Method for automatically generating vectorized house type planar graph

Abstract

The invention discloses a method for automatically generating a vectorized house type planar graph, which comprises the following steps of: 1, generating a node set for representing wall intersections in the house type planar graph by adopting a node generation model, each node comprising a position coordinate of the node and a semantic label of a surrounding room, meanwhile, the geometric quality of the generated node set is improved through a novel loss function item; 2, taking a candidate edge set formed by combining nodes in the node set obtained in the step 1 as input, predicting to obtain an edge set, and meanwhile, improving the geometric quality of the predicted edge set through a novel random self-supervised geometric enhancement method; and step 3, on the basis of semantic information of the rooms associated with the angular points, all minimum rings, namely all rooms, are obtained through operation of traversing the minimum rings of the graph. According to the invention, tedious manual design is reduced, and the intelligent degree is high.

Description

Technical Field

[0001] The invention relates to the field of intelligent design methods for floor plans, and in particular to a method for automatically generating vectorized floor plans. Background Art

[0002] In the field of computer graphics and vision, the automatic design of architectural floor plans has attracted widespread attention, because detailed architectural floor plans are essential for building houses, interior scene design, etc. Architects usually draw sketches manually, evaluate and adjust them iteratively to get a satisfactory design result. Unfortunately, this process is time-consuming and requires the participation of professional architects, which is costly.

[0003] The automatic generation of floor plans can enable the general public to participate in the design process, allowing them to customize floor plans according to their needs and preferences; in addition, it can also provide inspiration and reference for architects or designers to improve design efficiency. Deep learning is the ability of computers to achieve self-learning and improvement. Deep neural networks are the basis of deep learning. Deep layers make neural networks have stronger representation capabilities and stronger learning capabilities. The real world has accumulated a large number of high-quality floor plans. A direct idea is to discover the implicit design rules inherent in floor plans from the vector format of existing floor plans and realize the automatic generation of floor plans.

[0004] Therefore, it is necessary to propose a vectorized floor plan generation method based on deep learning neural network to solve the problems of architects' manual design of vector floor plans, which are time-consuming, inefficient, costly, and dependent on professional knowledge, while improving the geometric quality of the design results. Summary of the invention

[0005] The present invention provides a method for automatically generating vectorized floor plans to solve the problems of architects manually designing vector floor plans in the prior art, such as being time-consuming, inefficient, costly, and dependent on professional knowledge, while improving the design quality and enhancing the geometric quality of the generated results.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is: A method for automatically generating a vectorized floor plan includes the following steps: Step 1: using a node generation model based on a diffusion model, using random noise sampled from a standard Gaussian distribution as input of the node generation model, and generating a preliminary node set by the node generation model, each node in the preliminary node set represents a wall intersection, and each node has position coordinates, semantic information of surrounding rooms, and background attributes; Then, judging whether the corresponding node is a filler node according to the background attribute value of each node in the preliminary node set, and removing the filler nodes in the preliminary node set, and forming a node set of the plane graph with the remaining nodes; Step 2: Using an edge prediction model based on a neural network, a candidate edge set is constructed using candidate edges formed by any combination of two nodes in the node set of the plane graph obtained in step 1, and the candidate edge set is used as an input of the edge prediction model. The edge prediction model predicts whether there is an edge between each pair of nodes, thereby obtaining an edge set of the plane graph, where the edges in the edge set represent walls in the plane graph, and the node set of the plane graph is combined with the edge set, thereby obtaining a structural diagram of the plane graph; Step 3: traverse the structure diagram of the floor plan through the minimum ring operation, and obtain all the minimum rings, that is, all the rooms, thereby extracting a room set from the structure diagram; combine the structure diagram with the room set to obtain a vectorized floor plan.

[0007] Furthermore, the neural network included in the diffusion model based on the node generation model in step 1 uses a trained Transformer, which is recorded as a node Transformer.

[0008] Furthermore, a multi-base alignment loss function is used during the node transformer training, and the geometric alignment effect of the node transformer is enhanced by hybrid optimization of the alignment errors under multiple bases.

[0009] Furthermore, the neural network-based edge prediction model in step 2 uses a trained Transformer, which is denoted as an edge Transformer.

[0010] Furthermore, a random self-supervised geometric enhancement method is introduced during the edge transformer training. For each candidate edge, a point is randomly inserted on the edge, and the edge transformer is required to predict the random coefficient of the interpolation point, thereby strengthening the edge transformer's understanding of the overall geometric features of the edge.

[0011] Furthermore, in step 3, the operation process of traversing the minimum loop of the graph is: take each edge of the structure graph as the starting edge, specify a certain endpoint to another endpoint as the starting direction, iteratively search for the edge of all the edges in the next step, which is located on the left side of the current edge and has the smallest angle with the current edge, until the starting edge is found in the next step, forming a loop, and if it is a non-repeating minimum loop, it is recorded, and all minimum loops are obtained by traversing each edge of the structure graph.

[0012] Furthermore, step 3 also includes: for each room, selecting the room semantic information with the highest frequency in the room semantic information associated with all corner points constituting the room as the final room category; if the frequency of multiple categories is the highest, considering the scarcity factor of the category, the room type is determined in the following priority order: storage room>bathroom>kitchen>bedroom>balcony>living room.

[0013] Floor plans can be represented by graph structures, where the nodes of the graph represent the intersections of the walls in the floor plan, and the edges of the graph represent the walls in the floor plan. The semantic information of the rooms associated with the nodes is used as the attributes of the corresponding vertices. Floor plans can be represented as connected graph structures. In this way, the generation problem of vectorized floor plans can be transformed into the generation problem of graph structures. The essence of the generation problem is to learn the probability distribution of data and randomly sample from it. Due to the irregularity of the graph structure, its probability distribution is quite complex; however, the node set and edge set are regular, and it is feasible to decouple the graph structure into node set and edge set and process them separately.

[0014] The floor plan has a multi-level structure, and the generated node set basically determines the geometric structure of the entire floor plan. By considering the semantic information of the rooms related to the nodes as attributes of the nodes, the geometric and semantic dependencies of the floor plan can be learned at the same time.

[0015] In the present invention, the generation of the node set is achieved through a diffusion model, which defines a series of mathematical steps to generate data by gradually transferring noise to the target distribution, and uses a neural network to predict the coefficients of each step. Given the generated node set, the edge set is basically determined and can be simply predicted using a neural network. Therefore, the present invention proposes a novel generation model that divides the generation process of the structure graph into two stages: node generation and edge prediction. First, the probability distribution of the node set is learned through the generation model, and the node set is sampled and generated. Then, the edge set is predicted based on the generated node set to obtain a complete structure graph. Finally, all the minimum polygon rings (i.e., rooms) are extracted from the structure graph to obtain the final vectorized floor plan.

[0016] Neural networks are probabilistic in nature. In the above process, there may be some noise in the nodes generated in the node generation stage, such as the generated node coordinates are not aligned. In order to solve this problem, the present invention introduces an alignment loss function, which takes the "degree of misalignment" between nodes as the optimization object, thereby reducing the coordinate noise caused by the probabilistic nature of the neural network. At the same time, in the edge prediction stage, the neural network may output some unreasonable predictions, such as missing or false wall segments, resulting in geometric unreasonableness. In order to solve this problem, the present invention enhances the model's understanding of the overall geometric characteristics of the edge and improves the geometric quality by randomly inserting a third point on the edge and self-supervisingly requiring the model to predict the interpolation coefficient of the third point.

[0017] Compared with the prior art, the present invention has the following advantages: The present invention can automatically generate vectorized floor plans of apartment types, so that ordinary homebuyers can also participate in the design process and customize the floor plans according to their own needs and preferences; in addition, it can also provide inspiration and reference for architects or designers to improve design efficiency.

[0018] The method of the invention is simple, easy to deploy on a computer, has a high degree of automation, reduces tedious manual design, and the model based on deep neural network has sufficient and reliable theoretical support. It can solve the problems of architects' manual design of vector floor plans being time-consuming, inefficient, costly, and dependent on professional knowledge, while improving the geometric quality of the design results. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 A workflow diagram generated for a node in an embodiment of the present invention.

[0020] Figure 2 This is a technical solution diagram of the node Transformer in an embodiment of the present invention.

[0021] Figure 3 The figure is a workflow diagram of edge prediction according to an embodiment of the present invention.

[0022] Figure 4 The present invention is a flowchart of the minimum ring extraction process of a planar graph according to an embodiment of the present invention.

[0023] Figure 5 The following is a flowchart of an embodiment of the present invention. DETAILED DESCRIPTION

[0024] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0025] like Figure 1 , Figure 5 As shown, this embodiment discloses a method for automatically generating a vectorized floor plan, comprising the following steps: Step 1: Use a node generation model based on a diffusion model, and use random noise sampled from a standard Gaussian distribution as the input of the node generation model. The node generation model generates a preliminary node set, in which each node represents a wall intersection, and each node has position coordinates, semantic information of surrounding rooms, and background attributes. Then, determine whether the corresponding node is a filler node based on the background attribute value of each node in the preliminary node set, and remove the filler nodes in the preliminary node set, so as to form a node set of the plan view with the remaining nodes; In this embodiment, the node generation model is implemented based on the diffusion model. The diffusion model is a type of generative model that can generate samples that are highly similar to the original data from simple noise by learning the noise addition process of the data and reversing it. The diffusion model includes a forward process and a reverse process. The forward process adds noise to the original sample (corresponding to time step 0). As the time step advances from 0 to the maximum value T, the noise level in the sample continues to rise, and eventually becomes almost the same as Gaussian noise. The reverse process is the inverse operation of this process. Starting from the Gaussian white noise corresponding to the time step T, a neural network is used to gradually remove the noise and restore clear data; the neural network, as a component in the reverse process, is responsible for estimating the noise contained in the sample in each step t as part of the mean parameter of the Gaussian distribution corresponding to the next time step t-1.

[0026] In this embodiment, the node generation model based on the diffusion model uses Transformer as the neural network part, which is denoted as node Transformer. The node Transformer consists of a decoder and multiple downstream prediction heads, where the decoder consists of multiple decoder layers, and each decoder layer contains "instance normalization layer", "global self-attention layer", "residual connection", "instance normalization layer", "fully connected layer", and "residual connection" in sequence. The model captures the dependencies between nodes by stacking multiple layers of Transformer decoder layers. Each downstream prediction head is a fully connected layer structure.

[0027] In this embodiment, the node Transformer is pre-trained using the RPLAN dataset, which contains more than 80,000 densely annotated floor plans, each of which has pixel-by-pixel labels such as rooms and walls.

[0028] Specifically, in order to train the node Transformer, this embodiment performs a series of preprocessing steps on the RPLAN dataset to extract clean vectorized floor plans as training data. The preprocessing process is as follows: (1) The image is binarized, with the wall represented as white and other areas represented as black.

[0029] (2) Since the wall thickness varies, the thickness is normalized by repeated morphological operations (erosion, dilation) and template matching. The white pixels representing the wall are iteratively eroded until the next erosion results in a decrease in the number of connected regions in the image, indicating that some wall segments have been eroded to a thickness of 1 pixel.

[0030] (3) Apply a series of template matching operations, sliding a 3×3 window across the image. If the match is successful, the matching area is replaced with white. The template shape contains lines with a thickness of 1 pixel or defective local wall shapes. Through this process, the wall shape is gradually standardized and the thickness becomes more uniform.

[0031] The above two steps (2) and (3) form an iteration, and this iterative process continues until the thickness of all walls is eroded to 1 pixel: the judgment criterion is that the next erosion will cause all walls to disappear, making the image completely black. At this time, the wall intersections and wall segments can be obtained from the image, where the wall intersections contain the position coordinates, and the wall segments contain the indexes of the wall intersections at both ends, recording which two wall intersections are the endpoints of the wall segment.

[0032] (4) The semantic information of the surrounding room at each wall intersection is obtained from the four-channel images of the original RPLAN dataset.

[0033] All unprocessable data samples in the above steps are discarded.

[0034] After the node Transformer is trained using the data set, it is used as the node generation model described in this embodiment.

[0035] In this embodiment, Figure 1 Taking the node generation process shown in the figure as an example, the input of the node generation model is a set of random noise nodes sampled from a standard Gaussian distribution. Each node contains its position coordinates, the semantic labels of the surrounding rooms, and background attributes. The background attributes are generated because the diffusion model requires a fixed size of the node set in implementation. Therefore, the size of the node set is set to a larger fixed value, and the node set is filled to this size. At the same time, the background attributes are set to distinguish whether the node is real or a filler node. At this time, all attributes of all nodes (position coordinates, semantic labels of the surrounding rooms, background) are pure noise and have no information. At this time, the time step is T.

[0036] Next, the attributes of each input node are mapped to a high-dimensional representation as the input of the node transformer, where the position coordinates are mapped to a high-dimensional vector through triangular position encoding; the semantic labels and background attributes of the surrounding rooms are mapped to embedding vectors of the same dimension through a fully connected layer. In addition, the current time step is also represented in a high dimension through triangular position encoding + fully connected layer mapping. The high-dimensional representation of the above position coordinates, semantic labels and background attributes of the surrounding rooms, and time steps are added together to obtain the high-dimensional embedding of the node; the high-dimensional embeddings of all nodes form the high-dimensional embedding of the node set and are input into the node transformer.

[0037] like Figure 2As shown, the node Transformer contains the Transformer decoder and the downstream prediction head. The Transformer decoder processes the node set, and the decoder consists of multiple stacked layers, each of which contains "instance normalization layer", "global self-attention layer", "residual connection", "instance normalization layer", "fully connected layer", and "residual connection" in order.

[0038] Through the self-attention mechanism, the node transformer can capture the complex dependencies between nodes. The output of the transformer decoder is a high-dimensional embedding of the node set mapped by the transformer, which is input to the downstream prediction head for attribute prediction. The downstream prediction head is based on a fully connected layer, with a total of 2 dimensions, and the output dimensions are 2 and 8 respectively. The former is the noise estimate of the position coordinates, and the latter is the joint noise estimate of the semantic label of the surrounding room-background; the two are spliced ​​to form a noise estimate of the node set (10 dimensions). The noise estimate is used to calculate the mean parameter of the Gaussian distribution of the next time step T-1 (for each time step, the variance is a preset fixed value), and the Gaussian distribution of the next time step T-1 is obtained, from which sampling is obtained to obtain the node set at time T-1. These node sets will iteratively repeat the above process until the time step reaches 0, and a clean node set is obtained, which contains the precise position of each node, the semantic information of the surrounding room, and the background attributes. According to the background attribute value of each node, it is judged whether the node is a filler node, and the filler nodes in the preliminary node set are removed to obtain the node set of the floor plan.

[0039] The Transformer loss function includes the MSE regression loss between the noise estimate (10 dimensions) of the node set and the real noise (10 dimensions) added in the forward process. During the training process, in order to improve the geometric quality of the generated node set, this embodiment introduces a multi-base alignment loss function as the loss function for node Transformer training. The multi-base alignment loss function optimizes the node alignment error in different bases such as binary, quaternary, octal and hexadecimal. Specifically, the numerical value of the "alignment deviation" between the nodes is converted into a numerical representation of a different base, and the error is calculated bit by bit and accumulated as an item of the loss function, which prompts the nodes to be accurately aligned at multiple scales, thereby enhancing the geometric accuracy of the generated plane graph. At this time, the loss function includes the MSE regression loss and the multi-base alignment loss function between the noise estimate (10 dimensions) of the node set and the real noise (10 dimensions) added in the forward process.

[0040] Step 2: Use a neural network-based edge prediction model to construct a candidate edge set with candidate edges formed by any two-node combinations in the node set of the plane graph obtained in step 1, and use the candidate edge set as input to the edge prediction model. The edge prediction model predicts whether there is an edge between each pair of nodes, thereby obtaining the edge set of the plane graph. The edges in the edge set represent walls in the plane graph. The node set of the plane graph is combined with the edge set to obtain a structural diagram of the plane graph.

[0041] In this embodiment, the edge prediction model based on the neural network uses a trained Transformer, which is denoted as an edge Transformer.

[0042] by Figure 3 Taking the edge prediction shown in the figure as an example, the input of the edge prediction model is all node pairs formed by the combination of all nodes in the node set. Each pair of nodes is used as a candidate edge, and all candidate edges form a candidate edge set. For each candidate edge, the high-dimensional embeddings of the two nodes of the corresponding node pair are extracted, including the position coordinate encoding and the semantic label of the surrounding room - background embedding. All high-dimensional embeddings of each node are concatenated to obtain the high-dimensional embedding of each node, and then the high-dimensional embeddings of the two nodes are added to obtain the high-dimensional embedding of the candidate edge.

[0043] In order to enhance the model's understanding of edge geometric features, the edge transformer of this embodiment introduces a random self-supervised geometric enhancement method during the training process. Specifically, a point is randomly inserted on each candidate edge, and the position is determined by the random interpolation coefficient λ, that is, the point is located at a random position between the two endpoints. The position coordinate encoding of the interpolation point is extracted, and the semantic label-background of the room around the interpolation point is set to a zero vector to extract the corresponding embedding, and the two are added to obtain a high-dimensional embedding of the interpolation point. The high-dimensional embedding of the interpolation point is added to the high-dimensional embedding of the candidate edge. The high-dimensional embeddings of all candidate edges form a high-dimensional embedding of the candidate edge set, which is input into the edge transformer.

[0044] The structure of the edge transformer is similar to that of the node transformer, including a transformer decoder and a downstream prediction head. The transformer decoder processes the candidate edge set. The decoder consists of multiple stacked layers, each of which contains "instance normalization layer", "global self-attention layer", "residual connection", "instance normalization layer", "fully connected layer", and "residual connection" in order. Through the self-attention mechanism, the edge transformer can capture the complex dependencies between candidate edges. The output of the transformer decoder is a high-dimensional embedding of the candidate edge set mapped by the transformer, which is input to the downstream prediction head for attribute prediction. The downstream prediction head is based on a fully connected layer, with a total of 2, and the output dimensions are 2 and 1 respectively. The former is the two category probabilities of the true / false binary classification of the candidate edge, and the latter is the random interpolation coefficient λ.

[0045] Since the edge is undirected, in order to avoid the ambiguity introduced by directionality, when λ is greater than 0.5, it is replaced by 1-λ in this embodiment to ensure that the value range of λ is within [0, 0.5]. In this way, the edge transformer is forced to focus on the overall geometric characteristics of the edge during the learning process, rather than just relying on the information of the endpoints. The loss function consists of the cross entropy loss of edge classification and the L1 regression loss of the interpolation coefficient. By optimizing these two objectives at the same time, the edge transformer can more accurately judge the existence of the edge by having a deeper understanding of the geometric characteristics of the edge. The set of candidate edges that are judged to be true is obtained as the edge set. At this point, a structural graph consisting of a node set and an edge set is obtained, which contains the information of a complete apartment floor plan.

[0046] Step 3: Based on the structure diagram of the floor plan obtained in step 2, the structure diagram is traversed by traversing the minimum ring operation of the structure diagram of the floor plan, and all the minimum rings are obtained, that is, all the rooms, thereby extracting a room set from the structure diagram; the structure diagram is combined with the room set to obtain a vectorized floor plan.

[0047] Specifically, the minimum loop operation process of traversing the graph is to number each node of the structure graph in an arbitrary order in advance, and set all the edges of the structure graph as the edges to be visited; take each edge to be visited in the structure graph as the starting edge, specify the endpoint numbering from small to large as the starting direction, and iteratively search for the edge on the left side of the current edge and with the smallest angle with the current edge among all the edges of the next step, until the starting edge is found in the next step, forming a loop; if the rotation direction of the loop is counterclockwise, it is a minimum loop, record the minimum loop, and set all the edges visited in the process of forming the minimum loop from small to large in terms of endpoint numbering as visited; if the rotation direction of the loop is clockwise, it is a maximum loop, do not record the loop, and set all the edges visited in the process of forming the loop from small to large in terms of endpoint numbering as visited; iterate the above process until there are no edges to be visited, and obtain all minimum loops. Figure 4 Figure 2 shows an example of the formation of a minimal ring.

[0048] For each room, the room semantic information with the highest frequency among all the room semantic information associated with the corner points of the room (excluding the outdoor) is selected as the final room category; if multiple categories have the highest frequency, considering the scarcity factor of the category, the room type is determined in the following priority order: storage room > bathroom > kitchen > bedroom > balcony > living room. Thus, a vectorized floor plan is obtained.

[0049] The preferred embodiments of the present invention are described in detail above in conjunction with the accompanying drawings. The embodiments described in the present invention are merely descriptions of the preferred embodiments of the present invention, and do not limit the concept and scope of the present invention. The various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction, and such combinations should also be regarded as the contents disclosed in the present disclosure as long as they do not violate the concept of the present invention. In order to avoid unnecessary repetition, the present invention will not further describe various possible combinations.

[0050] The present invention is not limited to the specific details of the above-mentioned embodiments. Within the technical concept of the present invention and without departing from the design concept of the present invention, various modifications and improvements made to the technical solution of the present invention by technical personnel in this field should fall within the protection scope of the present invention. The technical contents for which protection is sought in the present invention have been fully recorded in the claims.

Claims

1. A method for automatically generating a vectorized floor plan, characterized in that: The following steps are involved: Step 1: using a node generation model based on a diffusion model, using random noise sampled from a standard Gaussian distribution as input of the node generation model, and generating a preliminary node set by the node generation model, each node in the preliminary node set represents a wall intersection, and each node has position coordinates, semantic information of surrounding rooms, and background attributes; Then, judging whether the corresponding node is a filler node according to the background attribute value of each node in the preliminary node set, and removing the filler nodes in the preliminary node set, and forming a node set of the plane graph with the remaining nodes; Step 2: Using an edge prediction model based on a neural network, a candidate edge set is constructed using candidate edges formed by any combination of two nodes in the node set of the plane graph obtained in step 1, and the candidate edge set is used as an input of the edge prediction model. The edge prediction model predicts whether there is an edge between each pair of nodes, thereby obtaining an edge set of the plane graph, where the edges in the edge set represent walls in the plane graph, and the node set of the plane graph is combined with the edge set, thereby obtaining a structural diagram of the plane graph; Step 3: traverse the structure diagram of the floor plan through the minimum ring operation, and obtain all the minimum rings, that is, all the rooms, thereby extracting a room set from the structure diagram; combine the structure diagram with the room set to obtain a vectorized floor plan.

2. The method for automatically generating a vectorized floor plan according to claim 1, characterized in that: The neural network included in the diffusion model based on the node generation model in step 1 uses a trained Transformer, which is denoted as a node Transformer.

3. The method for automatically generating a vectorized floor plan according to claim 2, characterized in that: The node transformer is trained by using a multi-base alignment loss function, and the geometric alignment effect of the node transformer is enhanced by hybrid optimization of the alignment errors under multiple bases.

4. The method for automatically generating a vectorized floor plan according to claim 1, characterized in that: The neural network-based edge prediction model described in step 2 uses a trained Transformer, denoted as an edge Transformer.

5. The method for automatically generating a vectorized floor plan according to claim 4, characterized in that: A random self-supervised geometric enhancement method is introduced during edge transformer training. For each candidate edge, a point is randomly inserted on the edge, and the edge transformer is required to predict the random coefficient of the interpolation point, thereby strengthening the edge transformer's understanding of the overall geometric features of the edge.

6. The method for automatically generating a vectorized floor plan according to claim 1, characterized in that: In step 3, the operation process of traversing the minimum loop of the graph is: take each edge of the structure graph as the starting edge, specify a certain endpoint to another endpoint as the starting direction, iteratively search for the edge of all the edges in the next step, which is located on the left side of the current edge and has the smallest angle with the current edge, until the starting edge is found in the next step, forming a loop, and if it is a non-repeating minimum loop, record it, and obtain all minimum loops by traversing each edge of the structure graph.

7. The method for automatically generating a vectorized floor plan according to claim 1, characterized in that: Step 3 also includes: for each room, selecting the room semantic information with the highest frequency among the room semantic information associated with all corner points constituting the room as the final room category; if multiple categories have the highest frequency, considering the scarcity factor of the category, determine the room type in the following priority order: storage room>bathroom>kitchen>bedroom>balcony>living room.

Images