Apparatus, program and method for generating a low-resolution mesh map that can be reconstructed into a high-resolution mesh map
The mesh map generation device uses stacked low-resolution maps and advanced loss functions to convert low-resolution mesh maps into high-resolution maps, addressing privacy concerns and enhancing map detail recognition.
Patent Information
- Application Number
- JP2022028408
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2022-02-25
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2042-02-25
AI Technical Summary
Existing super-resolution technologies struggle to accurately convert low-resolution mesh maps provided by mobile phone carriers into high-resolution mesh maps due to privacy concerns, making it difficult to recognize minute movements in urban areas.
A mesh map generation device that combines multiple low-resolution mesh maps into representative meshes, using a training stage with stacked low-resolution maps and a pseudo-high-resolution method, employing SRCNN and loss functions like MSE and PE to generate high-resolution maps.
Enables the reproduction of low-resolution mesh maps into high-resolution maps with high accuracy, allowing clients to recognize detailed population and category distributions.
Smart Images

Figure 0007577702000001 
Figure 0007577702000002 
Figure 0007577702000003
Abstract
Description
[Technical field]
[0001] The present invention relates to a super-resolution technique based on CV (Computer Vision) technology, and is particularly suitable for distribution maps of areas on maps. [Background technology]
[0002] Conventionally, a super-resolution technology is known as SRCNN (Super-Resolution Convolutional Neural Network), which regenerates a high-resolution mesh map from a low-resolution mesh map (see, for example, Non-Patent Document 1). According to this technology, during training, a high-resolution image X is converted into a low-resolution image Y, and then the high-resolution image is restored by a three-layer CNN (Convolutional Neural Network). The first layer inputs a low-resolution image Y and extracts a feature map by convolution. The second layer converts these feature maps into a high-resolution patch representation (feature vector). The third layer combines predictions from the feature vector to generate a high-resolution image X' = F(Y) based on a mapping function F. Between X' and X, the network weights are trained to minimize the error using the loss function MSE (Mean Squared Error). During training, the system learns from a large amount of teacher data, and during estimation, a high-resolution image X is output by inputting only a low-resolution image Y.
[0003] Furthermore, as a general existing service, mobile phone carriers provide big data based on regions, collected from a huge number of users' mobile terminals (see, for example, Non-Patent Document 2). From the viewpoint of privacy protection, the big data is provided to clients after being processed so that individual location information cannot be identified. For example, the data is processed as follows: The area is divided into meshes, and distribution trends are calculated for each mesh for each time period. - Eliminate movement history within an individual's specific living area. · Keep personal identifiers secret. - Eliminate movement histories of a small number of people to prevent the extraction of rare location information. [Prior art documents] [Non-patent literature]
[0004] [Non-Patent Document 1] Chao Dong, Chen Change Loy, Kaiming He, Xiaoou Tang. Learning a Deep Convolutional Network for Image Super-Resolution, Proc. of European Conference on Computer Vision (ECCV), 2014, [online], [Retrieved February 13, 2022], Internet<URL:https: / / link.springer.com / chapter / 10.1007 / 978-3-319-10593-2_13> [Non-Patent Document 2] KDDI, Utilizing "Location Information Big Data", [online], [Searched on February 13, 2022], Internet<URL:https: / / www.kddi.com / corporate / kddi / public / bigdata / > Summary of the Invention [Problem to be solved by the invention]
[0005] However, big data provided by mobile phone carriers is only provided in low-resolution mesh units (e.g., 250m square or 125m square) to prevent individuals from identifying their location. For example, in the case of population distribution, the numerical value for each mesh unit is the number of people (population), and in areas where large numbers of people move around, such as urban areas, it is difficult to recognize detailed human flows.
[0006] In response to this, the inventors of the present application considered whether it might be possible to recognize big data provided by mobile phone carriers as a high-resolution mesh map as much as possible on the client side, even if the big data is based on a low-resolution mesh map.
[0007] Therefore, according to the present invention, it is an object to provide an apparatus, a program, and a method that enable reproduction of a low-resolution mesh map based on big data from a business operator into a relatively accurate high-resolution mesh map. [Means for solving the problem]
[0008] According to the present invention, there is provided a mesh map generating device for generating a low-resolution mesh map from which a high-resolution mesh map can be reproduced, comprising: a first resolution reducing means for inputting a first high-resolution mesh map X1 in which the first information is distributed, and outputting a first low-resolution mesh map Y1 (=D1(X1)) in which a plurality of meshes are integrated into one representative mesh; a second resolution reduction means for inputting a second high-resolution mesh map X2 having the same range as the first high-resolution mesh map X1 and having second information distributed therein, and outputting a second low-resolution mesh map Y2 (=D2(X2)) in which a plurality of meshes are integrated into one representative mesh; a pseudo-high-resolution means for inputting a first low-resolution mesh map Y1 and a second low-resolution mesh map Y2 as teacher data in a layered manner (Y) and training the means to output a first high-resolution mesh map F(Y;θ) in a training stage, and inputting a first low-resolution mesh map Y1 and a second low-resolution mesh map Y2 as target data in an estimation stage in a layered manner (Y) and outputting a pseudo high-resolution mesh map F(Y;θ); The present invention is characterized by having the following.
[0009] According to another embodiment of the mesh map generating device of the present invention, The first information is an arbitrary numerical value for each mesh, The first resolution reduction means sets a representative value that is the sum of the numerical values of a plurality of meshes. It is also preferable.
[0010] According to another embodiment of the mesh map generating device of the present invention, The high-resolution mesh map and the low-resolution mesh map are based on the area on the map, The first high-resolution mesh map X1 in which the first information is distributed and the second high-resolution mesh map X2 in which the second information is distributed are in the same area range, The numbers in the first piece of information are the population of each grid on the map. It is also preferable.
[0011] According to another embodiment of the mesh map generating device of the present invention, The second information is an arbitrary category or a numerical value for each mesh, The second resolution reducing means sets a representative value derived based on a predetermined rule from the categories or values of the plurality of meshes. It is also preferable.
[0012] According to another embodiment of the mesh map generating device of the present invention, The second category of information is POI (Point of Interest) or coverage information for each mesh on the map. It is also preferable.
[0013] According to another embodiment of the mesh map generating device of the present invention, The pseudo-high-resolution method is SRCNN (Super-Resolution Convolutional Neural Network). It is also preferable.
[0014] According to another embodiment of the mesh map generating device of the present invention, The pseudo-resolution high-enhancement means sets a second loss function that reduces the error between the high-resolution mesh map X and the pseudo high-resolution mesh map F(Y;θ) output from the pseudo-resolution high-enhancement means. It is also preferable.
[0015] According to another embodiment of the mesh map generating device of the present invention, The second loss function is MSE (Mean Squared Error). It is also preferable.
[0016] According to another embodiment of the mesh map generating device of the present invention, a pseudo-high-resolution mesh map F(Y;θ) output from the pseudo-high-resolution means and a pseudo-low-resolution mesh map D1(F(Y;θ)) in which multiple meshes are combined into one representative mesh; Furthermore, The pseudo high-resolution means sets a first loss function that reduces the error between the low-resolution mesh map Y1=D1(X) and the pseudo low-resolution mesh map D1(F(Y;θ)). It is also preferable.
[0017] According to another embodiment of the mesh map generating device of the present invention, The first loss function is PE (Population Error). It is also preferable.
[0018] According to another embodiment of the mesh map generating device of the present invention, The pseudo-resolution improving means is set by multiplying the second loss function by a parameter α and the first loss function by a parameter 1-α. It is also preferable.
[0019] According to the present invention, there is provided a mesh map reproducing device for reproducing a pseudo high-resolution mesh map from the above-mentioned mesh map generating device, comprising: A pseudo-high resolution means for the mesh map generating device is provided, The pseudo high-resolution means inputs a first low-resolution mesh map Y1 and a second low-resolution mesh map Y2 in a layered manner (Y), and outputs a pseudo high-resolution mesh map F(Y;θ). It is also preferable.
[0020] According to the present invention, there is provided a program for causing a computer installed in a mesh map generating device to function, the program generating device generating a low-resolution mesh map capable of reproducing a high-resolution mesh map, the program comprising: a first resolution reducing means for inputting a first high-resolution mesh map X1 in which the first information is distributed, and outputting a first low-resolution mesh map Y1 (=D1(X1)) in which a plurality of meshes are integrated into one representative mesh; a second resolution reduction means for inputting a second high-resolution mesh map X2 having the same range as the first high-resolution mesh map X1 and having second information distributed therein, and outputting a second low-resolution mesh map Y2 (=D2(X2)) in which a plurality of meshes are integrated into one representative mesh; a pseudo-high-resolution means for inputting a first low-resolution mesh map Y1 and a second low-resolution mesh map Y2 as teacher data in a layered manner (Y) and training the means to output a first high-resolution mesh map F(Y;θ) in a training stage, and inputting a first low-resolution mesh map Y1 and a second low-resolution mesh map Y2 as target data in an estimation stage in a layered manner (Y) and outputting a pseudo high-resolution mesh map F(Y;θ); The present invention is characterized in that it causes a computer to function.
[0021] According to the present invention, there is provided a method for generating a high-resolution mesh map from a low-resolution mesh map, the method comprising the steps of: The device is a first step of inputting a first high-resolution mesh map X1 in which first information is distributed, and outputting a first low-resolution mesh map Y1 (=D1(X1)) in which multiple meshes are grouped together into one representative mesh, and inputting a second high-resolution mesh map X2 in which second information is distributed in the same range as the first high-resolution mesh map X1, and outputting a second low-resolution mesh map Y2 (=D2(X2)) in which multiple meshes are grouped together into one representative mesh; a training step in which a first low-resolution mesh map Y1 and a second low-resolution mesh map Y2 are input as teacher data in a layered configuration (Y) and trained to output a first high-resolution mesh map F(Y;θ); and a second step in which a first low-resolution mesh map Y1 and a second low-resolution mesh map Y2 are input as target data in a layered configuration (Y) and trained to output a pseudo high-resolution mesh map F(Y;θ) as an estimation step. The present invention is characterized by carrying out the following steps. Effect of the Invention
[0022] According to the mesh map generating device, program, and method of the present invention, even a low-resolution mesh map based on big data from a business operator can be regenerated into a relatively accurate high-resolution mesh map. [Brief description of the drawings]
[0023] [Figure 1] FIG. 1 is a system configuration diagram according to the present invention. [Diagram 2] FIG. 1 is an illustration showing multiple mesh maps based on geographical areas. [Diagram 3] FIG. 2 is a functional configuration diagram of a mesh map generating device according to the present invention. [Figure 4] FIG. 13 is an explanatory diagram showing a reduction in resolution for three mesh maps. [Diagram 5] FIG. 13 is an explanatory diagram illustrating a loss function. [Figure 6] FIG. 2 is a functional configuration diagram of a mesh map reproducing device according to the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0024] Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.
[0025] FIG. 1 is a diagram showing a system configuration according to the present invention.
[0026] The system in FIG. 1 includes a mesh map generating device 1 and a mesh map reproducing device 2. The mesh map generating device 1 and the mesh map reproducing device 2 include a mesh map generating device 1 and a mesh map reproducing device 2. [Mesh map generator 1] The mesh map generating device 1 may be a business facility that collects big data through a mobile phone business, processes the big data, and provides it to clients. Big data is, for example, location information based on region collected from a huge number of users' mobile terminals. This big data is collected regionally and temporally at a fine granularity and created as a high-resolution mesh map. Note that a mesh is a region on a map that is divided into four areas of a predetermined length. According to the present invention, a mesh map generating device 1 generates a pseudo low resolution mesh map from a region-based high resolution mesh map, which allows for a reproduction of the high resolution mesh map as close as possible. Then, the mesh map generating device 1 transmits the pseudo low resolution mesh map to the mesh map reproducing device 2.
[0027] [Mesh map playback device 2] The mesh map reproduction device 2 is managed by a client that uses the big data. According to the present invention, the mesh map reproduction device 2 receives a pseudo low resolution mesh map from the mesh map generation device 1 . Then, the mesh map reproducing device 2 reproduces the pseudo high resolution mesh map from the pseudo low resolution mesh map.
[0028] FIG. 2 is an explanatory diagram showing a number of mesh maps based on geographical areas.
[0029] According to FIG. 2, there are three mesh maps of the same area on the map, each based on a different perspective. First high-resolution mesh map X1: Population distribution map 21st high resolution mesh map X 21 :POI (Point Of Interest) distribution map 22nd high resolution mesh map X 22 :Coverage (map) map
[0030] In the first high-resolution mesh map X1, first information is distributed for each mesh (area range on a map). The first information is an arbitrary numerical value for each mesh, and may be, for example, the population (number of people) on the map. The population may be a normalized numerical value between 0 and 1. The first high-resolution mesh map may be created from point cloud data of user location information collected by a mobile phone company.
[0031] The second high-resolution mesh map X2 (the 21st high-resolution mesh map X 21 and 22nd high-resolution mesh map X 22 ) is the same area as the first high-resolution mesh map X1, and the second information is distributed therein. The second information is any category or numerical value for each mesh, and may be, for example, POI and / or coverage information on the map. A POI may be a category such as the architectural features of the area, for example, buildings (1), stores (2), parks (3), and shrines (4). The coverage information may be categories such as map features of the area, for example, rivers and lakes (1), residential areas (2), and road rights of way (3).
[0032] The relationship between the first high-resolution mesh map X1 and the second high-resolution mesh map X2 is expected to include semantic features such as "coverage information: in meshes that are rivers, the population is basically 0."
[0033] FIG. 3 is a functional configuration diagram of the mesh map generating device according to the present invention.
[0034] The mesh map generating device 1 generates a first high-resolution mesh map X1 and a second high-resolution mesh map X2 (the 21st high-resolution mesh map X 21 and 22nd high-resolution mesh map X 22 ) and output a pseudo low-resolution mesh map. 3, the mesh map generating device 1 has a first resolution reducing section 111, a second resolution reducing section 112 (a 21st resolution reducing section 1121 and a 22nd resolution reducing section 1122), a pseudo resolution high reducing section 12, and a pseudo resolution low reducing section 13. These functional components are realized by executing a program that causes a computer mounted in the mesh map generating device to function. The processing flow of these functional components can also be understood as a mesh map generating method.
[0035] FIG. 4 is an explanatory diagram showing the lowering of the resolution for three mesh maps.
[0036] The first resolution reduction section 111 and the second resolution reduction section 112 (the twenty-first resolution reduction section 1121 and the twenty-second resolution reduction section 1122) each reduce the resolution of numerical values and categories using different methods depending on the channel.
[0037] [First resolution reducing unit 111] The first resolution reducing unit 111 outputs a first low resolution mesh map Y1 (=D1(X1)) in which a plurality of meshes are combined into one representative mesh. First low-resolution mesh map Y1 First resolution reduction function D1 The first high-resolution mesh map X1 The first resolution reduction unit 111 sets a "representative value" that is the sum of the numerical values of multiple meshes. In the case of population distribution, the sum (total value) of the numerical values of multiple meshes in the high resolution mesh map is set as the representative value of one mesh of the low resolution mesh map. Of course, instead of using the total value, a value normalized between 0 and 1 may also be used.
[0038] [Second resolution reducing unit 112] (21st resolution reduction unit 1121 and 22nd resolution reduction unit 1122) The second resolution lowering section 112 also outputs a second low resolution mesh map Y2 (=D2(X2)) in which a plurality of meshes are combined into one representative mesh. Here, the second resolution lowering unit 112 sets a representative value derived based on a "predetermined rule" from the categories or numerical values of a plurality of meshes.
[0039] For example, when the POI distribution is reduced in resolution, categories may be prioritized. For example, the priorities may be set as follows: Building -> Store -> Other. In this case, if one of the four meshes in the high-resolution mesh map contains the category "Building", the category "Building" is always assigned to the one mesh compiled in the low-resolution mesh map. Also, if one mesh in the high-resolution mesh map does not contain the category "Building" but does contain the category "Store", the category "Store" is assigned to the one mesh compiled in the low-resolution mesh map.
[0040] For example, when reducing the resolution of coverage (map) information, categories may be prioritized. For example, the priorities may be set as follows: Rivers and lakes -> Road land -> Other. In this case, if one of the four meshes in the high-resolution mesh map contains the category "rivers", the one mesh compiled in the low-resolution mesh map will always be given the category "rivers". Also, if one mesh in the high-resolution mesh map does not contain the category "rivers" but does contain the category "buildings", the one mesh compiled in the low-resolution mesh map will be given the category "buildings".
[0041] Of course, even in the case of POI distribution and coverage information, it is not limited to prioritizing categories, and the most common category among the four meshes of the high-resolution mesh map may be assigned to one mesh grouped together in the low-resolution mesh map.
[0042] [Pseudo high resolution section 12] The pseudo-resolution increasing unit 12 may be the SRCNN described above as an existing technique (see, for example, Non-Patent Document 1). In addition, the pseudo-resolution increasing section 12 functions as follows. (Training stage) A first low resolution mesh map Y1 and a second low resolution mesh map Y2 are input in a layered configuration (Y) as teacher data, and training is performed to output a first high resolution mesh map F(Y;θ). F: Mapping function θ: weight parameter (Estimation stage) A first low-resolution mesh map Y1 and a second low-resolution mesh map Y2 are input as target data in a layered manner (Y), and a pseudo high-resolution mesh map F(Y;θ) is output.
[0043] In order to reproduce a pseudo population distribution as close as possible to the first high-resolution mesh map from the population distribution in the first low-resolution mesh map, a pseudo high-resolution mesh map is generated by layering not only the first low-resolution mesh map but also a wide variety of second low-resolution mesh maps covering the same area.
[0044] According to the pseudo-resolution enhancement unit 12 of the present invention, by stacking and simultaneously training not only a first low-resolution mesh map based on a first high-resolution mesh map, but also a second low-resolution mesh map based on a second high-resolution mesh map in the same range, it is expected that the super-resolution accuracy of the pseudo low-resolution mesh map can be improved to a pseudo high-resolution mesh map.
[0045] [Pseudo low resolution section 13] The pseudo low resolution unit 13 inputs the pseudo high resolution mesh map F(Y;θ) output from the pseudo high resolution unit 12, and outputs a pseudo low resolution mesh map D1(F(Y;θ)) that combines multiple meshes into one representative mesh. Resolution reduction function D1 (same as the resolution reduction function D1 of the first resolution reduction unit 111) The pseudo low resolution mesh map D1(F(Y;θ)) is transmitted to the mesh map reproduction device 2 so that a high resolution mesh map can be reproduced.
[0046] FIG. 5 is an explanatory diagram illustrating a loss function.
[0047] The pseudo high-resolution unit 12 sets a first loss function that reduces the error between the low resolution mesh map Y=D1(X) and the pseudo low resolution mesh map D1(F(Y;θ)). The first loss function may be, for example, PE (Population Error). PE(θ)=1 / n Σ i=1 n ||D1(F(Y i ;θ))-Y i || 2 θ: weight parameter at a given training point PE is a concept not found in existing CNNs. For example, unlike images, population distributions do not change in value before and after super-resolution. PE evaluates whether super-resolution has been achieved while preserving the original number of people (preserving the shape of the original population distribution). This provides feedback as to whether the pseudo low resolution mesh map D1(F(Y;θ)) has been estimated to be close to the multiple low resolution mesh maps Y.
[0048] Furthermore, the pseudo high resolution unit 12 sets a second loss function that reduces the error between the high resolution mesh map X and the pseudo high resolution mesh map F(Y;θ) output from the pseudo high resolution unit 12. The second loss function may be a Mean Squared Error (MSE). MSE(θ)=1 / n Σ i=1 n ||F(Y i ;θ)-X i || 2 θ: weight parameter at a given training point MSE is a loss function used in existing SRCNNs to evaluate whether fine meshes can be reproduced. This provides feedback as to whether the population distribution of the pseudo high-resolution mesh map has been estimated to be close to the population distribution of the high-resolution mesh map.
[0049] The pseudo-resolution improving unit 12 may set only the first loss function, may set only the second loss function, or may set both the first loss function and the second loss function. In this case, the pseudo-resolution improving unit 12 multiplies the second loss function by a parameter α and multiplies the first loss function by a parameter 1-α to set them together. L(θ)=α·MSE(θ)+(1-α)·PE(θ) α: The reflection ratio of the MSE loss function By combining PE, which evaluates a global area, with MSE, which evaluates a narrow area, it is expected that the population distribution in fine meshes can be accurately estimated while preserving the shape of the overall population distribution.
[0050] FIG. 6 is a functional block diagram of the mesh map reproducing device according to the present invention.
[0051] 6, the mesh map reproducing device 2 can reproduce a pseudo high resolution mesh map from the mesh map generating device. The mesh map reproducing device 2 has the pseudo high resolution unit 12 described above. The pseudo high-resolution unit 12 receives Y, which is a layered map of a low-resolution population distribution map, a POI distribution map, and a coverage map, and outputs a pseudo high-resolution mesh map F(Y;θ).
[0052] This means that even if the big data provided by a mobile phone carrier, for example, is a low-resolution mesh map of population distribution based on privacy considerations, the client can recognize it as a pseudo-high-resolution mesh map that is as close as possible to the original population distribution.
[0053] As described in detail above, according to the mesh map generating device, program, and method of the present invention, even a low-resolution mesh map based on big data from a business operator can be regenerated into a high-resolution mesh map with relatively high accuracy.
[0054] Furthermore, this will "promote the use of big data," which will make it possible to contribute to Goal 9 of the United Nations-led Sustainable Development Goals (SDGs), which is to "build resilient infrastructure, promote sustainable industrialization and foster innovation."
[0055] Regarding the various embodiments of the present invention described above, various changes, modifications, and omissions within the scope of the technical idea and perspective of the present invention can be easily made by those skilled in the art. The above description is merely an example and is not intended to be restrictive in any way. The present invention is limited only by the scope of the claims and their equivalents. [Explanation of symbols]
[0056] 1 Mesh map generator 111 First low-resolution section 112 Second low-resolution section 1121 21st low resolution section 1122 22nd low resolution section 12 Pseudo high resolution section 13 Pseudo low resolution section 14 Reverse pseudo low resolution section 2 Mesh map playback device
Claims
1. A mesh map generating device for generating a low-resolution mesh map from which a high-resolution mesh map can be reproduced, comprising: A first high-resolution mesh map X in which the first information is distributed 1 A first low-resolution mesh map Y 1 (=D 1 (X 1 ) a first resolution reducing means for outputting the image; First high-resolution mesh map X 1 A second high-resolution mesh map X in which the second information is distributed and which is the same range as 2 A second low-resolution mesh map Y is generated by combining multiple meshes into one representative mesh. 2 (=D 2 (X 2 ) a second resolution reducing means for outputting the second resolution reducing image; In the training stage, a first low-resolution mesh map Y 1 and a second low-resolution mesh map Y 2 and input them in a layered form (Y), and trained to output a first high-resolution mesh map F(Y;θ) and, in the estimation stage, a first low-resolution mesh map Y 1 and a second low-resolution mesh map Y 2 and a pseudo-high-resolution means for inputting the above in a layered manner (Y) and outputting a pseudo-high-resolution mesh map F(Y;θ). A mesh map generating device comprising:
2. The first information is an arbitrary numerical value for each mesh, The first resolution reduction means sets a representative value that is the sum of the numerical values of a plurality of meshes.
2. The mesh map generating device according to claim 1 .
3. The high-resolution mesh map and the low-resolution mesh map are based on the area on the map, A first high-resolution mesh map X in which the first information is distributed 1 and a second high-resolution mesh map X in which the second information is distributed. 2 means the same area; The first piece of information is the population of each grid on the map.
3. The mesh map generating device according to claim 2.
4. The second information is an arbitrary category or a numerical value for each mesh, The second resolution reducing means sets a representative value derived based on a predetermined rule from the categories or values of the plurality of meshes.
4. The mesh map generating device according to claim 1, wherein the mesh map generating device is a mesh map generating device.
5. The second category of information is POI (Point Of Interest) or coverage information for each mesh on the map.
5. The mesh map generating device according to claim 4.
6. The pseudo-resolution enhancement method is SRCNN (Super-Resolution Convolutional Neural Network).
6. The mesh map generating device according to claim 1,
7. The pseudo-resolution high-enhancing means sets a second loss function that reduces the error between the high-resolution mesh map X and the pseudo high-resolution mesh map F(Y;θ) output from the pseudo-resolution high-enhancing means.
7. The mesh map generating device according to claim 1,
8. The second loss function is the Mean Squared Error (MSE).
8. The mesh map generating device according to claim 7.
9. A pseudo high-resolution mesh map F(Y;θ) output from the pseudo high-resolution means is input, and a plurality of meshes are combined into one representative mesh to generate a pseudo low-resolution mesh map D. 1 A pseudo-resolution reducing means for outputting (F(Y;θ)) Furthermore, The pseudo-high resolution means is a low resolution mesh map Y 1 =D 1 (X) and pseudo low-resolution mesh map D 1 Set a first loss function that reduces the error between (F(Y;θ)) 9. The mesh map generating device according to claim 1,
10. The first loss function is PE (Population Error).
10. The mesh map generating device according to claim 9.
11. The pseudo-resolution increasing means multiplies the second loss function by a parameter α and multiplies the first loss function by a parameter 1-α to set them together.
11. The mesh map generating device according to claim 9 or 10.
12. A mesh map reproducing device for reproducing a pseudo high-resolution mesh map from the mesh map generating device according to any one of claims 1 to 11, comprising: A pseudo-high resolution means for the mesh map generating device is provided, The pseudo-high-resolution means is a first low-resolution mesh map Y 1 and a second low-resolution mesh map Y 2 and input them as a stack (Y), and output a pseudo high-resolution mesh map F(Y;θ). A mesh map reproducing device comprising:
13. A program for causing a computer installed in a mesh map generating device to function and generating a low-resolution mesh map capable of reproducing a high-resolution mesh map, comprising: A first high-resolution mesh map X in which the first information is distributed 1 A first low-resolution mesh map Y 1 (=D 1 (X 1 ) a first resolution reducing means for outputting the image; First high-resolution mesh map X 1 A second high-resolution mesh map X in which the second information is distributed and which is the same range as 2 A second low-resolution mesh map Y is generated by combining multiple meshes into one representative mesh. 2 (=D 2 (X 2 ) a second resolution reducing means for outputting the second resolution reducing image; In the training stage, a first low-resolution mesh map Y 1 and a second low-resolution mesh map Y 2 and input them in a layered form (Y), and trained to output a first high-resolution mesh map F(Y;θ) and, in the estimation stage, a first low-resolution mesh map Y 1 and a second low-resolution mesh map Y 2 and a pseudo-high-resolution means for inputting the above in a layered manner (Y) and outputting a pseudo-high-resolution mesh map F(Y;θ). A program for causing a computer to function.
14. 1. A method for generating a high-resolution mesh map from a low-resolution mesh map, comprising: The device is A first high-resolution mesh map X in which the first information is distributed 1 A first low-resolution mesh map Y 1 (=D 1 (X 1 )) and a first high-resolution mesh map X 1 A second high-resolution mesh map X in which the second information is distributed and which is the same range as 2 A second low-resolution mesh map Y is generated by combining multiple meshes into one representative mesh. 2 (=D 2 (X 2 )) as a first step; In the training stage, a first low-resolution mesh map Y 1 and a second low-resolution mesh map Y 2 and input them in a layered form (Y), and trained to output a first high-resolution mesh map F(Y;θ) and, in the estimation stage, a first low-resolution mesh map Y 1 and a second low-resolution mesh map Y 2 and a second step of inputting the above-mentioned data in a layered manner (Y) and outputting a pseudo high-resolution mesh map F(Y;θ). A method for generating a mesh map, comprising:
Citation Information
Patent Citations
Method, program and device for preparing detail statistical data
JP2004118853A
Image processing apparatus, image processing method, and program
JP2018084982A
Image processing apparatus and image processing method
JP2018195069A
