Building simplification method and system based on conditional graph diffusion model

Abstract

The invention relates to the comprehensive technical field of map making, in particular to a building simplification method and system based on a condition graph diffusion model, noise data are input into a pre-trained diffusion model, building graph structure data to be processed are used as generation conditions to be injected into the diffusion model, and the generation conditions are generated; enabling the diffusion model to carry out the denoising of the noise data based on the generation condition, and obtaining the simplified data of the structure of the building drawing, the diffusion model takes building graph structure data of a first scale as a generation condition and building graph structure data of a second scale as a generation target in a training process, correlation characteristics between noisy target data and condition data are learned through forward noise adding operation, and the first scale is larger than the second scale. According to the method, the generation type simplification is carried out by understanding the overall and detail morphological characteristics of the building based on the diffusion model, and the overall quality of building simplification in drawing is improved.

Description

Technical Field

[0001] This invention relates to the field of map production technology, and in particular to a method and system for simplifying buildings based on a conditional graph diffusion model. Background Technology

[0002] Existing building simplification techniques mainly include traditional simplification algorithms and deep learning algorithms. Traditional simplification algorithms include template matching, rectangle fitting, and four-nearest-neighbors, but these algorithms rely heavily on manual parameter adjustments, resulting in poor stability and low automation. Deep learning algorithms can be further divided into two technical paths. One is based on image maps, converting buildings into raster format, transforming the simplification task into an image transformation task. Methods based on Generative Adversarial Networks (GANs) and Backpropagation Neural Networks (BPNNs) have lower geometric accuracy and require post-processing to fully meet the requirements of vector data. The second approach is to directly simplify based on vector data, ensuring accuracy. The main approaches include: formally defining the simplification task as two related sub-tasks, "node removal" and "node movement," and utilizing multi-task graph convolutional neural networks. However, these methods essentially rely on simplification through local geometric operations, lacking perception and planning of the overall building form. Another approach is to simplify the building coordinate sequence using the Transformer model, but this sequence modeling method loses the two-dimensional structural information of the building. Moreover, due to its unsupervised nature, the model can only achieve simplified reconstruction of the input data, making it difficult to imitate the visual thinking relied upon by cartographic experts during the simplification process. This makes it impossible to effectively maintain the hierarchy of morphological features, such as typicalizing continuous concave and convex structures or reasonably exaggerating key features. Summary of the Invention

[0003] To address the problem of unsatisfactory building simplification in existing integrated drawing, this invention provides a building simplification method and system based on a conditional graph diffusion model. By utilizing the conditional graph diffusion model to achieve building simplification, this invention solves the problem of visually generative simplification of buildings.

[0004] According to the design scheme provided by the present invention, on the one hand, a building simplification method based on a conditional graph diffusion model is provided, comprising:

[0005] Acquire the structural data and noise data of the building to be processed;

[0006] Noisy data is input into a pre-trained diffusion model, and the structural data of the building map to be processed is injected into the diffusion model as a generation condition. The diffusion model then denoises the noisy data based on the generation condition to obtain simplified structural data of the building map. During the training process, the diffusion model uses structural data of the building map at a first scale as the generation condition and structural data of the building map at a second scale as the generation target. It also learns the correlation features between the noisy target data and the condition data through a forward noise addition operation. The first scale is larger than the second scale.

[0007] As a building simplification method based on the conditional graph diffusion model of the present invention, the diffusion model further includes: a time encoder for providing time guidance for embedding features in the model denoising operation; a conditional encoder for feature encoding and mapping the structural data of the second scale building map; a feature encoder for feature encoding the noisy input data; a feature stitching layer for stitching features; a fusion layer for fusing local geometric features and global structural features; and an output layer for predicting noise based on the fused features and outputting the result. The fusion layer adopts an alternating stacked structure of ring convolutional blocks and attention blocks.

[0008] As a building simplification method based on the conditional graph diffusion model of the present invention, the time encoder further employs sine and cosine position encoding for time embedding; the annular convolutional block employs multiple parallel feature extraction paths to extract geometric features at different scales; and the attention block employs a standard multi-head attention mechanism and dynamically adjusts the influence intensity between vertices of the building graph structure according to geometric importance to capture long-range dependencies.

[0009] As a building simplification method based on the conditional graph diffusion model of this invention, the training process of the diffusion model further includes:

[0010] Construct a building drawing structure data sample, which includes a first-scale building drawing structure data sample and a second-scale building drawing structure data sample obtained by manually simplifying the first-scale building drawing structure data sample.

[0011] The building structure data in the sample is preprocessed to obtain unified building structure data. The preprocessing includes: for buildings with fewer contour nodes than the target length, inserting new nodes between adjacent vertices of the building contour using a uniform interpolation method based on arc length so that the building contour nodes meet the target length; for buildings with more contour nodes than the target length, performing Douglas simplification on the building contours so that the number of contour nodes of each building is uniform.

[0012] Using preprocessed first-scale building map structural data samples as input and corresponding preprocessed second-scale building map structural data samples as output, the probability distribution of the transformation from the first-scale building map structural geometry to the second-scale building map structural morphology is learned through a noise prediction task and based on minimizing the noise prediction error.

[0013] As a building simplification method based on the conditional graph diffusion model of the present invention, further, during the diffusion model training, Gaussian noise is gradually added to the structural data samples of the building map at the second scale, and the probability distribution of the noise data obtained after T steps of noise addition is obtained based on the Markov chain, wherein a random time step sampling strategy is used to process the noise level of each time step.

[0014] As a building simplification method based on the conditional graph diffusion model of the present invention, the random time step sampling strategy further includes: for each pair of training data, randomly sampling a time step, using reparameterization to obtain the noise data at the corresponding time step, and causing the model to perform a forward propagation once to predict the added noise.

[0015] As a building simplification method based on a conditional graph diffusion model of the present invention, further, the structural data of the building graph to be processed is injected into the diffusion model as a generation condition, so that the diffusion model denoises the noisy data based on the generation condition, including:

[0016] Starting with noisy data, based on the structural data of the first-scale building map, the structural data of the second-scale building map is gradually recovered through a Markov process. The noisy data is pure noise data, and the denoising process in each step learns a conditional Gaussian distribution.

[0017] Furthermore, this invention also provides a building simplification system based on a conditional graph diffusion model, comprising: a data acquisition module and a simplification output module, wherein,

[0018] The data acquisition module is used to acquire the structural data and noise data of the building drawings to be processed;

[0019] The simplified output module is used to input noisy data into a pre-trained diffusion model, and to inject the building map structure data to be processed as the generation condition into the diffusion model, so that the diffusion model can denoise the noisy data based on the generation condition to obtain simplified building map structure data. In the training process, the diffusion model uses building map structure data at a first scale as the generation condition and building map structure data at a second scale as the generation target, and learns the correlation features between the noisy target data and the condition data through forward noise addition operation. The first scale is larger than the second scale.

[0020] The beneficial effects of this invention are:

[0021] This invention converts both large-scale and small-scale buildings into graph structure data, constructs a conditional graph diffusion model, and trains it using large-scale buildings as conditional data and small-scale buildings as target data. When a new large-scale building is input (as a condition), the trained model generates a corresponding simplified small-scale building. This overcomes the shortcomings of existing algorithms, such as the inability to simplify based on the overall shape, loss of two-dimensional structural information, and difficulty in imitating the visual thinking relied upon by cartographic experts during the simplification process, thereby improving the overall quality of building simplification in cartography. Attached image description:

[0022] Figure 1 This is a schematic diagram of the building simplification process based on the conditional graph diffusion model in the embodiment;

[0023] Figure 2 This is a schematic diagram of the building simplification algorithm framework based on the conditional graph diffusion model in the embodiment;

[0024] Figure 3 This is a schematic representation of the building structure in the embodiment;

[0025] Figure 4 This is a schematic diagram of the diffusion process in the conditional diagram of the embodiment;

[0026] Figure 5 This is a schematic diagram illustrating the working principle of the multi-scale circular convolutional block and the global attention block in the embodiment;

[0027] Figure 6 This is an example of circular convolution in the embodiments;

[0028] Figure 7 This is one of the simplified building results shown in the examples;

[0029] Figure 8 This is the second illustration of the simplified building results in the example. Detailed implementation method:

[0030] To make the objectives, technical solutions, and advantages of this invention clearer and more understandable, the invention will be further described in detail below with reference to the accompanying drawings and technical solutions.

[0031] For the simplification of buildings, see the embodiments of the present invention. Figure 1 As shown, a building simplification method based on a conditional graph diffusion model is provided, which specifically includes the following:

[0032] S101. Obtain the structural data and noise data of the building to be processed;

[0033] S102. Input the noise data into the pre-trained diffusion model, and inject the building map structure data to be processed as the generation condition into the diffusion model, so that the diffusion model can denoise the noise data based on the generation condition to obtain simplified building map structure data. In the training process, the diffusion model uses the building map structure data at a first scale as the generation condition, the building map structure data at a second scale as the generation target, and learns the correlation features between the noisy target data and the condition data through forward noise addition operation. The first scale is larger than the second scale.

[0034] Specifically, the diffusion model can be designed to include: a temporal encoder for providing time guidance for embedding features in the model denoising operation; a conditional encoder for mapping the structural data of the second scale building map using feature encoding; a feature encoder for encoding the noisy input data using features; a feature stitching layer for stitching features together; a fusion layer for fusing local geometric features and global structural features; and an output layer for predicting noise based on the fused features and outputting the result. The fusion layer adopts an alternating stacked structure of ring convolutional blocks and attention blocks.

[0035] The time encoder uses sine and cosine position encoding for time embedding; the annular convolutional block uses multiple parallel feature extraction paths to extract geometric features at different scales; and the attention block uses a standard multi-head attention mechanism and dynamically adjusts the influence intensity between vertices of the architectural drawing structure according to geometric importance to capture long-range dependencies.

[0036] like Figure 2 As shown, using building structure data as input data ensures two-dimensional information; using manually simplified small-scale data as the target for model training enables the model to learn operations based on visual thinking; the diffusion model is a powerful generative model that can perform generative simplification while taking into account visual thinking by understanding the overall and detailed morphological features of buildings.

[0037] The training process of the diffusion model can be designed to include:

[0038] Construct a building drawing structure data sample, which includes a first-scale building drawing structure data sample and a second-scale building drawing structure data sample obtained by manually simplifying the first-scale building drawing structure data sample.

[0039] The building structure data in the sample is preprocessed to obtain unified building structure data. The preprocessing includes: for buildings with fewer contour nodes than the target length, inserting new nodes between adjacent vertices of the building contour using a uniform interpolation method based on arc length so that the building contour nodes meet the target length; for buildings with more contour nodes than the target length, performing Douglas simplification on the building contours so that the number of contour nodes of each building is uniform.

[0040] Using preprocessed first-scale building map structural data samples as input and corresponding preprocessed second-scale building map structural data samples as output, the probability distribution of the transformation from the first-scale building map structural geometry to the second-scale building map structural morphology is learned through a noise prediction task and based on minimizing the noise prediction error.

[0041] Each building polygon can be represented as an undirected cyclic graph, whose adjacency matrix exhibits a typical circular structure. In this embodiment, the node features can be set to 3, with each vertex v i The eigenvectors are defined as follows:

[0042] f i =(l prev ,l next ,θ i )

[0043] Among them, l prev For the current vertex v i to the previous vertex v i-1 Euclidean distance between them; l next For the current vertex v i To the next vertex v i-1 The Euclidean distance between them; θ i This represents the rotation angle at the current vertex, with a value ranging from -180° to 180°. The sign depends on the geometric direction, i.e., clockwise is positive and counterclockwise is negative. For example... Figure 3 In this case, α1 is -90°, while α2 is 90°.

[0044] Due to model structure limitations, the input data needs to have a fixed number of nodes. To address the issue of inconsistent node counts in the original building polygons, a uniform interpolation method based on arc length can be used for data preprocessing. For building outlines of target lengths with insufficient node counts, new nodes are inserted between adjacent vertices. For example... Figure 3As shown, vertices v4, v5, v6, and v7 are new nodes introduced through interpolation. These interpolation points lie on the straight lines of the original edge segments, therefore their corner features are marked as -180°, indicating that they are flat angles and the direction is counterclockwise. For building outlines with more nodes than the target length, the outlines are first simplified using Douglas. Through this preprocessing method, the number of nodes for all input buildings is unified to a fixed value, which satisfies the model input requirements while preserving the integrity of the original geometric information to the greatest extent.

[0045] A conditional graph diffusion model is employed to simplify buildings across multiple scales. The model takes the original large-scale building outline as input and the corresponding simplified small-scale building outline as output. The core objective is to learn the probability distribution that governs the reasonable transformation from complex geometry to simplified form. This model indirectly learns this distribution through a noise prediction task, the core of which is training a graph neural network to accurately predict the noise added during the forward pass. Specifically, the training objective is to minimize the noise prediction error.

[0046]

[0047] Where ∈ represents real noise, ∈ θ (S t ,t,C) Neural network predicts noise.

[0048] The complex generation task is decomposed into multiple progressive denoising steps. By gradually optimizing the geometry, it better adapts to the fine-grained generation requirements of building outlines. The specific implementation includes two phases: training and inference (generation). In the training phase, such as... Figure 4 The pink part in the diagram illustrates how the model constructs training samples through a forward noise-adding process and learns the deep correlation features between the noisy target data and the conditional input. Its core task is to predict the added noise. During the inference phase, the trained model, based on the conditional input, such as... Figure 4 The blue part in the image is used to gradually recover the simplified geometry from random noise through a reverse denoising process.

[0049] The conditional graph diffusion model network architecture is crucial for noise prediction. This part achieves accurate estimation of the current noise by jointly extracting the structured correlation between conditional data and the noisy target, thereby driving the generation process towards a reasonable simplification. The network architecture is as follows: Figure 5 As shown.

[0050] In the conditional diffusion model, temporal embedding provides crucial temporal guidance to the denoising network, directing the model to perform corresponding denoising operations at different noise levels. This design ensures that the model can adaptively adjust the denoising strategy according to the current time step t. Sine and cosine positional encoding can be used for temporal embedding, and the mathematical expression of the vector is:

[0051]

[0052] Where, ω k For the frequency parameter, the formula is:

[0053]

[0054] In the formula, d is the embedding dimension and i is the position of the embedding dimension.

[0055] In the conditional graph diffusion model, the conditional mechanism is the core component for achieving accurate conversion of buildings from large scale to small scale. A multi-level conditional fusion mechanism can be used to ensure a high degree of semantic correlation between the generated results and the input conditions.

[0056] First, the input large-scale building data is mapped to a high-dimensional feature space using a conditional encoder.

[0057] Then, at the model input, the encoded conditional features are initially fused with the noisy input buildings using channel concatenation. This establishes a basic semantic association between the generation process and the conditional information, ensuring that the model clearly defines its goal of "what to generate" from the initial stage, that is, determining the basic framework for small-scale representation based on the geometric features of large-scale buildings.

[0058] Secondly, at the deep feature level of the network, multi-scale circular convolutional blocks are used to further integrate conditional information through conditional fusion (MLP). Its core function is to maintain deep semantic consistency, ensuring that key conditional information is not lost during propagation within the deep network. This mechanism allows the model to always remember "what should be done" during the generation process, maintaining a reflection of the original building features at every stage of the denoising chain.

[0059] This dual-condition fusion mechanism exhibits good feature preservation, process stability, and semantic consistency in architectural drafting and generalization tasks.

[0060] To address the ring-shaped structural characteristics of polygonal buildings, a multi-scale ring convolution module is used with a ring-filling mechanism to ensure geometric continuity and topological integrity. Traditional convolution operations suffer from boundary information loss when processing polygonal sequences, while ring filling, by connecting the beginning and end of the sequence, ensures geometric continuity and topological integrity. Figure 6 As shown, this defect is effectively avoided, and the inherent closed property of polygons is maintained.

[0061] The multi-scale circular convolution module contains four parallel feature extraction paths, such as... Figure 5 As shown, each path is optimized for geometric patterns at different scales:

[0062] 1) Small-scale local feature paths, such as Figure 6 (a) uses a smaller convolution kernel to specifically capture the subtle geometric relationships between adjacent nodes.

[0063] 2) Mesoscale neighborhood feature paths, such as Figure 6 (b) In this example, using a medium-sized convolutional kernel, the path is able to capture local geometric patterns composed of multiple consecutive nodes.

[0064] 3) Large-scale regional characteristic paths, such as Figure 6 (c) Using a larger convolution kernel, regional shape features of buildings are identified.

[0065] 4) Global Context Feature Path. This path consists of two stacked convolutional layers, such as... Figure 5 As shown, the overall shape characteristics of a building are modeled using deep receptive fields.

[0066] This multi-scale design enables the model to capture comprehensive geometric information from local details to global structure simultaneously, effectively maintaining the geometric integrity of the polygon.

[0067] The global attention mechanism adopts the standard multi-head attention mechanism, which establishes fully connected relationships between vertices to achieve direct interaction between any vertices that are not limited by distance in terms of information transmission. Its calculation process can be represented as follows:

[0068]

[0069] In the formula, α ij h represents the attention weight of vertex i to vertex j. i W represents the characteristics of vertex i. Q For the query vector, W K For key vectors, This is a scaling factor to prevent the dot product from becoming too large. Feature update based on attention weights:

[0070]

[0071] In the formula, For the updated vertex features, W V This is a value vector. This mechanism can dynamically adjust the influence strength between vertices based on geometric importance, effectively capturing long-range dependencies.

[0072] The model employs an architecture that alternates between stacked ring-shaped convolutional blocks and attention blocks to effectively fuse local geometric features with global structural information.

[0073] To ensure the training stability of deep networks, residual connection mechanisms are used within each module:

[0074] Hl+1 =H l +F(H l Θ l )

[0075] In the formula, H l Let F be the feature output of the l-th layer, and F be the transformation function of the current layer. l These are the corresponding parameters. Essentially, it uses skip connections to directly add the output of the previous layer to the output of the next layer, mitigating the gradient vanishing problem and promoting the effective propagation of deep features.

[0076] In the training of the diffusion model, Gaussian noise is progressively added to the structural data samples of the second-scale building map. The probability distribution of the noise data after T steps of noise addition is obtained based on a Markov chain. A random time-step sampling strategy is used to process the noise level at each time step. For each pair of training data, a time step is randomly sampled, and reparameterization is used to obtain the noise data at the corresponding time step. The model then performs one forward propagation to predict the added noise.

[0077] The forward noise addition process progressively adds Gaussian noise to the target small-scale building data; this process is defined as a Markov chain. Specifically, the diffusion process at each step can be described as follows:

[0078]

[0079] Where q(S) 1:T |S0) represents the noise data S obtained after T steps of noise addition, starting from the original data S0. T The probability distribution of the single-step diffusion process is defined by a Gaussian distribution:

[0080]

[0081] In the formula, S t This is the current data; the mean is... β t Variance is the noise scheduling parameter that controls the amount of noise added in each step; I is the identity matrix.

[0082] It should be noted that, during training, in order to efficiently train a model capable of handling all noise levels, we employ a random time-step sampling strategy. Specifically, for each pair of training data (S0, condition), a random time step t is sampled, and the corresponding noisy data S is directly calculated using reparameterization techniques. t Subsequently, the model performs a forward propagation to predict the added noise. This approach avoids running a complete T-step forward propagation process for each sample, greatly ensuring training efficiency, while enabling the model to learn the behavior of the entire denoised sequence through numerous iterations.

[0083] The reverse denoising process starts from pure noise S T Starting with large-scale building condition data C, the target small-scale building is gradually recovered through a Markov process:

[0084]

[0085] Each step of the denoising process learns a conditional Gaussian distribution:

[0086]

[0087] In the formula, the mean is μ θ .

[0088] The structure of a building can be simplified using the trained model, as follows: Figure 7 and 8 As shown, a diffusion model is introduced into building simplification, and generative simplification is performed by understanding the overall and detailed morphological features of the building.

[0089] Furthermore, based on the above method, this embodiment of the invention also provides a building simplification system based on a conditional graph diffusion model, comprising: a data acquisition module and a simplification output module, wherein,

[0090] The data acquisition module is used to acquire the structural data and noise data of the building drawings to be processed;

[0091] The simplified output module is used to input noisy data into a pre-trained diffusion model, and to inject the building map structure data to be processed as the generation condition into the diffusion model, so that the diffusion model can denoise the noisy data based on the generation condition to obtain simplified building map structure data. In the training process, the diffusion model uses building map structure data at a first scale as the generation condition and building map structure data at a second scale as the generation target, and learns the correlation features between the noisy target data and the condition data through forward noise addition operation. The first scale is larger than the second scale.

[0092] Unless otherwise specifically stated, the relative steps, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention.

[0093] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0094] The units and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations are not considered to be beyond the scope of this invention.

[0095] Those skilled in the art will understand that all or part of the steps in the above methods can be implemented by a program instructing related hardware, and the program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk. Optionally, all or part of the steps in the above embodiments can also be implemented using one or more integrated circuits. Accordingly, each module / unit in the above embodiments can be implemented in hardware or as a software functional module. This invention is not limited to any particular combination of hardware and software.

[0096] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A building simplification method based on a conditional graph diffusion model, characterized in that, Include: Acquire the structural data and noise data of the building to be processed; Noisy data is input into a pre-trained diffusion model, and the structural data of the building map to be processed is injected into the diffusion model as a generation condition. The diffusion model then denoises the noisy data based on the generation condition to obtain simplified structural data of the building map. During the training process, the diffusion model uses structural data of the building map at a first scale as the generation condition and structural data of the building map at a second scale as the generation target. It also learns the correlation features between the noisy target data and the condition data through a forward noise addition operation. The first scale is larger than the second scale.

2. The building simplification method based on the conditional graph diffusion model according to claim 1, characterized in that, The diffusion model includes: a temporal encoder for providing time guidance for embedding features in the model denoising operation; a conditional encoder for mapping the structural data of the second scale building map using feature encoding; a feature encoder for encoding the noisy input data using features; a feature stitching layer for stitching features together; a fusion layer for fusing local geometric features and global structural features; and an output layer for predicting noise based on the fused features and outputting the result. The fusion layer adopts an alternating stacked structure of ring convolutional blocks and attention blocks.

3. The building simplification method based on the conditional graph diffusion model according to claim 2, characterized in that, The time encoder uses sine and cosine position encoding for time embedding; the annular convolutional block uses multiple parallel feature extraction paths to extract geometric features at different scales; the attention block uses a standard multi-head attention mechanism and dynamically adjusts the influence intensity between vertices of the architectural drawing structure according to geometric importance to capture long-range dependencies.

4. The building simplification method based on the conditional graph diffusion model according to claim 1, characterized in that, The training process of the diffusion model includes: Construct a building drawing structure data sample, which includes a first-scale building drawing structure data sample and a second-scale building drawing structure data sample obtained by manually simplifying the first-scale building drawing structure data sample. The building diagram structure data in the sample is preprocessed to obtain unified building diagram structure data. The preprocessing includes: for buildings whose outline node number is less than the target length, new nodes are inserted between adjacent vertices of the building outline using a uniform interpolation method based on arc length so that the building outline node meets the target length. For buildings with more outline nodes than the target length, Douglas simplification is performed on the building outlines to make the number of outline nodes uniform for each building. Using preprocessed first-scale building map structural data samples as input and corresponding preprocessed second-scale building map structural data samples as output, the probability distribution of the transformation from the first-scale building map structural geometry to the second-scale building map structural morphology is learned through a noise prediction task and based on minimizing the noise prediction error.

5. The building simplification method based on the conditional graph diffusion model according to claim 1 or 4, characterized in that, In the training of the diffusion model, Gaussian noise is gradually added to the structural data samples of the second scale building map, and the probability distribution of the noise data after T steps of noise addition is obtained based on the Markov chain. The noise level of each time step is processed by a random time step sampling strategy.

6. The building simplification method based on the conditional graph diffusion model according to claim 5, characterized in that, The random time step sampling strategy includes: for each pair of training data, randomly sampling a time step, using reparameterization to obtain the noise data at the corresponding time step, and causing the model to perform a forward propagation once to predict the added noise.

7. The building simplification method based on the conditional graph diffusion model according to claim 1, characterized in that, The diffusion model denoises the noise data based on the generation conditions, comprising: Starting with noisy data, based on the structural data of the first-scale building map, the structural data of the second-scale building map is gradually recovered through a Markov process. The noisy data is pure noise data, and the denoising process in each step learns a conditional Gaussian distribution.

8. A building simplification system based on a conditional graph diffusion model, characterized in that, It includes: a data acquisition module and a simplified output module, wherein, The data acquisition module is used to acquire the structural data and noise data of the building drawings to be processed; The simplified output module is used to input noisy data into a pre-trained diffusion model, and to inject the building map structure data to be processed as the generation condition into the diffusion model, so that the diffusion model can denoise the noisy data based on the generation condition to obtain simplified building map structure data. In the training process, the diffusion model uses building map structure data at a first scale as the generation condition and building map structure data at a second scale as the generation target, and learns the correlation features between the noisy target data and the condition data through forward noise addition operation. The first scale is larger than the second scale.

9. An electronic device, characterized in that, include: At least one processor, and a memory coupled to said at least one processor; The memory stores a computer program that can be executed by the at least one processor to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed, enables the implementation of the method as described in any one of claims 1 to 7.

Images