CN111027574A - Building mode identification method based on graph convolution - Google Patents
Building mode identification method based on graph convolution Download PDFInfo
- Publication number
- CN111027574A CN111027574A CN201911251240.8A CN201911251240A CN111027574A CN 111027574 A CN111027574 A CN 111027574A CN 201911251240 A CN201911251240 A CN 201911251240A CN 111027574 A CN111027574 A CN 111027574A
- Authority
- CN
- China
- Prior art keywords
- building
- graph convolution
- buildings
- graph
- recognition method
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
- G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
- G06V10/464—Salient features, e.g. scale invariant feature transforms [SIFT] using a plurality of salient features, e.g. bag-of-words [BoW] representations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/285—Selection of pattern recognition techniques, e.g. of classifiers in a multi-classifier system
Abstract
The invention provides a building mode identification method based on graph convolution, which comprises the following steps: dividing building data into preset blocks, and establishing an adjacency relation in the preset blocks by utilizing a Delaunay triangulation network; calculating a building feature description index set to obtain a building feature description index value set matrix; respectively reducing the dimension of the description index characteristics to obtain a building characteristic value vector; and marking a building mode label, and adopting a semi-supervised graph convolutional network identification mode to obtain a result. The building pattern recognition method based on graph convolution only needs less labeled data to accurately find the building group pattern, effectively overcomes the defects of difficult parameter setting, unstable recognition result and low precision of the current building group recognition method, and can be popularized and applied to the fields of map synthesis, space data updating, city planning, landscape design and the like.
Description
Technical Field
The invention relates to the field of cartography and geographic information systems, in particular to a building mode identification method based on graph convolution
Background
The building group model can be used for spatial data updating, city planning, city landscape design and map drawing synthesis. How to efficiently and accurately identify the building mode is a current urgent need.
The current methods for identifying building groups can be divided into four categories, namely: a natural principles based approach, a partition based approach, a graph based approach and a region merging approach.
The method based on natural principles comprises the following steps: the distance between buildings is usually measured as the minimum distance, and the grouping operation is guided by natural principles, defining "physical limits" as the "minimum visible target" related to the resolution of the human eye. The size of the "smallest visible target" is then used as a "separation threshold" to measure the proximity between buildings, and adjacent buildings that cannot be visually separated when scaled down will be merged into a group. The natural principles-based approach only considers the proximity of the trellis grouping criterion and appears too one-sided.
The partition-based method represents buildings by building center of mass, and the distance between buildings is the distance between centers of mass. A building complex is defined as a series of buildings, each of which is closer to the central building representing the building complex than to the central buildings of the other building complexes. The partition-based method uses the centroid to represent the buildings, only the centroid distance between the buildings is considered, and the attributes of the buildings such as the shape and the size are ignored, and the partition-based method is too unilateral.
The graph-based approach first detects the neighborhood between buildings using neighborhood graphs such as minimum binary trees, delaunay triangulation, nearest neighbor graphs, etc. Two strategies are then employed to detect the building groups. The strategy is a separation method, edges with the length larger than the average value of the connection lengths in the graph are separated, and the rest graph is the building group. The second strategy is a merging method, wherein two buildings are merged according to the length (only when the proximity is considered) between the buildings in the building adjacency graph included in the building neighborhood graph. Graph-based approaches take into account other constraints besides proximity, but the choice of the separation threshold size remains a problem.
The zone merging based approach takes into account the proximity, similarity and continuity of buildings. Each building is first treated as a single group, and then whether the single group and the adjacent group can be combined is judged, and whether the metric values determined by the similarity, the continuity and the proximity among the buildings meet the set threshold value is the principle of combination. Although the method based on the region merging takes into consideration the adjacency, similarity, and continuity of buildings, there is no objective unified standard for setting the threshold value and determining the weight of each property.
Disclosure of Invention
The invention provides a building mode identification method based on graph convolution, and aims to overcome the defects of difficulty in parameter setting, unstable identification result and low precision of the conventional building group identification method.
In order to achieve the above object, an embodiment of the present invention provides a building pattern recognition method based on graph convolution, including:
step 3, respectively reducing the dimension of the description index features to obtain building feature value vectors;
and 4, marking the building mode label, and adopting a semi-supervised graph convolutional network identification mode to obtain a result.
Wherein, the step 1 specifically comprises:
dividing buildings by using an urban road network and rivers to obtain disconnected blocks;
inserting points at approximate equal intervals on the vector edge of the building, and respectively constructing Delaunay triangulation network for the points in turn in the disconnected blocks.
Wherein the step of constructing the delaunay triangulation network comprises:
and removing the triangulation network connecting the building and the block edge, the triangulation network connecting the block edge, the triangulation network inside the building and the triangulation network connecting more than two buildings, and reserving the triangulation network only connecting the two buildings to obtain the adjacency relation of the buildings.
Wherein, the step 2 specifically comprises:
and calculating the corresponding values of 23 indexes in the total set of all the building characteristic description indexes.
Wherein, the step 3 specifically comprises:
and respectively performing feature dimensionality reduction on the multi-dimensional description indexes under each type of feature by adopting a principal component analysis method from the three aspects of shape, area and direction, and respectively selecting the first three components with the largest contribution to form a feature value vector of the building.
Wherein, the step 4 specifically comprises:
marking a training data set of linear, grid and high-density building group modes based on lattice tower grouping cognition;
linear, grid and high-density group modes are discovered from a spatial graph structure constructed by a building by adopting a semi-supervised graph convolution network.
Wherein, the input of the semi-supervised graph convolutional network comprises a characteristic value vector of the buildings, the adjacency relation of each building and a mode label corresponding to a single building.
The scheme of the invention has the following beneficial effects:
the building pattern recognition method based on graph convolution in the embodiment of the invention can accurately discover the building group pattern only by needing less labeled data, effectively overcomes the defects of difficult parameter setting, unstable recognition result and low precision of the current building group recognition method, and can be popularized and applied to the fields of map synthesis, space data updating, city planning, landscape design and the like.
Drawings
FIG. 1 is a flow chart of an embodiment of the building pattern extraction method based on graph convolution according to the present invention;
FIG. 2 is a diagram of the original Delaunay triangulation network;
FIG. 3 is a Delaunay triangulation network containing adjacency relationships.
Detailed Description
In order to make the technical problems, technical solutions and advantages of the present invention more apparent, the following detailed description is given with reference to the accompanying drawings and specific embodiments.
The invention provides a building mode identification method based on graph convolution, aiming at the problems of difficult parameter setting, unstable identification result and low precision of the existing building group identification method.
As shown in fig. 1, an embodiment of the present invention provides a building pattern recognition method based on graph convolution, including:
step 3, respectively reducing the dimension of the description index features to obtain building feature value vectors;
and 4, marking the building mode label, and adopting a semi-supervised graph convolutional network identification mode to obtain a result.
Wherein, the step 1 specifically comprises:
dividing buildings by using an urban road network and rivers to obtain disconnected blocks;
inserting points at approximate equal intervals on the vector edge of the building, and respectively constructing Delaunay triangulation network for the points in turn in the disconnected blocks.
Wherein the step of constructing the delaunay triangulation network comprises:
and removing the triangulation network connecting the building and the block edge, the triangulation network connecting the block edge, the triangulation network inside the building and the triangulation network connecting more than two buildings, and reserving the triangulation network only connecting the two buildings to obtain the adjacency relation of the buildings.
The method of the present invention divides buildings into disconnected blocks by using city roads and river networks, wherein the triangulation network is established by inserting points on the sides of the vector data buildings at approximate equal intervals by means of interpolation, and then constructing Delaunay triangulation networks for the points, as shown in FIG. 2, ① represents triangulation networks for connecting block boundaries and buildings, ② represents triangulation networks in building outlines, ③ represents triangulation networks for connecting more than two buildings, ④ represents triangulation networks for connecting two buildings, ⑤ represents triangulation networks for connecting block boundaries, but not all triangulation networks can be used for expressing spatial adjacency relations between buildings.
Wherein, the step 2 specifically comprises:
and calculating the corresponding values of 23 indexes in the total set of all the building characteristic description indexes.
Wherein, the step 3 specifically comprises:
and respectively performing feature dimensionality reduction on the multi-dimensional description indexes under each type of feature by adopting a principal component analysis method from the three aspects of shape, area and direction, and respectively selecting the first three components with the largest contribution to form a feature value vector of the building.
The method described in the above embodiment of the present invention calculates all values of 23 indexes of 3 aspects of features of all buildings to obtain a building feature description index value set matrix, which is used as an input of a network and needs to have a feature value vector describing spatial features of the buildings. A single building has own spatial characteristics and can be described by different indexes, often the same spatial characteristics can be described by different indexes, for example, the size characteristics of the building can be represented by the indexes such as perimeter, area, circumscribed rectangle area and the like. A building feature description index set is summarized from the conventional literature about describing the spatial features of buildings, and comprises 23 indexes including 3 features of the size, the shape and the orientation of the building. However, the characteristics of the building described by all indexes are redundant, so the indexes are subjected to dimension reduction by a principal component analysis method. The principal component analysis converts a group of variables with correlation into a group of linear uncorrelated variables, effectively reduces the dimension of the building space characteristic index, and determines the principal component according to the accumulation of the contribution rate of the principal component. Of all the indexes describing the 3 features of the building, there are 5 indexes describing the size feature, 15 indexes describing the shape feature, and 3 indexes describing the direction feature, as shown in table 1. And respectively carrying out principal component analysis on the indexes of the three characteristics, and respectively selecting the first three principal components to be combined together to be used as a characteristic value vector for describing the characteristics of the building.
TABLE 1
Wherein, the step 4 specifically comprises:
marking a training data set of linear, grid and high-density building group modes based on lattice tower grouping cognition;
linear, grid and high-density group modes are discovered from a spatial graph structure constructed by a building by adopting a semi-supervised graph convolution network.
The method according to the above embodiment of the present invention summarizes the building graph relationship as follows: adjacency, similarity, continuity, association, connectivity, and common area, as described in detail below: adjacency: the close objects are more likely to belong to a group. Similarity: buildings of similar size, shape, orientation are more likely to belong to a group. Continuity: similarity building series are more likely to belong to a group. Relevance: buildings that are close together or overlap are more likely to belong to a group. Connectivity: a building that is part of a closed graph is more likely to belong to a group. Public area: buildings in the same area are more likely to belong to a group. Since roads and rivers have divided buildings into blocks. Relevance and connectivity can be replaced by proximity, just with a focus on buildings within the same neighborhood. Therefore, proximity, similarity, continuity are the most dominant principles of grouping. Previous pattern acquisition methods relied on various parameters and algebraic operations under the packet format tower described above. A large number of building group modes can be marked in advance based on the recognition of the format tower of the building group, and then the identification rule is implicitly learned by utilizing a semi-supervised graph convolution network.
The invention divides regular building group modes into a linear mode and a grid mode, and a high-density mode belongs to an irregular mode. A linear pattern is typically a set of buildings that are similar in shape, size, and principal direction, and are visually aligned. The grid patterns are arranged in a combination of linear patterns which are parallel to each other in two groups and perpendicular to each other between the groups. The high density model usually occurs in the urban village of the Chinese urbanization process, which has numerous buildings with different sizes, shapes and directions and scattered randomly. These three modes are assigned to different tags in the present invention to mark buildings.
Wherein, the input of the semi-supervised graph convolutional network comprises a characteristic value vector of the buildings, the adjacency relation of each building and a mode label corresponding to a single building.
The semi-supervised graph convolutional network according to the above embodiment of the present invention mainly aims at graph structure data, and the building arrangement is a network structure, that is, a topological graph in graph theory. The topological graph has the following characteristics: node characteristics: each node has its own characteristics; the structure is characterized in that: the nodes connected in the graph data have certain contact with the nodes; the graph convolution neural network can automatically learn node characteristics and association information between nodes, each node is a building, each edge contains the topological relation of every two buildings, and the core idea of graph convolution is to use the information of the edges to aggregate node information to extract the spatial characteristics of a topological graph.
The core part of the graph convolution network is as follows: the building graph structure has N nodes, each node has its own features, the features of the nodes form a matrix X with dimensions of N X D, and then the relationship between the nodes also forms an adjacent matrix A with dimensions of N X N. X and A are the inputs to our model. The propagation mode formula (1) between the graph convolution network layers:
H(l+1)＝σ(B-1/2CB-1/2H(l)W(l)) (1)
wherein:
c ═ C + E, E is the identity matrix;
b is the degree matrix of C, which is related to formula (2):
Bii＝∑jCii(2)
h is a characteristic of each layer, and H is X for the input layer;
and D is the eigenvalue vector of the selected building.
While the foregoing is directed to the preferred embodiment of the present invention, it will be understood by those skilled in the art that various changes and modifications may be made without departing from the spirit and scope of the invention as defined in the appended claims.
Claims (7)
1. A building pattern recognition method based on graph convolution is characterized by comprising the following steps:
step 1, dividing building data into preset blocks, and establishing an adjacency relation in the preset blocks by utilizing a Delaunay triangulation network;
step 2, calculating a building feature description index set to obtain a building feature description index value set matrix;
step 3, respectively reducing the dimension of the description index features to obtain building feature value vectors;
and 4, marking the building mode label, and adopting a semi-supervised graph convolutional network identification mode to obtain a result.
2. The graph convolution-based building pattern recognition method according to claim 1, wherein the step 1 specifically includes:
dividing buildings by using an urban road network and rivers to obtain disconnected blocks;
inserting points at approximate equal intervals on the vector edge of the building, and respectively constructing Delaunay triangulation network for the points in turn in the disconnected blocks.
3. The graph convolution-based building pattern recognition method of claim 2, wherein the step of constructing a delaunay triangulation network includes:
and removing the triangulation network connecting the building and the block edge, the triangulation network connecting the block edge, the triangulation network inside the building and the triangulation network connecting more than two buildings, and reserving the triangulation network only connecting the two buildings to obtain the adjacency relation of the buildings.
4. The graph convolution-based building pattern recognition method according to claim 2, wherein the step 2 specifically includes:
and calculating the corresponding values of 23 indexes in the total set of all the building characteristic description indexes.
5. The graph convolution-based building pattern recognition method according to claim 4, wherein the step 3 specifically includes:
and respectively performing feature dimensionality reduction on the multi-dimensional description indexes under each type of feature by adopting a principal component analysis method from the three aspects of shape, area and direction, and respectively selecting the first three components with the largest contribution to form a feature value vector of the building.
6. The graph convolution-based building pattern recognition method according to claim 5, wherein the step 4 specifically includes:
marking a training data set of linear, grid and high-density building group modes based on lattice tower grouping cognition;
linear, grid and high-density group modes are discovered from a spatial graph structure constructed by a building by adopting a semi-supervised graph convolution network.
7. The graph convolution-based building pattern recognition method of claim 6, wherein the input of the semi-supervised graph convolution network comprises a feature value vector of buildings, each building adjacency and a pattern label corresponding to a single building.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911251240.8A CN111027574A (en) | 2019-12-09 | 2019-12-09 | Building mode identification method based on graph convolution |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911251240.8A CN111027574A (en) | 2019-12-09 | 2019-12-09 | Building mode identification method based on graph convolution |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN111027574A true CN111027574A (en) | 2020-04-17 |
Family
ID=70205014
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201911251240.8A Pending CN111027574A (en) | 2019-12-09 | 2019-12-09 | Building mode identification method based on graph convolution |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111027574A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112002012A (en) * | 2020-08-26 | 2020-11-27 | 中南大学 | Visibility analysis method for urban area |
| CN112632535A (en) * | 2020-12-18 | 2021-04-09 | 中国科学院信息工程研究所 | Attack detection method and device, electronic equipment and storage medium |
| CN112632652A (en) * | 2020-12-31 | 2021-04-09 | 上海孚典智能科技有限公司 | Building CAD model analysis method based on data mining |
| CN113486135A (en) * | 2021-07-27 | 2021-10-08 | 中南大学 | Building comprehensive method based on deep learning network |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108305136A (en) * | 2017-01-13 | 2018-07-20 | 周现荣 | A kind of Network trade platform of combination numerical map |
-
2019
- 2019-12-09 CN CN201911251240.8A patent/CN111027574A/en active Pending
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108305136A (en) * | 2017-01-13 | 2018-07-20 | 周现荣 | A kind of Network trade platform of combination numerical map |
Non-Patent Citations (2)
| Title |
|---|
| XIONGFENG YANG 等: "A graph convolutional neural network for classification of building patterns using spatial vector data" * |
| 孙建东 等: ""《Semi-Supervised Classification with Graph Convolutional Networks》阅读笔记" * |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112002012A (en) * | 2020-08-26 | 2020-11-27 | 中南大学 | Visibility analysis method for urban area |
| CN112002012B (en) * | 2020-08-26 | 2022-07-08 | 中南大学 | Visibility analysis method for urban area |
| CN112632535A (en) * | 2020-12-18 | 2021-04-09 | 中国科学院信息工程研究所 | Attack detection method and device, electronic equipment and storage medium |
| CN112632535B (en) * | 2020-12-18 | 2024-03-12 | 中国科学院信息工程研究所 | Attack detection method, attack detection device, electronic equipment and storage medium |
| CN112632652A (en) * | 2020-12-31 | 2021-04-09 | 上海孚典智能科技有限公司 | Building CAD model analysis method based on data mining |
| CN113486135A (en) * | 2021-07-27 | 2021-10-08 | 中南大学 | Building comprehensive method based on deep learning network |
| CN113486135B (en) * | 2021-07-27 | 2024-04-12 | 中南大学 | Building comprehensive method based on deep learning network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111027574A (en) | Building mode identification method based on graph convolution | |
| CN107657226B (en) | People number estimation method based on deep learning | |
| Jiang et al. | Saliency detection via absorbing markov chain | |
| CN108428015B (en) | Wind power prediction method based on historical meteorological data and random simulation | |
| CN112257597B (en) | Semantic segmentation method for point cloud data | |
| CN106600617A (en) | Method of extracting building contour line from Lidar point cloud data based on curvature | |
| CN110008602B (en) | Road network selection method considering multi-feature coordination under large scale | |
| CN108961403A (en) | A method of major trunk roads are extracted with open source street map | |
| CN103136540B (en) | A kind of Activity recognition method based on implicit structure reasoning | |
| CN105046714A (en) | Unsupervised image segmentation method based on super pixels and target discovering mechanism | |
| CN112085072A (en) | Cross-modal retrieval method of sketch retrieval three-dimensional model based on space-time characteristic information | |
| CN107818338B (en) | Method and system for building group mode identification facing map synthesis | |
| CN111858810A (en) | Modeling elevation point screening method for road DEM construction | |
| CN112270266B (en) | Multi-feature constrained mesh river mainstream identification method | |
| CN104361135A (en) | Image search method | |
| Hu et al. | Geometric feature enhanced line segment extraction from large-scale point clouds with hierarchical topological optimization | |
| CN115757604A (en) | GDP (generalized projection) space-time evolution analysis method based on noctilucent image data | |
| CN113988558B (en) | Power grid dynamic security assessment method based on blind area identification and electric coordinate system expansion | |
| CN111667373A (en) | Evolution community discovery method based on neighbor subgraph social network dynamic increment | |
| CN112116709B (en) | Terrain characteristic line processing method for improving terrain expression precision | |
| AU2021102429A4 (en) | Method for selecting roads in a small-mesh accumulation area | |
| CN106611418A (en) | Image segmentation algorithm | |
| CN109033746A (en) | A kind of protein complex recognizing method based on knot vector | |
| CN110728688B (en) | Energy optimization-based three-dimensional mesh model segmentation method and system | |
| CN111291662B (en) | Vehicle-mounted point cloud rod-shaped object extraction method based on plane selection |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |