BIM model generation method based on AI image recognition

Abstract

The application relates to the field of building engineering digitization and computer vision technology, in particular to a BIM model generation method based on AI image recognition, which comprises the following steps: collecting a building scene image sequence, using a convolutional neural network based on a feature pyramid network to extract multi-scale visual features and generate a visual semantic density map; generating a binary mask based on the semantic density map, extracting a component instance aggregated feature vector; generating initial parameters through a parameterized inference function and combining the initial parameters into an initial BIM component parameter set; finally, constructing a total optimization objective function containing topological consistency loss and building specification compliance evaluation, optimizing the initial parameter set to output a final BIM model parameter set, the application constitutes a full-automatic and high-fidelity generation path from an original image to a compliant BIM model, ensures that the output final BIM model is not only geometrically accurate, but also logically self-consistent and compliant with engineering specifications, and significantly improves the intelligent level of building engineering digitization.

Description

Technical Field

[0001] This invention relates to the fields of digital building engineering and computer vision technology, specifically to a BIM model generation method based on AI image recognition. Background Technology

[0002] Currently, the creation of Building Information Models (BIMs) mainly relies on manual modeling using specialized software or reverse fitting based on 3D point clouds. Manual modeling is inefficient, and point cloud methods are essentially geometry-driven, making it difficult to automatically infer the implicit logical topological relationships between building components and the semantic types of the components. Existing technologies generally process visual recognition and BIM parameter generation in separate stages, leading to information transmission distortion and failing to utilize the inherent logical constraints of the building to guide the visual recognition process. Therefore, directly and intelligently generating structured, parametric BIM models that conform to building codes from unstructured site images is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0003] The purpose of this invention is to provide a BIM model generation method based on AI image recognition, which solves the problems existing in the background technology.

[0004] To address the aforementioned technical problems, this invention provides a BIM model generation method based on AI image recognition, comprising the following steps:

[0005] S1. Collect image sequences containing architectural scenes; process the image sequences using a convolutional neural network based on a feature pyramid network to extract multi-scale visual features; and generate a visual semantic density map based on the multi-scale visual features.

[0006] S2. Based on the visual semantic density map and a preset threshold, generate a binary mask representing potential building component instances; and based on the binary mask, extract the aggregated feature vector of each component instance from multi-scale visual features.

[0007] S3. Input the aggregated feature vector of each component instance into the parameterized inference function to generate the corresponding initial parameters; and combine the initial parameters of all component instances to form the initial BIM component parameter set.

[0008] S4. Construct a general optimization objective function consisting of topology consistency loss and building code compliance assessment; and optimize the initial BIM component parameter set by minimizing the general optimization objective function to output the final BIM model parameter set.

[0009] Preferably, the step of generating the visual semantic density map in S1 specifically includes:

[0010] Learnable weights are assigned to each level of multi-scale visual features; local neighborhood normalization is performed on the feature activation values ​​of each level; the normalized multi-scale features are weighted and summed according to the learnable weights, and then processed by the Sigmoid function to generate a visual semantic density map.

[0011] Preferably, the step of extracting the aggregated feature vector in S2 specifically includes: using the region of interest alignment technique to accurately pool the features corresponding to the binarized mask; and processing the pooled features through a fully connected layer or a convolutional layer to compress and generate an aggregated feature vector of fixed dimensions.

[0012] Preferably, the step of generating initial parameters in S3 specifically includes:

[0013] The classification branch of the parameterized inference function generates class probabilities for component instances; the geometric regression branch of the parameterized inference function generates geometric parameters for component instances; and the class probabilities and geometric parameters are combined to construct initial parameters.

[0014] Preferably, the geometric regression branch performs a linear transformation on the aggregated feature vector using the learned weight matrix and bias vector to achieve the mapping from dimensionless features to geometric parameters with physical dimensions.

[0015] Preferably, the topology consistency loss in the overall optimization objective function is calculated as follows:

[0016] Calculate the actual spatial relationship between components; determine whether the actual spatial relationship exists in the predefined set of valid relationships; if the actual spatial relationship does not exist in the set of valid relationships, generate a preset non-zero penalty value; if the actual spatial relationship exists in the set of valid relationships, generate a zero penalty value.

[0017] Preferably, the building code compliance assessment in the overall optimization objective function is calculated in the following manner:

[0018] The building design code is quantified into computable constraint rules that include statutory thresholds; the difference between the model parameters and the statutory thresholds is calculated; if the difference indicates a violation of the constraint rules, a non-zero penalty value related to the difference is generated through the ReLU function; if the difference indicates no violation of the constraint rules, a zero penalty value is generated.

[0019] Preferred options also include:

[0020] During the model training phase, the gradient of the overall optimization objective function is backpropagated to update the learnable parameters inside the parameterized inference function, thereby achieving joint optimization of the visual feature extraction process and the parameter generation process.

[0021] Preferred, including:

[0022] The visual feature generation module is used to extract multi-scale visual features from image sequences containing architectural scenes and generate visual semantic density maps.

[0023] The instance feature extraction module is used to determine the candidate regions of component instances based on the visual semantic density map and extract the aggregated feature vector of each component instance from the multi-scale visual features.

[0024] The initial parameter generation module is used to process aggregated feature vectors, generate initial parameters for each component instance that include class probabilities and geometric parameters, and combine all initial parameters to generate an initial BIM component parameter set;

[0025] The model optimization module is used to construct the overall optimization objective function and optimize the initial BIM component parameter set by minimizing the overall optimization objective function to generate the final BIM model parameter set.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] (1) Through integrated process design, the problem of information distortion and error accumulation caused by staged processing in the existing technology is solved. By generating visual semantic density map, extracting aggregated feature vector, and using parameterized inference function to generate initial BIM component parameter set, and finally optimizing the model, a closed loop with seamless connection and continuous data flow is formed. This end-to-end architecture ensures that the mapping from original pixels to final BIM parameters is direct and high-fidelity, thereby improving the overall accuracy of the generated model.

[0028] (2) By introducing the concept of multi-scale visual semantic density, the accuracy and robustness of building component recognition are improved. By adaptively weighting and summing multi-scale visual features from different network layers, the model can intelligently combine macroscopic structural information with microscopic detail information. The learned weights make this fusion dynamic and targeted. This invention generates feature maps with richer information content and more accurately reflect the probability of component existence, laying a solid foundation for subsequent instance segmentation and parameter extraction.

[0029] (3) This method constructs a total optimization objective function that includes topological consistency and building code compliance, giving the generated BIM model architectural logic and engineering compliance. This is a fundamental transcendence of the existing geometry-driven method, which transforms abstract engineering knowledge into computable mathematical constraints. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a logic block diagram of the system of the present invention;

[0032] Figure 2 This is a logic block diagram generated from the visual semantic density map of the present invention;

[0033] Figure 3 The logic block diagram generated for the initial parameters of this invention;

[0034] Figure 4 This is a logic block diagram for calculating the topology consistency loss in this invention. Detailed Implementation

[0035] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0036] Example 1

[0037] Please see Figure 1 This invention provides a BIM model generation method based on AI image recognition, comprising the following steps:

[0038] S1. Collect image sequences containing architectural scenes; process the image sequences using a convolutional neural network based on a feature pyramid network to extract multi-scale visual features; and generate a visual semantic density map based on the multi-scale visual features.

[0039] S2. Based on the visual semantic density map and a preset threshold, generate a binary mask representing potential building component instances; and based on the binary mask, extract the aggregated feature vector of each component instance from multi-scale visual features.

[0040] S3. Input the aggregated feature vector of each component instance into the parameterized inference function to generate the corresponding initial parameters; and combine the initial parameters of all component instances to form the initial BIM component parameter set.

[0041] S4. Construct a general optimization objective function consisting of topology consistency loss and building code compliance assessment; and optimize the initial BIM component parameter set by minimizing the general optimization objective function to output the final BIM model parameter set.

[0042] This embodiment discloses a BIM model generation method based on AI image recognition, aiming to solve the problems of low model generation efficiency, information transmission distortion, and lack of logical constraints in existing technologies. The method constructs an end-to-end processing framework. Its process begins by using a convolutional neural network to extract multi-scale visual features containing rich spatial hierarchical information from image sequences collected from the building site, and then fusing them into a high-information-content visual semantic density map, laying the data foundation for subsequent accurate identification of components. Based on this density map, the method further locates and segments independent building component instances, and extracts a highly condensed aggregated feature vector for each instance. This step completes... The process involves a crucial transformation from unstructured pixels to structured features. Subsequently, a specially designed parametric inference function directly maps the feature vector to initial BIM parameters containing category and geometric information, enabling direct inference from vision to parameters. The final step of the process introduces a total optimization objective function quantified by building topology logic and industry standards to iteratively optimize the initial parameter set. This complete process constitutes a fully automated, high-fidelity generation path from the original image to a compliant BIM model, ensuring that the final output BIM model is not only geometrically accurate but also logically consistent and compliant with engineering standards, significantly improving the intelligent level of digitalization in building engineering.

[0043] Example 2

[0044] Please see Figure 2 The steps for generating the visual semantic density map in S1 specifically include: setting learnable weights for each level of multi-scale visual features; performing local neighborhood normalization on the feature activation values ​​of each level; processing the normalized features of each level with the Sigmoid function, and then performing a weighted summation of the processed results based on the learnable weights to generate the visual semantic density map.

[0045] The steps for extracting aggregated feature vectors in S2 specifically include: using region of interest alignment technology to accurately pool and binarize the features corresponding to the mask; and processing the pooled features through fully connected layers or convolutional layers to compress and generate aggregated feature vectors of fixed dimensions.

[0046] In this embodiment, the generation step of the visual semantic density map and the extraction step of the aggregated feature vector are closely coupled, together forming the core of the transformation from raw pixels to instance-level features. The technical motivation for generating the visual semantic density map is to overcome the limitation of traditional single-scale features in taking into account both the macroscopic structure and microscopic details of buildings. To achieve this goal, the method calculates the visual semantic density map using the following formula:

[0047]

[0048] Where p is any pixel in the image;

[0049] D sem (p) represents the final visual semantic density value of point p;

[0050] L is a set of predefined convolutional neural network feature layers, such as {C2, C3, C4, C5} layers from fine to coarse, to represent different receptive fields and semantic levels;

[0051] l represents a specific level within this set;

[0052] w l These are weights corresponding to the l-th level, which are automatically learned during model training and enable the model to adaptively assign higher contributions to the feature scales most critical to the identification of specific components.

[0053] A l (i) is the feature activation value at position i on the feature map of layer l, which represents the network's response strength to the local image pattern at that position;

[0054] N l (p) is the local neighborhood centered at point p on the feature map of the l-th layer;

[0055] |N l (p)| is the total number of elements in the neighborhood, which is used to normalize the sum of the activation values ​​in the neighborhood;

[0056] σ(·) is the Sigmoid function, which maps the weighted fusion result to the ((0,1) interval, making it a probability heatmap with dense information.

[0057] The visual semantic density map D generated based on the above calculations sem The step of extracting aggregated feature vectors can then proceed; this step first applies a preset threshold to D. sem The image is binarized, and this threshold is determined experimentally on a validation set to maximize segmentation accuracy, thereby generating multiple binarized masks representing potential building component instances. Subsequently, for each component instance mask, the system employs a region of interest alignment technique to precisely pool the features strictly corresponding to that instance from the multi-scale feature map generated in the previous steps. This technique ensures the accuracy of feature extraction and avoids the quantization errors of traditional pooling operations. Finally, these pooled multi-scale features are compressed through a fully connected layer or convolutional layer to generate a fixed-dimensional aggregated feature vector f. sem,kThis vector encapsulates the high-order semantic information of component instances, providing direct and high-quality input for subsequent parametric reasoning.

[0058] Example 3

[0059] Please see Figure 3 The steps for generating initial parameters in S3 specifically include: generating class probabilities for component instances through the classification branch of the parameterized inference function; generating geometric parameters for component instances through the geometric regression branch of the parameterized inference function; and combining the class probabilities and geometric parameters to construct the initial parameters.

[0060] The geometric regression branch performs a linear transformation on the aggregated feature vectors using the learned weight matrix and bias vector to achieve a mapping from dimensionless features to geometric parameters with physical dimensions.

[0061] In this embodiment, the core of the initial parameter generation step is a parameterized inference function F with shared weights. infer This function is designed to innovatively extend the instance segmentation task from traditional two-dimensional mask prediction to direct regression of three-dimensional building component parameters, thereby avoiding the accumulation of errors in multi-stage processes. The function receives the aggregated feature vector f generated in the previous step. sem,k and output the initial parameter set P. init,k Its mathematical expression is:

[0062] P init,k =F infer (f sem,k ;Θ)={Prob cls,k Param geom,k}

[0063] Among them, P init,k These are the initial parameters generated for the k-th component instance;

[0064] f sem,k It is the aggregated feature vector of the input;

[0065] Θ represents the function F infer The set of all learnable parameters internally;

[0066] The function is internally deconstructed into two parallel processing branches: one is a classification branch, whose calculation formula is Prob cls,k =Softmax(W cls ·f sem,k ), where W cls It is a learnable classification weight matrix, and this branch outputs a probability distribution vector Prob. cls,k The first branch represents the probability that a component belongs to a predefined category (such as wall, column, beam, window, etc.); the second branch is geometric regression, whose calculation formula is Param...geom,k =W geom ·f sem,k +b geom W geom and b geom These are the learnable geometric regression weight matrix and bias vector, respectively. This branch directly regresses the geometric parameters Param of the component. geom,k Such as the coordinates of the center point in three-dimensional space and dimensions such as length, width, and height.

[0067] Therefore, a key technical issue is ensuring dimensional consistency; the input feature vector f sem,k It is dimensionless, while the output geometric parameters Param geom,k It then has physical dimensions; this method solves this problem through a supervised learning mechanism: during the training phase, the model's output parameters are... geom,k Compared to BIM data with real physical units, the loss function drives the model to learn a specific set of W. geom and b geom Therefore, the learned parameter matrix implicitly contains the scaling and transformation factors required to map dimensionless features to geometric parameters with correct physical dimensions. This design realizes direct, end-to-end regression from abstract visual semantics to specific engineering parameters, which is the core manifestation of the intelligence of this invention.

[0068] Example 4

[0069] Please see Figure 4 The topology consistency loss in the overall optimization objective function is calculated as follows: calculate the actual spatial relationship between components; determine whether the actual spatial relationship exists in the predefined set of valid relationships; if the actual spatial relationship does not exist in the set of valid relationships, a preset non-zero penalty value is generated; if the actual spatial relationship exists in the set of valid relationships, a zero penalty value is generated.

[0070] The building code compliance assessment in the overall optimization objective function is calculated as follows: the building design code is quantified into computable constraint rules containing statutory thresholds; the difference between the model parameters and the statutory thresholds is calculated; if the difference indicates a violation of the constraint rules, a non-zero penalty value related to the difference is generated through the ReLU function; if the difference indicates no violation of the constraint rules, a zero penalty value is generated.

[0071] In this embodiment, the core of model optimization lies in constructing and minimizing a total optimization objective function L. final The technical motivation behind this function is to transform abstract architectural prior knowledge and industry design standards into machine-computable and optimizable mathematical constraints, thereby guiding the automatic correction of model parameters; its expression is:

[0072] L final(P)=α topo L topo (P)+α code E code ((P)

[0073] Where P is the set of parameters of the BIM model to be optimized, and its initial value is P. init ;α topo and α code It is a hyperparameter used to balance the importance of the two loss terms, and its value can be determined by cross-validation on the validation set.

[0074] The first component of this function is the topology consistency loss L. topo Used to punish spatial relationships between components that do not conform to architectural logic:

[0075]

[0076] In the formula, v i and v j Represents any two distinct component instances in the model, whose types are T and T, respectively. i and T j ;

[0077] Rel(v i ,v j () is the actual spatial relationship calculated based on its current geometric parameters;

[0078] ValidRels(T i ,T j ) is a predefined set of valid relations, stored in the form of a data structure built upon prior architectural knowledge. For example, for a T of type window... i T and the wall j Its valid relation only includes those contained in;

[0079] I(·) is an indicator function. Its value is 1 when the actual relationship between components does not belong to the set of valid relationships, and 0 otherwise.

[0080] C(T i ,T j ) is a preset, dimensionless penalty cost value whose magnitude reflects the severity of violating a specific topology rule;

[0081] The second component of this function is the building code compliance assessment E. code Used to quantify the degree of violation of building design codes:

[0082]

[0083] Where K is the total number of building codes being coded;

[0084] It is a computational function that extracts physical quantities related to the k-th rule from the model parameters P, such as the width of the door; τ k This is the legal threshold for the rule, and its source is publicly available standard documents such as building design codes;

[0085] δ k The value can be +1 or -1, which is used to define whether the rule is an upper limit constraint or a lower limit constraint.

[0086] The ReLU(·) function ensures that a non-zero penalty is only generated when a rule is violated; to guarantee dimensional consistency, when the difference within the parentheses has physical dimensions, the weighting coefficient β... k It must have its reciprocal dimension to ensure that the final E code The penalty value is dimensionless;

[0087] In applications, the system uses P init Starting with L, we iteratively solve the problem using optimization algorithms such as gradient descent to minimize L. final This process establishes a mathematically optimized landscape, where any unreasonable layout or non-compliant design will lead to L... final The value is increased, thereby driving the optimization algorithm to automatically adjust the parameters of relevant components to reduce the value, ensuring that the final generated BIM model is internally logically consistent and externally conforms to the specifications.

[0088] Example 5

[0089] During the model training phase, the gradient of the overall optimization objective function is backpropagated to update the learnable parameters inside the parameterized inference function, thereby achieving joint optimization of the visual feature extraction process and the parameter generation process.

[0090] This embodiment illustrates the closed-loop mechanism of the end-to-end learning framework of the present invention; during the model training phase, the overall optimization objective function L... final Its core function is to guide the parameter updates of the entire deep learning model; its inherent logic lies in L final It is a direct measure of the model parameter P, while P (initially P) init F is a parameterized inference function. infer The direct output, which in turn depends on its internal learnable parameters Θ (i.e., the weight matrix W). cls and W geom (etc.); therefore, the system can use the chain rule to calculate L. final For the gradient of parameter Θ, i.e. This gradient signal is then backpropagated to update F. inferThe internal parameters; this mechanism forms a powerful feedback loop: if a generated component instance is found to violate topological logic or building codes during the optimization phase, the resulting gradient signal will propagate back and precisely correct the parameterized inference function F. infer This allows the system to infer more reasonable parameters from similar visual features in subsequent iterations. This process achieves deep coupling and joint optimization between the visual feature extraction process and the parameterization and compliance generation process, forming a core of the end-to-end learning framework of this invention.

[0091] Example 6

[0092] A BIM model generation system based on AI image recognition includes:

[0093] The visual feature generation module is used to extract multi-scale visual features from image sequences containing architectural scenes and generate visual semantic density maps.

[0094] The instance feature extraction module is used to determine the candidate regions of component instances based on the visual semantic density map and extract the aggregated feature vector of each component instance from the multi-scale visual features.

[0095] The initial parameter generation module is used to process aggregated feature vectors, generate initial parameters for each component instance that include class probabilities and geometric parameters, and combine all initial parameters to generate an initial BIM component parameter set;

[0096] The model optimization module is used to construct the overall optimization objective function and optimize the initial BIM component parameter set by minimizing the overall optimization objective function to generate the final BIM model parameter set.

[0097] This embodiment provides a BIM model generation system based on AI image recognition. Through a series of clearly defined and highly collaborative modular designs, it achieves an automated process from raw images to compliant BIM models. The system's visual feature generation module integrates a convolutional neural network based on a feature pyramid network to process the input building scene image, extract multi-scale visual features, and calculate and generate a visual semantic density map. Based on this density map, the instance feature extraction module determines candidate regions for components through thresholding and uses region of interest alignment technology to extract aggregated feature vectors for each instance from the multi-scale visual features. Subsequently, the initial parameter generation module receives these aggregated feature vectors. The core parameterized inference function of the eigenvector performs classification and regression tasks in parallel for each vector, generating an initial BIM component parameter set containing categories and precise geometric parameters. Finally, the model optimization module refines and validates this initial parameter set. This module has a built-in overall optimization objective function defined by building topology logic and industry design specifications. By running optimization algorithms, iteratively adjusting parameters, it minimizes the function value. This process can automatically correct unreasonable spatial relationships between components and dimensions that violate building codes, ultimately outputting a geometrically accurate, topologically valid, and compliant final BIM model parameter set, thus forming a self-consistent and self-optimizing intelligent BIM generation system.

[0098] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A BIM model generation method based on AI image recognition, characterized in that, Includes the following steps: S1. Collect image sequences containing architectural scenes; use a convolutional neural network based on a feature pyramid network to process the image sequences in order to extract multi-scale visual features; And based on multi-scale visual features, a visual semantic density map is generated; S2. Based on the visual semantic density map and a preset threshold, generate a binary mask representing potential building component instances; and based on the binary mask, extract the aggregated feature vector of each component instance from multi-scale visual features. S3. Input the aggregated feature vector of each component instance into the parameterized inference function to generate the corresponding initial parameters; and combine the initial parameters of all component instances to form the initial BIM component parameter set. S4. Construct the overall optimization objective function consisting of topology consistency loss and building code compliance assessment; The initial BIM component parameter set is optimized by minimizing the overall optimization objective function to output the final BIM model parameter set. The specific steps for generating the visual semantic density map in S1 include: Learnable weights are set for each level of multi-scale visual features; local neighborhood normalization is performed on the feature activation values ​​of each level; the normalized multi-scale features are weighted and summed according to the learnable weights, and then processed by the Sigmoid function to generate a visual semantic density map. The specific steps for generating initial parameters in S3 include: The classification branch of the parameterized inference function generates class probabilities for component instances; the geometric regression branch of the parameterized inference function generates geometric parameters for component instances; the class probabilities and geometric parameters are combined to construct initial parameters. The geometric regression branch performs a linear transformation on the aggregated feature vectors using the learned weight matrix and bias vector to achieve a mapping from dimensionless features to geometric parameters with physical dimensions.

2. The BIM model generation method based on AI image recognition according to claim 1, characterized in that, The steps for extracting aggregated feature vectors in S2 specifically include: using region of interest alignment techniques to accurately pool and binarize the features corresponding to the mask; and processing the pooled features through fully connected layers or convolutional layers to compress and generate aggregated feature vectors of fixed dimensions.

3. The BIM model generation method based on AI image recognition according to claim 1, characterized in that, The topology consistency loss in the overall optimization objective function is calculated as follows: Calculate the actual spatial relationships between components; determine whether the actual spatial relationships exist in the predefined set of valid relationships; If the actual spatial relation does not exist in the set of valid relations, a preset non-zero penalty value is generated; If the actual spatial relation exists in the set of valid relations, then a zero penalty value is generated.

4. The BIM model generation method based on AI image recognition according to claim 1, characterized in that, The building code compliance assessment in the overall optimization objective function is calculated as follows: The building design code is quantified into computable constraint rules that include statutory thresholds; the difference between the model parameters and the statutory thresholds is calculated; if the difference indicates a violation of the constraint rules, a non-zero penalty value related to the difference is generated through the ReLU function; if the difference indicates no violation of the constraint rules, a zero penalty value is generated.

5. The BIM model generation method based on AI image recognition according to claim 1, characterized in that, Also includes: During the model training phase, the gradient of the overall optimization objective function is backpropagated to update the learnable parameters inside the parameterized inference function, thereby achieving joint optimization of the visual feature extraction process and the parameter generation process.

6. A BIM model generation system based on AI image recognition, based on the BIM model generation method based on AI image recognition as described in any one of claims 1 to 5, characterized in that, include: The visual feature generation module is used to extract multi-scale visual features from image sequences containing architectural scenes and generate visual semantic density maps. The instance feature extraction module is used to determine the candidate regions of component instances based on the visual semantic density map and extract the aggregated feature vector of each component instance from the multi-scale visual features. The initial parameter generation module is used to process aggregated feature vectors, generate initial parameters for each component instance that include class probabilities and geometric parameters, and combine all initial parameters to generate an initial BIM component parameter set; The model optimization module is used to construct the overall optimization objective function and optimize the initial BIM component parameter set by minimizing the overall optimization objective function to generate the final BIM model parameter set.

Images