pre-conversion image I1. For example, as shown on the left side of Figure 6, if the area of ​​region R1 showing the first part P1 in the input image I2 is smaller than the areas of regions R2 and R3 showing the second part P2 and third part P3, which have lower permeability coefficients than the first part P1, then, as shown on the right side of Figure 6, the regions R1 to R3 may be geometrically transformed so that they become congruent. It is known that the pressure head within the soil structure S changes abruptly at the boundary of the region with a high permeability coefficient. For this reason, if the region with a relatively high permeability coefficient in the input image I2 is narrow, the generator G, which will be described later, may not be able to capture the change in pressure head. However, by doing as described above, the generator G can more easily detect changes in pressure head. Furthermore, the extent to which the area representing the region with a relatively high permeability coefficient is increased during the geometric transformation can be appropriately adjusted according to the value of the permeability coefficient.

[0027] The above geometric transformations can be performed, for example, using the second transformation unit 134 (see Figure 7) of the image processing device 1, which will be described later. Note that the second transformation step S14 may be performed before the first transformation step S12.

[0028] (First output step) After the geometric transformation, the process moves to the first output step S15. In the first output step S15, the input image I2 is output (in a format usable for learning and estimation). The output can be performed, for example, using the first output unit 12 and the first output control unit 135 (see Figure 7) of the image processing device 1, which will be described later. Note that if it is not necessary to output the input image I2 (for example, if the image processing method has a generated image output step S22, an analysis step S23, or an estimation step S32 of the generator construction method, which will be described later, and the input image I2 generated in the second transformation step S14 is used as is in the generated image output step S22, analysis step S23, or estimation step S32), then this first output step S15 does not need to be performed.

[0029] [Effects of image processing methods] As described above, the image processing method according to this embodiment allows for the conversion of physical quantities to discrete values ​​and the rectangularization of images by geometric transformation. In particular, the conversion to discrete values ​​allows for the color tone of the image to be adjusted to capture the characteristics of the image in the construction of the generator described later. This makes it possible to obtain an input image I2 used when estimating the distribution of other physical quantities different from one physical quantity related to a soil structure S by image transformation (a method of inputting an input image I2 to a generator G and generating an estimated image I6). Using this input image I2, it becomes possible to estimate a two-dimensional distribution that includes the vertical direction as well as the horizontal direction. Furthermore, if the input image I2 shows a hydraulic conductivity distribution, it becomes possible to estimate a pressure head distribution that includes not only the saturated region but also the unsaturated region.

[0030] <Second Invention Embodiment> Next, an embodiment of the second invention (image processing device 1) will be described.

[0031] [Image Processing Unit Configuration] As shown in Figure 7, the image processing device 1 comprises a first input unit 11, a first output unit 12, and a first control unit 13.

[0032] [First Input Section] The first input unit 11 consists of at least one of the following: a communication module that receives data and signals from other devices (for example, a measuring device that measures physical quantities, a storage device that stores physical quantities, etc.), terminals that connect to other devices, and a drive that reads information from a recording medium.

[0033] [First output section] The first output unit 12 consists of at least one of the following: a communication module for transmitting data and signals to other devices (for example, the generator construction device 2 and estimation device 3 described later), terminals for connecting to other devices, and a drive for writing information to a recording medium.

[0034] [First Control Unit] The first control unit 13 according to this embodiment includes a first acquisition unit 131, a first conversion unit 132, a generation unit 133, a second conversion unit 134, and a first output control unit 135.

[0035] (First Acquisition Department) The first acquisition unit 131 controls the first input unit 11 to acquire physical quantities related to the soil structure S to be analyzed. As a result, if the first input unit 11 is a communication module or terminal, the first input unit 11 receives physical quantity data from other devices. If the first input unit 11 is a drive, the first input unit 11 reads physical quantity data from a recording medium.

[0036] (First conversion section) The first conversion unit 132 converts the physical quantity acquired by the first acquisition unit 131 into discrete values ​​corresponding to the characteristics of the physical quantity. Specifically, the first conversion unit 132 performs the same actions as described in the first conversion step S12 (see Figure 2) in the image processing method (first invention).

[0037] (Generation part) The generation unit 133 generates a pre-conversion image I1 that shows the distribution of discrete values ​​within the soil structure S and has the same shape as the cross-section of the soil structure S. The generation unit 133 according to this embodiment includes a partitioning unit 133a and an extraction unit 133b. The partitioning unit 133a partitions the pre-conversion image I1 into a plurality of minute regions. Specifically, the partitioning unit 133a performs the same actions as described in the partitioning step of the image processing method described above. The extraction unit 133b extracts one pixel p from each minute region. Specifically, the extraction unit 133b performs the same actions as described in the extraction step of the image processing method described above.

[0038] (Second conversion section) The second transformation unit 134 geometrically transforms the pre-transformation image I1 into a rectangular input image I2. Specifically, the second transformation unit 134 performs the same actions as described in the second transformation step S14 (see Figure 2) of the image processing method described above.

[0039] (First Output Control Unit) The first output control unit 135 controls the first output unit 12 to output the input image I2. In addition, the first output control unit 135 in this embodiment controls the first output unit 12 to output the pre-conversion image I1 as well. As a result, if the first output unit 12 is a communication module or terminal, the first output unit 12 transmits the data of the pre-conversion image I1 and the input image I2 to another device. If the first output unit 12 is a drive, the first output unit 12 writes the data of the pre-conversion image I1 and the input image I2 to a recording medium. Note that if it is not necessary to output the pre-conversion image I1 and the input image I2 (for example, if the image processing device 1 is equipped with the generated image output unit 242, analysis unit 243, or estimation unit 342 of the generator construction device 2 described later, and the generated image output unit 242, analysis unit 243, or estimation unit 342 uses the input image I2 generated by the second conversion unit 134 as is), the first control unit 13 does not need to be equipped with this first output control unit 135.

[0040] [Effects and Effects of Image Processing Devices] According to the image processing apparatus 1 of this embodiment described above, similar to the image processing method described above, by converting physical quantities to discrete values ​​and rectangularizing the image by geometric transformation, and in particular by converting to discrete values, the color tone of the image is set to a color tone that can capture the characteristics of the image in the construction of the generator described later, thereby obtaining an input image I2 that can be used when estimating the distribution of a physical quantity different from one physical quantity by image transformation from an image showing the distribution of one physical quantity related to a soil structure S. Using this input image I2, it is possible to estimate a two-dimensional distribution that includes the vertical direction as well as the horizontal direction. Furthermore, if the input image I2 shows the hydraulic conductivity distribution, it is possible to estimate the pressure head distribution that includes not only the saturated region but also the unsaturated region.

[0041] <Embodiment of the Third Invention> Next, an embodiment of the third invention (generator construction method and generator G) will be described.

[0042] [Generator Construction Procedure] The generator construction method according to this embodiment includes, for example, a second acquisition step S21, a generated image output step S22, an analysis step S23, an identification result output step S24, a judgment step S25, a learning step S26, and a second output step S27, as shown in Figure 8.

[0043] (Second acquisition step) In the initial second acquisition step S21, the input image I2 is acquired. In addition, the pre-conversion image I1 is also acquired in the second acquisition step S21 according to this embodiment. The acquisition of the input image I2 and the pre-conversion image I1 can be performed, for example, using the second acquisition unit 241 (see Figure 11) of the generator construction device 2 (fourth invention) described later. The input image I2 and the pre-conversion image I1 may be generated by the image processing device 1, or they may be generated by a device other than the image processing device 1. Furthermore, if it is not necessary to acquire the input image I2 and the pre-conversion image I1 (for example, if the generator construction method has a generation step S13 and a second conversion step S14 of the image processing method, and the pre-conversion image I1 generated in the generation step S13 and the input image I2 generated in the second conversion step S14 are used directly in the generated image output step S22 or analysis step S23 described later), then this second acquisition step S21 may not be performed.

[0044] (Generated image output step) After acquiring the input image I2, the process moves to the generated image output step S22. In the generated image output step S22, as shown in Figure 9, the input image I2 acquired in the second acquisition step S21 is input to a generator that is either untrained or in the process of being trained (hereinafter referred to as the untrained generator g) or a trained generator G, and a generated image I3 showing the distribution of a second physical quantity is output. The first physical quantity is either the hydraulic conductivity or the pressure head, and the second physical quantity is the other. If the first physical quantity is the hydraulic conductivity, a generated image I3 showing the distribution of the pressure head can be obtained. On the other hand, if the first physical quantity is the pressure head, a generated image I3 showing the distribution of the hydraulic conductivity can be obtained. The output of this generated image I3 can be performed, for example, using the generated image output unit 242 (see Figure 11) of the generator construction device 2 described later.

[0045] (Analysis step) After outputting the generated image I3, the process moves to the analysis step S23. In the analysis step S23, the pre-transformation image I1 is analyzed to obtain the ground truth image I5, which shows the distribution of the second physical quantity. The ground truth image I5 is obtained by geometrically transforming the pre-transformation analysis image I4 (see Figure 9), which was obtained by analyzing the pre-transformation image I1, so that it becomes a rectangle. This analysis of the pre-transformation image I1 can be performed, for example, using the analysis unit 243 (see Figure 11) of the generator construction device 2, which will be described later. Alternatively, the analysis of the pre-transformation image I1 can also be performed using conventional methods. Note that the analysis step S23 may be performed before the generated image output step S22, or it may be performed in parallel with the generated image output step S22.

[0046] (Identification result output step) After obtaining the generated image I3 and the ground truth image I5, the process moves to the identification result output step S24. In the identification result output step S24, the generated image I3 and the ground truth image I5 are input to the classifier D, and the identification result of whether the generated image I3 is genuine or not is output. This identification result output can be performed, for example, using the identification result output unit 244 (see Figure 11) of the generator construction device 2, which will be described later. In addition, in the identification result output step S24, depending on the structure of the untrained generator g, the pair of generated image I3 and input image I2, and the pair of ground truth image I5 and input image I2 may be input to the classifier D.

[0047] (Decision-making step) After outputting the identification result, the process moves to the judgment step S25. In the judgment step S25, the identification result of whether the generated image I3 is genuine or fake is determined. This determination of the identification result can be performed, for example, using the determination unit 245 (see Figure 11) of the generator construction device 2, which will be described later. In addition, in the judgment step S25, other items besides genuine or fake (such as whether the untrained generator g has been trained a predetermined number of times) may also be determined. If it is not necessary to determine the identification result (for example, if the identification result is output only in the identification result output step S24 if it is either "genuine" or "fake"), then this judgment step S25 does not need to be performed.

[0048] (Learning Steps) If the identification result in the judgment step S25 is determined to be "fake" (step S25: NO), the process proceeds to the learning step S26 as shown in Figure 8. In the learning step S26, at least the unlearned generator g is trained (the unlearned generator g and the unlearned discriminator d may be trained separately). The unlearned generator g trained in the learning step S26 according to this embodiment, and the generator G constructed by training, are U-Net. In this way, for example, as shown in Figure 10, at least one layer of the encoder of the unlearned generator g or the generator G is skip-connected to the corresponding layer of the decoder. Then, as the encoder performs max pooling, the position information of the feature quantities extracted by either layer is transmitted to the corresponding layer of the decoder via the skip connection. As a result, a generated image I3 with local features preserved can be obtained. The training of this unlearned generator g can be performed, for example, using the learning unit 246 of the generator construction device 2 described later (see Figure 11).

[0049] After performing the learning step S26, as shown in Figure 8, the process returns to step S22 using the trained untrained generator g or generator G. That is, the generated image output step S22, the identification result output step S24, the decision step S25, and the learning step S26 are repeated until the identification result is determined to be "true" in the judgment step S25. With each repetition of these steps, the estimation accuracy of the untrained generator g improves.

[0050] (Second output step) If the identification result is determined to be "true" in the judgment step S25 (step S25: YES), the process proceeds to the second output step S27. In the second output step S27, the generator G is output. This output of the generator G can be performed, for example, using the second output unit 22 and the second output control unit 247 (see Figure 11) of the image processing device 1, which will be described later. Note that if it is not necessary to output the generator G (for example, if the generator construction method has an estimation step S32 of the estimation method described later, and the generator G constructed in the learning step S26 is used directly in the estimation step S32), this second output step S27 does not need to be performed.

[0051] [Generator construction methods and other information] Furthermore, if the first physical quantity is pressure head and the second physical quantity is the permeability coefficient, the generator construction method may further include a step of obtaining the permeability coefficient of at least a portion of the soil structure S as prior information before performing the learning step S26. Then, in the learning step S26, prior information may be provided to the unlearned generator g when training it. It is known that when attempting to obtain a generated image I3 showing the distribution of permeability coefficients from an input image I2 showing the distribution of pressure head, it is more difficult to obtain a highly accurate generated image I3 compared to obtaining a generated image I3 showing the distribution of pressure heads from an input image I2 showing the distribution of permeability coefficients. However, by providing prior information to the unlearned generator g when constructing the generator G, the accuracy of the generated image I3 (estimated image I6) can be improved.

[0052] [Generator] The generator G (trained model) constructed by the generator construction method according to this embodiment is constructed using a generative adversarial network. The generator G is constructed by inputting the input image I2 into an untrained generator g, outputting the generated image I3, inputting the generated image I3 and the ground truth image I5 into a classifier D, outputting the result of the authenticity classification of the generated image I3, and training at least the untrained generator g if the classification result is false. When the generator G constructed in this way receives the input image I2, it generates an estimated image I6. The estimated image I6 refers to the generated image I3 that is generated by the trained generator G and not identified as false by the classifier D.

[0053] [Effects of the generator construction method] According to the generator construction method of this embodiment described above, by using an input image I2 in which physical quantities are converted to discrete values ​​and the shape is geometrically transformed into a rectangle, and repeating the generated image output step S22, the identification result output step S24, and the learning step S26 (by a generative adversarial network), the accuracy of the generator G is improved, and an estimated image I6 showing a two-dimensional distribution including the vertical direction, not just the horizontal direction, can be obtained. Furthermore, if the input image I2 shows a hydraulic conductivity distribution, an estimated image I6 showing a pressure head distribution including not only the saturated region but also the unsaturated region can be obtained.

[0054] <Fourth Invention Embodiment> Next, an embodiment of the fourth invention (generator construction device 2) will be described.

[0055] [Configuration of the generator construction device] As shown in Figure 11, the generator construction device 2 according to this embodiment includes a second input unit 21, a second output unit 22, a storage unit 23, and a second control unit 24.

[0056] [Second Input Section] The second input unit 21 consists of at least one of the following: a communication module that receives data and signals from other devices (for example, the image processing device 1), terminals that connect to other devices, and a drive that reads information from a recording medium.

[0057] [Second Output Section] The second output unit 22 consists of at least one of the following: a communication module for transmitting data and signals to other devices (for example, the estimation device 3 described later), terminals for connecting to other devices, and a drive for writing information to a recording medium.

[0058] [Storage section] The memory unit 23 stores unlearned generators g. The memory unit 23 is also configured to store learned generators G.

[0059] [Second Control Unit] The second control unit 24 includes a second acquisition unit 241, a generated image output unit 242, an analysis unit 243, an identification result output unit 244, a judgment unit 245, a learning unit 246, and a second output control unit 247.

[0060] (Second Acquisition Department) The second acquisition unit 241 controls the second input unit 21 to acquire the input image I2. If the second input unit 21 is a communication module or terminal, the second input unit 21 receives the input image I2 data from another device. If the second input unit 21 is a drive, the second input unit 21 reads the input image I2 data from the recording medium. Note that if there is no need to acquire the input image I2 (for example, if the generator construction device 2 is equipped with the second conversion unit 134 of the image processing device 1, and the input image I2 generated by the second conversion unit 134 is used directly by the generated image output unit 242 or analysis unit 243 described later), the second control unit 24 does not need to include this second acquisition unit 241.

[0061] (Generated image output section) The generated image output unit 242 inputs the input image I2 acquired by the second acquisition unit 241 to the untrained generator g or generator G, and outputs the generated image I3. Specifically, the generated image output unit 242 performs the same actions as described in the generated image output step S22 (see Figure 8) in the generator construction method (third invention).

[0062] (Analysis Department) The analysis unit 243 analyzes the pre-conversion image I1 and obtains the ground truth image I5. Specifically, the analysis unit 243 performs the same actions as described in analysis step S23 (see Figure 8) in the generator construction method described above.

[0063] (Identification result output unit) The identification result output unit 244 inputs the generated image I3 and the ground truth image I5 to the classifier D and outputs the identification result of whether the generated image I3 is genuine or not. Specifically, the identification result output unit 244 performs the same actions as described in the identification result output step S24 (see Figure 8) in the above-described generator construction method.

[0064] (Judgment Department) The determination unit 245 determines the authenticity of the generated image I3. Specifically, the identification result output unit 244 performs the same actions as described in the determination step S25 (see Figure 8) in the generator construction method described above. Note that if it is not necessary to determine the identification result (for example, if the identification result output unit 244 outputs the identification result only if it is either "true" or "fake"), the second control unit 24 does not need to include this determination unit 245.

[0065] (Learning Department) If the determination unit 245 determines that the identification result is false, the learning unit 246 trains at least the untrained generator g. Specifically, the learning unit 246 performs the same actions as described in the learning step S26 (see Figure 8) of the generator construction method.

[0066] (Second Output Control Unit) The second output control unit 247 controls the second output unit 22 to output the generator G. As a result, if the second output unit 22 is a communication module or terminal, the second output unit 22 transmits the data of the generator G to another device. If the second output unit 22 is a drive, the second output unit 22 writes the data of the generator G to a recording medium. Note that if it is not necessary to output the generator G (for example, if the generator construction device 2 is equipped with an estimation unit 342 of the estimation device 3 described later, and the estimation unit 342 uses the generator G constructed by the learning unit 246 as is), the second control unit 24 does not need to be equipped with this second output control unit 247.

[0067] [Generator construction equipment and other devices] Furthermore, if the first physical quantity is the pressure head and the second physical quantity is the permeability coefficient, the generator construction device 2 may further include a pre-information acquisition unit that obtains the permeability coefficient of at least a portion of the soil structure S as prior information. The learning unit may also be configured to provide the pre-information to the unlearned generator g when training the unlearned generator g.

[0068] [Effects and Effects of Generator Construction Device] According to the generator construction apparatus 2 of this embodiment described above, by using an input image I2 in which physical quantities are converted to discrete values ​​and the shape is geometrically transformed into a rectangle, the generated image output unit 242, the identification result output unit 244, the judgment unit 245, and the learning unit 246 repeat processing (by a generative adversarial network), similar to the generator construction method described above, the accuracy of the generator G is improved, and an estimated image I6 showing a two-dimensional distribution including the vertical direction, not just the horizontal direction, can be obtained. Furthermore, if the input image I2 shows a hydraulic conductivity distribution, an estimated image I6 showing a pressure head distribution including not only the saturated region but also the unsaturated region can be obtained.

[0069] <Embodiment of the Fifth Invention> Next, an embodiment of the fifth invention (estimation method) will be described.

[0070] [Estimation Method Flow] The estimation method according to this embodiment includes, for example, a third acquisition step S31, an estimation step S32, a third conversion step S33, and a third output step S34, as shown in Figure 12.

[0071] (Third acquisition step) In the initial third acquisition step S31, the input image I2 is acquired. This input image I2 can be acquired, for example, using the third acquisition unit 341 (see Figure 14) of the estimation device 3 (sixth invention) described later. The input image I2 may be generated by the image processing device 1, or it may be generated by a device other than the image processing device 1. In addition, the generator G may be acquired in the third acquisition step S31. Furthermore, if it is not necessary to acquire the input image I2 (for example, if the estimation method has a second conversion step S14 of the image processing method, and the input image I2 generated in the second conversion step S14 is used as is in the estimation step S32 described later), this second acquisition step S21 may be omitted.

[0072] (Estimated step) After acquiring the input image I2, the process moves to the estimation step S32. In the estimation step S32, the input image I2 acquired in the third acquisition step S31 is input to the generator G to generate the estimated image I6. This estimated image I6 can be generated, for example, using the estimation unit 342 (see Figure 14) of the estimation device 3, which will be described later.

[0073] (Third transformation step) After generating the estimated image I6, the process moves to the third transformation step S33. In the third transformation step S33, as shown in Figure 13, the rectangular estimated image I6 is geometrically transformed into an output image I7 of the cross-sectional shape (trapezoid) of the earth structure S. Multiple pixels p located within the rectangle are moved so that they are uniformly distributed within the shape (trapezoid) congruent to the pre-transformation image I1. In other words, this third transformation step S33 performs the reverse of what is done in the second transformation step in the first invention (image processing method) described above. This geometric transformation can be performed, for example, using the third transformation unit 343 (see Figure 14) of the estimation device 3, which will be described later.

[0074] (Third output step) After generating the estimated image I6, the process moves to the third output step S34, as shown in Figure 12. In the third output step S34, the output image I7 is output. This output image I7 can be output, for example, using the third output unit 32 and the third output control unit 344 (see Figure 14) of the estimation device 3, which will be described later.

[0075] [Effects of the estimation method] According to the estimation method of this embodiment described above, by using a generator G constructed using a generative adversarial network, it is possible to obtain an estimated image I6 (a generated image I3 that is not identified as fake by the classifier D) that shows a two-dimensional distribution including the vertical direction, not just the horizontal direction. Furthermore, if the input image I2 shows a hydraulic conductivity distribution, it is possible to obtain an estimated image I6 that shows a pressure head distribution including not only the saturated region but also the unsaturated region.

[0076] <Sixth Invention Embodiment> Next, an embodiment of the sixth invention (estimation device) will be described.

[0077] [Configuration of the estimation device] As shown in Figure 14, the estimation device 3 according to this embodiment includes a third input unit 31, a third output unit 32, a storage unit 33, and a third control unit 34.

[0078] [Third input section] The third input unit 31 consists of at least one of the following: a communication module that receives data and signals from other devices (for example, the image processing device 1, the generator construction device 2, etc.), a terminal that connects to other devices, and a drive that reads information from a recording medium.

[0079] [Third output section] The third output unit 32 consists of at least one of the following: a communication module for transmitting data and signals to other devices (such as a display device for displaying the output image I7, or a storage device capable of storing the data of the output image I7); terminals for connecting to other devices; a drive for writing information to a recording medium; and a display for displaying the image.

[0080] [Storage section] The memory unit 33 stores the generator G. The generator G may be constructed by the generator construction device 2, or it may be constructed by another device.

[0081] [Third Control Unit] The third control unit 34 comprises a third acquisition unit 341, an estimation unit 342, a third conversion unit 343, and a third output control unit 344.

[0082] (Third Acquisition Department) The third acquisition unit 341 controls the second input unit 21 to acquire the input image I2. If the second input unit 21 is a communication module or terminal, the second input unit 21 receives the input image I2 data from another device. If the second input unit 21 is a drive, the second input unit 21 reads the input image I2 data from the recording medium. The third acquisition unit 341 may also be configured to acquire the generator G. Furthermore, if there is no need to acquire the input image I2 (for example, if the estimation device 3 includes the second conversion unit 134 of the image processing device 1, and the estimation unit 342, described later, uses the input image I2 generated by the second conversion unit 134), the third control unit 34 does not need to include this third acquisition unit 341.

[0083] (Estimation Department) The estimation unit 342 generates an estimated image I6 by inputting the input image I2 acquired by the third acquisition unit 341 into the storage unit 33 or into the generator G acquired by the third acquisition unit 341. Specifically, the estimation unit 342 performs the same actions as described in estimation step S32 (see Figure 12) in the estimation method (fifth invention).

[0084] (Third conversion section) The third transformation unit 343 geometrically transforms the estimated image I6 into an output image I7 of the cross-sectional shape (trapezoidal) of the earth structure S. Specifically, the third transformation unit 343 performs the same actions as described in the third transformation step S33 (see Figure 12) of the estimation method described above.

[0085] (Third output control unit) The third output control unit 344 controls the third output unit to output the output image I7. If the third output unit 32 is a communication module or terminal, the third output unit 32 transmits the data of the output image I7 to another device. If the third output unit 32 is a drive, it writes the data of the output image I7 to a recording medium. If the third output unit 32 is a display, it displays the output image I7.

[0086] [Effects and Effects of Estimation Devices] As described above, the estimation device 3 according to this embodiment can be used by employing a generator G constructed using a generative adversarial network to obtain an estimated image I6 (a generated image I3 that is not identified as fake by the classifier D) that shows a two-dimensional distribution including the vertical direction, not just the horizontal direction, similar to the estimation method described above. Furthermore, if the input image I2 shows a hydraulic conductivity distribution, an estimated image I6 can be obtained that shows a pressure head distribution including not only the saturated region but also the unsaturated region.

[0087] <First, Third, Fifth Inventions, and Others> Furthermore, the above image processing method may include at least one of the steps in the above generator construction method and the above estimation method. Also, the above generator construction method may include at least one of the steps in the above image processing method and the above estimation method. Furthermore, the above estimation method may include at least one of the steps in the above image processing method and the above generator construction method.

[0088] <Second, Fourth, Sixth Inventions, and Others> Furthermore, the image processing device 1 may include at least one control block of the generator construction device 2 and the estimation device 3. Furthermore, the generator construction device 2 may include at least one control block of the image processing device 1 and the estimation device 3. Furthermore, the estimation device 3 may include at least one control block of the image processing device 1 and the generator construction device 2.

[0089] Furthermore, the functions of at least one of the image processing apparatus 1, generator construction apparatus 2, and estimation apparatus 3 can be realized by programs (image processing program, generator construction program, estimation program) that cause a computer to function as the apparatus, and which cause a computer to function as each control block of the apparatus (in particular, each part included in at least one of the first control unit 13, second control unit 24, and third control unit 34). In this case, the device includes a computer having at least one control device (e.g., a processor) and at least one storage device (e.g., memory) as hardware for executing the program. By executing the information processing program using this control device and storage device, the functions described in each of the embodiments are realized. The above program may be recorded on one or more computer-readable recording media, not temporary ones. These recording media may or may not be provided by the above device. In the latter case, the program may be supplied to the above device via any wired or wireless transmission medium.

[0090] Furthermore, some or all of the functions of each of the above control blocks can also be realized by logic circuits. For example, an integrated circuit in which logic circuits functioning as each of the above control blocks are formed is also included in the scope of the present invention. In addition, it is also possible to realize the functions of each of the above control blocks by, for example, a quantum computer.

[0091] It should be noted that the present invention is not limited to the embodiments described above, and various modifications are possible within the scope of the claims. Embodiments obtained by appropriately combining the technical means disclosed in different embodiments are also included within the technical scope of the present invention. [Examples]

[0092] Next, embodiments of the first to sixth inventions described above will be explained.

[0093] First, using the above image processing method or the above image processing device 1, 15 3 An input image I2 showing the distribution of individual hydraulic conductivity coefficients was generated. Next, the above 15 3 The input images I2, and the 15 obtained by sequential analysis of each input image I2. 3 A set of individual correct images I5 was used as training data, and a generator G was constructed using the generator construction method described above, or using the generator construction device 2 described above. Next, the constructed generator G was input with an input image I2 showing the hydraulic conductivity distribution, and an estimated image I6 showing the pressure head distribution was generated. This was then geometrically transformed into the output image I7 (forward). Next, the SSIM (Similarity Score to the pixel values ​​of the ground truth image I5) for each pixel p in the output image I7 was calculated.

[0094] Furthermore, using the above image processing method or the above image processing device 1, 15 3 An input image I2 showing the pressure head distribution was generated. Next, the above 15 3 The learning input images I2, and the 15 obtained by inverse analysis of each input image I2. 3 A set of individual correct images I5 was used as training data, and a generator G was constructed using the generator construction method described above, or using the generator construction device 2 described above. Next, 343 verification images showing pressure head distributions (prediction input images I2) were input to the constructed generator G, and an estimated image I6 showing the hydraulic conductivity distribution was generated. This was then geometrically transformed into the output image I7 (inversion). Next, the SSIM of the output image I7 was calculated respectively.

[0095] Next, using the calculated SSIM, a histogram (a superimposed display of forward and inversion) with the SSIM on the horizontal axis and the frequency (number of images) on the vertical axis, as shown in FIG. 15, was created. From the histogram, it can be seen that the SSIM of most of the output images I7 (forward) shows a value of 0.98 or more. This indicates that the output image I7 (forward) according to this embodiment has an accuracy level (similarity to the correct image I5) that cannot be distinguished from the correct image I5 by the discriminator D of the adversarial generation network. On the other hand, it can be seen that the SSIM of many output images I7 (inversion) is less than 0.98.

[0096] Next, the above 15 3 input images I2 showing the piezometric head distributions, and 15 3 pairs of correct images I5 obtained by inverse analysis of each input image I2 were used as learning data, and another generator G was constructed by the above generator construction method or using the above generator construction device 2. At that time, the prior information obtained in advance (the permeability coefficient of the downstream part (third part P3) of the embankment corresponding to each input image I2) was given to the unlearned generator g. Next, the input image I2 showing the piezometric head distribution was input into the constructed another generator G, and the estimated image I6 showing the permeability coefficient distribution was generated. Then, it was geometrically transformed into the output image I7 (inversion). Next, the error of the output image I7 with respect to the correct image I5 was calculated.

[0097] Next, using the calculated error, a histogram (a superimposed display of with and without prior information during learning) with the error on the horizontal axis and the frequency (number of images) on the vertical axis, as shown in FIG. 16, was created. From the histogram, it can be seen that the error of most of the output images I7 (with prior information) is 10 1The following can be observed. This indicates that by using prior information in conjunction with the construction of generator G, the accuracy of the output image I7 (inversion) in this embodiment is greatly improved. [Explanation of symbols]

[0098] 1 Image Processing Device 11 First Input Section 12 First output section 13 First Control Unit 131 First Acquisition Department 132 First Conversion Unit 133 Generation part 134 Second Conversion Section 135 First Output Control Unit 2 Generator construction device 21 Second Input Section 22 Second Output Section 23 Memory section 24 Second Control Unit 241 Second Acquisition Department 242 Generated Image Output Unit 243 Analysis Department 244 Identification Result Output Unit 245 Judgment Department 246 Learning Department 247 Second Output Control Unit 3 Estimation device 31 Third Input Section 32 Third output section 33 Storage section 34 Third Control Unit 341 Third Acquisition Department 342 Estimation Department 343 Third Conversion Unit 344 Third Output Control Unit D Discriminator d Unlearned classifier G generator g Unlearned Generator I1 image before conversion Region indicating the first part of R1 Region indicating the second part of R2 R3 indicates the third region. I2 Each input image I3 generated image I4 pre-conversion analysis image I5 Correct Image I6 Estimated Image Image for I7 output L is the line indicating the infiltration surface. S earth structure P1 first part P2 second part P3 third part p pixel

Claims

1. The process involves converting a first physical quantity related to the soil structure to be analyzed into discrete values ​​according to its characteristics, generating a pre-conversion image with the same shape as the cross-section of the soil structure showing the distribution of the discrete values, inputting a rectangular input image generated by geometrically transforming the pre-conversion image into the generator, and outputting a generated image showing the distribution of a second physical quantity. The steps include analyzing the pre-conversion image to obtain a ground truth image showing the distribution of a second physical quantity, The steps include inputting the generated image and the correct image into a classifier and outputting a result of determining whether the generated image is genuine or not, If the identification result is false, at least the step of training the generator, A method for constructing a generator having the following characteristics.

2. The first physical quantity is one of the hydraulic conductivity and the pressure head. The second physical quantity is the other, The method for constructing a generator according to claim 1.

3. If the first physical quantity is the pressure head and the second physical quantity is the hydraulic conductivity, The method further includes the step of obtaining in advance the permeability coefficient of at least a portion of the soil structure as prior information, In the learning step, when training the generator, provide the generator with prior information. The method for constructing a generator according to claim 2.

4. The generator is U-Net. A method for constructing a generator according to any one of claims 1 to 3.

5. A generated image output unit converts a first physical quantity relating to the soil structure to be analyzed into discrete values ​​according to its characteristics, generates a pre-conversion image with the same shape as the cross-section of the soil structure showing the distribution of the discrete values, inputs a rectangular input image generated by geometrically transforming the pre-conversion image into a generator, and outputs a generated image showing the distribution of a second physical quantity. An analysis unit analyzes the pre-conversion image to obtain a ground truth image showing the distribution of a second physical quantity, An identification result output unit inputs the generated image and the correct image to a classifier and outputs the identification result of whether the generated image is genuine or not, A learning unit that causes the generator to learn if the identification result is false, A generator construction device equipped with the following features.

6. A generator construction program for causing a computer to function as a generator construction device according to claim 5, wherein the computer functions as the generated image output unit, the analysis unit, the identification result output unit, and the learning unit.

7. A generator constructed using a generative adversarial network, The system is constructed by converting a first physical quantity related to the soil structure under analysis into discrete values ​​according to its characteristics, generating a pre-conversion image with the same shape as the cross-section of the soil structure showing the distribution of the discrete values, inputting a rectangular input image generated by geometrically transforming the pre-conversion image into a generator, outputting a generated image showing the distribution of a second physical quantity, inputting a ground truth image showing the distribution of the second physical quantity obtained by analyzing the generated image and the pre-conversion image into a classifier, outputting a genuine / fake identification result of the generated image, and training the generator if the identification result is fake. The system converts a first physical quantity related to the target soil structure into discrete values ​​according to its characteristics, generates a pre-conversion image with the same shape as the cross-section of the soil structure showing the distribution of the discrete values, and when an input image generated by geometrically transforming the pre-conversion image is input, it generates an estimated image showing the distribution of a second physical quantity. generator.

8. An estimation method comprising the steps of generating an estimated image by converting a first physical quantity relating to a soil structure to be analyzed into discrete values ​​according to its characteristics, and generating a rectangular input image by geometrically transforming an image of the soil structure's cross-sectional shape before conversion that shows the distribution of the discrete values, and then inputting this rectangular input image into a generator constructed using a generative adversarial network, which generates an estimated image showing the distribution of a second physical quantity when the input image showing the distribution of the first physical quantity is input.

9. An estimation device comprising an estimation unit that generates an estimated image by converting a first physical quantity relating to a soil structure to be analyzed into discrete values ​​according to its characteristics, and geometrically transforming an image of the soil structure's cross-sectional shape before conversion showing the distribution of the discrete values ​​to generate a rectangular input image, which is constructed using a generative adversarial network and, upon receiving the input image showing the distribution of the first physical quantity, generates an estimated image showing the distribution of a second physical quantity, thereby generating the estimated image.

10. An estimation program for causing a computer to function as an estimation device according to claim 9, wherein the computer functions as the estimation unit.

Images