A method and system for automatically generating building layout schemes based on GauGAN

Abstract

The present invention provides a method and system for automatically generating a building layout scheme based on GauGAN. The method comprises: obtaining a plot boundary, surrounding information and planning conditions of a to-be-generated scheme of a target area, wherein the surrounding information comprises distribution information of buildings, roads, rivers and green spaces around the target area, and the planning conditions comprises building density, volume ratio and building height limit required by the target area; inputting the plot boundary, the surrounding information and the planning conditions into a preset building layout scheme generation model, so that the building layout scheme generation model automatically generates a corresponding initial building layout scheme according to the surrounding information and the planning conditions; optimizing the layout of the initial building layout scheme based on an attraction-repulsion model, adjusting the distance between each building in the initial building layout scheme, obtaining a building layout scheme, and improving the efficiency and accuracy of automatic generation of building layout schemes.

Description

Technical Field

[0001] The present invention relates to the fields of urban design and computers, and in particular to a method and system for automatically generating a building layout scheme based on GauGAN. Background Art

[0002] As my country develops towards three-dimensional development, the layout of buildings plays a decisive role in the overall construction of the city. It not only shapes the physical structure of the building form, but also affects the interaction mode between people and the environment. How to improve the efficiency of building complex design has become a difficult problem. In recent years, academia and industry have conducted extensive research on building layout generation, which uses computer-aided generation methods to automatically explore design solutions that meet various design requirements. It can not only provide a preview of various high-quality layout solutions, but also support interactive exploration and informed decision-making in the early conceptualization stage, thereby significantly improving design efficiency and bringing practical benefits to urban design projects.

[0003] There are many ways to generate solutions, which have achieved certain results, but there are several major problems. Using optimization algorithms, such as genetic algorithms and rule-based methods, requires complex rules to be formulated; contains many restrictions and can only generate specific types of buildings; requires a lot of calculations, which is very time-consuming; and relies on the initial layout. The quality of the initial input is crucial to whether a successful layout solution can be obtained.

[0004] The generative adversarial learning methods currently used mostly learn image features from the perspective of images; they lack the input of the site environment and only consider the characteristics of the plot itself, ignoring the influence of the surrounding buildings, roads, and urban style of the plot; at the same time, they do not impose constraints on design conditions, making the results lack controllability and operability; at the same time, the generated results are mostly 2D, that is, they focus on generating building footprints, which is not conducive to performance analysis and leads to insufficient accuracy of the generated model. Summary of the invention

[0005] In view of the above problems, the present invention provides a method and system for automatically generating building layout plans based on GauGAN. A plan generation model is constructed based on GauGAN, and corresponding building layout plans are automatically generated according to user requirements, thereby improving the efficiency and accuracy of automatic generation of building layout plans.

[0006] In a first aspect, an embodiment of the present invention provides a method for automatically generating a building layout solution based on GauGAN, comprising:

[0007] Obtaining the plot boundary, surrounding information and planning conditions of the to-be-generated scheme of the target area, wherein the surrounding information includes the distribution information of buildings, roads, rivers and green spaces around the target area, and the planning conditions include the building density, volume ratio and building height limit required for the target area;

[0008] Inputting the plot boundary, the surrounding information and the planning conditions into a preset building layout scheme generation model, so that the building layout scheme generation model automatically generates a corresponding initial building layout scheme according to the surrounding information and the planning conditions, wherein the building layout scheme generation model is constructed based on GauGAN and obtained by training according to a preset case sample library;

[0009] The initial building layout scheme is optimized based on the attraction-repulsion model, and the distances between the buildings in the initial building layout scheme are adjusted to obtain a building layout scheme.

[0010] The embodiment of the present invention provides a method for automatically generating a building layout scheme based on GauGAN, by constructing a building layout scheme generation model based on GauGAN and training it according to a preset case sample library, then inputting relevant information of a target area into the building layout scheme generation model, thereby generating an initial building layout scheme that meets the user's requirements, and finally using an attraction-repulsion model to optimize the layout of the initial building layout scheme to obtain a final building layout scheme. In the scheme generation process, the embodiment of the present invention utilizes relevant information such as the plot boundary, surrounding information and planning conditions of the scheme to be generated in the target area, so that the scheme generated by the model can meet the basic planning conditions and can also coordinate with the surrounding buildings and natural environment, thereby improving the accuracy of automatic generation of building layout schemes; constructing a building layout scheme generation model based on GauGAN, compared with the existing genetic algorithm and rule-based method, does not require complex rules and calculations, and has higher scheme generation efficiency; finally, based on the attraction-repulsion model, further optimizes the layout of the initial scheme generated by the building layout scheme generation model, so that the final building layout scheme is more in line with the actual situation, thereby improving the rationality and practicality of automatic generation of building layout schemes.

[0011] In a possible implementation, the building layout solution generation model is constructed based on GauGAN and trained according to a preset case sample library, including:

[0012] An initial building layout generation model is constructed based on the network features of GauGAN and StyleGAN2;

[0013] Obtaining the building layout plans of each area and constructing the case sample library;

[0014] Input each building layout scheme in the case sample library into the initial building layout scheme generation model for training to obtain the building layout scheme generation model.

[0015] The embodiment of the present invention provides a method for constructing and training a model. First, an initial building layout scheme generation model is constructed based on the network features of GauGAN and StyleGAN2. Based on GauGAN, by drawing on the network features of StyleGAN2, an initial building layout scheme generation model with multi-channel input and multi-condition and multi-scale injection is formed. Since the model supports multi-channel input and multi-condition and multi-scale injection, it can accept more detailed building layout rules and related information, so that the model can generate a more accurate layout scheme, thereby improving the accuracy of automatic generation of building layout schemes. In addition, by obtaining the building layout schemes of each area, constructing the case sample library, and using the case sample library to train the initial building layout scheme generation model, the initial building layout scheme generation model learns the layout style of each area, thereby further improving the accuracy of automatic generation of building layout schemes.

[0016] Furthermore, the process of inputting the first building layout scheme in the case sample library into the initial building layout scheme generation model for training includes:

[0017] Inputting the building information, first surrounding information, first plot boundary and first planning condition corresponding to the first building layout scheme into the initial building layout scheme generation model through different channels, respectively, wherein the building information includes the location, length, width and height of each building in the first building layout scheme;

[0018] Parsing the building information, the first surrounding information, and the first plot boundary to obtain an original semantic map;

[0019] Performing multi-scale encoding on the planning conditions to obtain a condition feature map;

[0020] Performing feature fusion on the original semantic map and the conditional feature map to obtain a fused feature map;

[0021] The image authenticity and conditional rationality of the fused feature map are evaluated respectively by a preset global discriminator and an attribute discriminator, thereby optimizing and adjusting the initial building layout scheme generation model.

[0022] The embodiment of the present invention further illustrates the process of training the initial building layout scheme generation model using a case sample library, wherein, after feature extraction of the first building layout scheme, the corresponding original semantic map and conditional feature map are respectively obtained, and then the original semantic map and the conditional feature map are feature fused to obtain a fused feature map, which is the building layout scheme generated after the model learns the first building layout scheme, and then a global discriminator and an attribute discriminator are used to evaluate the image authenticity and conditional rationality of the fused feature map, and then the initial building layout scheme generation model is optimized and adjusted according to the evaluation results, and the accuracy of automatic generation of the building layout scheme can be further improved by training the initial building layout scheme generation model multiple times.

[0023] In a possible implementation, the inputting the land parcel boundary, the surrounding information and the planning conditions into a preset building layout scheme generation model so that the building layout scheme generation model automatically generates a corresponding initial building layout scheme according to the surrounding information and the planning conditions includes:

[0024] Calculating the Hamming distance between the target area and each plot in the case sample library according to the plot boundary, and obtaining the morphological similarity between the target area and each plot in the case sample library;

[0025] Calculating the similarity between the target area and each plot in the case sample library according to the planning conditions and a preset formula;

[0026] The morphological similarity and the conditional similarity are combined to obtain the comprehensive similarity between the target area and each plot in the case sample library;

[0027] When the comprehensive similarity of a certain plot in the case sample library exceeds a first preset value, the plot boundary, surrounding information and planning conditions of the to-be-generated scheme of the target area are input into the building layout scheme generation model to obtain a corresponding initial building layout scheme;

[0028] Otherwise, a preset number of plots are extracted from the case sample library in descending order of the comprehensive similarity to fine-tune the building layout scheme generation model, and then the plot boundaries, surrounding information and planning conditions of the to-be-generated scheme of the target area are input into the building layout scheme generation model to obtain the corresponding initial building layout scheme.

[0029] The embodiment of the present invention provides a method for generating an initial building layout plan based on a building layout plan generation model. First, the morphological similarity and conditional similarity between a target area and each plot in a case sample library are calculated and integrated into a comprehensive similarity. Then, according to the comprehensive similarity, plots similar to the target area are retrieved from the case sample library. If there are plots very similar to the target area in the case sample library, there is no need to perform additional training on the building layout plan generation model. The corresponding information of the target area is directly input into the building layout plan generation model to generate an initial building layout plan. If there are no plots very similar to the target area in the case sample library, a certain number of plots are selected from the case sample library according to the similarity from large to small and input into the building layout plan generation model for fine-tuning. The model is adjusted in a targeted manner to further improve the accuracy of automatic generation of building layout plans.

[0030] In a possible implementation, the optimization of the initial building layout scheme based on the attraction-repulsion model, adjusting the distances between the buildings in the initial building layout scheme, and obtaining the building layout scheme includes:

[0031] Performing vectorization processing on the initial building layout plan to obtain a corresponding vectorized building layout plan;

[0032] Calculating the distance between each building in the vectorized building layout scheme and comparing it with the preset standard distance, and converting it into the attraction and repulsion between each building;

[0033] Based on the attraction-repulsion model, each building in the vectorized building layout solution is moved according to the attraction-repulsion between the buildings, so that each building in the vectorized building layout solution reaches a balanced state, and the building layout solution is obtained.

[0034] The embodiment of the present invention provides a method for optimizing an initial building layout scheme. Since the initial building layout scheme exists in the form of a picture, the initial building layout scheme is firstly vectorized so as to subsequently calculate the distance between each building in the scheme; then the distance between each building in the scheme is compared with the standard distance to obtain the attraction and repulsion between each building, and each building is moved based on the attraction and repulsion model, so that the distance between each building is more reasonable and standardized, thereby improving the rationality and practicality of automatic generation of the building layout scheme.

[0035] Furthermore, vectorizing the initial building layout plan to obtain a corresponding vectorized building layout plan includes filtering, vectorizing, clustering height values, and replacing building outlines on the initial building layout plan.

[0036] The embodiment of the present invention provides a method for vectorizing an initial building layout plan. Due to some residual noise in the generated image, the initial building layout plan is firstly filtered to eliminate some small pseudo buildings and improve the accuracy of the image; then the generated plan image is vectorized to obtain the building vector boundary, which is convenient for subsequent distance calculation and result display; the pixel values ​​within the building boundary are clustered for height values ​​to obtain more accurate building height values; finally, the most similar building type vector boundary is found from a commonly used building type library to replace the building type outline in the initial building layout plan, so that the initial building layout plan is more beautiful and accurate.

[0037] In a second aspect, accordingly, an embodiment of the present invention provides a system for automatically generating a building layout plan based on GauGAN, including an acquisition module, a plan generation module, and a plan optimization module;

[0038] The acquisition module is used to acquire the land boundary, surrounding information and planning conditions of the to-be-generated scheme of the target area, wherein the surrounding information includes the distribution information of buildings, roads, rivers and green spaces around the target area, and the planning conditions include the building density, volume ratio and building height limit required by the target area;

[0039] The scheme generation module is used to input the plot boundary, the surrounding information and the planning conditions into a preset building layout scheme generation model, so that the building layout scheme generation model automatically generates a corresponding initial building layout scheme according to the surrounding information and the planning conditions, and the building layout scheme generation model is constructed based on GauGAN and obtained by training according to a preset case sample library;

[0040] The scheme optimization module is used to optimize the initial building layout scheme based on the attraction-repulsion model, adjust the distance between each building in the initial building layout scheme, and obtain the building layout scheme.

[0041] In a possible implementation, the building layout scheme automatic generation system further includes a modeling module, which is used to construct an initial building layout scheme generation model based on GauGAN and to obtain the building layout scheme generation model by training according to a preset case sample library, and includes a model construction unit, a sample library construction unit, and a training unit;

[0042] The model building unit is used to build an initial building layout scheme generation model based on the network features of GauGAN and StyleGAN2;

[0043] The sample library construction unit is used to obtain the building layout plan of each area and construct the case sample library;

[0044] The training unit is used to input each building layout scheme in the case sample library into the initial building layout scheme generation model for training to obtain the building layout scheme generation model.

[0045] Furthermore, the training unit inputs the first building layout scheme in the case sample library into the initial building layout scheme generation model for training, including:

[0046] Inputting the building information, first surrounding information, first plot boundary and first planning condition corresponding to the first building layout scheme into the initial building layout scheme generation model through different channels, respectively, wherein the building information includes the location, length, width and height of each building in the first building layout scheme;

[0047] Parsing the building information, the first surrounding information, and the first plot boundary to obtain an original semantic map;

[0048] Performing multi-scale encoding on the planning conditions to obtain a condition feature map;

[0049] Performing feature fusion on the original semantic map and the conditional feature map to obtain a fused feature map;

[0050] The image authenticity and conditional rationality of the fused feature map are evaluated respectively by a preset global discriminator and an attribute discriminator, thereby optimizing and adjusting the initial building layout scheme generation model.

[0051] In a possible implementation, the solution generation module includes a morphological similarity calculation unit, a conditional similarity calculation unit, a comprehensive similarity calculation unit, and a solution generation unit;

[0052] The morphological similarity calculation unit is used to calculate the Hamming distance between the target area and each plot in the case sample library according to the plot boundary, so as to obtain the morphological similarity between the target area and each plot in the case sample library;

[0053] The condition similarity calculation unit is used to calculate the condition similarity between the target area and each plot in the case sample library according to the planning conditions and a preset formula;

[0054] The comprehensive similarity calculation unit is used to combine the morphological similarity and the conditional similarity to obtain the comprehensive similarity between the target area and each plot in the case sample library;

[0055] The solution generation unit is used for inputting the boundary of the land parcel, the surrounding information and the planning conditions of the to-be-generated solution of the target area into the building layout solution generation model to obtain the corresponding initial building layout solution when the comprehensive similarity of a land parcel in the case sample library exceeds a first preset value;

[0056] Otherwise, a preset number of plots are extracted from the case sample library in descending order of the comprehensive similarity to fine-tune the building layout scheme generation model, and then the plot boundaries, surrounding information and planning conditions of the to-be-generated scheme of the target area are input into the building layout scheme generation model to obtain the corresponding initial building layout scheme.

[0057] In a possible implementation, the scheme optimization module includes a vector processing unit, an attraction-repulsion force calculation unit, and an optimization unit:

[0058] The vector processing unit is used to perform vector processing on the initial building layout scheme to obtain a corresponding vectorized building layout scheme;

[0059] The attraction-repulsion calculation unit is used to calculate the distance between each building in the vectorized building layout scheme and compare it with the preset standard distance, and convert it into the attraction-repulsion between each building;

[0060] The optimization unit is used to move each building in the vectorized building layout solution based on the attraction-repulsion model according to the attraction-repulsion force between the buildings, so that each building in the vectorized building layout solution reaches a balanced state, and the building layout solution is obtained.

[0061] Furthermore, the vectorization processing unit performs vectorization processing on the initial building layout plan to obtain a corresponding vectorized building layout plan, including filtering processing, vectorization operation, height value clustering, and building outline replacement on the initial building layout plan. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 : A flow chart of an embodiment of a method for automatically generating a building layout plan based on GauGAN provided by the present invention.

[0063] Figure 2 : A schematic diagram of the process of training an initial building layout scheme generation model in a building layout scheme automatic generation method based on GauGAN provided by the present invention.

[0064] Figure 3 : A network structure diagram of an initial building layout scheme generation model in a building layout scheme automatic generation method based on GauGAN provided by the present invention.

[0065] Figure 4 : A schematic diagram of the process of feature fusion in the automatic generation method of building layout solutions based on GauGAN provided by the present invention.

[0066] Figure 5 : A schematic diagram of a process of generating an initial building layout plan through a building layout plan generation model in a method for automatically generating building layout plans based on GauGAN provided by the present invention.

[0067] Figure 6 : A schematic diagram of the flow of optimizing the layout of an initial building layout plan in a method for automatically generating a building layout plan based on GauGAN provided by the present invention.

[0068] Figure 7 : This is a rendering of the effect of optimizing the initial building layout plan in the automatic generation method of building layout plans based on GauGAN provided by the present invention.

[0069] Figure 8 : A structural schematic diagram of an embodiment of a building layout scheme automatic generation system based on GauGAN provided by the present invention.

[0070] Fig. 9 : A structural schematic diagram of a modeling module in a building layout scheme automatic generation system based on GauGAN provided by the present invention.

[0071] Fig.10 : A structural schematic diagram of a solution generation module in a building layout solution automatic generation system based on GauGAN provided by the present invention.

[0072] Fig.11 : A structural schematic diagram of an optimization module in a building layout plan automatic generation system based on GauGAN provided by the present invention. DETAILED DESCRIPTION

[0073] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0074] It should be noted that the step numbers in the text are only for the convenience of explaining the specific embodiments and are not used to limit the order in which the steps are executed.

[0075] Throughout this specification, the GauGAN described in this specification is a generative adversarial algorithm developed by NVIDIA. The function of the GauGAN algorithm is to process the input label map (semantic maps) through the generative model G to obtain a synthetic image with real texture and semantic information. Different from the existing image synthesis GAN in the past, which simply stacks the convolution layer, normalization layer, and nonlinear layer together to form a generative model, GauGAN introduces a new normalization layer, uses the semantic map to adjust the result of the normalization output, so that it has better semantic information, and the semantic information runs through the entire network. And its method can cope with various generation tasks using semantic maps. In addition, StyleGAN2 described in this specification is a generative adversarial network that generates realistic images by learning a large amount of image sample data. The core principle of StyleGAN2 is to use a generator network and a discriminator network for game optimization. The generator network receives a noise vector as input and generates images through multiple layers; the goal of the discriminator network is to distinguish between generated images and real images. The generator network and the discriminator network continuously compete with each other during the training process, thereby optimizing the model and enabling the model to generate more realistic images.

[0076] Embodiment 1:

[0077] like Figure 1 As shown, embodiment 1 provides a method for automatically generating a building layout solution based on GauGAN, including steps S1-S3:

[0078] Step S1, obtaining the land boundary, surrounding information and planning conditions of the to-be-generated scheme of the target area, wherein the surrounding information includes the distribution information of buildings, roads, rivers and green spaces around the target area, and the planning conditions include the building density, volume ratio and building height limit required by the target area;

[0079] Step S2, inputting the plot boundary, the surrounding information and the planning conditions into a preset building layout scheme generation model, so that the building layout scheme generation model automatically generates a corresponding initial building layout scheme according to the surrounding information and the planning conditions, wherein the building layout scheme generation model is constructed based on GauGAN and obtained by training according to a preset case sample library;

[0080] Step S3: Optimize the initial building layout plan based on the attraction-repulsion model, adjust the distance between each building in the initial building layout plan, and obtain a building layout plan.

[0081] The embodiment of the present invention provides a method for automatically generating a building layout scheme based on GauGAN, by constructing a building layout scheme generation model based on GauGAN and training it according to a preset case sample library, then inputting relevant information of a target area into the building layout scheme generation model, thereby generating an initial building layout scheme that meets the user's requirements, and finally using an attraction-repulsion model to optimize the layout of the initial building layout scheme to obtain a final building layout scheme. In the scheme generation process, the embodiment of the present invention utilizes relevant information such as the plot boundary, surrounding information and planning conditions of the scheme to be generated in the target area, so that the scheme generated by the model can meet the basic planning conditions and can also coordinate with the surrounding buildings and natural environment, thereby improving the accuracy of automatic generation of building layout schemes; constructing a building layout scheme generation model based on GauGAN, compared with the existing genetic algorithm and rule-based method, does not require complex rules and calculations, and has higher scheme generation efficiency; finally, based on the attraction-repulsion model, further optimizes the layout of the initial scheme generated by the building layout scheme generation model, so that the final building layout scheme is more in line with the actual situation, thereby improving the rationality and practicality of automatic generation of building layout schemes.

[0082] In a preferred embodiment, in step S1, the land boundary of the target area, the road grade, road width, river grade, green space category and other information of the target area and the surrounding area are obtained through Baidu Maps to form a base library of the city, and the surrounding information of the land in the target area is quickly obtained by using a buffer zone. The land arrangement condition restrictions of the target area are extracted from the upper planning file, including building density, volume ratio, number of buildings, building height limit, etc.

[0083] In a possible implementation, the building layout solution generation model is constructed based on GauGAN and trained according to a preset case sample library, such as Figure 2 As shown, steps S401-S403 are included:

[0084] Step S401, constructing an initial building layout scheme generation model based on the network features of GauGAN and StyleGAN2;

[0085] Step S402: Obtain the building layout plan of each area and construct the case sample library;

[0086] Step S403: input each building layout scheme in the case sample library into the initial building layout scheme generation model for training to obtain the building layout scheme generation model.

[0087] Specifically, in step S401, based on GauGAN, by referring to the network features of StyleGAN2, an initial building layout scheme generation model MC-GauGAN with multi-channel input and multi-condition and multi-scale injection is formed. The network structure of the model is as follows: Figure 3 shown.

[0088] In step S402, a case sample library is constructed for cities across the country by category and region. The location of the plots is divided into first-tier, second-tier, third-tier and fourth-tier cities according to the city level; the same city is divided into core areas and peripheral areas; the building age, category, height and other information are added to the buildings in the plan, and topological errors are resolved; and the style, parameters and other descriptions are added to each layout plan.

[0089] The embodiment of the present invention provides a method for constructing and training a model. First, an initial building layout scheme generation model is constructed based on the network features of GauGAN and StyleGAN2. Based on GauGAN, by drawing on the network features of StyleGAN2, an initial building layout scheme generation model with multi-channel input and multi-condition and multi-scale injection is formed. Since the model supports multi-channel input and multi-condition and multi-scale injection, it can accept more detailed building layout rules and related information, so that the model can generate a more accurate layout scheme, thereby improving the accuracy of automatic generation of building layout schemes. In addition, by obtaining the building layout schemes of each area, constructing the case sample library, and using the case sample library to train the initial building layout scheme generation model, the initial building layout scheme generation model learns the layout style of each area, thereby further improving the accuracy of automatic generation of building layout schemes.

[0090] Furthermore, in step S403, the process of inputting the first building layout scheme in the case sample library into the initial building layout scheme generation model for training includes:

[0091] Inputting the building information, first surrounding information, first plot boundary and first planning condition corresponding to the first building layout scheme into the initial building layout scheme generation model through different channels, respectively, wherein the building information includes the location, length, width and height of each building in the first building layout scheme;

[0092] Parsing the building information, the first surrounding information, and the first plot boundary to obtain an original semantic map;

[0093] Performing multi-scale encoding on the planning conditions to obtain a condition feature map;

[0094] Performing feature fusion on the original semantic map and the conditional feature map to obtain a fused feature map;

[0095] The image authenticity and conditional rationality of the fused feature map are evaluated respectively by a preset global discriminator and an attribute discriminator, thereby optimizing and adjusting the initial building layout scheme generation model.

[0096] Specifically, in a preferred embodiment, the surrounding information, land boundary and building information corresponding to the building layout scheme used for training are first input as a channel respectively. In order to make the generated building result have height information, the building height is encoded at the time of input, and a pixel value is corresponding to every 0.5m at an interval of 20 to 240. The height of buildings over 100 meters is adjusted to 100 meters. At the same time, the input of GauGAN is modified, and the single channel is changed to multiple channels. The model extracts features from the input building information surrounding information and land boundary to generate the original semantic map.

[0097] Then, the input multiple planning conditions are multi-scale encoded. Formally, given a vector A indicating a list of attributes, a multi-condition encoding module is formed by a set of sub-encoders Enci(A). Each sub-encoder consists of 3 fully connected layers. Through a series of affine transformations, the transformed vector t is obtained. i , after transformation t i The SPADE module injected into ResBlk forms C-SPADE. The scale i of each C-SPADE module is also different. The expression of multi-scale encoding is as follows:

[0098]

[0099] Among them, t i is the vector obtained by multi-scale encoding of multiple planning conditions, is the preset sub-encoder, and i represents different scales.

[0100] After multi-scale encoding of multiple planning conditions, the conditional feature map corresponding to the building layout scheme is obtained. The original semantic map and the conditional feature map are fused. The fusion method is as follows: Figure 4 As shown in the figure, after the conditional feature map is operated by Mod and Demod, it acts on the convolution kernel, and the conditional information is applied to each position of the original semantic map in a convolutional manner to obtain the final fused feature map. Mod stands for weight modulated convolution, that is, the weight size is adjusted at different spatial positions of the weight; Demod stands for demodulation operation, which is to normalize the weight of the weight convolution kernel as a whole, and then integrate the processed results into the subsequent convolution 3*3Conv operation as a weight to adjust the features of the image convolution, so as to achieve the fusion of image features and conditional features.

[0101] Specifically, since the original semantic map s input into each SPADE module is not the same size as the conditional feature map x, s is first scaled to be the same size as x. Then it passes through the convolution layer and activation layer to an intermediate layer s', and then s' passes through two different convolution layers conv γ and conv β Get the corresponding γ and β, and use them as normalization coefficients and biases to bn(x in ), the specific expression is as follows:

[0102] s′=ReLU(conv(Resize(s)))

[0103] x out =bn(x in )·conv γ (s′)+conv β (s′)

[0104] Among them, x in is the conditional feature map input to the SPADE module, s is the original semantic map, s′ is the intermediate variable, and x out It is the feature map output by the SPADE module, Resize() represents the scaling operation, conv() represents the operation in the convolutional layer, ReLU() represents the operation in the activation layer, and bn() represents the batch normalization operation.

[0105] Mod operations are:

[0106] W′ ijk =S i ·W ijk

[0107] Where S i The output of the corresponding conditional encoding module, W ijk is the weight of the convolution kernel, where i represents the channel index of the input feature map, j represents the channel index of the output feature map, k represents the spatial position index of the convolution kernel, and W′ ijk is the updated weight.

[0108] The Demod operation is:

[0109]

[0110] Among them, W″ ijk The normalized convolutional layer weights can make the training more stable.

[0111] Finally, the image authenticity and conditional rationality of the fused feature map are evaluated by the preset global discriminator and attribute discriminator, and then the initial building layout plan generation model is optimized and adjusted. The two discriminators share network features in the early stage and extract feature maps, which reduces the computational burden and improves the model efficiency, avoids overfitting, and improves the generalization performance of the model.

[0112] The embodiment of the present invention further illustrates the process of training the initial building layout scheme generation model using a case sample library, wherein, after feature extraction of the first building layout scheme, the corresponding original semantic map and conditional feature map are respectively obtained, and then the original semantic map and the conditional feature map are feature fused to obtain a fused feature map, which is the building layout scheme generated after the model learns the first building layout scheme, and then a global discriminator and an attribute discriminator are used to evaluate the image authenticity and conditional rationality of the fused feature map, and then the initial building layout scheme generation model is optimized and adjusted according to the evaluation results, and the accuracy of automatic generation of the building layout scheme can be further improved by training the initial building layout scheme generation model multiple times.

[0113] In a possible implementation, in step S2, the plot boundary, the surrounding information and the planning conditions are input into a preset building layout scheme generation model, so that the building layout scheme generation model automatically generates a corresponding initial building layout scheme according to the surrounding information and the planning conditions, such as Figure 5 As shown, steps S201-S204 are included:

[0114] Step S201, calculating the Hamming distance between the target area and each plot in the case sample library according to the plot boundary, and obtaining the morphological similarity between the target area and each plot in the case sample library;

[0115] Step S202: Calculate the similarity between the target area and each plot in the case sample library according to the planning conditions and a preset formula;

[0116] Step S203, combining the morphological similarity and the conditional similarity to obtain a comprehensive similarity between the target area and each plot in the case sample library;

[0117] Step S204: When the comprehensive similarity of a certain plot in the case sample library exceeds a first preset value, the plot boundary, surrounding information and planning conditions of the to-be-generated scheme of the target area are input into the building layout scheme generation model to obtain a corresponding initial building layout scheme; otherwise, a preset number of plots are extracted from the case sample library in descending order of the comprehensive similarity to fine-tune the building layout scheme generation model, and then the plot boundary, surrounding information and planning conditions of the to-be-generated scheme of the target area are input into the building layout scheme generation model to obtain a corresponding initial building layout scheme.

[0118] Specifically, in step S201, the similarity of the plot morphology is first obtained by calculating the Hamming distance; the Hamming distance uses numerical differences to measure the distance between two vectors, which is then converted into the morphological similarity between each plot in the case sample library and the plot in the target area.

[0119] Then in step S202, the conditional similarity between each plot in the case sample library and the plot in the target area is calculated, and the calculation formula is as follows:

[0120]

[0121] Among them, each x, y and z represent the number of buildings, volume ratio and building density of the input plot respectively. Respectively represent the average number of buildings, volume ratio, and building density of each plot in the case sample library. x ,t y and t z are the number of buildings, volume ratio, and building density in the user's target area, σ x , σ y , σ z are the standard deviations of the corresponding values ​​respectively. The algorithm normalizes the block information data into a normal distribution. Then, the Euclidean distance between the standardized user input value and the standardized block information is obtained as the conditional similarity.

[0122] Furthermore, the morphological similarity and the conditional similarity are combined to obtain a comprehensive similarity, and the target area is compared with each plot in the case sample library. If there are cases of similar plot types in the case sample library, no additional training is required, and the initial building layout plan is directly generated according to the corresponding information of the target area; otherwise, the top 2000 plots with the highest comprehensive similarity in the case sample library are selected as similar plots, the model is fine-tuned, and then the initial building layout plan is generated according to the corresponding information of the target area.

[0123] The embodiment of the present invention provides a method for generating an initial building layout plan based on a building layout plan generation model. First, the morphological similarity and conditional similarity between a target area and each plot in a case sample library are calculated and integrated into a comprehensive similarity. Then, according to the comprehensive similarity, plots similar to the target area are retrieved from the case sample library. If there are plots very similar to the target area in the case sample library, there is no need to perform additional training on the building layout plan generation model. The corresponding information of the target area is directly input into the building layout plan generation model to generate an initial building layout plan. If there are no plots very similar to the target area in the case sample library, a certain number of plots are selected from the case sample library according to the similarity from large to small and input into the building layout plan generation model for fine-tuning. The model is adjusted in a targeted manner to further improve the accuracy of automatic generation of building layout plans.

[0124] In a possible implementation, in step S3, the initial building layout scheme is optimized based on the attraction-repulsion model, and the distances between the buildings in the initial building layout scheme are adjusted to obtain a building layout scheme, such as Figure 6 As shown, steps S301-S303 are included:

[0125] Step S301, performing vectorization processing on the initial building layout scheme to obtain a corresponding vectorized building layout scheme;

[0126] Step S302: Calculate the distance between each building in the vectorized building layout solution and compare it with the preset standard distance, and convert it into the attraction and repulsion between each building;

[0127] Step S303: moving each building in the vectorized building layout solution according to the attraction and repulsion force between the buildings based on the attraction and repulsion force model, so that each building in the vectorized building layout solution reaches a balanced state, and obtaining the building layout solution.

[0128] In a preferred embodiment, in steps S302-S303, by combing through the specification documents, sorting out the setbacks, spacing and sunlight distances, building a rule library, using the attraction and repulsion model, by calculating the distance between buildings, comparing it with the specification distance, converting it into attraction and repulsion, and by calculating the resultant force between the buildings, determining the movement direction of the buildings, and finally obtaining a reasonable building layout plan.

[0129] Specifically, the FR algorithm is adopted. This is a classic spring system that can generate an aesthetically pleasing and compliant building layout. All buildings are regarded as electrons. Each electron node is subject to the repulsive force from other nodes and the attractive force of the community boundary on the point. Under the interaction of forces, the entire layout will eventually reach a balanced state, thereby obtaining the optimization of the building layout under multiple constraints. By combining the specifications, the optimal moving distance between nodes is determined after repeated tests. When real≥std:

[0130]

[0131] When real < std:

[0132]

[0133] Here, real is the actual distance between two buildings, std is the specification requirement between two buildings, and f (real) represents the magnitude of the repulsive and attractive force. The magnitude of the repulsive and attractive force between each two buildings can be calculated through the above formula. However, there are multiple buildings around each building. The resultant force of the repulsive and attractive forces of all the surrounding buildings can be calculated to finally determine the magnitude and direction of the movement. Figure 7 is the effect diagram of distance adjustment by FR in this embodiment. Through Figure 7 it can be seen that the distance of the generated scheme is adjusted, and the distance between each building after adjustment meets the requirements of specification and aesthetics.

[0134] The embodiment of the present invention provides a method for optimizing an initial building layout scheme. Since the initial building layout scheme exists in the form of a picture, the initial building layout scheme is first vectorized to facilitate subsequent calculation of the distance between each building in the scheme; then the distance between each building in the scheme is compared with the standard distance to obtain the repulsive and attractive forces between each building, and each building is moved based on the repulsive and attractive force model, making the distance between each building more reasonable and standardized, and improving the rationality and practicability of the automatic generation of the building layout scheme.

[0135] Furthermore, the vectorization process of the initial building layout scheme to obtain the corresponding vectorized building layout scheme includes filtering processing, vectorization operation, height value clustering, and building type contour replacement of the initial building layout scheme.

[0136] Specifically, due to some residual noise in the generated image, the filtering process adds a median filter to eliminate some small pseudo-buildings; then the generated scheme image is vectorized to obtain the building vector boundary, which is convenient for subsequent distance calculation and result display; the pixel values ​​within the building boundary are clustered using the DBSCAN method to obtain more accurate building height values; finally, although it has been vectorized, the boundary is still irregular, and buildings have common building types. By sorting out the common building types, a building type library is constructed, and the building form is calculated to find the closest building type vector boundary from the building type library for replacement. The replacement operation not only maintains the learned arrangement pattern, but also increases practicality.

[0137] The embodiment of the present invention provides a method for vectorizing an initial building layout plan. Due to some residual noise in the generated image, the initial building layout plan is firstly filtered to eliminate some small pseudo buildings and improve the accuracy of the image; then the generated plan image is vectorized to obtain the building vector boundary, which is convenient for subsequent distance calculation and result display; the pixel values ​​within the building boundary are clustered for height values ​​to obtain more accurate building height values; finally, the most similar building type vector boundary is found from a commonly used building type library to replace the building type outline in the initial building layout plan, so that the initial building layout plan is more beautiful and accurate.

[0138] Embodiment 2:

[0139] Correspondingly, such as Figure 8 As shown, an embodiment of the present invention provides a building layout scheme automatic generation system based on GauGAN, including an acquisition module 10, a scheme generation module 20 and a scheme optimization module 30;

[0140] The acquisition module 10 is used to acquire the land boundary, surrounding information and planning conditions of the to-be-generated scheme of the target area, wherein the surrounding information includes the distribution information of buildings, roads, rivers and green spaces around the target area, and the planning conditions include the building density, volume ratio and building height limit required by the target area;

[0141] The solution generation module 20 is used to input the plot boundary, the surrounding information and the planning conditions into a preset building layout solution generation model, so that the building layout solution generation model automatically generates a corresponding initial building layout solution according to the surrounding information and the planning conditions, and the building layout solution generation model is constructed based on GauGAN and obtained by training according to a preset case sample library;

[0142] The scheme optimization module 30 is used to optimize the initial building layout scheme based on the attraction-repulsion model, adjust the distance between each building in the initial building layout scheme, and obtain a building layout scheme.

[0143] In one possible implementation, Fig. 9 As shown, the automatic building layout scheme generation system further includes a modeling module 40, which is used to construct an initial building layout scheme generation model based on GauGAN and to obtain the building layout scheme generation model through training according to a preset case sample library, including a model construction unit 401, a sample library construction unit 402 and a training unit 403;

[0144] The model building unit 401 is used to build an initial building layout scheme generation model based on the network features of GauGAN and StyleGAN2;

[0145] The sample library construction unit 402 is used to obtain the building layout plan of each area and construct the case sample library;

[0146] The training unit 403 is used to input each building layout scheme in the case sample library into the initial building layout scheme generation model for training to obtain the building layout scheme generation model.

[0147] Furthermore, the training unit 403 inputs the first building layout scheme in the case sample library into the initial building layout scheme generation model for training, including:

[0148] Inputting the building information, first surrounding information, first plot boundary and first planning condition corresponding to the first building layout scheme into the initial building layout scheme generation model through different channels, respectively, wherein the building information includes the location, length, width and height of each building in the first building layout scheme;

[0149] Parsing the building information, the first surrounding information, and the first plot boundary to obtain an original semantic map;

[0150] Performing multi-scale encoding on the planning conditions to obtain a condition feature map;

[0151] Performing feature fusion on the original semantic map and the conditional feature map to obtain a fused feature map;

[0152] The image authenticity and conditional rationality of the fused feature map are evaluated respectively by a preset global discriminator and an attribute discriminator, thereby optimizing and adjusting the initial building layout scheme generation model.

[0153] In one possible implementation, Fig.10As shown, the solution generation module 20 includes a morphological similarity calculation unit 201, a conditional similarity calculation unit 202, a comprehensive similarity calculation unit 203 and a solution generation unit 204;

[0154] The morphological similarity calculation unit 201 is used to calculate the Hamming distance between the target area and each plot in the case sample library according to the plot boundary, so as to obtain the morphological similarity between the target area and each plot in the case sample library;

[0155] The condition similarity calculation unit 202 is used to calculate the condition similarity between the target area and each plot in the case sample library according to the planning conditions and a preset formula;

[0156] The comprehensive similarity calculation unit 203 is used to combine the morphological similarity and the conditional similarity to obtain the comprehensive similarity between the target area and each plot in the case sample library;

[0157] The solution generation unit 204 is used to input the boundary of the target area, the surrounding information and the planning conditions of the solution to be generated into the building layout solution generation model to obtain the corresponding initial building layout solution when the comprehensive similarity of a certain plot in the case sample library exceeds a first preset value;

[0158] Otherwise, a preset number of plots are extracted from the case sample library in descending order of the comprehensive similarity to fine-tune the building layout scheme generation model, and then the plot boundaries, surrounding information and planning conditions of the to-be-generated scheme of the target area are input into the building layout scheme generation model to obtain the corresponding initial building layout scheme.

[0159] In one possible implementation, Fig.11 As shown, the scheme optimization module 30 includes a vectorization processing unit 301, an attraction-repulsion force calculation unit 302 and an optimization unit 303:

[0160] The vectorization processing unit 301 is used to perform vectorization processing on the initial building layout scheme to obtain a corresponding vectorized building layout scheme;

[0161] The attraction-repulsion calculation unit 302 is used to calculate the distance between each building in the vectorized building layout solution and compare it with the preset standard distance to convert it into the attraction-repulsion between each building;

[0162] The optimization unit 303 is used to move each building in the vectorized building layout solution based on the attraction-repulsion model according to the attraction-repulsion between the buildings, so that each building in the vectorized building layout solution reaches a balanced state, and obtain the building layout solution.

[0163] Furthermore, the vectorization processing unit 301 performs vectorization processing on the initial building layout plan to obtain a corresponding vectorized building layout plan, including filtering processing, vectorization operation, height value clustering, and building outline replacement on the initial building layout plan.

[0164] The embodiment of the present invention provides a building layout scheme automatic generation system based on GauGAN, by constructing a building layout scheme generation model based on GauGAN and training it according to a preset case sample library, then inputting the relevant information of the target area into the building layout scheme generation model, thereby generating an initial building layout scheme that meets the user's requirements, and finally using the attraction and repulsion model to optimize the layout of the initial building layout scheme to obtain the final building layout scheme. In the scheme generation process, the embodiment of the present invention utilizes the relevant information such as the plot boundary, surrounding information and planning conditions of the scheme to be generated in the target area, so that the scheme generated by the model can meet the basic planning conditions and can also coordinate with the surrounding buildings and natural environment, thereby improving the accuracy of the automatic generation of the building layout scheme; the building layout scheme generation model is constructed based on GauGAN, and compared with the existing genetic algorithm and rule-based method, complex rules and calculations are not required, and the scheme generation efficiency is higher; finally, the initial scheme generated by the building layout scheme generation model is further optimized based on the attraction and repulsion model, so that the final building layout scheme is more in line with the actual situation, thereby improving the rationality and practicality of the automatic generation of the building layout scheme.

[0165] The more detailed working principle and step flow of this embodiment can refer to, but are not limited to, the relevant records of Embodiment 1.

[0166] The specific embodiments described above further illustrate the purpose, technical solutions and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. It is particularly pointed out that for those skilled in the art, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for automatically generating building layout solutions based on GauGAN, characterized in that: include: Obtaining the plot boundary, surrounding information and planning conditions of the to-be-generated scheme of the target area, wherein the surrounding information includes the distribution information of buildings, roads, rivers and green spaces around the target area, and the planning conditions include the building density, volume ratio and building height limit required for the target area; The plot boundary, the surrounding information and the planning conditions are input into a preset building layout scheme generation model, so that the building layout scheme generation model automatically generates a corresponding initial building layout scheme according to the surrounding information and the planning conditions, including: calculating the Hamming distance between the target area and each plot in the case sample library according to the plot boundary, and obtaining the morphological similarity between the target area and each plot in the case sample library; calculating the conditional similarity between the target area and each plot in the case sample library according to the planning conditions and a preset formula; and combining the morphological similarity and the conditional similarity to obtain the target area and the case sample library. The comprehensive similarity of each plot in the sample library; when the comprehensive similarity of a plot in the case sample library exceeds a first preset value, the plot boundary, surrounding information and planning conditions of the to-be-generated scheme of the target area are input into the building layout scheme generation model to obtain the corresponding initial building layout scheme; otherwise, a preset number of plots are extracted from the case sample library in descending order of the comprehensive similarity to fine-tune the building layout scheme generation model, and then the plot boundary, surrounding information and planning conditions of the to-be-generated scheme of the target area are input into the building layout scheme generation model to obtain the corresponding initial building layout scheme; The building layout scheme generation model is constructed based on GauGAN and obtained by training according to a preset case sample library, including: constructing an initial building layout scheme generation model based on the network features of GauGAN and StyleGAN2; obtaining the building layout schemes of each area and constructing the case sample library; inputting each building layout scheme in the case sample library into the initial building layout scheme generation model for training, and obtaining the building layout scheme generation model; The initial building layout scheme is optimized based on the attraction-repulsion model, and the distance between each building in the initial building layout scheme is adjusted to obtain the building layout scheme, including: vectorizing the initial building layout scheme to obtain a corresponding vectorized building layout scheme; calculating the distance between each building in the vectorized building layout scheme and comparing it with a preset standard distance, and converting it into the attraction-repulsion force between each building; based on the attraction-repulsion model, each building in the vectorized building layout scheme is moved according to the attraction-repulsion force between each building, so that each building in the vectorized building layout scheme reaches a balanced state, and the building layout scheme is obtained.

2. The method for automatically generating a building layout scheme based on GauGAN according to claim 1, characterized in that: The process of inputting the first building layout scheme in the case sample library into the initial building layout scheme generation model for training includes: Inputting the building information, first surrounding information, first plot boundary and first planning condition corresponding to the first building layout scheme into the initial building layout scheme generation model through different channels, respectively, wherein the building information includes the location, length, width and height of each building in the first building layout scheme; Parsing the building information, the first surrounding information, and the first plot boundary to obtain an original semantic map; Performing multi-scale encoding on the planning conditions to obtain a condition feature map; Performing feature fusion on the original semantic map and the conditional feature map to obtain a fused feature map; The image authenticity and conditional rationality of the fused feature map are evaluated respectively by a preset global discriminator and an attribute discriminator, thereby optimizing and adjusting the initial building layout scheme generation model.

3. The method for automatically generating a building layout scheme based on GauGAN according to claim 1, characterized in that: The vectorizing the initial building layout plan to obtain the corresponding vectorized building layout plan includes filtering, vectorizing, height value clustering and building outline replacement on the initial building layout plan.

4. A GauGAN-based building layout automatic generation system, characterized in that: It includes acquisition module, solution generation module, modeling module and solution optimization module; The acquisition module is used to acquire the land boundary, surrounding information and planning conditions of the to-be-generated scheme of the target area, wherein the surrounding information includes the distribution information of buildings, roads, rivers and green spaces around the target area, and the planning conditions include the building density, volume ratio and building height limit required by the target area; The scheme generation module is used to input the plot boundary, the surrounding information and the planning conditions into a preset building layout scheme generation model, so that the building layout scheme generation model automatically generates a corresponding initial building layout scheme according to the surrounding information and the planning conditions. The scheme generation module includes a morphological similarity calculation unit, a conditional similarity calculation unit, a comprehensive similarity calculation unit and a scheme generation unit; wherein the morphological similarity calculation unit is used to calculate the Hamming distance between the target area and each plot in the case sample library according to the plot boundary, and obtain the morphological similarity between the target area and each plot in the case sample library; the conditional similarity calculation unit is used to calculate the conditional similarity between the target area and each plot in the case sample library according to the planning conditions and a preset formula. The comprehensive similarity calculation unit is used to combine the morphological similarity and the conditional similarity to obtain the comprehensive similarity between the target area and each plot in the case sample library; the solution generation unit is used to input the plot boundary, surrounding information and planning conditions of the to-be-generated solution of the target area into the building layout solution generation model to obtain the corresponding initial building layout solution when the comprehensive similarity of a certain plot in the case sample library exceeds a first preset value; otherwise, a preset number of plots are extracted from the case sample library in descending order of the comprehensive similarity to fine-tune the building layout solution generation model, and then the plot boundary, surrounding information and planning conditions of the to-be-generated solution of the target area are input into the building layout solution generation model to obtain the corresponding initial building layout solution; The modeling module is used to construct an initial building layout scheme generation model based on GauGAN and train it according to a preset case sample library to obtain the building layout scheme generation model, including a model construction unit, a sample library construction unit and a training unit; wherein the model construction unit is used to construct the initial building layout scheme generation model based on the network features of GauGAN and StyleGAN2; the sample library construction unit is used to obtain the building layout schemes of each area and construct the case sample library; the training unit is used to input each building layout scheme in the case sample library into the initial building layout scheme generation model for training to obtain the building layout scheme generation model; The scheme optimization module is used to optimize the layout of the initial building layout scheme based on the repulsive force model, adjust the distance between each building in the initial building layout scheme, and obtain the building layout scheme. The scheme optimization module includes a vectorization processing unit, a repulsive force calculation unit and an optimization unit: wherein the vectorization processing unit is used to vectorize the initial building layout scheme to obtain a corresponding vectorized building layout scheme; the repulsive force calculation unit is used to calculate the distance between each building in the vectorized building layout scheme and compare it with the preset standard distance, and convert it into the repulsive force between each building; the optimization unit is used to move each building in the vectorized building layout scheme according to the repulsive force between each building based on the repulsive force model, so that each building in the vectorized building layout scheme reaches a balanced state, and obtains the building layout scheme.

5. The automatic generation system of building layout scheme based on GauGAN as claimed in claim 4, characterized in that: The process of inputting the first building layout scheme in the case sample library into the initial building layout scheme generation model for training by the training unit includes: Inputting the building information, first surrounding information, first plot boundary and first planning condition corresponding to the first building layout scheme into the initial building layout scheme generation model through different channels, respectively, wherein the building information includes the location, length, width and height of each building in the first building layout scheme; Parsing the building information, the first surrounding information, and the first plot boundary to obtain an original semantic map; Performing multi-scale encoding on the planning conditions to obtain a condition feature map; Performing feature fusion on the original semantic map and the conditional feature map to obtain a fused feature map; The image authenticity and conditional rationality of the fused feature map are evaluated respectively by a preset global discriminator and an attribute discriminator, thereby optimizing and adjusting the initial building layout scheme generation model.

6. The automatic generation system of building layout scheme based on GauGAN as claimed in claim 4, characterized in that: The vectorization processing unit performs vectorization processing on the initial building layout plan to obtain a corresponding vectorized building layout plan, including filtering processing, vectorization operation, height value clustering, and building outline replacement on the initial building layout plan.

Images