Algorithm for automatically identifying and generating house type image AI based on home decoration

Abstract

The invention discloses a house type image AI automatic identification generation algorithm based on home decoration, and relates to the field of house type image automatic identification, and the algorithm comprises the following steps: inputting a house type image, carrying out the denoising of the image through Markov random field modeling, enhancing the edge and texture of the image through constructing a differential equation, carrying out the conditional generation of an adversarial network, and outputting a standardized floating-point array. The method comprises the following steps: creating a house type graph region-of-interest model, analyzing to obtain a final target graph cutting region, constructing a random graph model, updating node features layer by layer through a geometric graph convolutional network, then generating a probability door and window bounding box by utilizing Gaussian process regression, and constructing a random graph model through a structured coding floating point matrix of a multi-dimensional floating point matrix. The floating point matrix is converted into structured point, line and hole data through a broken line simplification formula and an orthogonalization constraint formula, polygon triangularization processing is conducted through a wall thickness mapping formula and a floor generation formula, construction of walls, doors, windows, roofs and floors is completed through analysis, and it is guaranteed that the generated house type image is accurate.

Description

Technical Field

[0001] The present invention relates to the field of automatic recognition of floor plans, and specifically to an algorithm for automatic recognition and generation of floor plans based on AI for home decoration. Background Art

[0002] With the continuous advancement of intelligent technology and the digitization of the home improvement industry, accurate and efficient automatic recognition and generation of floor plans has become a key requirement in fields such as digital real estate management and interior design. Traditional manual floor plan drawing is labor-intensive and often influenced by individual subjective thinking, resulting in large errors and low efficiency. Therefore, algorithms for automatic AI recognition and generation of floor plans for home improvement have emerged.

[0003] The existing AI-based automatic recognition and generation algorithm for floor plans based on home decoration has difficulty in balancing the preservation of image noise and edge texture during operation, and it is also difficult to ensure accurate recognition of key areas such as walls, doors and windows. As a result, structured data often deviates and is lost during the conversion process, ultimately resulting in a large gap between the generated results and reality.

[0004] In order to solve the above-mentioned defects, a technical solution is now provided. Summary of the Invention

[0005] In order to solve the technical problems raised by the above background technology, the present invention is proposed. The embodiments of the present invention provide an algorithm for automatically identifying and generating floor plans based on AI for home decoration.

[0006] The purpose of the present invention can be achieved by the following technical solution: an algorithm for automatically identifying and generating a home improvement floor plan based on AI includes the following steps:

[0007] Step S101: Input the floor plan, perform image denoising through Markov random field modeling, enhance image edges and textures by constructing differential equations, conditional generative adversarial networks, and output a standardized floating-point array;

[0008] Step S102: creating a floor plan area of ​​interest model according to the standardized floating-point array, and obtaining the final target cut area through the floor plan area of ​​interest model;

[0009] Step S103: Based on the generated final target cut-out area, a random graph model is constructed, and node features are updated layer by layer through a geometric graph convolutional network. Subsequently, probabilistic door and window bounding boxes are generated using Gaussian process regression. A floating-point matrix containing wall connection points, line segment parameters, and door and window bounding boxes is obtained through structured encoding of a multi-dimensional floating-point matrix.

[0010] Step S104: converting the floating-point matrix into structured point, line, and hole data through a broken line simplification formula and an orthogonal constraint formula;

[0011] Step S105: Based on the point, line, and hole data in the floor plan, key points, walls with thickness, and doors and windows are generated using the wall thickness mapping formula. Then, based on the closed area formed by the walls, polygon triangulation is performed using the floor generation formula to determine the area of ​​the area, and then the floor is generated. At the same time, the corresponding roof is generated in combination with the wall layout to complete the construction of walls, doors and windows, roof, and floor.

[0012] Furthermore, the standardized floating-point array analysis step:

[0013] The house type image with denoised image and enhanced image edges and textures is mapped into a standardized floating point array through the conditional generative adversarial network. The conditional generative adversarial network includes a generator and a discriminator, wherein the generator includes an input layer, an output layer and a structure layer. The house type image with denoised image and enhanced image edges and textures is input into the input layer. The structure layer includes a U-Net architecture. The output layer outputs a standardized floating point array I sd ∈[0,1] H×W , [0,1] H×W It is a two-dimensional matrix with a height of H and a width of W, and the element values ​​are normalized between 0 and 1. The discriminator includes an input layer, an output layer, and a structure layer. sd or I real Real labeled data I rea输入 The structure layer includes a multi-layer convolutional network that gradually downsamples to extract global features, and the output layer outputs the discrimination results. During the training process, the loss function combines adversarial loss and L1 reconstruction loss.

[0014] Furthermore, the steps of analyzing the apartment image after noise reduction and enhancing the image edges and textures are as follows:

[0015] A Markov random field model is established based on the apartment layout image to enhance the spatial correlation between pixels, suppress noise while retaining the edge structure to perform denoising on the image. The Markov random field model includes:

[0016]

[0017] Where P(I|I0) is the probability distribution of the denoised image I under the condition of the known noisy image I0, δ is the set of all pixels in the image, (x, y) is the coordinate of a pixel in the image, I is the pixel value of the denoised image, I0 is the pixel value of the original noisy image, λ1 is the weight parameter of the data fidelity term, λ2 is the weight parameter of the smoothing term, ɑ(x, y) is the neighborhood pixel set of pixel (x, y), ζ(*) is the potential function, (x / ,y / ) is the coordinate of the neighboring pixel of pixel (x, y), exp is the exponential function with the natural constant e as the base, and ∝ is the proportional sign;

[0018] The specific analysis of the potential function ζ(*) is:

[0019] According to the Huber function, the potential function ζ(*) is obtained. Where Δ is the difference in the values ​​of the neighboring pixels, specifically I(x, y)-I0(x / ,y / ), γ is the threshold for distinguishing small difference noise from large difference edges;

[0020] By constructing differential equations to enhance the edges and textures of the house image and avoid over-smoothing, the differential equation is: in Is the diffusion coefficient function, controlling the diffusion intensity, the specific diffusion coefficient function Where div(*) is the divergence operator, is the gradient vector of image I, is the gradient amplitude, indicating the strength of the edge, and L is the gradient threshold.

[0021] Furthermore, the final target cut image area analysis step:

[0022] Creating a floor plan region of interest model based on a standardized floating-point array, wherein the floor plan region of interest model includes: detecting wall segments in the image using the Hough transform line detection formula ρ = xcosθ + ysinθ, where θ is the angle between the line and the x-axis and ρ is the distance from the line to the origin;

[0023] Also includes: contour area filtering formula if Area(K i ) <C area ,then reject K i , where C area is the set area threshold, Area(K i ) is the contour K i The area of ​​the image is calculated, the contours with an area smaller than the area threshold are removed, key areas such as walls, doors and windows are located, and the target cutting area is finally obtained.

[0024] Furthermore, a random graph model is constructed based on a standardized floating-point array. The model includes: Where H(0) is the initial floor plan structure, H(t) is the floor plan structure at time t, μ H is the deterministic evolution rate, σ H is the topological diffusion coefficient, which controls the randomness of the connection point generation, dW(s) is the Brownian motion term, ds is the differential of the integral variable s, and its value ranges from 0 to t;

[0025] The node features are updated layer by layer through the graph convolution operator in the geometric graph convolution network. The graph convolution operator is in represents the feature vector of node u (wall endpoint) at layer v, e pu represents the edge feature, N(u) represents the set of neighbor nodes of node u, ψ represents the node update function, φ represents the learnable edge message function, and θ represents the parameters of the learnable edge message function φ. Represents the feature vector of node p in layer v;

[0026] The door and window positions are predicted by Gaussian process regression, and the output is the mean and covariance.

[0027] Furthermore, the structured coding analysis step of the multi-dimensional floating point matrix is:

[0028] The channels encode wall connection points, line segment parameters, and door and window bounding boxes respectively. All data are normalized to the range of [0,1]. The connection point information is identified by the mask channel, where 1 indicates existence and 0 indicates non-existence. Its precise coordinates are recorded through the incremental channel. The line segment parameters are stored by independent channels to store the starting and ending coordinates of the line segment and the confidence level. The door and window bounding boxes encode multi-channel data of the center position and rotation angle.

[0029] Furthermore, the structured point, line, and hole data analysis steps are:

[0030] The line simplification formula is used to determine whether to segment by judging the relationship between the vertical distance from the point to the line segment and the threshold value. The line simplification formula is: if d(j,j1j m )>∈, then split, where d represents point j to line segment j1j n The vertical distance of , ∈ is the simplified threshold, and the wall segment angle is forced to align with the coordinate axis through the orthogonal constraint formula to output structured point, line, and hole data. The orthogonal constraint formula is θ1 = arctan(y2-y1) / (x2-x1)stθ1∈{0°,90°,180°,270°}, stθ1∈{0°,90°,180°,270°} means that the calculated θ1 is constrained to 0°,90°,180°,270°, arctan is the inverse tangent function, (x1,y1) and (x2,y2) are the coordinates of the two endpoints of the wall segment.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] 1. The present invention denoises the image by inputting a floor plan through Markov random field modeling, enhances the image edge and texture by constructing differential equations, and generates a conditional adversarial network to output a standardized floating-point array. A floor plan region of interest model is created according to the standardized floating-point array, and the final target cutting area is obtained through the floor plan region of interest model. According to the generated final target cutting area, a random graph model is constructed, and node features are updated layer by layer through a geometric graph convolutional network. Subsequently, a probabilistic door and window bounding box is generated using Gaussian process regression. A floating-point matrix containing wall connection points, line segment parameters, and door and window bounding boxes is obtained through structured coding of a multi-dimensional floating-point matrix. This method can preserve both the noise and edge texture of the image and ensure accurate recognition of key areas such as walls, doors, and windows. In addition, through multi-step innovative processing such as Markov random field modeling for denoising, geometric graph convolutional network analysis of spatial relationships, and Gaussian process regression prediction of door and window positions, the bottleneck of existing technologies is effectively overcome, providing a reliable solution for the intelligent processing of home decoration floor plans.

[0033] 2. The present invention converts a floating-point matrix into structured point, line, and hole data through a broken line simplification formula and an orthogonalization constraint formula. Based on the point, line, and hole data in the floor plan, the wall thickness mapping formula is used to generate key points, walls with thickness, and doors and windows. Then, based on the closed area formed by the walls, polygon triangulation is performed using the floor generation formula to determine the area of ​​the area, and then the floor is generated. At the same time, the corresponding roof is generated in combination with the wall layout, completing the construction of walls, doors and windows, roofs, and floors, ensuring that the structured data will not be deviated or lost during the conversion process. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. The following drawings are not intentionally scaled to the actual size, and the focus is on illustrating the main purpose of the present invention.

[0035] Figure 1 is a flow chart of the method of the present invention;

[0036] Figure 2 This is the AI ​​model analysis flow chart of the present invention. DETAILED DESCRIPTION

[0037] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts also fall within the scope of protection of the present invention.

[0038] like Figure 1-Figure 2As shown in the figure, the algorithm for automatic recognition and generation of floor plans based on home decoration AI includes the following steps:

[0039] Step S101: Input the floor plan, perform image denoising through Markov random field modeling, enhance image edges and textures by constructing differential equations, conditional generative adversarial networks, and output a standardized floating-point array;

[0040] Step S102: creating a floor plan area of ​​interest model according to the standardized floating-point array, and obtaining the final target cut area through the floor plan area of ​​interest model;

[0041] Step S103: Based on the generated final target cut-out area, a random graph model is constructed, and node features are updated layer by layer through a geometric graph convolutional network. Subsequently, probabilistic door and window bounding boxes are generated using Gaussian process regression. A floating-point matrix containing wall connection points, line segment parameters, and door and window bounding boxes is obtained through structured encoding of a multi-dimensional floating-point matrix.

[0042] Step S104: converting the floating-point matrix into structured point, line, and hole data through a broken line simplification formula and an orthogonal constraint formula;

[0043] Step S105: Based on the point, line, and hole data in the floor plan, key points, walls with thickness, and doors and windows are generated using the wall thickness mapping formula. Then, based on the closed area formed by the walls, polygon triangulation is performed using the floor generation formula to determine the area of ​​the area, and then the floor is generated. At the same time, the corresponding roof is generated in combination with the wall layout to complete the construction of walls, doors and windows, roof, and floor.

[0044] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0045] A Markov random field model is established based on the apartment layout image to enhance the spatial correlation between pixels, suppress noise while retaining the edge structure to perform denoising on the image. The Markov random field model includes:

[0046]

[0047] Where P(I|I0) is the probability distribution of the denoised image I under the condition of the known noisy image I0, δ is the set of all pixels in the image, specifically the entire area of ​​the image, (x, y) is the coordinate of a pixel in the image, I is the pixel value of the denoised image, I0 is the pixel value of the original noisy image, λ1 is the weight parameter of the data fidelity term, λ2 is the weight parameter of the smoothing term, ɑ(x, y) is the neighborhood pixel set of pixel (x, y), ζ(*) is the potential function, (x / ,y / ) is the coordinate of the neighboring pixel of pixel (x, y), exp is the exponential function with the natural constant e as the base, and ∝ is the proportional sign;

[0048] The specific analysis of the potential function ζ(*) is:

[0049] According to the Huber function, the potential function ζ(*) is obtained. Where Δ is the difference in the values ​​of the neighboring pixels, specifically I(x, y)-I0(x / ,y / ), γ is the threshold for distinguishing small difference noise from large difference edges;

[0050] By constructing differential equations to enhance the edges and textures of the house image and avoid over-smoothing, the differential equation is: in Is the diffusion coefficient function, controlling the diffusion intensity, the specific diffusion coefficient function Where div(*) is the divergence operator, is the gradient vector of image I, is the amplitude of the gradient, indicating the strength of the edge, and L is the gradient threshold, which determines the boundary between the edge and the flat area;

[0051] The house image with denoised image edges and textures enhanced is mapped to a standardized floating-point array through a conditional generative adversarial network. The conditional generative adversarial network includes a generator and a discriminator, wherein the generator includes an input layer, an output layer and a structure layer. The house image with denoised image edges and textures enhanced is input to the input layer. The structure layer includes a U-Net architecture, which includes jump connections to retain details. The output layer outputs a standardized floating-point array I sd ∈[0,1] H×W , [0,1] H×W The output is a two-dimensional matrix with a height of H and a width of W. Each element value is normalized between 0 and 1. The discriminator includes an input layer, an output layer, and a structure layer. sd or I real Real labeled data I rea输入 The structure layer includes a multi-layer convolutional network that gradually downsamples to extract global features, and the output layer outputs the discrimination results. During the training process, the loss function combines the adversarial loss and the L1 reconstruction loss. The specific loss function is:

[0052] L cGAN =E Isd [log D(I sd )]+E Ireal [log1-D(G(I den ))]+λ3×||I sd -I real ||1, where LcGAN is the total loss function of the conditional generation adversarial network, D is the discriminator, G is the discriminator, E is the expected value, log D(I sd ) is the logarithm of the discriminator’s judgment on the generated data, log1-DG(I den ) is the negative logarithmic loss of the discriminator for the generated data, λ3 is the weight parameter of the reconstruction loss, ||I sd -I real ||1 is the L1 norm loss, which measures the pixel-level difference between the generated data and the real annotation;

[0053] Specifically, the U-Net architecture is a convolutional neural network designed specifically for image segmentation tasks. Through the encoder-decoder structure and jump connections, it significantly improves the detail quality of generated images in generative adversarial networks and is suitable for tasks that require retaining input structural information.

[0054] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0055] Creating a floor plan region of interest model based on a standardized floating-point array, wherein the floor plan region of interest model includes: detecting wall segments in the image using the Hough transform line detection formula ρ = xcosθ + ysinθ, where θ is the angle between the line and the x-axis and ρ is the distance from the line to the origin;

[0056] Also includes: contour area filtering formula if Area(K i ) <C area ,then reject K i , where C area is the set area threshold, Area(K i ) is the contour K i The area of ​​​​the object is obtained by removing the contours with an area smaller than the area threshold and filtering out irrelevant small areas, thereby locating key areas such as walls, doors and windows, and finally obtaining the target cutting area.

[0057] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0058] First, a random graph model is constructed based on a standardized floating-point array. The model includes:

[0059] Where H(0) is the initial floor plan structure, H(t) is the floor plan structure at time t, μ H is the deterministic evolution rate, σ H is the topological diffusion coefficient, which controls the randomness of the connection point generation, dW(s) is the Brownian motion term, which simulates the drawing scanning error or design variation, and ds is the differential of the integral variable s, which ranges from 0 to t;

[0060] Specifically, by fusing deterministic rules and randomness into a random graph model, we can evolve a floor plan structure that is closer to the actual situation over time from the initial floor plan structure.

[0061] The node features are updated layer by layer through the graph convolution operator in the geometric graph convolution network. The graph convolution operator is in represents the feature vector of node u (wall endpoint) at layer v, e pu Represents edge features, specifically segment length, angle, and curvature. N(u) represents the set of neighbor nodes of node u. ψ represents the node update function. φ represents a learnable edge message function. θ represents the parameters of the learnable edge message function φ. Represents the feature vector of node p in layer v. Specifically, through geometric convolution operation, the spatial relationship of wall segments is analyzed hierarchically;

[0062] The node update function ψ is specifically: Where ReLU is the activation function, Concat is the concatenation operation, W is the learnable parameter matrix, α pu is the attention weight, m pu is a message from node p to u, generated by the edge message function. Among them, MLP is a multi-layer perceptron, which is responsible for fusing the geometric features after splicing. It is a vector concatenation operation;

[0063] The door and window positions are predicted by Gaussian process regression, and the output is the mean and covariance. The Gaussian process regression prediction is p(BBox|H(t))~N(μ(H(t)),K(G(t),G(t)), where p(BBox|H(t)) is the probability distribution of the door and window bounding box (BBox) under the condition of known floor plan structure H(t), ~ is the obedience sign, N is the normal distribution, μ(H(t)) is the mean of the bounding box center coordinates, and K(G(t),G(t)) is the covariance function;

[0064] By performing random graph modeling, graph convolution feature updating, and Gaussian process regression prediction operations on the floor plan, the wall structure, node features, and door and window position related data are obtained, and structured encoding of the multi-dimensional floating-point matrix is ​​performed to obtain a floating-point matrix containing wall connection points, line segment parameters, and door and window bounding boxes:

[0065] The channels encode wall connection points, line segment parameters, and door and window bounding boxes respectively. All data is normalized to the range of [0,1]. The connection point information is identified by the mask channel, specifically 1 indicates existence and 0 indicates non-existence, and its precise coordinates are recorded through the bias increment channel. The line segment parameters are stored in independent channels to store the starting and ending coordinates of the line segment and the confidence level. The door and window bounding boxes are encoded with multi-channel data of the center position and rotation angle to fully describe the characteristics of the door and window bounding boxes.

[0066] Specifically, a random graph model is constructed to dynamically generate the structure of the floor plan, providing a basic topological framework for the wall connection points and line segment parameters. The graph convolution operator updates the node features to analyze the spatial relationship of the wall segments, aggregates the neighborhood information to update the node features (including direction, coordinates, etc.), and clarifies the precise position of the wall connection points and the geometric constraints of the segments. Gaussian process regression is used to predict the positions of doors and windows, output the mean (center coordinates) and covariance (position uncertainty) of the door and window bounding boxes, and determine the position and size of the doors and windows in the floor plan. The data sources for the floating-point matrix are provided from three key dimensions: floor plan structure construction, spatial relationship analysis, and door and window position prediction.

[0067] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0068] The line simplification formula is used to determine whether to segment by judging the relationship between the vertical distance from the point to the line segment and the threshold value. The specific line simplification formula is: if d(j,j1j m )>∈, then split, where d represents point j to line segment j1j n The vertical distance of , ∈ is the simplified threshold, and the orthogonal constraint formula is used to force the wall segment angle to align with the coordinate axis, and output structured point, line, and hole data. The specific orthogonal constraint formula is θ1 = arctan(y2-y1) / (x2-x1)stθ1∈{0°,90°,180°,270°}, stθ1∈{0°,90°,180°,270°} means that the calculated θ1 is constrained to 0°,90°,180°,270°, arctan is the inverse tangent function, (x1,y1) and (x2,y2) are the coordinates of the two endpoints of the wall segment.

[0069] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0070] The wall thickness mapping formula is W3D = r × W2D + f, where W2D is the line width in the 2D drawing, r is the scaling factor, and f is the offset. The floor generation formula is Where (x i ,y i) are the vertex coordinates of the closed polygon of the room, Area is the area of ​​the floor area, i is the sequence number of the vertex coordinate, and the maximum value is n1.

[0071] The above is an illustration of the present invention and should not be considered as limiting thereof. Although several exemplary embodiments of the present invention have been described, it will be readily understood by those skilled in the art that many modifications may be made to the exemplary embodiments without departing from the novel teachings and advantages of the present invention. Therefore, all such modifications are intended to be included within the scope of the present invention as defined by the claims. It should be understood that the above is an illustration of the present invention and should not be considered as being limited to the specific embodiments disclosed, and modifications to the disclosed embodiments and other embodiments are intended to be included within the scope of the appended claims. The present invention is defined by the claims and their equivalents.

Claims

1. The algorithm for automatic recognition and generation of floor plans based on home decoration is characterized by: The following steps are involved: Step S101: Input the floor plan, perform image denoising through Markov random field modeling, enhance image edges and textures by constructing differential equations, conditional generative adversarial networks, and output a standardized floating-point array; Step S102: creating a floor plan area of ​​interest model according to the standardized floating-point array, and obtaining the final target cut area through the floor plan area of ​​interest model; Step S103: Based on the generated final target cut-out area, a random graph model is constructed, and node features are updated layer by layer through a geometric graph convolutional network. Subsequently, probabilistic door and window bounding boxes are generated using Gaussian process regression. A floating-point matrix containing wall connection points, line segment parameters, and door and window bounding boxes is obtained through structured encoding of a multi-dimensional floating-point matrix. Step S104: converting the floating-point matrix into structured point, line, and hole data through a broken line simplification formula and an orthogonal constraint formula; Step S105: Based on the point, line, and hole data in the floor plan, key points, walls with thickness, and doors and windows are generated using the wall thickness mapping formula. Then, based on the closed area formed by the walls, polygon triangulation is performed using the floor generation formula to determine the area of ​​the area, and then the floor is generated. At the same time, the corresponding roof is generated in combination with the wall layout to complete the construction of walls, doors and windows, roof, and floor.

2. The algorithm for automatically identifying and generating floor plans based on home decoration according to claim 1 is characterized in that: The standardized floating-point array analysis steps are: The house type image with denoised image and enhanced image edges and textures is mapped into a standardized floating point array through the conditional generative adversarial network. The conditional generative adversarial network includes a generator and a discriminator, wherein the generator includes an input layer, an output layer and a structure layer. The house type image with denoised image and enhanced image edges and textures is input into the input layer. The structure layer includes a U-Net architecture. The output layer outputs a standardized floating point array I sd ∈[0,1] H×W , [0,1] H×W It is a two-dimensional matrix with a height of H and a width of W, and the element values ​​are normalized between 0 and 1. The discriminator includes an input layer, an output layer, and a structure layer. sd or I real Real labeled data I rea The input layer and the structural layer include a multi-layer convolutional network that gradually downsamples to extract global features. The output layer outputs the discrimination results. During the training process, the loss function combines adversarial loss and L1 reconstruction loss.

3. The algorithm for automatically identifying and generating floor plans based on home decoration according to claim 2 is characterized in that: The steps of analyzing the apartment image after noise reduction and enhancing the image edges and textures are as follows: A Markov random field model is established based on the apartment layout image to enhance the spatial correlation between pixels, suppress noise while retaining the edge structure to perform denoising on the image. The Markov random field model includes: ζ(I(x,y)-I0(x / ,y / )))),where P(I|I0) is the probability distribution of the denoised image I given the known noisy image I0, δ is the set of all pixels in the image, (x, y) is the coordinate of a pixel in the image, I is the pixel value of the denoised image, I0 is the pixel value of the original noisy image, λ1 is the weight parameter of the data fidelity term, λ2 is the weight parameter of the smoothing term, ɑ(x, y) is the neighborhood pixel set of pixel (x, y), ζ(*) is the potential function, (x / ,y / ) is the coordinate of the neighboring pixel of pixel (x, y), exp is the exponential function with the natural constant e as the base, and ∝ is the proportional sign; The specific analysis of the potential function ζ(*) is: According to the Huber function, the potential function ζ(*) is obtained. Where Δ is the difference in the values ​​of the neighboring pixels, specifically I(x, y)-I0(x / ,y / ), γ is the threshold for distinguishing small difference noise from large difference edges; By constructing differential equations to enhance the edges and textures of the house image and avoid over-smoothing, the differential equation is: in Is the diffusion coefficient function, controlling the diffusion intensity, the specific diffusion coefficient function Where div(*) is the divergence operator, is the gradient vector of image I, is the gradient amplitude, indicating the strength of the edge, and L is the gradient threshold.

4. The algorithm for automatically identifying and generating floor plans based on home decoration according to claim 1 is characterized in that: The final target cut-image area analysis steps: Creating a floor plan region of interest model based on a standardized floating-point array, wherein the floor plan region of interest model includes: detecting wall segments in the image using the Hough transform line detection formula ρ = xcosθ + ysinθ, where θ is the angle between the line and the x-axis and ρ is the distance from the line to the origin; Also includes: contour area filtering formula if Area(K i ) <C area ,then reject K i , where C area is the set area threshold, Area(K i ) is the contour K i The area of ​​​​the image is obtained, the contours with an area smaller than the area threshold are removed, the key areas are located, and finally the target cutting area is obtained.

5. The algorithm for automatically identifying and generating floor plans based on home decoration according to claim 1 is characterized in that: The random graph model, updated node features and probabilistic door and window bounding box analysis steps: First, a random graph model is constructed based on a standardized floating-point array. The model includes: Where H(0) is the initial floor plan structure, H(t) is the floor plan structure at time t, μ H is the deterministic evolution rate, σ H is the topological diffusion coefficient, which controls the randomness of the connection point generation, dW(s) is the Brownian motion term, ds is the differential of the integral variable s, and its value ranges from 0 to t; The node features are updated layer by layer through the graph convolution operator in the geometric graph convolution network. The graph convolution operator is in Represents the feature vector of node u in layer v, e pu represents the edge feature, N(u) represents the set of neighbor nodes of node u, ψ represents the node update function, φ represents the learnable edge message function, and θ represents the parameters of the learnable edge message function φ. Represents the feature vector of node p in layer v; The door and window positions are predicted by Gaussian process regression, and the output is the mean and covariance.

6. The algorithm for automatically identifying and generating floor plans based on home decoration according to claim 1 is characterized in that: The structured coding analysis steps of the multi-dimensional floating point matrix are as follows: The channels encode wall connection points, line segment parameters, and door and window bounding boxes respectively. All data are normalized to the range of [0,1]. The connection point information is identified by the mask channel, where 1 indicates existence and 0 indicates non-existence. Its precise coordinates are recorded through the incremental channel. The line segment parameters are stored by independent channels to store the starting and ending coordinates of the line segment and the confidence level. The door and window bounding boxes encode multi-channel data of the center position and rotation angle.

7. The algorithm for automatically identifying and generating floor plans based on home decoration according to claim 1 is characterized in that: The structured point, line and hole data analysis steps are as follows: The line simplification formula is used to determine whether to segment by judging the relationship between the vertical distance from the point to the line segment and the threshold value. The line simplification formula is: if d(j,j1j m )>∈, then split, where d represents point j to line segment j1j n The vertical distance of , ∈ is the simplified threshold, and the wall segment angle is forced to align with the coordinate axis through the orthogonal constraint formula to output structured point, line, and hole data. The orthogonal constraint formula is θ1 = arctan(y2-y1) / (x2-x1)stθ1∈{0°,90°,180°,270°}, stθ1∈{0°,90°,180°,270°} means that the calculated θ1 is constrained to 0°,90°,180°,270°, arctan is the inverse tangent function, (x1,y1) and (x2,y2) are the coordinates of the two endpoints of the wall segment.

Images