Intelligent generation system of huizhou architecture components based on non-heritage pattern recognition

Abstract

The application relates to the technical field of building digitization, and particularly discloses a system for intelligently generating a building component of a Huizhou style based on non-heritage pattern recognition, which collects and analyzes non-heritage component images and semantic descriptions, constructs a pattern narrative knowledge set containing pattern semantic entities and relation predicates, receives target cultural implications and component form constraints, generates a semantic relation topological structure through intelligent reasoning, then instantiates the topological structure into a preliminary geometric contour according to a set of geometric form rules, then performs non-heritage visual stylization enhancement processing on the contour, finally performs multidimensional compliance verification and iterative optimization on the stylized contour, and outputs a digital building component model meeting comprehensive requirements of culture, form, aesthetics and engineering; the application solves the key technical problem of disconnection between artistic conception narration and algorithm form in a traditional generation method, and realizes intelligent and faithful innovative transformation of non-heritage patterns from cultural semantics to constructible geometry.

Description

Technical Field

[0001] This invention relates to the field of digital building technology, specifically to an intelligent generation system for Hui-style architectural components based on intangible cultural heritage pattern recognition. Background Technology

[0002] The Hui-style architecture, especially the ancient wood carvings represented by Wuyuan, features rich auspicious patterns engraved on its doors, windows, beams, and other components, which are important intangible cultural heritage. These patterns are not simply decorations, but rather carry profound cultural connotations, ethical concepts, and regional aesthetics. Their composition follows a rigorous narrative logic and organizational principle that "every image must have a meaning."

[0003] Existing technologies suffer from the following shortcomings: There is an inherent conflict between "algorithm-driven formal optimization" and "narrative composition driven by the artistic conception of intangible cultural heritage patterns" in existing intelligent generation technologies. Traditional methods can only handle the visual style and geometric rules of patterns, failing to understand the dynamic narrative spatial relationships (such as the contact and posture echo between "magpie" and "plum blossom") necessary for expressing specific auspicious meanings (e.g., "joy on the eyebrows") between the generated pattern elements, resulting in a formally complete but artistically weak generated product. This invention, through the pioneering use of a "semantic relationship graph" as the core intermediate representation, first decodes the cultural connotations into a network of entity relationships, then instantiates them into geometric forms. This algorithmically guarantees the integrity and logic of the inherent cultural narrative structure of the generated components, achieving a qualitative leap from "formal resemblance" to "spiritual resemblance." Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent generation system for Hui-style architectural components based on the recognition of intangible cultural heritage patterns, so as to solve the problems mentioned above.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A smart generation system for Hui-style architectural components based on intangible cultural heritage pattern recognition includes:

[0007] The multimodal data acquisition and narrative knowledge construction module is used to collect image data and semantic description data of intangible cultural heritage architectural components. Based on the pattern elements and spatial relationships in the image data, combined with the corresponding semantic descriptions, it constructs a pattern narrative knowledge set containing pattern semantic entities and relational predicates through intelligent parsing.

[0008] The semantic relation topology reasoning module is used to receive the target cultural connotation text and the form constraint parameters of the target architectural components. Based on the pattern narrative knowledge set, it generates a semantic relation topology structure that matches the cultural connotation and form constraint through intelligent reasoning.

[0009] The parameterized geometry instantiation module instantiates and converts the semantic relation topology according to a predefined, parameterized geometry shape rule set, the geometry shape rule set defines geometry shape parameters corresponding to different semantic entities and geometry constraint conditions triggered by relation edges, and a preliminary component geometry contour is generated by solving parameters satisfying all geometry constraint conditions;

[0010] The non-heritage visual stylization enhancement module applies visual feature enhancement processing conforming to an artistic style to the preliminary component geometry contour, so that the preliminary component geometry contour has preset non-heritage visual style features on the basis of preserving narrative geometry relations;

[0011] The multi-constraint compliance verification and output module performs multi-dimensional compliance verification and optimization adjustment on the component geometry contour after stylization processing, and outputs a digital building component model meeting cultural implications, shape constraints, artistic styles and engineering feasibility requirements.

[0012] As a further scheme of the present application, the construction process of the pattern narrative knowledge set is:

[0013] The image data is subjected to multi-scale feature extraction and spatial coding processing, and the geometric feature vector of each pattern element and the spatial relation matrix composed of the relative positions between elements are output;

[0014] The semantic description data is subjected to syntax dependency analysis and entity relation extraction processing, and a semantic relation graph with pattern entities as nodes and relation descriptions as edges is output;

[0015] The geometric feature vector and the spatial relation matrix are matched with the nodes and edges in the semantic relation graph for feature fusion and matching, the corresponding relation between the pattern semantic entities and the relation predicates is analyzed, and the pattern narrative knowledge set is constructed.

[0016] As a further scheme of the present application, the semantic relation topology matched with the cultural implications and the shape constraints is generated through intelligent reasoning, specifically including:

[0017] The target cultural implication text is subjected to entity deconstruction and relation extraction based on the pattern narrative knowledge set, and an initial semantic entity node set and a potential relation edge set are obtained;

[0018] The shape constraint parameters of the target building component are converted into spatial layout boundary conditions, the entity nodes under the spatial layout boundary conditions are simulated and preliminarily arranged through introduction of a random walk process, and an entity node set with position attributes is formed;

[0019] A probabilistic graph network is constructed based on a set of entity nodes with location attributes and a set of potential relation edges. Through multiple rounds of message passing and probabilistic reasoning, the probability of the existence of each potential relation edge is calculated, and stable relation edges are selected according to the probability threshold. These edges are combined to generate a semantic relation topology that matches both cultural connotations and morphological constraints.

[0020] As a further aspect of the present invention: the formation of a set of entity nodes with positional attributes specifically includes:

[0021] Based on the component outline and internal functional zoning defined by the morphological constraint parameters, a two-dimensional discretized mesh representation space is generated, and the discretized mesh representation space is marked as the boundary condition region that allows node distribution.

[0022] Each node in the entity node set is assigned an initial random location seed point within the boundary condition region, and the walk step size and direction are controlled based on preset weight parameters related to the node semantic type, and multiple rounds of parallel random walk simulation are performed.

[0023] During the random walk, the relative distance and repulsive force between nodes are dynamically calculated. When the cumulative movement distance of all nodes is lower than the preset threshold and the basic distribution equilibrium condition is met, the simulation stops and the final stable position coordinates of each node are output, forming a set of entity nodes with position attributes.

[0024] As a further aspect of the present invention: the generation of the preliminary component geometric profile specifically includes:

[0025] Based on the semantic relation topology, traverse each relation edge and the entity node connected by the relation edge, and retrieve and instantiate the corresponding parameterized geometric primitives and geometric constraint equations from the set of geometric morphology rules.

[0026] By combining all the instantiated geometric constraint equations, a global nonlinear constraint solution problem is constructed with the morphological parameters of each geometric primitive as variables.

[0027] The constraint problem is transformed into an energy minimization problem. By iteratively adjusting the values ​​of various morphological parameters, the total energy of the system is gradually reduced until it converges to a steady state.

[0028] The converged morphological parameter values ​​are assigned to the corresponding geometric primitives, and Boolean operations and surface stitching between primitives are completed to output the preliminary geometric contour of the component.

[0029] As a further aspect of the present invention: the construction of a global nonlinear constraint solution problem with the morphological parameters of each geometric primitive as variables specifically includes:

[0030] According to the correlation between variables and equations in the instantiated geometric constraint equation set, a bidirectional constraint network graph is constructed, nodes in the graph represent morphological parameter variables and constraint equations, and edges represent the participation of variables in equations;

[0031] Perform hierarchical topological sorting and strongly connected component division on the constraint network graph, decompose the global solving problem into a sequence of subproblems with sequential dependencies, and identify the coupled constraint groups containing circular dependencies;

[0032] Introduce a virtual relaxation variable to the identified coupled constraint group, and reconstruct the original constraint equation set into a hierarchical cascading optimization problem containing equality and inequality relationships, to form a structured global nonlinear constraint solving problem statement.

[0033] As a further scheme of the present application: the visual feature enhancement processing conforming to the artistic style applied to the preliminary component geometric profile specifically includes:

[0034] The surface gradient domain of the preliminary component geometric profile is decomposed to separate low-frequency geometric basic information representing the core narrative structure and high-frequency surface detail information to be enhanced;

[0035] Access the pre-constructed visual style feature library of non-heritage, retrieve the corresponding stroke, texture and decoration unit feature data according to the target style, and adapt them to the topological framework defined by the geometric basic information;

[0036] Through the differential feature fusion process, the adapted style feature data and the high-frequency surface detail information are weighted and synthesized to generate an enhanced geometric surface that maintains the original profile narrative structure and superimposes the target non-heritage visual style.

[0037] As a further scheme of the present application: the adaptation to the topological framework defined by the geometric basic information specifically includes:

[0038] The topological connection relationship between the surface key points and feature lines of the geometric basic information is calculated, and an associated mapping matrix corresponding thereto is generated;

[0039] According to the geometric properties of the associated mapping matrix, the stroke, texture and decoration unit feature data retrieved are parameterized and morphed to be resampled, so that the spatial distribution of the feature data is consistent with the geometric trend of the topological framework;

[0040] Through the distributed weighted fusion operation based on key points and feature lines, the morphed and resampled feature data is accurately aligned and attached to the corresponding position of the topological framework, completing the geometric adaptation of the style feature and the narrative structure.

[0041] As a further scheme of the present application: the multi-dimensional compliance verification and optimization adjustment of the component geometric profile after the style processing specifically includes:

[0042] A multi-objective compliance evaluation index system is established, and cultural implications, form constraints, artistic styles and engineering feasibility requirements are quantified as computable semantic consistency parameters, geometric deviation parameters, style similarity parameters and mechanical safety parameters;

[0043] According to the evaluation index system, a hierarchical and cascaded parallel verification operation is performed on the geometric contour of the style processing component, and a compliance diagnosis report recording the deviation degree of each parameter is generated;

[0044] Based on the compliance diagnosis report, a multi-objective gradient guided parameter space optimization method is used to iteratively fine-tune the form and style parameters of the geometric contour, until all evaluation parameters reach the preset compliance threshold, and finally output the optimized digital building component model.

[0045] As a further scheme of the present application: the digital building component model specifically comprises:

[0046] A multi-level data structure, the top layer of the multi-level data structure is a semantic relationship topology structure, the middle layer is a geometric primitive set carrying programmable parameters, and the bottom layer is a high-precision curved surface grid fused with style features;

[0047] A parameterized interface associated with the geometric primitive set, the parameterized interface records the final values and adjustment history of all form parameters and style parameters during the generation process;

[0048] A semantic association matrix, the semantic association matrix stores the mapping relationship between each geometric feature on the curved surface grid and the corresponding semantic entity and relationship predicate in the pattern narrative knowledge set in a retrievable form.

[0049] The present application has the following advantages:

[0050] (1) The present application realizes the leap of intangible cultural heritage patterns from "visual form imitation" to "cultural semantic understanding and generation". Traditional digital methods only collect and style copy patterns at the image level, and cannot understand and generate the internal narrative logic and cultural implications. The present application constructs a structured "pattern narrative knowledge set", and uses "semantic relationship topology structure" as the intermediate blueprint of generation, and converts "joy on the eyebrows" and "asking children under the pine tree" abstract implications into entity relationship networks that can be reasoned and executed by machines. This makes the final generation result not a simple splicing of patterns, but ensures that the pattern elements have dynamic association and spatial narrative that conform to traditional cultural logic, solves the core problems of "loss of artistic conception" and "lack of narrative", realizes the accurate inheritance and living expression of intangible cultural heritage connotation.

[0051] (2) The application establishes a multi-constraint collaborative generation and verification mechanism that integrates culture, aesthetics and engineering. Existing technologies often separate the association between cultural design, artistic expression and engineering manufacturing, resulting in difficulties in implementing designs. The application unifies cultural implications, local artistic styles, specific component form constraints (such as size, light transmittance) and engineering feasibility (such as structural safety, manufacturability) into an solvable optimization model through parameterized geometric instantiation and multi-objective compliance verification. The system can automatically adjust iteratively until the optimal solution that meets all dimensional requirements is output. It provides a complete technical closed loop for the high-quality innovative application of non-heritage patterns in modern architecture. BRIEF DESCRIPTION OF DRAWINGS

[0052] The application will be further described below with reference to the accompanying drawings.

[0053] Figure 1 is a system block diagram of the application. DETAILED DESCRIPTION

[0054] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the application.

[0055] Please refer to Figure 1 The application is a local building component intelligent generation system based on non-heritage pattern recognition, which comprises:

[0056] A multi-modal data acquisition and narrative knowledge construction module is used to acquire image data and semantic description data of non-heritage building components. For pattern element and spatial relationship in the image data, the corresponding semantic description is combined to construct a pattern narrative knowledge set containing pattern semantic entities and relationship predicates through intelligent analysis;

[0057] A semantic relationship topology reasoning module is used to receive target cultural implication text and form constraint parameters of target building components, generate a semantic relationship topology structure matched with cultural implications and form constraints through intelligent reasoning based on the pattern narrative knowledge set;

[0058] A parameterized geometric instantiation module is used to instantiate and convert the semantic relationship topology structure according to a predefined, parameterized geometric form rule set. The geometric form rule set defines geometric form parameters corresponding to different semantic entities and geometric constraint conditions triggered by relationship edges. A preliminary component geometric contour is generated by solving parameters that meet all geometric constraint conditions;

[0059] The non-heritage visual stylization enhancement module applies visual feature enhancement processing conforming to the artistic style to the preliminary component geometric contour, so that the preliminary component geometric contour has the preset non-heritage visual style feature on the basis of retaining the narrative geometric relationship;

[0060] The multi-constraint compliance verification and output module performs multi-dimensional compliance verification and optimization adjustment on the component geometric contour subjected to the stylization processing, and outputs a digital building component model meeting the cultural implication, morphological constraint, artistic style and engineering feasibility requirements.

[0061] In the multi-modal data acquisition and narrative knowledge construction module, first, data acquisition and image processing are performed. A large-format platform scanner (such as Creo, i2S series products) is used to digitally acquire the wood carving components of Huizhou architecture selected from the Wuyuan area, and obtain their two-dimensional orthographic images. At the same time, a structured semantic description text corresponding to each wood carving image is recorded, which clearly describes the pattern scene in the subject-predicate-object sentence form, for example, “magpie stands on the branch of the plum blossom”. The image data collected is subjected to multi-scale feature extraction and spatial coding processing. Specifically, a feature extraction network containing 3 different scale convolution layers is used to process the image: the first layer of convolution kernel size is 3 by 3, and the step is 1, which is used to extract the basic features of the pattern edge; the second layer of convolution kernel size is 5 by 5, and the step is 2, which is used to extract the medium-scale pattern element features; the third layer of convolution kernel size is 7 by 7, and the step is 2, which is used to extract the large-scale combined pattern features. The feature maps extracted at the three scales are spliced and reduced in dimension, and a multi-layer perceptron containing 512 neurons is used to generate a 512-dimensional geometric feature vector for each independent pattern element (such as “magpie” and “plum branch”) identified and segmented in the image. At the same time, the centroid coordinates of all identified pattern elements in the image plane are calculated, and a spatial relationship matrix is constructed, in which each element value is calculated by the inverse tangent function from the relative distance and angle between the centroid coordinates of the corresponding two pattern elements.

[0062] Secondly, the semantic data is structured. The semantic description text of the aforementioned record is processed by syntax dependency analysis and entity relationship extraction. Specifically, a transition-based syntax analyzer is used to parse the text, identifying the core nouns in the sentence as entity nodes (such as "magpie" and "plum branch"), and identifying verbs and prepositions as potential relationship edges (such as "stands on"). According to the dependency relationships between words, such as modification, subject-predicate, and verb-object, a directed graph is constructed with pattern entities as nodes and word-to-word grammatical relationships as preliminary relationship edges, i.e., a semantic relationship graph. On this basis, a set of predefined rules (for example, if the verb "stands on" connects two nouns, the first noun node is connected to the second noun node through the "stands on" relationship edge) are used to clean and classify the preliminary relationship edges, forming the final semantic relationship graph.

[0063] Finally, cross-modal feature fusion and knowledge set construction are performed. The geometric feature vector of each pattern element obtained in the first step is aligned with the corresponding entity node in the semantic relationship graph. This process is achieved by calculating the cosine similarity between the geometric feature vector and the semantic node vector (initialized by the word embedding vector of the node name), and associating items with a similarity higher than a pre-set threshold (e.g., 0.85). At the same time, the spatial relationship matrix obtained in the first step is matched with the relationship edges in the semantic relationship graph. The matching basis is whether the positional relationship of the two entity nodes corresponding to the relationship edge in the spatial relationship matrix is consistent with the semantics of the relationship edge (such as "stands on" which usually requires a vertical spatial proximity). Through iterative optimization, a record is created in the pattern narrative knowledge set for each successfully matched "entity-relationship-entity" combination, linking the semantic entity, relationship predicate, corresponding geometric feature vector template, and typical spatial relationship constraints, thus completing the construction of the knowledge set.

[0064] In the semantic relationship topology reasoning module, first, the input target cultural implication text is semantically deconstructed based on the pattern narrative knowledge set. Specifically, the implication text (such as "joy on the brow") is segmented and tagged with part-of-speech, and then the key words in it are matched with the semantic entity names defined in the knowledge set by string matching or semantic similarity calculation (for example, using word vector to calculate cosine similarity, with a threshold of 0.8). The matched words are mapped to the corresponding semantic entity nodes, forming an initial semantic entity node set. At the same time, the verb and preposition structures in the text are parsed and compared with the relationship predicates defined in the knowledge set, and potential relationship hypotheses connecting these entity nodes are extracted, forming a set of potential relationship edges. For example, from "joy on the brow", the entities "magpie" and "plum branch" can be deconstructed, as well as the potential "stands on" relationship.

[0065] Secondly, the above-mentioned entity nodes are preliminarily spatially laid out in combination with the morphological constraint parameters of the target architectural component. The morphological constraint parameters include the outline boundary of the component (such as a 2000 mm by 800 mm rectangle) and the internal restricted area (such as a mounting hole position that needs to be reserved). Discretize this outline into a two-dimensional grid composed of square cells with a side length of 10 mm, mark the available and unavailable areas, and form the boundary conditions of the spatial layout. Randomly place each semantic entity node obtained in the first step in the grid cells allowed by the boundary conditions as its initial position seed. Assign each node a moving weight related to its semantic type, for example, the weight of the "main body" role (such as a magpie) is higher, meaning that the range of its single random walk step is larger (such as 3 to 8 grid cells), and the weight of the "set scene" role (such as cloud pattern) is lower, and the step range is smaller (such as 1 to 3 grid cells). Let all nodes perform multiple rounds (for example, 1000 rounds) of random movement in the boundary simultaneously, the moving direction in each round is randomly selected within 0 to 360 degrees, and the step length is randomly determined within its weight range. After each round of movement, calculate the Euclidean distance between any two nodes, if the distance is less than a preset repulsion radius (for example, 50 mm), then apply a repulsion force vector to the two nodes, the size of which is inversely proportional to the distance, and the direction is away from each other, and adjust the moving direction of the next round of the nodes according to it. When the total distance of the position movement of all nodes in 10 consecutive rounds of simulation is less than a preset threshold (for example, 5 mm), stop the simulation, and record the grid coordinates of each node at this time as its final position, thereby forming a set of entity nodes with position attributes.

[0066] Finally, based on the nodes with position attributes and potential relation edges, a probabilistic graph is constructed and solved to determine the final stable relation structure. The node positions obtained in the previous step are taken as observation evidence. For each potential relation edge, an existence probability of 0.5 is initialized. Multiple rounds of message passing and probabilistic inference are performed: in each round, according to the actual spatial distance between two nodes, the relative azimuth angle, and against the spatial relation constraint templates recorded in the pattern narrative knowledge set for this type of relation edge, a likelihood probability is calculated. For example, for the "standing on" relation, the knowledge set may stipulate that the two entities should be close in the vertical direction and there is a possibility of contact; the vertical coordinate difference and centroid distance of node A (magpie) and node B (plum branch) are calculated, and if they meet the "close" condition, the likelihood probability is high. At the same time, the semantic compatibility of the node pair is considered (obtained from the knowledge set). These information is iteratively updated through a probability updating formula (which is essentially a Bayesian update that combines the prior probability with the new likelihood probability to calculate the posterior probability) to update the existence probability of each edge. After multiple rounds (e.g. 20 rounds) of iteration, the edge probability tends to be stable. All relation edges with existence probability higher than a pre-set threshold (e.g. 0.7) are retained, and edges with probability lower than the threshold are deleted. Finally, the graph composed of the retained nodes and relation edges is the semantic relation topology structure that satisfies both the cultural connotation logic and the component spatial form constraint, providing accurate guidance for subsequent geometric generation. The connotation logic refers to the narrative association between the cultural symbols and auspicious meanings carried in the intangible cultural heritage pattern, which goes beyond the basic linguistic semantic relation and is a deep expression system in the cultural context.

[0067] In the parametric geometric instantiation module, first, according to the semantic relation topology structure, each relation edge and the entity nodes connected by the edge are traversed, and the corresponding parametric geometric primitive and geometric constraint equation set are retrieved and instantiated from the geometric form rule set. The geometric form rule set is a pre-defined database, in which each record maps a specific semantic entity (such as "magpie") to a parametric geometric primitive (such as a closed B-spline curve contour defined by 20 control points), and maps a specific relation predicate (such as "standing on") to a set of geometric constraint conditions (such as contact point continuity constraint, normal alignment constraint). Instantiation means creating an instance of the B-spline curve primitive for the specific "magpie" node in the current topology structure, with the control point coordinates becoming the morphological parameter variables to be solved; at the same time, for the "magpie standing on plum branch" relation edge, an instance of its corresponding geometric constraint equation set is created, and the variables in the equation set refer to the specific morphological parameters of the "magpie" and "plum branch" two primitives.

[0068] Secondly, all the geometry constraint equations generated by the instantiation are solved simultaneously to form a global nonlinear constraint problem with all the geometric primitive morphological parameters as variables. In the implementation, a bidirectional constraint network graph is constructed according to the relationship between the variables and equations in the constraint equations. In the graph, two types of nodes are set up: one represents a morphological parameter variable (e.g. "X coordinate of the 3rd control point of the magpie primitive"), and the other represents a constraint equation (e.g. "contact point curvature equality equation"). A connection edge is established between the two types of nodes if and only if the variable appears in the corresponding constraint equation. In this way, the coupling relationship between all the variables and equations is visualized and structured.

[0069] Then, the hierarchical topological sorting and strongly connected component partitioning are performed on the constructed constraint network graph to decompose the global solving problem. Tarjan algorithm based on depth-first search is used to analyze the network graph. The core of the Tarjan algorithm is to calculate the "depth index" and "low link value" of each node, and track the nodes by using a stack, and finally classify all the mutually reachable nodes into a strongly connected component. The specific process is as follows: start depth-first traversal from an arbitrary unvisited node, assign it an increasing index value, and initialize its "low link value" as the index value, then recursively visit its adjacent nodes; in the backtracking process, update the "low link value" of the current node as the minimum value of its own initial value and the "low link values" of all adjacent nodes; when the "depth index" of a node is equal to its "low link value", all the nodes popped from the stack until the node itself are classified into a strongly connected component. Through this algorithm, the network graph can be decomposed into multiple subgraphs. Among them, the part that does not form a circular dependency and has a one-way dependency relationship can be sorted into a linear solving sequence by topological sorting. Those node sets that have circular dependency relationship in the interior are identified as coupled constraint groups. For example, the control point parameters of multiple primitives are linked together through a set of mutually referenced equations (such as "the profile of A depends on the position of B, and the profile of B depends on the curvature of A"), forming a strongly connected component, i.e. a coupled constraint group.

[0070] Then, each identified coupled constraint group is processed to transform it into a form that can be efficiently solved. For each coupled constraint group, a set of additional virtual relaxation variables is introduced, the number of which is equal to the number of independent constraint equations in the group. Let the original coupled constraint group contain equations, denoted as: where is the vector of all morphological parameter variables in the group, represents the th equation. Introduce the relaxation variable vector: . The original equality constraint is reconstructed as the sum of two terms: one is the original constraint function , another is the corresponding slack variable multiplied by a preset positive penalty coefficient . Specifically, a part of the new objective function (energy function) is constructed as follows: ;

[0071] At the same time, inequality constraints are added to the slack variable ≥ 0. In this way, the original coupled set of equality constraints is transformed into an optimization sub-problem with inequality constraints on the variable and . The principle is that by minimizing , the variable is driven to 0, and at the same time, due to the positive penalty of the term, the variable is also pressed to 0, thus approaching the satisfaction of the original equality constraint at the optimal solution. For the non-coupled, ordered constraint sequence, its original equality form is maintained.

[0072] Finally, all the processed sub-problems, i.e. the serialized independent constraints and the transformed coupled constraint optimization sub-problems, are organized into a hierarchical cascading structure according to the dependency relationship. This structured representation constitutes the global nonlinear constraint solving problem. The solving order between levels is determined by the topological sorting, and the solving of coupled sub-problems within the same level is completed by minimizing its corresponding energy function. By decomposing the complex, possibly conflicting global constraint system into this structured, clear, and sequentially and parallel alternately solvable hierarchical problem, the efficiency and stability of the solution are improved, laying a computational foundation for generating component profiles that conform to both complex narrative logic and geometric consistency.

[0073] In the visual style enhancement module of intangible cultural heritage, first, the surface gradient domain decomposition is performed on the preliminary component geometric profile to separate the geometric information of different spatial frequencies. Specifically, the surface of the component profile is represented by a dense triangular mesh. For the three-dimensional coordinates of each vertex on the mesh, a filtering process based on the discrete Laplacian operator is applied. The core of this process is to calculate the difference between the weighted average value of the coordinates of each vertex and all its adjacent vertices, which reflects the high-frequency details of the local area. By setting a preset cutoff frequency value (for example, corresponding to details with a spatial wavelength less than 2 mm), the calculated high-frequency components can be separated from the original vertex coordinates. The result obtained by subtracting the high-frequency components from the original coordinates is the low-frequency geometric basic information representing the relationship between the overall form of the component and the main narrative structure; while the separated part is the high-frequency surface detail information containing surface micro-relief, which has not yet been stylized.

[0074] Secondly, a pre-built visual style feature library of intangible cultural heritage is accessed, and the retrieved style features are fitted into the topology framework defined by the aforementioned low-frequency geometric base information. The library is built by collecting typical physical objects or high-precision digital maps of target genres (such as emblematic genres), and pre-processing them to decouple features such as brush stroke units, wood textures, and repetitive decorative patterns into independent, parameterized data blocks. The first step in the fitting process is to analyze the mesh surface corresponding to the low-frequency geometric base information. The vertices with significant curvature changes on the mesh are extracted as key points, and adjacent key points are connected to form feature lines reflecting the main geometric trends. According to the connection relationship between these points and lines, a correlation mapping matrix is constructed. This matrix is a two-dimensional table, with rows representing sampling points in the pre-processed style feature data blocks and columns representing key points and feature line segments on the target topology framework; the value of each element in the matrix represents the similarity between the sampling point and the geometric element represented by the column in the local normal and tangent directions.

[0075] Then, the deformation and resampling of style features are guided by the correlation mapping matrix. For each segment of brush stroke or texture feature data retrieved from the feature library, a deformation transformation parameter is calculated based on its similarity to each part of the target topology framework. This parameter controls the degree of stretching, compression, or bending of the feature data, so that its spatial distribution pattern can conform to the curvature of the feature lines and the distribution density of the key points in the target framework as much as possible. For example, a straight line of data representing "axe-cut carving" brush strokes will be automatically resampled into a corresponding curved shape when fitted onto a curved ivy pattern feature line.

[0076] Finally, the accurate attachment and synthesis of style features are completed through a differentiable weighted fusion operation. The resampled style feature data is placed in the corresponding position of the target topology framework. In this process, each style feature data point is assigned a weight value, which depends on its distance to the nearest key point or feature line segment on the target framework, with closer distances resulting in higher weights. At the same time, the high-frequency surface detail information separated in the first step is preserved. Through a weighted synthesis calculation, the fitted style feature data (representing the target visual style) is fused with the original high-frequency detail information. This calculation is performed on the new coordinates of each vertex of the mesh, and the result is a new, higher-precision triangular mesh surface. This new surface strictly maintains the topology framework and narrative structure defined by the original low-frequency geometric base information, but its surface details are replaced by new geometric surfaces with specific artistic qualities, which are a mixture of target intangible cultural heritage style features and partial original details, thus completing the visual enhancement.

[0077] In the multi-constraint compliance verification and output module, first, a multi-objective compliance evaluation index system is established, and the abstract requirements of "cultural implications, form constraints, artistic styles and engineering feasibility" are converted into four quantifiable core calculation parameters. One is the semantic consistency parameter. The calculation method is: extract the key geometric features (such as the contour line of a specific area, the carving depth) on the current geometric contour surface mesh, find its corresponding semantic entity and relationship predicate by querying the "semantic association matrix", and then perform semantic matching degree analysis with the "target cultural implication text" input initially. The matching degree is expressed in percentage, and the target value is not less than 95%. Two is the geometric deviation parameter. The calculation method is: compare the boundary and control point coordinates of the current geometric contour with the ideal boundary and allowable tolerance range defined in the "target building component form constraint parameter", calculate the maximum distance deviation and the average distance deviation, and the unit is millimeter. The target value is that the maximum deviation is less than 5 millimeters. Three is the style similarity parameter. The calculation method is: extract the visual feature vectors such as stroke direction distribution and texture roughness spectrum from the enhanced surface of the current geometric contour, and calculate the cosine similarity with the reference feature vector of the target style in the "non-heritage visual style feature library". The similarity value is between 0 and 1, and the target value is greater than 0.85. Four is the mechanical safety parameter. The calculation method is: according to the specified material properties (such as the elastic modulus of wood) of the component, perform a simplified static finite element analysis on the current geometric contour, calculate the ratio of the maximum stress value under standard load (such as self-weight) to the allowable stress, i.e. the safety factor, and the target value is greater than 2.0.

[0078] Secondly, according to the above evaluation index system, the component geometric contour after style processing is executed in a hierarchical and cascading parallel verification operation. The verification operation is layered according to the calculation dependency of the parameters: first, the geometric deviation parameter and the style similarity parameter are calculated in parallel, because these two parameters do not depend on other results; then, the mechanical safety parameter is calculated based on the stable geometric contour; finally, after the geometric, style and mechanical data are ready, the comprehensive semantic consistency parameter evaluation is performed. After each parameter calculation is completed, a sub-report will be generated to record the calculation value, target value and deviation degree (the difference or ratio of the calculation value and the target value). After all the sub-reports are collected, a complete compliance diagnosis report is formed, which clearly points out in which specific parameters the current model does not meet the preset compliance threshold, and the size of the deviation value.

[0079] Next, based on the generated compliance diagnosis report, a multi-objective gradient-guided parameter space optimization method is initiated to perform the optimization adjustment. The core of this method is to simultaneously adjust the underlying "form parameters" and "style parameters" related to the problem parameters. A parameter influence relationship table is established, which defines the influence weight and direction (positive or negative influence) of each adjustable parameter (such as the control point coordinates of a certain geometric primitive, the intensity coefficient of a certain style stroke) on the above four evaluation parameters. The optimization process is iterative: in each iteration, first read the compliance diagnosis report to find out the parameters that do not meet the standard; then, according to the parameter influence relationship table, for each parameter that does not meet the standard, calculate which underlying parameters should be adjusted and the gradient direction (i.e. should the parameter value be increased or decreased); then, by a weighted summation algorithm, a set of comprehensive adjustment amounts of the underlying parameters that can make the overall compliance improve the fastest is calculated; finally, apply this set of adjustment amounts to modify the model parameters, generate a new geometric contour, and re-execute the verification operation in the second step to generate a new diagnosis report. This cycle continues until the latest diagnosis report shows that all evaluation parameters meet the preset compliance threshold, and the iteration stops.

[0080] Finally, the optimized final digitalized building component model that meets all requirements is output. The model is composed of three parts associated in the data layer. The first part is a multi-level data structure. The top layer is the "semantic relationship topology structure" that drives the entire generation, stored in graph data format; the middle layer is the "geometric primitive set carrying programmable parameters", which stores the types, IDs and final parameter values of all parameterized geometric elements in the current model; the bottom layer is the "high-precision curved surface mesh after fusion of style features", i.e. the final visual three-dimensional model, stored in common formats such as OBJ or STL. The three layers of data are linked by pointers or unique identifiers. The second part is a parameterized interface closely associated with the geometric primitive set in the middle layer. This interface is actually a structured log file or database table that records all versions of each form parameter and style parameter of each geometric primitive from the initial generation to the final optimization adjustment, as well as the reason for each adjustment (associated with the compliance diagnosis report entry of a certain iteration). The third part is a semantic association matrix. The matrix is a two-dimensional lookup table, whose row index corresponds to a specific face or feature group of the bottom layer curved surface mesh, and whose column index corresponds to the ID of a specific semantic entity and the ID of a relationship predicate in the pattern narrative knowledge set. The value in the matrix identifies the mapping strength between the geometric feature and the cultural semantic element. This matrix enables the geometric form of any local part of the model to be traced back to its cultural implications, achieving deep fusion and retrievability of geometric information and semantic information.

[0081] The working principle of the present application is as follows: firstly, through the multi-modal data acquisition and narrative knowledge construction module, the image and semantic description data of the building component of Huizhou architecture are collected, and the structured pattern narrative knowledge set is constructed through multi-scale feature extraction, syntax analysis and cross-modal fusion; secondly, the semantic relationship topology reasoning module receives the target cultural implication and component form constraint, performs entity decomposition and relationship extraction based on the knowledge set, and generates a semantic relationship topology structure that meets the implication logic and spatial layout by combining random walk simulation and probability graph reasoning; then, the parameterized geometric instantiation module instantiates the topology structure into parameterized geometric primitives according to the predefined geometric form rules, generates the preliminary component geometric contour by constructing and solving the hierarchical constraint optimization problem with the introduction of slack variables; then, the non-heritage visual style enhancement module performs gradient domain decomposition on the contour, retrieves the target style data from the style feature library, and gives it non-heritage visual characteristics on the premise of preserving the narrative structure through morphing resampling and weighted fusion; finally, the multi-constraint compliance verification and output module establishes a multi-objective evaluation index system, performs hierarchical verification and gradient-guided iterative optimization on the stylized contour, and outputs the digital building component model integrated with semantic topology, parameterized geometric primitives, high-precision stylized surface and complete semantic correlation matrix.

[0082] The above describes one embodiment of the present application in detail, but the content described is only the preferred embodiment of the present application, and cannot be considered as limiting the scope of the present application. Any equivalent changes and improvements made according to the scope of the present application should still belong to the scope of the present application.

Claims

1. A system for intelligent generation of components of a local style of architecture based on non-heritage pattern recognition, characterized in that, Comprise: A multi-modal data acquisition and narrative knowledge construction module for acquiring image data and semantic description data of non-heritage building components, and for constructing a pattern narrative knowledge set containing pattern semantic entities and relation predicates through intelligent analysis based on pattern elements and spatial relationships in the image data and corresponding semantic descriptions; The construction process of the pattern narrative knowledge set is: Multi-scale feature extraction and spatial coding processing are performed on the image data to output geometric feature vectors of each pattern element and a spatial relationship matrix composed of relative positions between elements; The semantic description data is subjected to syntax dependency analysis and entity relationship extraction processing to output a semantic relationship graph with pattern entities as nodes and relationship descriptions as edges; The geometric feature vectors and the spatial relationship matrix are subjected to feature fusion and matching with the nodes and edges in the semantic relationship graph to analyze the corresponding relationship between the pattern semantic entities and the relation predicates, and to construct the pattern narrative knowledge set; A semantic relationship topology reasoning module for receiving a target cultural connotation text and a form constraint parameter of a target building component, generating a semantic relationship topology structure matching the cultural connotation and the form constraint through intelligent reasoning based on the pattern narrative knowledge set, specifically comprising: Performing entity deconstruction and relationship extraction on the target cultural connotation text based on the pattern narrative knowledge set to obtain an initial semantic entity node set and a potential relationship edge set; Transforming the form constraint parameter of the target building component into a spatial layout boundary condition, simulating the spatial accessibility of the entity nodes under the spatial layout boundary condition through a random walk process, and forming an entity node set with position attributes; Constructing a probabilistic graph network based on the entity node set with position attributes and the potential relationship edge set, calculating the probability of existence of each potential relationship edge through multiple rounds of message passing and probability reasoning, and filtering out stable relationship edges according to a probability threshold to generate a semantic relationship topology structure matching both the cultural connotation and the form constraint; A parameterized geometric instantiation module for instantiating and converting the semantic relationship topology structure according to a predefined, parameterized geometric form rule set, which defines geometric form parameters corresponding to different semantic entities and geometric constraint conditions triggered by relationship edges, and generates a preliminary component geometric contour by solving parameters that satisfy all geometric constraint conditions; A non-heritage visual style enhancement module for applying visual feature enhancement processing conforming to the artistic style to the preliminary component geometric contour, so that the preliminary component geometric contour has the preset non-heritage visual style features while retaining the narrative geometric relationships; A multi-constraint compliance verification and output module for performing multi-dimensional compliance verification and optimization adjustment on the component geometric contour after style processing, and outputting a digital building component model meeting the requirements of cultural connotation, form constraint, artistic style and engineering feasibility.

2. The non-heritage pattern recognition-based system for intelligent generation of components of a style of architecture according to claim 1, wherein, The formation of the entity node set with position attributes specifically comprises: Generating a two-dimensional discretized grid representation space according to the component contour and internal functional zoning defined by the form constraint parameter, and marking the discretized grid representation space as a boundary condition region allowing node distribution; Assigning an initial random position seed point to each node in the set of entity nodes within the boundary condition region, and controlling the walk step and direction based on a preset weight parameter related to the semantic type of the node, performing multiple rounds of parallel random walk simulation; During the random walk process, the relative distance and repulsive force between nodes are dynamically calculated, and when the cumulative moving distance of all nodes is lower than the preset threshold and the basic distribution balance condition is met, the simulation is stopped and the final stable position coordinates of each node are output, forming a set of entity nodes with position attributes. 3.The non-heritage pattern recognition-based system for intelligent generation of components of a style of architecture, as claimed in claim 1, wherein, The generating the preliminary component geometric profile specifically comprises: According to the semantic relationship topology structure, traversing each relationship edge and the entity nodes connected by the relationship edge, retrieving and instantiating the corresponding parameterized geometric primitive and geometric constraint equation set from the geometric shape rule set; Solving the global nonlinear constraint problem with the geometric primitive shape parameters as variables by simultaneously solving all the instantiated geometric constraint equation sets; Converting the constraint solving problem into an energy minimization problem, and gradually reducing the total system energy by iteratively adjusting the numerical values of the shape parameters until convergence to a stable state is achieved; Assigning the shape parameter values obtained after convergence to the corresponding geometric primitives, and completing the Boolean operation between the primitives and the surface stitching, outputting the preliminary component geometric profile.

4. The non-heritage pattern recognition-based system for intelligent generation of components of a style of architecture according to claim 3, wherein, The generating the preliminary component geometric profile specifically comprises: According to the association relationship between variables and equations in the instantiated geometric constraint equation set, a bidirectional constraint network graph is constructed, in which nodes represent shape parameter variables and constraint equations, and edges represent the participation relationship of variables in equations; Performing hierarchical topological sorting and strongly connected component partitioning on the constraint network graph, decomposing the global solving problem into a sequence of sub-problems with a prior relationship, and identifying the coupled constraint groups containing circular dependencies; Introducing a virtual relaxation variable to the identified coupled constraint groups, reconstructing the original constraint equation set into a hierarchical cascading optimization problem containing equality and inequality relationships, and forming a structured global nonlinear constraint solving problem representation.

5. The non-heritage pattern recognition based smart generation system of architectural components of a style, according to claim 1, wherein, The applying visual feature enhancement processing conforming to the artistic style to the preliminary component geometric profile specifically comprises: Performing surface gradient domain decomposition on the preliminary component geometric profile to separate low-frequency geometric basic information representing the core narrative structure and high-frequency surface detail information to be enhanced; Accessing a pre-constructed non-heritage visual style feature library, retrieving corresponding stroke, texture and decoration unit feature data according to the target style, and adapting them to the topological framework defined by the geometric basic information; Through a differentiable feature fusion process, the adapted style feature data and the high-frequency surface detail information are weighted and synthesized to generate an enhanced geometric surface that maintains the original profile narrative structure and adds the target non-heritage visual style.

6. The non-heritage pattern recognition-based system for intelligent generation of components of a style of architecture according to claim 5, wherein, The applying visual feature enhancement processing conforming to the artistic style to the preliminary component geometric profile specifically comprises: Calculating the topological connection relationship of the surface key points and feature lines of the geometric basic information, and generating an associated mapping matrix corresponding thereto; According to the geometric properties of the association mapping matrix, the retrieved stroke, texture and decoration unit feature data are parameterized and morphed for resampling, so that the spatial distribution of the feature data is consistent with the geometric trend of the topological framework; Through distributed weighted fusion operation based on key points and feature lines, the morphed feature data is accurately aligned and attached to the corresponding position of the topological framework, completing the geometric adaptation of style features and narrative structure.

7. The non-heritage pattern recognition based smart generation system of architectural components of a style, according to claim 1, wherein, The multi-dimensional compliance verification and optimization adjustment of the geometric profile of the component processed by the stylization, specifically includes: Establish a multi-objective compliance evaluation index system, and quantify cultural implications, morphological constraints, artistic styles and engineering feasibility requirements into computable semantic consistency parameters, geometric deviation parameters, style similarity parameters and mechanical safety parameters; According to the evaluation index system, perform hierarchical cascading parallel verification operation on the component geometric profile processed by the stylization, and generate a compliance diagnosis report recording the deviation degree of each parameter in detail; Based on the compliance diagnosis report, a multi-objective gradient guided parameter space optimization method is used to iteratively fine-tune the morphological and style parameters of the component geometric profile until all evaluation parameters reach the preset compliance threshold, and finally output the optimized digital building component model.

8. The non-heritage pattern recognition-based system for intelligent generation of components of a style of architecture according to claim 7, wherein, The digital building component model specifically includes: A multi-level data structure, the top layer of the multi-level data structure is a semantic relationship topological structure, the middle layer is a geometric primitive set carrying programmable parameters, and the bottom layer is a high-precision curved surface grid fused with style features; A parameterized interface associated with the geometric primitive set, which records the final values and adjustment history of all morphological parameters and style parameters in the generation process; A semantic association matrix, which stores the mapping relationship between each geometric feature on the high-precision curved surface grid and the corresponding semantic entities and relationship predicates in the pattern narrative knowledge set in a retrievable form.

Images