Multi-task learning building vector data simplification method

Abstract

The invention provides a multi-task learning building vector data simplification method, and relates to the field of map making synthesis, and the method comprises the steps: dividing a building simplification task into a node removal classification task and a node movement prediction task; making a building node data set; according to the building node data set and the point convolutional neural network, obtaining a removal classification result of each node of the building; for the nodes of which the removal classification results are mobile labels, predicting the movement of the nodes by using a multi-layer perceptron network to obtain a movement prediction result; and combining a movement prediction result with a removal classification result, and processing the vector data of the original building to obtain a final building simplification result. According to the method, the retention constraint condition of building simplification can be comprehensively considered while the building is simplified by using the network, so that the simplification accuracy is effectively improved.

Description

Technical Field

[0001] The present application relates to the field of map synthesis, and in particular to a multi-task learning method for simplifying building vector data. Background Art

[0002] Cartographic generalization is the process of abstracting and generalizing data when reducing large-scale spatial data to small-scale spatial data. It is an important method for spatial data conversion, integration and analysis. Buildings are an important part of cartographic generalization data. In the process of abstracting and generalizing building data, especially from large scale to medium scale, simplification of building outlines is considered the main operation. This process not only involves simplifying the shape of buildings, but also requires retaining their main features to ensure that the map is still accurate and readable at a smaller scale.

[0003] In order to simplify buildings, many algorithms were proposed in the early days of cartographic generalization, such as simplification algorithms based on least squares adjustment, simplification algorithms based on template matching, simplification methods based on local structure processing, etc., which take into account the geometric characteristics of building shapes, such as side length, number of sides, and building area. Although these algorithms are simple, intuitive, and easy to understand, and perform well in simplifying buildings, there is still a major problem: the threshold parameters of the algorithm need to be manually set, which leads to professional barriers in traditional algorithms, and users who lack cartographic experience find it difficult to use them to achieve simplification.

[0004] Therefore, more and more studies have begun to focus on building simplification tasks based on deep learning methods. These methods are more accurate and automated than traditional methods. For example, methods based on U-net, residual U-net, GAN network, etc. are used to process rasterized building data, and methods based on graph neural networks are used to directly process building vector data. Although these methods have good performance and higher automation, there are still some problems. First, the simplification method based on raster images will introduce additional noise when representing the edges of buildings, resulting in reduced simplification accuracy. Secondly, the simplification method based on graph neural networks is mostly to construct autoencoders and perform self-supervised learning, which has relatively low accuracy. Therefore, how to add supervised learning to obtain better accuracy when directly using vector building data for processing is a problem that still needs to be solved. Thirdly, the simplification method based on graph neural networks needs to construct a graph structure for each building and preprocess the input to extract node features, so it is necessary to consider streamlining the overall process.

[0005] Therefore, in the building vector data simplification technology, how to improve the simplification accuracy and simplify the entire simplification process while ensuring the degree of automation is a technical problem that needs to be solved urgently. Summary of the invention

[0006] The purpose of the present invention is to provide a multi-task learning method for simplifying building vector data in order to solve the technical problem of how to improve the accuracy of building simplification while ensuring the degree of automation, effectively utilize prediction point information, and simplify the entire simplification process.

[0007] The above-mentioned purpose of the present application is achieved through the following technical solutions: S1: Split the building simplification task into a node removal classification task and a node movement prediction task; S2: Obtain the vector data of the original building and the simplified vector data of the target building, and create a building node dataset; S3: Construct a point convolutional neural network; based on the building node dataset and the point convolutional neural network, obtain the removal classification results of each building node; S4: For each node of the building, remove the nodes whose classification results are mobile labels, use the multi-layer perceptron network to predict their movement, and obtain the movement prediction result; S5: Combine the movement prediction result with the removal classification result, process the vector data of the original building, and obtain the final building simplification result.

[0008] Optionally, step S1 includes: Through the node removal classification task, the building nodes of the original building are divided into three categories; the three categories include: removal, movement or keeping still; The node movement prediction task is used to predict building nodes that are classified as moving. Axis direction and The displacement prediction in the axial direction is expressed as and .

[0009] Optionally, step S2 includes: Align the vector data of the original building and the vector data of the target building through an iterative closest point registration algorithm, and calculate the distance between the mapped building nodes; According to the set distance threshold and the calculated distance, the building node removal classification dataset and the building node movement prediction dataset are obtained and merged into the building node dataset.

[0010] Optionally, step S2 includes: Align the vector data of the original building and the vector data of the target building through an iterative closest point registration algorithm, and calculate the distance between the mapped building nodes; According to the set distance threshold and the calculated distance, the building node removal classification dataset and the building node movement prediction dataset are obtained and merged into the building node dataset.

[0011] Optionally, step S3 includes: The building node dataset is input into the point convolutional neural network and is processed as follows: The coordinate data of building nodes were normalized using the Min-Max method; Extract each building node, the two nodes before the building node, and the two nodes after the building node to form a point list ; In the point list, the normalized coordinate data of each building node and the vector formed by the two points before and after it are added to construct the first input data; The point convolutional neural network includes: a shared MLP layer, a Max pooling layer, a FC layer and a softmax function; Normalization layer and ReLU activation function are added between each hidden layer of the shared MLP layer and each FC layer; The first input data is input into the point convolutional neural network to obtain the probability scores of the three categories of nodes; According to the possibility score, the category of the node is determined, that is, the classification results of each building node are removed.

[0012] Optionally, step S4 includes: The process of node movement prediction by multi-layer perceptron network is self-supervised learning. The loss function is constructed by using the three conservative constraints of building simplification as follows: The shape loss, size loss, position loss and total loss are calculated as:

[0013]

[0014]

[0015]

[0016] in, is the shape loss, Represents the current point number of the building. is the total number of nodes of the building, which is the point With point , +1 is the angle of the angle formed; Indicate point With point , +1 is the angle of the angle formed; is the size loss, and Respectively represent the building area before and after simplification; is the position loss, Represents the point number of the current movement of the building, is the total number of nodes moving in the building, and Respectively represent Points in Axis direction and The displacement distance of the axial movement; is the total loss of the point movement prediction task in the multi-layer perceptron network; , and Represents weight.

[0017] Optionally, step S5 includes: The final building simplification result is a three-dimensional vector; The first dimension of the three-dimensional vector is the removal classification result of each node of the building, and the removal classification results include: removal, movement or keeping still; The second and third dimensions of the three-dimensional vector are the results of the motion prediction task, which include: and .

[0018] An electronic device comprises a processor, a memory, a user interface and a network interface, wherein the memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory so that the electronic device performs a multi-task learning method for simplifying building vector data.

[0019] A computer-readable storage medium stores instructions, and when the instructions are executed, a multi-task learning building vector data simplification method is performed.

[0020] The beneficial effects of the technical solution provided by this application are: 1. The building simplification task is divided into two subtasks: node removal classification and node movement prediction. Both subtasks are supervised by examples. Therefore, through the building simplification task segmentation method, supervised learning is introduced on the premise of directly processing the building vector data, so that the model learns from examples to better simplify the building.

[0021] 2. The point convolutional neural network focuses on the features of the classified nodes and the points in their second-order neighborhood, including the information of the adjacent points related to the node classification results, while not causing the reduction of accuracy due to information redundancy, and effectively utilizing the context information of the classified nodes. Moreover, it directly convolves the features of the building-related points in units of points, reducing the dependence on the graph structure and avoiding the information loss problem caused by the preprocessing of a unified number of nodes, making node classification more accurate and efficient.

[0022] 3. Multilayer Perceptron Network The multilayer perceptron is constructed with a simple symmetrical structure, which can efficiently predict the movement of nodes; the loss function is constructed according to the preservation constraints of building simplification, maintaining the orthogonal characteristics of the building as much as possible, keeping the building area unchanged, and ensuring that the simplified building is close to its original position, so as to make the node movement prediction more accurate and reasonable. This technology can effectively integrate map making experience with deep learning networks, making the simplified building shape more in line with the needs of map making.

[0023] 4. The technology of simplifying building vector data based on point convolutional neural network and multi-layer perceptron network introduces supervised learning under the premise of directly processing vector data, takes into account map making experience, simplifies the overall network structure, supplements the shortcomings of existing deep learning technology applied to building vector simplification, simplifies building vector data more efficiently and accurately, and provides algorithm reference for automatic mapping comprehensive work in the region. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The present application will be further described below with reference to the accompanying drawings and embodiments, in which: Figure 1 is a step diagram in an embodiment of the present application; Figure 2 is an overall network structure diagram in an embodiment of the present application; Figure 3 It is a structural diagram of a point convolutional neural network in an embodiment of the present application; Figure 4 It is a structural diagram of a point convolutional neural network in an embodiment of the present application; Figure 5 This is a simplified result diagram of some buildings in the embodiment of the present application; Figure 6 It is a schematic diagram of the structure of an electronic device in an embodiment of the present application. DETAILED DESCRIPTION

[0025] In order to have a clearer understanding of the technical features, purposes and effects of the present application, the specific implementation methods of the present application are now described in detail with reference to the accompanying drawings.

[0026] An embodiment of the present application provides a multi-task learning method for simplifying building vector data.

[0027] Please refer to Figure 1 , Figure 1 : is a step diagram of a multi-task learning building vector data simplification method in an embodiment of the present application, comprising: S1: Split the building simplification task into a node removal classification task and a node movement prediction task; S2: Obtain the vector data of the original building and the simplified vector data of the target building, and create a building node dataset; S3: Construct a point convolutional neural network; based on the building node dataset and the point convolutional neural network, obtain the removal classification results of each building node; S4: For each node of the building, remove the nodes whose classification results are mobile labels, use the multi-layer perceptron network to predict their movement, and obtain the movement prediction result; S5: Combine the movement prediction result with the removal classification result, process the vector data of the original building, and obtain the final building simplification result.

[0028] In one embodiment, Figure 2 The overall network structure of the model of this method is shown.

[0029] Step S1 includes: Through the node removal classification task, the building nodes of the original building are divided into three categories; the three categories include: removal, movement or keeping still; The node movement prediction task is used to predict building nodes that are classified as moving. Axis direction and The displacement prediction in the axial direction is expressed as and .

[0030] As an embodiment, the node removal classification task is defined as determining the state of each node of the original building polygon and classifying it into one of the following three types: removal, movement, or remaining stationary, and the three states are represented by labels 0, 1, and 2, respectively; the node movement prediction task is defined as predicting the displacement of nodes with a classification label of 1 (i.e., need to be moved), and determining their displacements Δx and Δy in the x-axis and y-axis directions, both of which can be positive or negative. A positive value indicates that the node is predicted to move in the positive direction along the coordinate axis, and a negative value indicates that the node is predicted to move in the reverse direction along the coordinate axis.

[0031] Step S2 includes: Align the vector data of the original building and the vector data of the target building through an iterative closest point registration algorithm, and calculate the distance between the mapped building nodes; According to the set distance threshold and the calculated distance, the building node removal classification dataset and the building node movement prediction dataset are obtained and merged into the building node dataset.

[0032] As an embodiment, the building node dataset is derived from the original buildings and simplified building vector data of the selected study area. First, the buildings whose number of vertices in the original buildings is within the top 1% and their corresponding simplified buildings are eliminated. Some buildings that have not changed before and after simplification are randomly excluded to ensure that their number is balanced and close to 1:1 with the number of buildings that have changed after simplification. The building node dataset is derived from the original buildings and simplified building vector data of the selected study area, and the original building and the target building are aligned through the iterative closest point (ICP) registration algorithm, and the distance between the mapped nodes is calculated. The distance threshold is set to 0.1m, and nodes less than the threshold are marked as stationary, the remaining mapped nodes are marked as moving, and the unmapped nodes are marked as deleted. According to the set distance threshold, the building node removal classification dataset and the building node movement prediction dataset are obtained and merged into the building node dataset.

[0033] Step S3 includes: The building node dataset is input into the point convolutional neural network and is processed as follows: The coordinate data of building nodes were normalized using the Min-Max method; Extract each building node, the two nodes before the building node, and the two nodes after the building node to form a point list ; In the point list, the normalized coordinate data of each building node and the vector formed by the two points before and after it are added to construct the first input data; The present application provides an embodiment as follows: point to a node in the list , whose characteristics include , is the normalized coordinate data, For Node With Node The vectors and nodes formed and The vector composed.

[0034] The point convolutional neural network includes: a shared MLP layer, a Max pooling layer, a FC layer and a softmax function; Normalization layer and ReLU activation function are added between each hidden layer of the shared MLP layer and each FC layer; The first input data is input into the point convolutional neural network to obtain the probability scores of the three categories of nodes; According to the possibility score, the category of the node is determined, that is, the classification results of each building node are removed.

[0035] The present application provides an embodiment as follows: the network structure of the point convolutional neural network is as follows Figure 3 As shown in the figure, the data features contained in the dataset are sequentially input into the shared MLP layer, Max pooling layer, FC layer and softmax function of the point convolutional neural network. Normalization layer and ReLU activation function are added between each hidden layer of the shared MLP layer and each FC layer. The network is specifically divided into a deep point encoding module, a feature extraction module and a point type recognition module. Among them, the deep point encoding module introduces a shared MLP layer to embed all points in the point list into a high-dimensional space; the feature extraction module uses a maximum pooling layer to perform dimensionality reduction operations on each dimension of the five points, and aggregates the features of all points in the point list to obtain the deep feature representation of the point; the point type recognition module consists of an FC layer and a softmax function. The FC layer converts the high-dimensional point feature representation into 3 dimensions. Finally, the softmax function performs a softmax operation on the 3-dimensional representation to obtain the probability scores of the three categories of nodes to determine the category of the node.

[0036] The loss function of the point convolutional neural network is the cross entropy loss function:

[0037] in, is the loss of the classification task in the model, i represents the current point number, N is the total number of training points, K is the number of classification categories, is the true probability that point i belongs to category j, is the predicted probability that point i belongs to category j, and S is the softmax function.

[0038] In this example, in order to test the performance of the proposed building vector data simplification model, the building data of Stuttgart, Germany was selected as the object to illustrate the performance of the model and the simplification results. In order to comprehensively analyze the sensitivity of each structure of the point convolutional neural network and build the optimal network structure to achieve accurate classification, four evaluation methods were used to perform sensitivity analysis on each structure. They are Accuracy, Precision, Recall and F1 score.

[0039] First, the positive sample P (Positive) is recorded as 1, and the negative sample (Negative) is recorded as 0. TP (True Positive) means that the predicted value and the actual value are the same, both 1, that is, the classification is correct; FP (False Positive) means that the predicted value is 1 and the actual value is 0, that is, the prediction is wrong; FN (False Negative) means that the predicted value is 0 and the actual value is 1, that is, the prediction is wrong; TN (True Negative) means that the predicted value and the actual value are both 0, that is, the prediction is correct.

[0040] Accuracy is the probability of correct classification.

[0041]

[0042] Precision is the probability of correct classification among all samples classified as positive.

[0043]

[0044] Recall is the probability of correct classification among all positive samples.

[0045]

[0046] F1 score: Consider both Precision and Recall at the same time, and make them reach the highest level at the same time to find a balance.

[0047]

[0048] The sensitivity analysis of the receptive field of the point convolutional neural network is conducted from two aspects: the number of nodes in the input point list and the range of the vectors contained in each point in the point list. The results are shown in Table 1 and Table 2. The sensitivity analysis results of the convolution operator in the network are shown in Table 3. The point convolutional network has the best performance in the structure: the point list contains 5 nodes, the range of the vector contained in each point in the point list is 1, and the MLP convolution operator is used. Its accuracy reaches 88.13% and the F1 score also reaches 88.20%.

[0049] Table 1 Sensitivity analysis of the number of nodes in the point list

[0050] Table 2 Sensitivity analysis of vector range in point list

[0051] Table 3 Sensitivity analysis of convolution operator

[0052] Step S4 includes: The process of node movement prediction by multi-layer perceptron network is self-supervised learning. The loss function is constructed by using the three conservative constraints of building simplification as follows: The shape loss, size loss, position loss and total loss are calculated as:

[0053]

[0054]

[0055]

[0056] in, is the shape loss, Represents the current point number of the building. is the total number of nodes of the building, which is the point With point , +1 is the angle of the angle formed; Indicate point With point , +1 is the angle of the angle formed; is the size loss, and Respectively represent the building area before and after simplification; is the position loss, Represents the point number of the current movement of the building, is the total number of nodes moving in the building, and Respectively represent Points in Axis direction and The displacement distance of the axial movement; is the total loss of the point movement prediction task in the multi-layer perceptron network; , and Represents weight.

[0057] As an embodiment, the multi-layer perceptron network is a self-encoder-like structure, and its network structure is as follows: Figure 4 As shown in the figure. For nodes classified as moving in the node removal classification task, their two-dimensional coordinates are used as input data features. The data features are passed through the non-shared and symmetrically structured MLP layer. The two-dimensional coordinate features of the points are first encoded and upgraded to 1024 dimensions, and then decoded and reduced back to 2 dimensions to obtain the predicted displacement Δx in the x-axis direction and the predicted displacement Δy in the y-axis direction. Normalization layers and ReLU activation functions are added between each hidden layer.

[0058] As an example, in order to comprehensively analyze the sensitivity of the MLP layer of the multi-layer perceptron network and build the optimal network structure to achieve accurate prediction, the The average processing time of each building in the network is used as the evaluation criteria to conduct a sensitivity analysis on the number of hidden layers of the MLP layer, and the results are shown in Table 4. It can be seen that when the number of hidden layers of the MLP layer is 9, the network has the lowest time complexity while having a relatively excellent performance.

[0059] Table 4 Sensitivity analysis of convolution operator

[0060] Step S5 includes: The final building simplification result is a three-dimensional vector; The first dimension of the three-dimensional vector is the removal classification result of each node of the building, and the removal classification results include: removal, movement or keeping still; The second and third dimensions of the three-dimensional vector are the results of the motion prediction task, which include: and .

[0061] As an embodiment, in order to evaluate the performance of the proposed building vector data simplification model, three evaluation indicators are introduced: HD, TFD and IoU, which respectively evaluate the position error, shape difference and area difference between the simplified result of the proposed model and the target simplified result.

[0062] HD is the Hausdorff distance:

[0063] in, represents the HD between A and B, where a is the node belonging to building A and b is the node belonging to building B. is the Euclidean distance between nodes ab.

[0064] TFD is the Turning Function distance:

[0065] in, represents the TFD between A and B, is the Turning Function of Building A, It is the Turning Function of building B.

[0066] IoU is the intersection over union ratio:

[0067] in, is the IoU between A and B, is the area of ​​the intersection of A and B, is the area of ​​the union of A and B.

[0068] The performance evaluation results of the model method of the present invention are shown in Table 5. The average values ​​of HD and TFD are only 0.860 and 0.347 respectively, while the average value of IoU reaches 0.961, which proves that the model method of the present invention can well learn the building simplification knowledge and has excellent simplification accuracy.

[0069] Table 5 Performance evaluation results of the model method of the present invention

[0070] The results of simplifying some buildings using the method of the present invention are as follows: Figure 5 As shown in Figure 2, introducing supervised learning into the building simplification method and taking into account map making experience can make the simplification of building vector data more reasonable and accurate.

[0071] The present application also discloses an electronic device. Figure 6 , Figure 6 The electronic device 500 may include: at least one processor 501 , at least one network interface 504 , a user interface 503 , a memory 505 , and at least one communication bus 502 .

[0072] The communication bus 502 is used to realize the connection and communication between these components.

[0073] The user interface 503 may include a display screen, and the optional user interface 503 may also include a standard wired interface or a wireless interface.

[0074] The network interface 504 may optionally include a standard wired interface or a wireless interface (such as a WI-FI interface).

[0075] The present application also discloses a computer-readable storage medium storing a plurality of instructions suitable for a processor to load to execute the above-mentioned multi-task learning building vector data simplification method.

[0076] The above are only exemplary embodiments of the present disclosure and cannot be used to limit the scope of the present disclosure. That is, any equivalent changes and modifications made according to the teachings of the present disclosure are still within the scope of the present disclosure.

[0077] This application is intended to cover any variation, use or adaptation of the present disclosure, which follows the general principles of the present disclosure and includes common knowledge or customary technical means in the art not described in the present disclosure. The description and examples are to be regarded as exemplary only, and the scope and spirit of the present disclosure are defined by the claims.

Claims

1. A multi-task learning building vector data simplification method, characterized in that: The method comprises the following steps: S1: Split the building simplification task into a node removal classification task and a node movement prediction task; S2: Obtain the vector data of the original building and the simplified vector data of the target building, and create a building node dataset; S3: Construct a point convolutional neural network; According to the building node dataset and point convolutional neural network, the removal classification results of each building node are obtained; S4: For each node of the building, remove the nodes whose classification results are mobile labels, use the multi-layer perceptron network to predict their movement, and obtain the movement prediction result; S5: Combine the movement prediction result with the removal classification result, process the vector data of the original building, and obtain the final building simplification result.

2. A multi-task learning building vector data simplification method as claimed in claim 1, characterized in that: Step S1 includes: Through the node removal classification task, the building nodes of the original building are divided into three categories; the three categories include: removal, movement or keeping still; The node movement prediction task is used to predict building nodes that are classified as moving. Axis direction and The displacement prediction in the axial direction is expressed as and .

3. A multi-task learning building vector data simplification method as claimed in claim 1, characterized in that: Step S2 includes: Align the vector data of the original building and the vector data of the target building through an iterative closest point registration algorithm, and calculate the distance between the mapped building nodes; According to the set distance threshold and the calculated distance, the building node removal classification dataset and the building node movement prediction dataset are obtained and merged into the building node dataset.

4. A multi-task learning building vector data simplification method as claimed in claim 1, characterized in that: Step S3 includes: The building node dataset is input into the point convolutional neural network and is processed as follows: The coordinate data of building nodes were normalized using the Min-Max method; Extract each building node, the two nodes before the building node, and the two nodes after the building node to form a point list ; In the point list, the normalized coordinate data of each building node and the vector formed by the two points before and after it are added to construct the first input data; The point convolutional neural network includes: a shared MLP layer, a Max pooling layer, a FC layer and a softmax function; Normalization layer and ReLU activation function are added between each hidden layer of the shared MLP layer and each FC layer; The first input data is input into the point convolutional neural network to obtain the probability scores of the three categories of nodes; According to the possibility score, the category of the node is determined, that is, the classification results of each building node are removed.

5. A multi-task learning building vector data simplification method as claimed in claim 1, characterized in that: Step S4 includes: The process of node movement prediction by multi-layer perceptron network is self-supervised learning. The loss function is constructed by using the three conservative constraints of building simplification as follows: The shape loss, size loss, position loss and total loss are calculated as: in, is the shape loss, Represents the current point number of the building. is the total number of nodes of the building, which is the point With point , +1 is the angle of the angle formed; Indicate point With point , +1 is the angle of the angle formed; is the size loss, and Respectively represent the building area before and after simplification; is the position loss, Represents the point number of the current movement of the building, is the total number of nodes moving in the building, and Respectively represent Points in Axis direction and The displacement distance of the axial movement; is the total loss of the point movement prediction task in the multi-layer perceptron network; , and Represents weight.

6. A multi-task learning building vector data simplification method as claimed in claim 1, characterized in that: Step S5 includes: The final building simplification result is a three-dimensional vector; The first dimension of the three-dimensional vector is the removal classification result of each node of the building, and the removal classification results include: removal, movement or keeping still; The second and third dimensions of the three-dimensional vector are the results of the motion prediction task, which include: and .

7. An electronic device, characterized in that: The electronic device comprises a processor (501), a memory (505), a user interface (503) and a network interface (504), wherein the memory (505) is used to store instructions, the user interface (503) and the network interface (504) are used to communicate with other devices, and the processor (501) is used to execute the instructions stored in the memory (505) so that the electronic device executes the method according to any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores instructions, and when the instructions are executed by a computer, the method according to any one of claims 1 to 6 is executed.

Images