Boundary extraction device, boundary extraction method, and program

Abstract

A boundary extraction device, a boundary extraction method, and a program are provided that are capable of extracting the boundary lines of topographical features with higher accuracy. [Solution] A boundary extraction device that analyzes a microtopography representation and extracts the boundary of an area where a predetermined topographical feature appears. The boundary extraction device includes a storage means for storing a trained model that has been trained in advance to output a boundary area in which the line width of the boundary of an area where a topographical feature appears in the microtopography representation in response to an input of the microtopography representation, an acquisition means for inputting the microtopography representation to be analyzed into the trained model and acquiring the boundary area output from the trained model, and an extraction means for thinning the boundary area acquired by the acquisition means and extracting the boundary of the area where a topographical feature appears in the microtopography representation to be analyzed.

Description

[Technical field]

[0001] The present invention relates to a boundary line extraction device, a boundary line extraction method, and a program. [Background technology]

[0002] Information on landslide areas in mountainous regions is used for forestry conservation measures. Traditionally, landslide areas have been identified by engineers from map data that emphasizes the topography, such as microtopographical representation maps. Landslide areas usually occur in U-shaped areas, and their boundaries tend to appear on map data. Therefore, in the process of identifying the boundaries of the U-shape, the boundary line is identified (Non-Patent Document 1). Such identification requires skill and a great deal of effort. Therefore, assistance through automatic extraction is required. [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] "Lecture 3. Extracting disaster factors through topographical interpretation," Collection of materials related to the Road Disaster Prevention Inspection Techniques Seminar, National Federation of Geological Survey Associations, published May 27, 2022 https: / / www.zenchiren.or.jp / geocenter / lec-road / docs / lecture3.pdf Summary of the Invention [Problem to be solved by the invention]

[0004] However, in the case of the boundary line of a topographical feature with a small width in an image, the correct answer area occupies a small proportion of the entire image. As a result, it is difficult to obtain appropriate learning results in the training of a machine learning model for automatically extracting the boundary line of the topographical feature, and there is a problem that the boundary line of the topographical feature cannot be extracted with high accuracy.

[0005] An object of the present invention is to provide a boundary extraction device, a boundary extraction method, and a program that are capable of extracting the boundary of a topographical feature with higher accuracy. [Means for solving the problem]

[0006] In order to achieve the above object, the present invention provides A boundary extraction device that analyzes a microtopographical representation map to extract a boundary line of an area in which a predetermined topographical feature appears, A storage means for storing a trained model that has been trained in advance to output a boundary area in which the line width of the boundary line of the area in which the topographical feature appears in the microtopographical representation map is expanded in response to the input of the microtopographical representation map; An acquisition means for inputting a microtopography representation of an analysis target into the trained model and acquiring the boundary area output from the trained model; an extraction means for thinning the boundary area acquired by the acquisition means and extracting a boundary line of an area in which the topographical feature appears in the microtopographical representation map of the analysis target; The boundary extraction device includes: Effect of the Invention

[0007] According to the present invention, it is possible to extract the boundary lines of topographical features with higher accuracy. [Brief description of the drawings]

[0008] [Figure 1] 1 is a block diagram showing a functional configuration of an information processing device that is a boundary extraction device according to an embodiment of the present invention. [Diagram 2] FIG. 13 is a diagram illustrating a procedure for detecting a collapsed area. [Diagram 3] FIG. 1 shows an example of a microtopographic representation and the resulting boundary region. [Figure 4] FIG. 13 is a diagram showing an example of boundary line data obtained by thinning the obtained boundary region image. [Diagram 5] 13 is a flowchart showing a control procedure for a boundary detection process. [Figure 6] 10 is a flowchart showing a control procedure for a wire connection process. [Figure 7] 13 is a flowchart showing a control procedure of a learning data generation process. [Figure 8] FIG. 13 is a diagram illustrating an example of generating correct answer data. [Figure 9] 13 is a flowchart showing a control procedure for a model learning process. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0009] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing the functional configuration of an information processing device 1 which is a boundary extraction device according to this embodiment.

[0010] The information processing device 1 includes a control unit 11, a storage unit 12 as a storage means, an input / output interface 13 (I / F), a display unit 14, an operation receiving unit 15, and the like.

[0011] The control unit 11 performs overall control of the operation of the information processing device 1. The control unit 11 has a processor that performs arithmetic processing. The processor may be a single general-purpose CPU (Central Processing Unit), or may have multiple CPUs that perform arithmetic processing in parallel or independently depending on the application. The processor may include a processor specialized for a specific arithmetic processing or image processing. The control unit 11 performs various control processing by reading and executing a program 120 from the storage unit 12.

[0012] The storage unit 12 has a RAM (Random Access Memory) and a non-volatile memory, and stores various data. The RAM provides a working memory space for the control unit 11 and stores temporary data. The non-volatile memory stores and holds the program 120, the machine learning model 122, setting data, and the like. The non-volatile memory is, for example, a flash memory or a HDD (Hard Disk Drive), but is not limited to these. The storage unit 12 may have a ROM (Read Only Memory). An initial control program and the like may be stored in the ROM. The program 120 includes a control program related to a boundary detection process, a learning data generation process, and a model learning process, which will be described later. The program 120 includes a trained model 121. The trained model 121 is a machine learning model 122 that has been trained. Once the trained model 121 is obtained, the machine learning model 122 may not be stored in the storage unit 12.

[0013] The input / output interface 13 inputs and outputs data between the information processing device 1 and the outside (including peripheral devices). The input / output interface 13 has a connection terminal 131 and a communication unit 132. The connection terminal 131 includes, for example, a Universal Serial Bus (USB) terminal and a Local Area Network (LAN) connector. The communication unit 132 controls communication in accordance with a communication rule (protocol) related to the LAN, such as TCP / IP.

[0014] The peripheral devices serving as the learning data storage means of this embodiment may include a database device 21 which is an auxiliary storage device, and an optical reader 22 which reads portable storage media (optical disks) such as CD-ROM, DVD, and Blu-ray (registered trademark). The portable storage media may include a magnetic tape, and the peripheral devices may include a reader which reads the magnetic tape.

[0015] Data that the information processing device 1 can acquire from the outside via the input / output interface 13 includes microtopography representation data 201, learning boundary data 202, and learning boundary area data 203. The microtopography representation data 201 stores one or more microtopography representations (learning microtopography representations) used for learning the machine learning model 122, a microtopography representation (analysis target microtopography representation) that is an analysis target by the learned model 121, and the correspondence between each of the microtopography representations and the geographic coordinates. The microtopography representation is an image that has been emphasized so that the relief of the terrain can be more easily read from a normal topography map. Examples of microtopography representations include CS (Curvature / Slope) 3D maps, red 3D maps, and the topography maps described in JP 2021-149059 A.

[0016] The microtopography representation is a pixmap (raster) image represented by a set of pixels arranged two-dimensionally in a grid pattern vertically and horizontally (x and y directions). The image may be any image capable of expressing the range of colors required to represent the microtopography representation. The micro-topography representation data 201 stores the XY coordinates of a plane rectangular coordinate system (the center of gravity of the geographical area represented by each pixel) represented by the center of each of the two diagonal pixels of the micro-topography representation in association with each micro-topography representation. This makes it possible to specify the correspondence between the position of each pixel of the micro-topography representation and the geographical coordinates. The geographical coordinates may be latitude and longitude.

[0017] The learning boundary data 202 is linked to the learning microtopography representation map and stores data on the boundary line of the collapsed area (learning boundary line) that appears in the learning microtopography representation map. In other words, the learning boundary line indicates the boundary line of the area in which topographical features appear in the learning microtopography representation map. Each line may be multi-line data that is vector-expressed in geographic coordinates. The learning boundary data 202 is generated manually by an engineer or the like while referring to the microtopography representation map data and, if necessary, aerial photographs, etc. Also, known collapsed area data, etc. may be used. Note that, if a trained model with a certain degree of accuracy is already available, a candidate extraction operation may be performed as preprocessing using the trained model before the engineer's manual operation.

[0018] A learning boundary area generated by expanding the line width of a learning boundary line is linked to a learning microtopography representation and stored in the learning boundary area data 203. Each area may be polygon data represented as a vector in a geographic coordinate system. The training boundary area is correct answer data of the boundary area for the training microtopography representation map, and the set of the training microtopography representation map and the training boundary area is training data for the machine learning model 122.

[0019] The display unit 14 performs display on the display screen based on the control of the control unit 11. The display screen is, for example, a liquid crystal display or an organic EL (Electro-Luminescent) display, but is not limited to these.

[0020] The operation reception unit 15 receives an input operation from the outside, and outputs an operation signal corresponding to the input operation to the control unit 11. The operation reception unit 15 includes, for example, a keyboard and a pointing device. The pointing device may be a mouse. The display unit 14 and / or the operation reception unit 15 may be peripheral devices of the information processing device 1. In other words, they may be attached to the main body (computer) of the information processing device 1 including the control unit 11, the storage unit 12, and the input / output interface 13.

[0021] Next, as a boundary extraction method according to this embodiment, an operation for detecting a collapsed area as an area where a predetermined topographical feature appears will be described. A collapse refers to the phenomenon whereby layers of rock collapse en masse, mainly on steep slopes, and a collapsed area is the area where such a collapse has occurred. In collapsed areas, the boundary of the collapse is generally clear, and a linear discontinuous surface appears. Collapses often occur on the uppermost slope of a valley, and as a result, a U-shaped boundary surrounding the valley is likely to appear. On the other hand, rocks and soil from the collapsed slope tend to spread and accumulate on a gentle or nearly flat slope to fill the valley. That is, a collapsed area is typically identified as a combination of a steeply U-shaped discontinuous surface surrounded by this and a relatively gentle area that partially spreads downstream. Note that vegetation may grow on the sediments that have been there for some time since the collapse. In such cases, it is often difficult to see the traces of the collapse inside the collapsed area from aerial photographs, etc., compared to the boundary where the steep slope is likely to remain clearly.

[0022] FIG. 2 is a diagram for explaining a procedure for detecting a collapsed area. The trained model 121 is stored in advance in the storage unit 12 (storage step). As shown in FIG. 2(a), in this embodiment, the trained model 121 extracts relatively clear boundaries of collapsed areas from the microtopography representation map to be analyzed. This trained model 121 outputs a boundary region having an area, as described below (P1; acquisition step). The boundary region acquired as the output of the trained model 121 is changed to a boundary line by thinning processing (P2; extraction step), and the final result is obtained.

[0023] The trained model 121 and the machine learning model 122 can be neural network models having a network structure related to image segmentation. For example, they may be modeled as Res U-Net. The trained model 121 and the machine learning model 122 can obtain the probability of each pixel being a boundary region. By binarizing the probability value with a predetermined reference value, an image indicating whether or not a pixel is a boundary region can be obtained.

[0024] FIG. 3 shows an example of a microtopography representation and the resulting boundary region. In FIG. 3(a), a CS 3D map, which is a microtopography representation map, is shown in grayscale. This CS 3D map is a part of the Shizuoka Prefecture CS 3D map (created by Shizuoka Prefecture, Shizuoka Prefecture Forest Planning Division, 2017) that has been converted to grayscale. The CS 3D map is a map in which the elevation, slope, and curvature are expressed in different colors and other gradations and superimposed. This map makes it easier to identify collapsed areas from a slope to a valley. Here, the dark colored parts correspond to the valley, and the U-shaped parts that suddenly cut deeply near the source of the valley generally correspond to collapsed areas. This CS 3D map is input into the trained model 121, and an example of a boundary area image in which the obtained probability distribution is binarized is shown in FIG. 3(b). As described above, many collapsed areas have a U-shaped boundary. In some cases, multiple collapsed areas may be located in a row.

[0025] FIG. 4 is a diagram showing an example of boundary line data obtained by thinning the obtained boundary region image. A boundary region shown in FIG. 4(a) is a curve with a certain width as described above, and often has a characteristic U-shape. In the thinning process, this region is thinned, that is, a line image with a reduced width in a direction perpendicular to the extension direction is obtained, and this is then represented as a multipolygon, which is a series of multiple polygons. The thinning process may be performed by any conventionally known method. As shown in FIG. 4(b), a multipolygon is represented by combining only vertical and horizontal lines of pixels according to the image resolution. That is, each polygon may be a combination of shapes such as a rectangle, a hook shape, and an L-shape. By connecting representative positions in each polygon, for example, the center of gravity positions, a broken line-like boundary line as shown in FIG. 4(c) is obtained.

[0026] It is not easy to artificially determine correct data for such a boundary area, and usually a boundary line is given as correct data. However, if learning is done using only lines, the area that the "lines" occupy in the image is too small, making it difficult to obtain effective learning results. Therefore, as shown in Figure 2(b), the line width of the given boundary line (learning boundary line) is expanded and thickened, and then converted into an image of a boundary area (learning boundary area) with width, thereby obtaining a correct area image (P11; expansion means). The correct area image and the original microtopography representation are matched to obtain learning data.

[0027] The machine learning model 122 is trained by the obtained training data (P12; training means). As a result, the trained model 121 is obtained.

[0028] 5 is a flowchart showing a control procedure of the boundary detection process executed by the information processing device 1. This boundary detection process is triggered by, for example, the operation receiving unit 15 or the communication unit 132 receiving a predetermined input operation by the user together with the designation of the analysis target microtopographic representation map.

[0029] As shown in FIG. 5(a), the control unit 11 as an acquisition means acquires the micro-topography representation map to be analyzed from the micro-topography representation map data 201 and inputs it to the trained model 121 (S11). The size of the image that can be input to the trained model 121 (input size) is predetermined. This size is smaller than the micro-topography representation map to be analyzed. The acquisition means sequentially inputs divided images obtained by dividing the micro-topography representation map to be analyzed into images of the input size to the trained model 121. Note that if the size of the micro-topography representation map to be analyzed is not a multiple of the input size, some of the divided images may be set to have overlapping ranges.

[0030] The acquisition means operates the trained model 121 to acquire a probability distribution map output from the trained model 121 (S12). In response to the input of the divided image in the process S11, the trained model 121 outputs a probability distribution map of the same size as the divided image. The acquisition means acquires a probability distribution map of the same size as the analysis target microtopography representation map by arranging the output probability distribution map in the same position as the corresponding divided image. In this case, the value of a pixel that overlaps between the multiple divided images may be the maximum value of the value of the pixel. Alternatively, the pixel value may be the average value of the values ​​obtained from the multiple divided images. The value of each pixel in the analysis target microtopography representation map represents the probability that it is a boundary area of ​​the collapsed area. The boundary area is an area obtained by expanding the width of the boundary line of the collapsed area. The acquisition means binarizes the probability distribution map using a predetermined reference value to identify the boundary area of ​​the collapsed area (S13). In addition, the trained model 121 may include a binarization process and output a binarized image. The control unit 11, which serves as an extraction means, executes a boundary line identification process based on the binarized image (S14).

[0031] As shown in Fig. 5(b), in the boundary line identification process, the extraction means thins the boundary area (S41). For example, the extraction means reduces the width perpendicular to the extension direction of the boundary area to a minimum reference value. The minimum reference value may be, for example, one pixel.

[0032] The extraction means converts the line image obtained by thinning the boundary area into polygons (S42). That is, as described above, the extraction means generates a polygon representing a geographical area corresponding to each group of pixels connected in the x direction and / or y direction in the line image. A rectangular polygon is generated in a portion of the line image where pixels are connected only in the x direction or y direction. An L-shaped or hook-shaped polygon is generated in a portion of the line image where a group of pixels in the x direction and a group of pixels in the y direction are connected at a common pixel.

[0033] The extraction means counts the number of polygons for each group of polygons adjacent to each other in a chain. The extraction means excludes a boundary area in which the number of polygons is less than three from subsequent processing, assuming that the boundary area is a detection noise that is not U-shaped (S43). If there is no group in which the number of polygons is three or more, the extraction means ends the boundary line identification processing without performing the processing in S44 and subsequent steps.

[0034] The extraction means extracts a representative position, here a center of gravity position, within each polygon that was not excluded in process S43 (S44). The coordinate system of the representative position is a geographic coordinate system. The extraction means executes a connection process that connects the extracted center of gravity positions (S45). Then, the control unit 11 ends the boundary line specification process and returns the process to the boundary detection process.

[0035] FIG. 6 is a flowchart showing a control procedure for the wire connection process. The extraction means counts the number of other polygons adjacent to each polygon (S401). The extraction means determines whether there is one polygon with an adjacent number of "1" (S402). A polygon with an adjacent number of "1" is a polygon at the end of a boundary line. In a normal boundary line, there are two polygons with an adjacent number of "1". If it is determined that there is one polygon with an adjacent number of "1" (S402; Y), the extraction means sets the polygon with an adjacent number of "1" as the reference polygon (S403). Then, the processing of the extraction means proceeds to step S405.

[0036] If it is determined that there is not one polygon with an adjacent number of "1" (S402; N), the polygon with an adjacent number of "1" that is closest to a predetermined position, for example, the upper left corner of the analysis target microtopography representation map, is set as the reference polygon (S404). Then, the processing of the extraction means proceeds to step S405.

[0037] When the process proceeds to S405, the extraction means additionally sets the representative position (center of gravity position) of the reference polygon to a point list that lists points on the boundary line (S405). The extraction means specifies the polygon that has the closest center of gravity position among the polygons that have not been set as the reference polygon from the center of gravity position of the reference polygon (S406). At this time, if there are no polygons that have not been set, the extraction means advances the process to S411.

[0038] The extraction means judges whether the distance between the center of gravity of the reference polygon and the center of gravity of the specified polygon is equal to or greater than a reference (S407). If it is judged that the distance is not equal to or greater than the reference (S408; N), the extraction means determines that the specified polygon is located on the same boundary line as the reference polygon, and sets the specified polygon as the next reference polygon (S408). Then, the processing of the extraction means returns to processing S405.

[0039] If it is determined that the distance is equal to or greater than the reference value (S408; Y), the extraction means determines whether or not the distance to the end point (usually the starting point) of another connected boundary line in step S413 is equal to or greater than the reference value (S411). If it is determined that there is a connected boundary line whose distance is not equal to or greater than the reference value (less than the reference value) (S411; N), the extraction means determines that the connected boundary line is part of the boundary line represented by the point list being processed, and adds and merges the points of the connected boundary line in order to the point list (S412). The extraction means deletes the added and merged connected boundary line from the connected data. Then, the processing of the extraction means proceeds to step S413. If it is determined that the distance to the end points of the other connected boundary lines are all equal to or greater than the reference value (S411; Y), the processing of the extraction means proceeds to step S413, determining that the point list being processed is for a new boundary line. Incidentally, even if there is no data of other connected boundary lines, the process of the extraction means proceeds to step S413.

[0040] When the process proceeds to step S413, the extraction means connects the points stored in the point list in a set order to identify the boundary line (S413). That is, the boundary line is obtained as vector data connecting coordinates. The extraction means stores the data of the point list in the storage unit 12 as data of the connected boundary line, and then erases the contents of the point list.

[0041] The extraction means determines whether there are any remaining polygons that have not been set as reference polygons (S414). If it is determined that there are any remaining polygons (S414; Y), the processing by the extraction means returns to processing S402. If it is determined that there are no remaining polygons (S414; N), the extraction means ends the line connection processing and returns the processing to the boundary line identification processing. Through the above processing, the boundary line of the collapsed area is extracted from the microtopography representation map.

[0042] Fig. 7 is a flowchart showing a control procedure for a learning data generation process for learning the machine learning model 122. Fig. 8 is a diagram showing an example of generation of correct answer data.

[0043] The control unit 11 selects a learning microtopographic representation map to be used as learning data from the microtopographic representation map data 201 (S31). The control unit 11 sets a learning boundary line by line data (S32). As shown in FIG. 8(a), the boundary line may be a broken line. Alternatively, the boundary line may be drawn freehand as a free curve. The boundary line may be set by a pointing device of the operation reception unit 15 on an image displayed on the display screen of the display unit 14. The control unit 11 may also obtain data of a learning boundary line that has been set externally via the communication unit 132. The line data set here may be data expressed as a vector in geographical coordinates.

[0044] The control unit 11, which serves as the expansion means, thickens the set learning boundary line (S33). Thickening in the vector data may be performed simply by changing the line width information. The expansion width is set in advance within a range in which the thickness in the microtopography representation map is two pixels or more and the boundary area is sufficiently thinner than the collapsed area. The expansion width may be set in actual size, or the width in the image data to be expanded may be determined according to the scale of the learning microtopography representation map so that the width of the boundary line is one meter, for example. As shown in FIG. 8(b), a boundary line that is simply thicker than that in FIG. 8(a) is obtained.

[0045] The expansion means identifies the position of the thick line and converts the vector data into a bitmap (S34). That is, as shown in FIG. 8(c), the control unit 11 identifies pixel positions where the thick line overlaps in a bitmap image having the same resolution as the learning micro-topography representation, and sets a binary value according to whether or not there is an overlap with the thick line. The control unit 11 combines the selected learning micro-topography representation with the generated bitmap data to generate learning data (S35). The generated bitmap data is linked to the learning micro-topography representation and registered in the learning boundary area data 203.

[0046] The control unit 11 judges whether the selection of the learning microtopography representation map is completed (S36). If it is judged that the selection of the learning microtopography representation map is not completed (S36; N), the process of the control unit 11 returns to the process S31. If it is judged that the selection of the learning microtopography representation map is completed (S36; Y), the control unit 11 ends the learning data generation process.

[0047] 9 is a flowchart showing a control procedure of the model learning process executed by the information processing device 1. The control unit 11 executes the model learning process by specifying the machine learning model 122 to be learned and the learning data to be used for learning.

[0048] The control unit 11 as a learning means selects one of the learning microtopography representation maps included in the microtopography representation map data 201, cuts out one image of the above-mentioned input size from the learning microtopography representation map, and inputs this to the machine learning model 122 (S21). The size of the image that can be input to the machine learning model 122 is predetermined to the above-mentioned input size, and the size is smaller than the learning microtopography representation map. The learning means, for example, randomly selects the learning microtopography representation map based on a random number, cuts out an image of the input size from the learning microtopography representation map at a random position in the learning microtopography representation map based on the random number, and inputs the cut-out image to the machine learning model 122. The learning means acquires, as an output result (S22), the probability distribution map output from the machine learning model 122. In response to the input of the cut-out image in process S21, the machine learning model 122 outputs a probability distribution map of the same size as the cut-out image.

[0049] The learning means calculates the error of the probability distribution map for the correct answer data (S23). First, the learning means reads out the learning boundary area corresponding to the learning microtopography representation map selected in step S21 from the learning boundary area data 203 as the correct answer data. Next, the learning means calculates the error (intra-area error) in the boundary area (inside the learning boundary area) in the correct answer data and the error (out-area error) in the area that is not a boundary area (outside the learning boundary area) in the correct answer data. The learning means calculates the error, for example, by using cross entropy loss as the loss function. In this case, the probability at pixel x indicated by the probability distribution map, i.e., the probability that pixel x is a pixel in the boundary area, is denoted as q(x), and the probability of a pixel within the learning boundary area is denoted as x(x). pоs , pixels outside the learning boundary are denoted by x neg Then, the error in the region is -Σln{q(x pоs )}, and the out-of-domain error is -Σln{q(x neg )}.

[0050] The learning means divides the errors into in-region errors and out-of-region errors and weights the errors (S24). The learning means weights the errors differently between these regions. For example, the learning means weights the number of pixels N pos The ratio of R pos and the number of pixels outside the learning boundary region N neg The ratio of R neg The weighting is done according to the learning method. pos )=R neg =N neg / (N pos +N neg ) multiplied by the intra-area error, and (1-R neg )=R pos =N pos / (N pos +N neg The error may be calculated by multiplying the out-of-area error by N pos and the number of pixels N neg Alternatively, the learning means may calculate the number of pixels N by averaging the calculated values ​​of the correct answer data of the learning micro-topography representation map included in the micro-topography representation map data 201. pos and the number of pixels Nneg Also, the number of pixels N pos and the number of pixels N neg may be set in advance based on a micro-topography representation map that is not included in the micro-topography representation map data 201. As described above, since the number of pixels inside the learning boundary region is significantly smaller than the number of pixels outside the learning boundary region, a larger weight is assigned to the intra-region error, and each error is added up.

[0051] The learning means performs feedback on the parameters of the machine learning model 122 so as to minimize the weighted error, and updates the machine learning model 122 (S25). The method of feedback of the error may be a conventionally known method, and for example, the backpropagation method may be used.

[0052] The learning means determines whether the selection and input of the learning microtopography representation map is completed (S26). The learning means determines that the input of process S21 is completed when the input has been performed a predetermined number of times, or when the error calculated in process S24 is less than a predetermined reference value. If it is determined that the input is not completed (S26; N), the processing of the learning means returns to process S21. If it is determined that the input is completed (S26; Y), the learning means determines the trained machine learning model 122 as the trained model 121 (S27). Then, the learning means ends the model learning processing.

[0053] The present invention is not limited to the above-described embodiment, but may be modified in various ways. For example, the architecture of the machine learning model 122 may be related to image segmentation other than Res U-Net, such as PSPNet, DeepLabV3+, TransUnet, and SegFormer.

[0054] In the above, the cross entropy (CE) function is used as the loss function to weight the error separately, but this is not limited to this. Focal loss, which takes weighting into account from the beginning, may also be used. Dice loss may also be used in addition to the CE function.

[0055] In the above, the boundary line is obtained by connecting the positions of the centers of gravity of the polygons, but this is not limited to the above. For example, depending on the shape of the polygon, the position of the center of gravity may be located outside the polygon. Taking such a case into consideration, or limiting it to such a case, a representative position (inner guarantee point) located inside the polygon may be appropriately determined. Furthermore, when thinning the boundary area and determining the position on the boundary line, it is not necessary to use a multipolygon.

[0056] Although the above description has been given with reference to an example in which the boundary line of a collapsed area is extracted, the boundary line is not limited to this, and may be other topographical features such as a zero-order valley.

[0057] In addition, in the above description, the control unit 11 executes all the processes within the information processing device 1, but this is not limited to the above. The processes may be divided among a plurality of information processes. Furthermore, the learning of the machine learning model 122 may be performed outside the information processing device 1. Only the trained model 121 may be copied to the information processing device 1 and used.

[0058] In the above description, the storage unit 12 including a non-volatile memory such as an HDD or a flash memory is used as an example of a computer-readable medium for storing the program 120 related to the control of boundary extraction and learning of a machine learning model of the present invention, but is not limited thereto. As other computer-readable media, other non-volatile memories such as MRAM and portable recording media such as CD-ROMs and DVD disks can be applied. In addition, a carrier wave is also applied to the present invention as a medium for providing data of the program according to the present invention via a communication line. In addition, the specific configurations, contents and procedures of the processing operations, etc. shown in the above embodiments can be modified as appropriate without departing from the spirit of the present invention. The scope of the present invention includes the scope of the invention described in the claims and its equivalents.

[0059] As described above, the information processing device 1 of this embodiment analyzes a microtopography representation and extracts a predetermined topographical feature, for example, a boundary line of an area where a collapsed area appears. The information processing device 1 includes a storage unit 12 that stores a trained model 121 that has been trained in advance to output a boundary area in which the line width of the boundary line of an area where a topographical feature appears in the microtopography representation in response to an input of the microtopography representation, and a control unit 11. The control unit 11, as an acquisition unit, inputs the microtopography representation to be analyzed to the trained model 121 and acquires the boundary area output from the trained model 121. The control unit 11, as an extraction unit, thins the acquired boundary area and extracts the boundary line of an area where a topographical feature appears in the microtopography representation to be analyzed. In this way, by using the two-stage process of extracting the boundary area using the trained model 121 and then thinning the boundary area to extract the boundary line, the information processing device 1 can extract the boundary line of the topographical features with higher accuracy.

[0060] The extraction means may extract a boundary line in a vector format along a line image obtained by thinning the boundary area acquired by the acquisition means. By extracting the boundary line as vector data, data of the boundary line that is independent of scale and resolution can be obtained.

[0061] The information processing device 1 may be connected or connectable to a database device 21 as a learning data storage means for storing a plurality of learning microtopography representations used for the learning and learning boundary lines indicating the boundary lines of the areas in which the topographical features appear in each of the learning microtopography representations, and an optical reading device 22 for reading a portable recording medium. The control unit 11 may be configured to generate a learning boundary area by performing a process for expanding the line width of each learning boundary line as an expansion means. The control unit 11 may be configured to generate a trained model 121 by updating the machine learning model 122 as a learning means so as to minimize the error between the output result when each learning microtopography representation is input to the machine learning model 122 and the learning boundary area corresponding to the learning microtopography representation. The update may be performed by updating parameters, for example, by an error backpropagation method using a gradient descent method. In this way, the information processing device 1 may learn the machine learning model 122 on its own device to obtain the trained model 121. In this case, contrary to when it is used, the boundary line is made thicker to be used as the correct answer data for learning, thereby allowing the machine learning model 122 to be trained appropriately.

[0062] The learning means may also weight an in-area error, which indicates that an output result corresponding to a pixel in the learning boundary area is not a pixel of the boundary area, more heavily than an out-of-area error, which indicates that an output result corresponding to a pixel outside the learning boundary area is a pixel of the boundary area, and calculate the error by adding the in-area error and the out-of-area error. Since the boundary area still occupies a significantly small area in the image even after being thickened, if the in-area error and the out-of-area error are evaluated with the same weight, the in-area error is likely to be underestimated. Therefore, by weighting the in-area error more heavily than the out-of-area error, the machine learning model 122 is trained to be able to extract the boundary area more efficiently and with higher accuracy.

[0063] Furthermore, the boundary of the area where the topographical feature appears may be a U-shaped boundary representing the boundary of the collapsed area. Compared to the boundary on the mountain side of the collapsed area, which tends to appear clearly on the microtopographical representation map, the boundary on the valley side of the collapsed area is unclear, making it difficult to identify the entire area. Furthermore, for disaster prevention and other measures, the upstream boundary is often particularly important. Therefore, it is very useful to be able to more easily identify the U-shaped boundary of the collapsed area.

[0064] Moreover, the boundary extraction method by the control unit 11 of this embodiment includes the following steps: (1) a storage step of storing a trained model 121 that has been trained in advance to output a boundary area in which the line width of the boundary of an area in which a topographical feature appears in a microtopographical representation map is expanded in response to an input of the microtopographical representation map; (2) an acquisition step of inputting the microtopographical representation map to be analyzed into the trained model 121 and acquiring a boundary area output from the trained model 121; (3) an extraction step of thinning the boundary area acquired in the acquisition step and extracting the boundary of an area in which a topographical feature appears in the microtopographical representation map to be analyzed. According to this boundary extraction method, the control unit 11 can utilize the capabilities of the trained model 121 to more accurately extract a desired boundary area.

[0065] Furthermore, by installing and executing the program 120 relating to the above-mentioned boundary extraction method on a computer, it is possible to obtain the boundary lines of topographical features with greater accuracy using the trained model 121 without requiring special hardware. [Explanation of symbols]

[0066] 1. Information processing device 11 Control section 12 Storage section 120 Programs 121 trained models 122 Machine Learning Models 13 Input / Output Interface 131 Connection terminal 132 Communications Department 14 Display section 15 Operation reception section 21 Database device 22 Optical reading device 201 Microtopographical representation map data 202 Learning boundary data 203 Boundary Region Data for Learning

Claims

1. A boundary extraction device that analyzes a microtopographical representation map to extract a boundary line of an area in which a predetermined topographical feature appears, A storage means for storing a trained model that has been trained in advance to output a boundary area in which the line width of the boundary line of the area in which the topographical feature appears in the microtopographical representation map is expanded in response to the input of the microtopographical representation map; An acquisition means for inputting a microtopography representation of an analysis target into the trained model and acquiring the boundary area output from the trained model; an extraction means for thinning the boundary area acquired by the acquisition means and extracting a boundary line of an area in which the topographical feature appears in the microtopographical representation map of the analysis target; A boundary extraction device comprising:

2. The extraction means extracts the boundary line in a vector format along a line image obtained by thinning the boundary area acquired by the acquisition means.

2. The boundary extraction device according to claim 1.

3. a learning data storage means for storing a plurality of learning microtopographic representation maps used in the learning and learning boundary lines indicating boundary lines of areas in which the topographical features appear in each of the learning microtopographic representation maps; an expansion means for expanding the line width of each of the learning boundaries to generate a learning boundary area; A learning means for updating the machine learning model so as to minimize an error between an output result when the learning microtopography representation map is input to the machine learning model and the learning boundary area corresponding to the learning microtopography representation map, thereby generating the trained model; The boundary extraction device according to claim 1 , comprising:

4. the learning means weights an in-area error, which indicates that the output result corresponding to a pixel within the learning boundary area is not a pixel of the boundary area, more heavily than an out-of-area error, which indicates that the output result corresponding to a pixel outside the learning boundary area is a pixel of the boundary area, and calculates the error by adding the in-area error and the out-of-area error.

4. The boundary extraction device according to claim 3.

5. 5. The boundary extraction device according to claim 1, wherein the boundary of the area in which the topographical feature appears is a U-shaped boundary representing a boundary of a collapsed area.

6. A boundary extraction method in which a control unit analyzes a microtopographic representation map and extracts a boundary line of an area in which a predetermined topographical feature appears, a storage step in which the control unit stores a trained model that has been trained in advance to output a boundary area in which the line width of the boundary line of the area in which the topographical feature appears in the microtopographical representation is expanded in response to the input of the microtopographical representation; an acquisition step in which the control unit inputs a microtopography representation of an analysis target into the trained model and acquires the boundary area output from the trained model; an extraction step in which the control unit thins the boundary area acquired in the acquisition step and extracts a boundary line of an area in which the topographical feature appears in the microtopographical representation map of the analysis target; A boundary extraction method comprising:

7. A computer that performs processing to analyze a microtopographic representation map and extract boundaries of areas in which predetermined topographical features appear, A storage means for storing a trained model that has been trained in advance to output a boundary area in which the line width of the boundary line of the area in which the topographical feature appears in the microtopographical representation map is expanded in response to the input of the microtopographical representation map; An acquisition means for inputting a microtopography representation of an analysis target into the trained model and acquiring the boundary area output from the trained model; an extraction means for thinning the boundary area acquired by the acquisition means and extracting a boundary line of an area in which the topographical feature appears in the microtopographical representation map of the analysis target; A program that functions as a

Images