An artificial intelligence-based building relief three-dimensional generation method

Abstract

The application discloses a kind of based on artificial intelligence's building relief three-dimensional generation method, comprising: three-dimensional scanning historical building obtains point cloud and texture data, constructs digital relief sample set;Extract sample feature and adopt generative adversarial network to train carving law, establish style database;Based on database rule analysis building geometric feature, match available area and generate preliminary three-dimensional relief design scheme;According to building style constraint, intelligent layout optimization is carried out to design scheme;From optimization distribution map, extract visual balance parameter, and conform to historical style three-dimensional relief model is generated through generative adversarial network iteration;Finally, texture is fused and is evaluated and style migration through deep learning, and the final result of coordination is output.The application realizes the whole process automation generation of relief from law learning, intelligent layout to closed-loop optimization, and greatly improves design efficiency, historical style restoration accuracy and overall coordination.

Description

Technical Field

[0001] This invention belongs to the field of digital architecture technology, and in particular relates to a method for generating three-dimensional architectural reliefs based on artificial intelligence. Background Technology

[0002] Architectural reliefs, as an important decorative form of building facades, directly influence the cultural expression and visual aesthetics of buildings. Their 3D generation technology plays a crucial role in the preservation of historical buildings, the design of new buildings, and the digitization of cultural heritage. With the rapid development of artificial intelligence technology, its application to the 3D generation of architectural reliefs has become an important direction, but existing methods still face significant limitations.

[0003] Current methods primarily rely on manual design or traditional computer-aided tools. Designers must manually adjust the pattern layout, carving depth, and style matching, a process that is time-consuming and highly dependent on personal experience. When building facades are complex, containing numerous windows, doorways, and load-bearing areas, manual layout can easily overlook spatial constraints, leading to uneven distribution of relief elements or visual clashes with the building's main structure. Furthermore, due to a lack of systematic study of historical relief design principles, the generated results often fail to achieve realistic carving effects in terms of undulation, line direction, and texture, and stylistic harmony is also difficult to guarantee.

[0004] The core of these limitations lies in the fact that the deep integration of spatial layout optimization and overall architectural style unity during the relief generation process has not yet been effectively achieved. The arrangement of elements in architectural reliefs must simultaneously meet geometric constraints, aesthetic rules, and structural limitations, but current technology struggles to automatically identify usable areas on the facade and intelligently optimize them according to visual balance and rhythm, making the layout prone to imbalances in hierarchy or rhythmic breaks. Furthermore, when it is necessary to ensure that the relief's stylistic language is consistent with the architectural era, columns, cornices, window frames, and other characteristics, the lack of an effective style constraint mechanism results in significant differences between the generated reliefs and the main building in terms of carving depth, surface treatment, and decorative themes.

[0005] For example, on the facade of a classical building, if the relief patterns are too densely arranged around the windows, it will disrupt the facade's symmetry and rhythm; if the depth of the carving does not match the decoration of the building's cornice, it will make the local decoration appear abrupt and disrupt the overall harmony. These contradictions have repeatedly occurred in actual projects.

[0006] Ultimately, the shortcomings of existing technologies lie in three aspects: First, they lack the ability to learn autonomously from historical data and quantify style patterns, resulting in the inability to generalize generation rules; second, they fail to establish a decision-making model that integrates architectural geometry, aesthetic rules, and style constraints, leading to a lack of systematic layout optimization; and third, the generation, evaluation, and optimization stages are fragmented, making it impossible to form a data-driven closed-loop iteration. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention proposes an artificial intelligence-based method for generating three-dimensional architectural reliefs, thereby resolving the issues present in the existing technologies.

[0008] To achieve the above objectives, the present invention provides a method for generating three-dimensional architectural reliefs based on artificial intelligence, comprising: Point cloud data and texture information are obtained from historical buildings and processed using 3D scanning equipment to obtain a set of digital relief samples. Feature vectors are extracted from the digital relief sample set, and the carving rules are trained using a generative adversarial network model to obtain the generation rules of relief patterns in the style database. By analyzing the architectural geometric features through the generation rules of relief patterns in the style database, if the matching degree between the geometric features and the carving rules is higher than a preset threshold, the distribution of usable areas is determined, and a preliminary three-dimensional relief design scheme is obtained. The layout of the preliminary three-dimensional relief design is optimized to obtain the constraints of the overall architectural style. If the constraints meet the coordination and matching requirements, the positions of the pattern elements are adjusted to obtain the optimized relief distribution map. Visual balance parameters are extracted from the optimized relief distribution map and iteratively updated through a generative adversarial network model to obtain a three-dimensional relief model that matches the historical building. Based on the texture information fused from the 3D relief model, a deep learning algorithm is used to evaluate the overall style consistency. If the consistency score is lower than a preset threshold, style migration processing is triggered to obtain the final coordinated 3D relief generation result.

[0009] Optionally, point cloud data and texture information are obtained from historical buildings and processed using 3D scanning equipment to obtain a set of digital relief samples, including: The surface of historical buildings is scanned using a 3D scanning device to obtain raw point cloud data and corresponding texture information. The raw point cloud data is then denoised and aligned to obtain clean point cloud data. Based on this clean point cloud data, geometric information is extracted through normal vector and curvature analysis to identify and determine the concave and convex contours of the relief area. According to the identified relief area, corresponding surface details are separated from the texture information to obtain texture detail data. Using a region-growing-based point cloud segmentation algorithm, the clean point cloud data is divided into multiple independent relief units based on the boundary information of the concave and convex contours and the curvature changes of the point cloud, resulting in segmented point cloud subsets. For each segmented point cloud subset, its corresponding texture detail data is mapped to establish a mapping relationship between point cloud and texture. Based on the point cloud subsets and their texture mapping relationship, a single digital relief model is reconstructed to obtain a complete relief sample containing geometric contours and surface textures. All complete relief samples are combined to form a digital relief sample set.

[0010] Optionally, the process of obtaining the generation rules for embossed patterns from the style database includes: From the set of digital relief samples, the geometric structure features and texture distribution features of each relief sample are extracted. The geometric structure features include depth maps, curvature distributions, and contour boundaries calculated based on concave and convex contour information. The geometric structure features and texture distribution features are fused and encoded, quantized into standardized feature vectors. A generative adversarial network model is constructed, using the feature vectors as training data. The generator learns to map from random noise to a distribution that conforms to the feature vectors of real reliefs, while the discriminator learns to distinguish between real and generated feature vectors. After iterative optimization through adversarial training, the network parameters of the generator are solidified and defined as generation rules that can capture and reproduce specific carving patterns. The generation rules are associated with corresponding architectural style tags and stored in a style database to form a queryable relief pattern generation rule library.

[0011] Optionally, by analyzing the architectural geometric features through the generation rules of relief patterns in the style database, if the matching degree between the geometric features and the carving rules is higher than a preset threshold, the distribution of usable areas is determined, and a preliminary three-dimensional relief design scheme is obtained, including: From the style database, retrieve relief pattern generation rules that match the target architectural style; obtain the architectural geometry data of the target building and calculate the geometric feature vectors of each section of its surface; calculate the matching degree between the geometric feature vectors of each section and the carving rules contained in the generation rules, wherein the carving rules include undulating shapes and depth variation constraints defined based on concave and convex contour information; if the matching degree of a single section is higher than a preset matching degree threshold, it is determined to be a usable area suitable for applying relief; combine all usable areas to determine their spatial distribution range on the building facade; within the distribution range of the usable areas, call the generation rules to automatically generate a planar layout of the relief pattern, combine the planar layout with the architectural geometry surface to generate a preliminary three-dimensional relief design scheme; wherein, the generated pattern must conform to the shape rules of concave and convex contours in the carving rules.

[0012] Optionally, aesthetic constraints corresponding to the overall style of the target building can be extracted from the style database. These aesthetic constraints include symmetry requirements, visual rhythm, and hierarchical distribution rules. Using the aforementioned aesthetic constraints as the optimization objective function, and simultaneously taking the morphological rules regarding the concave and convex contours in the carving principles as geometric constraints, the preliminary three-dimensional relief design scheme is optimized using multiple objectives. By iteratively adjusting the position, size, and rotation angle of the pattern elements, the optimized layout satisfies the aforementioned aesthetic constraints. Output the layout scheme that satisfies the constraints, i.e. the optimized relief distribution map.

[0013] Optionally, visual balance parameters are extracted from the optimized relief distribution map and iteratively updated using a generative adversarial network model to obtain a 3D relief model that conforms to the historical building, including: From the optimized relief distribution map, parameters for quantifying visual balance are extracted. These parameters include pattern density distribution, visual center of gravity position, and symmetry index. The calculation of the visual center of gravity position takes into account the influence of changes in the concave and convex contours on the lighting effect. The visual balance parameters and preset historical building style labels are used as input conditions to drive the trained conditional generative adversarial network model. Using the aforementioned generative adversarial network model, a two-dimensional distribution map is converted into a relief model with preliminary three-dimensional geometric information, wherein the generation of the three-dimensional geometric shape of the model must follow the generation rules of concave and convex contours in the aforementioned carving rules; The initial 3D relief model is compared with the historical building reference model in terms of style consistency. If the consistency score is lower than the preset style consistency threshold, the deviation information is fed back to the generative adversarial network model. The model parameters are iteratively updated based on the deviation information until the output model achieves the required consistency with the style of the historical building, thus obtaining the final three-dimensional relief model.

[0014] Optionally, based on the fusion of texture information from the 3D relief model, a deep learning algorithm is used to evaluate the overall style consistency. If the consistency score is lower than a preset threshold, style transfer processing is triggered to obtain the final coordinated 3D relief generation result, including: The geometric data of the three-dimensional relief model is fused and mapped with the original texture information of the historical building; A pre-trained deep learning style evaluation network is used to evaluate and score the overall style consistency of the fused model. The evaluation considers both geometric style consistency and texture style consistency. The evaluation dimensions of geometric style consistency include depth distribution, contour sharpness and undulation rhythm features calculated based on concave and convex contour information. If the evaluation score is lower than the preset final evaluation threshold, the texture style is determined to be inconsistent, and style migration processing is triggered. The style migration process transfers the texture style features of the target historical building to the current model texture while keeping its geometric structure unchanged. The migration results are evaluated a second time until the style consistency score meets the standard, and the final coordinated 3D relief generation result is output.

[0015] Optionally, the generative adversarial network model is a conditional generative adversarial network, with the generator adopting a U-Net structure and the discriminator adopting a PatchGAN structure. The generator takes random noise and architectural style labels as input and outputs an embossed feature vector.

[0016] Compared with the prior art, the present invention has the following advantages and technical effects: This invention autonomously learns carving patterns from digital relief samples using generative adversarial networks and constructs a queryable style database. This allows the system to break free from dependence on fixed templates, automatically extract and quantify the geometric structure and texture features of historical reliefs, and form generalizable generation rules, thereby creating relief patterns for different architectural carriers that both conform to historical style logic and adapt to new geometric forms.

[0017] In terms of multi-constraint fusion and intelligent layout, the system uniformly encodes architectural geometric features, aesthetic rules, and historical style constraints. Through matching degree calculation and multi-objective optimization algorithms, it automatically identifies usable areas on the facade and generates a harmonious relief layout. This method effectively resolves the contradictions between geometric feasibility, visual balance, and stylistic consistency in layout schemes, significantly reducing the subjectivity and trial-and-error costs of manual layout.

[0018] Regarding closed-loop iteration and quality optimization, this invention uses visual balance parameters to drive generative adversarial networks for model iteration and introduces deep learning evaluation and style transfer mechanisms to achieve a data-driven closed loop of "generation-evaluation-optimization". This process ensures that the final 3D relief is highly consistent with the historical building in terms of geometric details, texture, and overall style, improving the realism of the generated result and the accuracy of cultural heritage preservation.

[0019] This invention replaces the traditional fragmented work dominated by manual labor with fully automated generation. In scenarios such as historical building restoration, decoration of new cultural buildings, and digital archiving of cultural heritage, it can significantly improve design efficiency, ensure the accuracy of style restoration, and enhance the coordination and reusability of the overall solution. Attached Figure Description

[0020] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a method according to an embodiment of the present invention. Detailed Implementation

[0021] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0022] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0023] Example 1 like Figure 1 As shown, this embodiment provides a method for generating 3D architectural reliefs based on artificial intelligence, including: S1. Obtain point cloud data and texture information from historical buildings, and process them using a 3D scanning device to obtain a set of digital relief samples; The surface of historical buildings is scanned using a 3D scanning device to obtain raw point cloud data and corresponding texture information. The raw point cloud data is then denoised and aligned to obtain clean point cloud data. Based on this clean point cloud data, geometric information is extracted through normal vector and curvature analysis to identify and determine the concave and convex contours of the relief area. According to the identified relief area, corresponding surface details are separated from the texture information to obtain texture detail data. Using a point cloud segmentation algorithm, the clean point cloud data is divided into multiple independent relief units based on the boundary information of the concave and convex contours, resulting in segmented point cloud subsets. For each segmented point cloud subset, its corresponding texture detail data is mapped to establish a mapping relationship between point cloud and texture. Based on the point cloud subsets and their texture mapping relationships, a single digital relief model is reconstructed to obtain a complete relief sample containing geometric contours and surface textures. All complete relief samples are combined to form a digital relief sample set.

[0024] The specific process is as follows: First, a 3D scanning device is used to comprehensively scan the surface of the historical building, simultaneously collecting raw point cloud data and corresponding surface texture information. These two types of data are the foundation for restoring the geometric shape and visual details of the relief. Then, the raw point cloud data is preprocessed. Noise points mixed in during the scanning process are removed through denoising operations, and then coordinate alignment is performed to ensure that all point cloud data are in a unified spatial coordinate system, resulting in clean point cloud data.

[0025] Based on clean point cloud data, the core geometric information of the relief is extracted through normal vector and curvature analysis, accurately identifying the concave and convex contour boundaries of the relief area and clearly defining the boundary between the relief and the main building. Simultaneously, surface details corresponding to the relief area are separated from the collected texture information, forming texture detail data. Then, a point cloud segmentation algorithm based on region growing is used, with point cloud curvature changes and concave and convex contour boundaries as the core segmentation criteria, to divide the clean point cloud data into multiple independent relief units, obtaining segmented point cloud subsets.

[0026] Next, for each segmented point cloud subset, corresponding texture detail data is mapped one-to-one, establishing a precise correspondence between the point cloud geometry and texture details. Based on this mapping, a single relief unit is reconstructed in 3D, generating a digital relief model containing complete geometric contours and surface textures. Finally, all reconstructed individual digital relief models are integrated to form a standardized and structured set of digital relief samples, providing high-quality data support for the subsequent training of generative adversarial network models. The entire process achieves precise conversion from physical reliefs to digital samples, ensuring the geometric integrity and texture realism of the samples.

[0027] Taking the dragon relief sculpture in the Hall of Supreme Harmony in the Forbidden City as an example, a 3D laser scanning device was used to scan the dragon pattern area on the outer pillars of the hall, collecting the original point cloud and texture containing details of dragon scales and whiskers. After denoising to remove noise points caused by dust, the coordinates were aligned and the concave and convex contours of the dragon body were identified through curvature analysis. Then, a region growing algorithm was used to divide the dragon head, body, and tail into independent units, and the textures were mapped and reconstructed separately. Finally, the samples were integrated into a set of dragon relief sculpture samples from the Hall of Supreme Harmony.

[0028] S2. Extract feature vectors from the digital relief sample set, train the carving rules using a generative adversarial network model, and obtain the generation rules of relief patterns in the style database. From the set of digital relief samples, the geometric structure features and texture distribution features of each relief sample are extracted. The geometric structure features include depth maps, curvature distributions, and contour boundaries calculated based on concave and convex contour information. The geometric structure features and texture distribution features are fused and encoded, quantized into standardized feature vectors. A generative adversarial network model is constructed, using the feature vectors as training data. The generator learns to map from random noise to a distribution that conforms to the feature vectors of real reliefs, while the discriminator learns to distinguish between real and generated feature vectors. After iterative optimization through adversarial training, the network parameters of the generator are solidified and defined as generation rules that can capture and reproduce specific carving patterns. The generation rules are associated with corresponding architectural style tags and stored in a style database to form a queryable relief pattern generation rule library.

[0029] The specific process is as follows: First, from the constructed set of digital relief samples, two types of core features are extracted for each relief sample. One type is geometric structure features, which cover key morphological indicators such as depth map, curvature distribution and contour boundary calculated based on concave and convex contours; the other type is texture distribution features, which record the texture pattern and distribution rules of the relief surface.

[0030] These two types of features are then fused and encoded to transform them into standardized feature vectors, converting the complex relief morphology and texture information into numerical data that the model can recognize. Next, a conditional generative adversarial network (GAN) model is constructed. The generator uses a U-Net structure, and the discriminator uses a PatchGAN structure. The generator takes random noise and architectural style labels as input and outputs vectors that simulate realistic relief features. During training, the standardized feature vectors are used as training data, allowing the generator to learn the mapping from random noise to a distribution that conforms to the real relief feature vectors, while the discriminator learns to distinguish between real relief feature vectors and the fake vectors output by the generator.

[0031] Through multiple rounds of adversarial training and iterative optimization, the feature vectors generated by the generator become increasingly closer to real samples. Once the model performance reaches a stable state, the generator's network parameters are solidified. These parameters represent the generation rules that can capture and reproduce the patterns of specific architectural relief carvings. Finally, the generation rules are associated with corresponding architectural style tags, such as Gothic and Classical Chinese, and these rules are stored in a style database to construct a relief pattern generation rule library that can be queried by architectural style, providing core rule support for the subsequent generation of target architectural reliefs.

[0032] A sample set of Suzhou garden window reliefs in the classical Chinese style was selected. The geometric depth map (e.g., the protrusion depth of peony patterns), curvature distribution (curvature changes at the edges of petals), and texture distribution (the direction of leaf vein texture) of each relief were extracted and fused into feature vectors. These vectors were used to train a conditional generative adversarial network (GAN). The generator learned the patterns of Suzhou garden window reliefs: "symmetrical layout, rounded and protruding petals, and delicate leaf veins." After the parameters were solidified, the network was labeled "Suzhou garden window" and stored in a style database, forming the generation rules for this style.

[0033] S3. Analyze the architectural geometric features by analyzing the generation rules of relief patterns in the style database. If the matching degree between the geometric features and the carving rules is higher than the preset threshold, determine the distribution of usable areas and obtain a preliminary three-dimensional relief design scheme. From the style database, retrieve relief pattern generation rules that match the target architectural style; obtain the architectural geometry data of the target building and calculate the geometric feature vectors of each section of its surface; calculate the matching degree between the geometric feature vectors of each section and the carving rules contained in the generation rules, the carving rules including undulating shapes and depth variation constraints defined based on concave and convex contour information; if the matching degree of a certain section is higher than a preset first threshold, then the area is determined to be a usable area suitable for applying relief; combine all usable areas to determine their spatial distribution range on the building facade; within the distribution range of the usable areas, call the generation rules to automatically generate a planar layout of the relief pattern, wherein the generated pattern must conform to the shape rules of concave and convex contours in the carving rules; combine the planar layout with the architectural geometry surface to generate a preliminary three-dimensional relief design scheme.

[0034] The specific process is as follows: First, from the style database, based on the style positioning of the target building, retrieve the relief pattern generation rules that match it to ensure that the relief generated subsequently conforms to the style tone of the target building.

[0035] Subsequently, complete architectural geometry data of the target building is acquired, and geometric feature vectors of each section of the building surface are calculated using feature extraction algorithms. These vectors quantify key geometric attributes such as surface morphology and curvature changes of each section. Next, the geometric feature vectors of each section are compared with the carving rules contained in the retrieved generation rules to calculate the matching degree. The carving rules here mainly include the relief undulation shape and depth variation constraints defined based on the concave and convex contours. The core of the matching degree calculation is to determine whether the geometric conditions of the building section are suitable for supporting the relief of the corresponding style.

[0036] If the matching degree of a certain area is higher than a preset first threshold (range 0.65-0.75), then that area is determined to be a usable area suitable for applying relief. Areas with a matching degree lower than the threshold are excluded to avoid conflicts between the relief and the building structure. Then, all usable areas are combined to clarify their specific spatial distribution on the building facade, defining the boundaries of the relief design. Within the determined usable area distribution range, matching relief pattern generation rules are invoked to automatically generate the planar layout of the relief pattern. During the generation process, the morphological rules regarding concave and convex contours in carving principles are strictly followed to ensure the rationality of the relief form. Finally, the generated planar layout is fitted to the three-dimensional geometric surface of the target building, transforming the two-dimensional pattern into a three-dimensional form that adapts to the undulations of the building surface, thus obtaining a preliminary three-dimensional relief design scheme, laying the foundation for subsequent layout optimization.

[0037] The target building is a Gothic-style cathedral. Generation rules for "Gothic relief pointed arches, openwork patterns, and significant depth variations" were retrieved from a style database. After obtaining the cathedral's facade geometry data, the geometric vectors of the walls, lintels, and window frames were calculated. The matching degree between the lintel and window frame areas reached 0.72 (higher than the first threshold of 0.7), thus identifying them as usable areas. The rules were then used to generate a pointed arch relief planar layout, fitting the curved surface of the lintel and the window frame edges to form a preliminary 3D relief design.

[0038] S4. Extract the constraints of the overall architectural style from the style database, and optimize the layout of the preliminary three-dimensional relief design scheme with the constraints as the optimization target to obtain the optimized relief distribution map. Based on the style database, aesthetic constraints corresponding to the overall style of the target building are extracted, including symmetry requirements, visual rhythm, and hierarchical distribution rules. The preliminary 3D relief design scheme is analyzed to evaluate the coordination and matching degree between the current layout of its pattern elements and the aesthetic constraints. If the coordination and matching degree is lower than a preset second threshold, an optimization algorithm is initiated, using the aesthetic constraints as the objective function and the morphological rules regarding concave and convex contours in the carving rules as constraints. The position, size, and rotation angle of the pattern elements are iteratively adjusted through the optimization algorithm to improve the coordination and matching degree of the new layout to meet the requirements. The adjusted layout scheme, i.e., the optimized relief distribution map, is output.

[0039] The specific process is as follows: First, extract the aesthetic constraints corresponding to the overall style of the target building from the style database. These constraints are the aesthetic norms that have been formed over a long period of time for this type of building. They mainly include symmetry requirements, visual rhythm and hierarchical distribution rules. For example, Chinese architectural reliefs often emphasize symmetrical balance, while European architectural reliefs focus on the rhythm of hierarchical progression.

[0040] Subsequently, using these aesthetic constraints as standards, a comprehensive analysis of the preliminary 3D relief design scheme is conducted. The current layout of the relief pattern elements in the scheme is evaluated, along with their degree of harmony and matching with the aesthetic constraints. The position, size, and combination of the patterns are assessed to determine whether they conform to the style and aesthetics of the target building. If the evaluated degree of harmony and matching is lower than a preset second threshold (range 0.75-0.85), a layout optimization algorithm is initiated. During the optimization process, the aesthetic constraints are set as the objective function of the algorithm; that is, the final direction of optimization is to ensure that the relief layout meets these aesthetic requirements. Simultaneously, the morphological rules regarding the concave and convex contours in the carving principles are set as constraints to ensure that the core morphological characteristics of the relief are not destroyed during the optimization process.

[0041] The optimization algorithm iteratively adjusts the position, size, and rotation angle of the relief pattern elements. For example, it moves the pattern to meet symmetry requirements, adjusts the size ratio to enhance visual rhythm, and adjusts the rotation angle to adapt to the building surface orientation. After each adjustment, the coordination and matching degree between the layout and aesthetic constraints is recalculated until the matching degree exceeds a second threshold, satisfying the coordination and matching requirements. At this point, the adjusted layout scheme, i.e., the optimized relief distribution map, is output. This distribution map conforms to the overall architectural style and aesthetics while preserving the carving rules of the relief itself, achieving harmonious unity between the local relief and the overall building.

[0042] The target building is a traditional Chinese courtyard house with a hanging flower gate. The aesthetic constraint of "left-right symmetry and clear vertical hierarchy" was extracted. In the initial design, the relief patterns on the brackets on both sides of the hanging flower gate were misaligned, resulting in a matching degree of only 0.7 (below the second threshold of 0.8). An optimization algorithm was then initiated, aiming for symmetry. The positions of the reliefs on the brackets on both sides were adjusted to achieve perfect symmetry, and the size of the relief on the upper beam was reduced to enhance the hierarchy. After iteration, the matching degree reached 0.86, and the optimized relief distribution map was output.

[0043] S5. Extract visual balance parameters from the optimized relief distribution map, and iteratively update the model through a generative adversarial network to obtain a three-dimensional relief model that matches the historical building. From the optimized relief distribution map, parameters for quantifying visual balance are extracted. These parameters include pattern density distribution, visual center of gravity position, and symmetry index. The calculation of the visual center of gravity position considers the influence of changes in concave and convex contours on the lighting effect. Using the visual balance parameters and preset historical building style labels as input conditions, a trained conditional generative adversarial network (GAN) model is driven. The GAN model is used to convert the two-dimensional distribution map into a relief model with preliminary three-dimensional geometric information. The generation of the model's three-dimensional geometric shape must follow the rules for generating concave and convex contours in the carving principles. The preliminary three-dimensional relief model is compared with a historical building reference model for style consistency. If the consistency score is lower than a preset third threshold, deviation information is fed back to the GAN model. The model parameters are iteratively updated based on the deviation information until the output model achieves the required style consistency with the historical building, resulting in the final three-dimensional relief model.

[0044] The specific process is as follows: First, extract three sets of key parameters for quantifying visual balance from the optimized relief distribution map, namely pattern density distribution, visual center of gravity position and symmetry index. Among them, the calculation of the visual center of gravity position needs to take into special consideration the influence of the relief contour changes on the light and shadow effect, because the difference in light and shadow will directly affect the human eye's judgment of the visual center of gravity of the relief.

[0045] These visual balance parameters, along with preset historical architectural style labels, are then used as input to drive a pre-trained conditional generative adversarial network (GAN) model. Upon receiving the input conditions, the model initiates the generation process, transforming the two-dimensional relief distribution map into a relief model with preliminary three-dimensional geometric information. The generation process strictly adheres to the rules governing the generation of concave and convex contours in sculpting, ensuring that the model's three-dimensional geometry conforms to the core features of the target architectural relief.

[0046] After generating a preliminary 3D relief model, it is compared with a reference model of the historical building for stylistic consistency. Quantitative analysis is used to evaluate the similarity between the two models in terms of form, layout, and visual effects, resulting in a consistency score. If this score is lower than a preset third threshold (range 0.80-0.90), the current model is deemed not to meet stylistic requirements. Deviation information is then extracted, such as insufficient relief depth or uneven pattern density distribution, and this information is fed back to the conditional generative adversarial network (GAN) model. Based on the feedback, the model adjusts its internal network parameters and restarts the generation process to produce an optimized 3D relief model. This iterative process of style consistency comparison and parameter adjustment is repeated, with each iteration specifically correcting model deviations, until the style consistency score between the output 3D relief model and the historical building reference model exceeds the third threshold. The iteration process then terminates, and the resulting model is the final 3D relief model that conforms to the style of the historical building.

[0047] Targeting the Tang Dynasty style relief on the lintel of the Big Wild Goose Pagoda, parameters such as pattern density (the distribution ratio of the main pattern to the decorative patterns) and visual center of gravity (the centering of the main pattern) were extracted from the optimized distribution map and input into an adversarial network labeled "Tang Dynasty relief." The initially generated model had insufficient convexity depth of the main pattern, resulting in a consistency score of 0.82 (below the third threshold of 0.85). After feedback on the deviation, the parameters were adjusted, and a new model was iteratively generated. The main pattern depth met the requirements, and the center of gravity was centered under lighting, achieving a score of 0.88, thus obtaining the final 3D model.

[0048] S6. Based on the texture information fused in the three-dimensional relief model, a deep learning algorithm is used to evaluate the overall style consistency. If the consistency score is lower than a preset threshold, style migration processing is triggered to obtain the final coordinated three-dimensional relief generation result.

[0049] The geometric data of the 3D relief model is fused and mapped with the original texture information obtained in S1. A pre-trained deep learning style evaluation network is used to evaluate and score the overall style consistency of the fused model. The evaluation considers both geometric style consistency and texture style consistency. The evaluation dimensions of geometric style consistency include depth distribution, contour sharpness, and undulation rhythm features calculated based on concave and convex contour information. If the evaluation score is lower than a preset fourth threshold, the texture style is determined to be inconsistent, triggering style transfer processing. The style transfer processing transfers the texture style features of the target historical building to the current model texture while maintaining its geometric structure. The transfer result is evaluated a second time until the style consistency score reaches the standard, and the final coordinated 3D relief generation result is output.

[0050] The specific process is as follows: First, the geometric data of the three-dimensional relief model obtained in S5 is fused and mapped with the original texture information collected in S1, and the texture details are accurately attached to the geometric surface of the relief model to form a complete relief model that combines geometric shape and surface texture.

[0051] Subsequently, a pre-trained deep learning style evaluation network was used to evaluate and score the overall style consistency of the fused model. The evaluation dimensions were divided into two aspects: geometric style consistency and texture style consistency. The evaluation of geometric style consistency focused on the depth distribution, contour sharpness, and undulation rhythm features calculated based on the concave and convex contours to determine whether the model's geometric shape matched the relief style of historical buildings. Texture style consistency evaluated whether the pattern, color, and texture of the model's surface texture were consistent with the historical buildings. If the overall score obtained from the evaluation was lower than the preset fourth threshold (range 0.78-0.85), the model's texture style was determined to be unacceptable, and style transfer processing was triggered.

[0052] The core objective of style transfer processing is to transfer the texture style features of the target historical building to the texture of the current model, while strictly maintaining the established geometric structure of the model to avoid disrupting the three-dimensional form of the relief due to texture adjustments. During the transfer process, the deep learning model extracts the style features of the historical building's texture, such as line thickness, pattern repetition, and color saturation, and then injects these features into the texture of the current model to achieve precise texture style replacement. After style transfer is completed, the deep learning style evaluation network is called again to perform a secondary style consistency evaluation on the transferred model. If the score still does not meet the standard, the style transfer and secondary evaluation process is repeated until the overall style consistency score of the model exceeds the fourth threshold. At this point, the output model is the final three-dimensional relief generation result, in which both the geometric form and texture style are highly coordinated with the historical building.

[0053] The generated 3D model of the flying apsara relief from the Mogao Grottoes in Dunhuang was merged with the texture of the Mogao Grottoes murals collected by S1. Evaluation revealed that the model's texture colors were too bright, inconsistent with the "rustic, yellowish, and rugged" style of the Mogao Grottoes, scoring 0.76 (below the fourth threshold of 0.8). Style transfer was triggered, extracting the yellowish hue and rugged lines from the original Mogao Grottoes texture and transferring them to the model texture, preserving the concave-convex shape of the flying apsaras. A second evaluation scored 0.83, and the final, harmonious 3D relief result was output.

[0054] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for generating 3D architectural reliefs based on artificial intelligence, characterized in that, Includes the following steps: Point cloud data and texture information are obtained from historical buildings and processed using 3D scanning equipment to obtain a set of digital relief samples. Feature vectors are extracted from the digital relief sample set, and the carving rules are trained using a generative adversarial network model to obtain the generation rules of relief patterns in the style database. By analyzing the architectural geometric features through the generation rules of relief patterns in the style database, if the matching degree between the geometric features and the carving rules is higher than a preset threshold, the distribution of usable areas is determined, and a preliminary three-dimensional relief design scheme is obtained. The layout of the preliminary three-dimensional relief design is optimized to obtain the constraints of the overall architectural style. If the constraints meet the coordination and matching requirements, the positions of the pattern elements are adjusted to obtain the optimized relief distribution map. Visual balance parameters are extracted from the optimized relief distribution map and iteratively updated through a generative adversarial network model to obtain a three-dimensional relief model that matches the historical building. Based on the texture information fused from the 3D relief model, a deep learning algorithm is used to evaluate the overall style consistency. If the consistency score is lower than a preset threshold, style migration processing is triggered to obtain the final coordinated 3D relief generation result.

2. The method for generating three-dimensional architectural reliefs based on artificial intelligence according to claim 1, characterized in that, Point cloud data and texture information were obtained from historical buildings and processed using 3D scanning equipment to obtain a set of digital relief samples, including: The surface of historical buildings is scanned using a 3D scanning device to obtain raw point cloud data and corresponding texture information. The raw point cloud data is then denoised and aligned to obtain clean point cloud data. Based on this clean point cloud data, geometric information is extracted through normal vector and curvature analysis to identify and determine the concave and convex contours of the relief area. According to the identified relief area, corresponding surface details are separated from the texture information to obtain texture detail data. Using a region-growing-based point cloud segmentation algorithm, the clean point cloud data is divided into multiple independent relief units based on the boundary information of the concave and convex contours and the curvature changes of the point cloud, resulting in segmented point cloud subsets. For each segmented point cloud subset, its corresponding texture detail data is mapped to establish a mapping relationship between point cloud and texture. Based on the point cloud subsets and their texture mapping relationship, a single digital relief model is reconstructed to obtain a complete relief sample containing geometric contours and surface textures. All complete relief samples are combined to form a digital relief sample set.

3. The method for generating three-dimensional architectural reliefs based on artificial intelligence according to claim 1, characterized in that, The process of obtaining the generation rules for embossed patterns from the style database includes: From the set of digital relief samples, the geometric structure features and texture distribution features of each relief sample are extracted. The geometric structure features include depth maps, curvature distributions, and contour boundaries calculated based on concave and convex contour information. The geometric structure features and texture distribution features are fused and encoded, quantized into standardized feature vectors. A generative adversarial network model is constructed, using the feature vectors as training data. The generator learns to map from random noise to a distribution that conforms to the feature vectors of real reliefs, while the discriminator learns to distinguish between real and generated feature vectors. After iterative optimization through adversarial training, the network parameters of the generator are solidified and defined as generation rules that can capture and reproduce specific carving patterns. The generation rules are associated with corresponding architectural style tags and stored in a style database to form a queryable relief pattern generation rule library.

4. The method for generating three-dimensional architectural reliefs based on artificial intelligence according to claim 1, characterized in that, By analyzing the architectural geometric features using the generation rules of relief patterns in the style database, and determining if the matching degree between the geometric features and the carving rules is higher than a preset threshold, the distribution of usable areas is determined, resulting in a preliminary 3D relief design scheme, including: From the style database, retrieve relief pattern generation rules that match the target architectural style; obtain the architectural geometry data of the target building and calculate the geometric feature vectors of each section of its surface; calculate the matching degree between the geometric feature vectors of each section and the carving rules contained in the generation rules, wherein the carving rules include undulating shapes and depth variation constraints defined based on concave and convex contour information; if the matching degree of a single section is higher than a preset matching degree threshold, it is determined to be a usable area suitable for applying relief; combine all usable areas to determine their spatial distribution range on the building facade; within the distribution range of the usable areas, call the generation rules to automatically generate a planar layout of the relief pattern, combine the planar layout with the architectural geometry surface to generate a preliminary three-dimensional relief design scheme; wherein, the generated pattern must conform to the shape rules of concave and convex contours in the carving rules.

5. The method for generating three-dimensional architectural reliefs based on artificial intelligence according to claim 1, characterized in that, From the style database, the aesthetic constraints corresponding to the overall style of the target building are extracted. The aesthetic constraints include symmetry requirements, visual rhythm, and hierarchical distribution rules. Using the aforementioned aesthetic constraints as the optimization objective function, and simultaneously taking the morphological rules regarding the concave and convex contours in the carving principles as geometric constraints, the preliminary three-dimensional relief design scheme is optimized using multiple objectives. By iteratively adjusting the position, size, and rotation angle of the pattern elements, the optimized layout satisfies the aforementioned aesthetic constraints. Output the layout scheme that satisfies the constraints, i.e. the optimized relief distribution map.

6. The method for generating three-dimensional architectural reliefs based on artificial intelligence according to claim 1, characterized in that, Visual balance parameters are extracted from the optimized relief distribution map and iteratively updated using a generative adversarial network model to obtain a 3D relief model that conforms to the historical building, including: From the optimized relief distribution map, parameters for quantifying visual balance are extracted. These parameters include pattern density distribution, visual center of gravity position, and symmetry index. The calculation of the visual center of gravity position takes into account the influence of changes in the concave and convex contours on the lighting effect. The visual balance parameters and preset historical building style labels are used as input conditions to drive the trained conditional generative adversarial network model. Using the aforementioned generative adversarial network model, a two-dimensional distribution map is converted into a relief model with preliminary three-dimensional geometric information, wherein the generation of the three-dimensional geometric shape of the model must follow the generation rules of concave and convex contours in the aforementioned carving rules; The initial 3D relief model is compared with the historical building reference model in terms of style consistency. If the consistency score is lower than the preset style consistency threshold, the deviation information is fed back to the generative adversarial network model. The model parameters are iteratively updated based on the deviation information until the output model achieves the required consistency with the style of the historical building, thus obtaining the final three-dimensional relief model.

7. The method for generating three-dimensional architectural reliefs based on artificial intelligence according to claim 1, characterized in that, Based on the fusion of texture information from the 3D relief model, a deep learning algorithm is used to evaluate the overall style consistency. If the consistency score is lower than a preset threshold, style transfer processing is triggered to obtain the final coordinated 3D relief generation result, including: The geometric data of the three-dimensional relief model is fused and mapped with the original texture information of the historical building; A pre-trained deep learning style evaluation network is used to evaluate and score the overall style consistency of the fused model. The evaluation considers both geometric style consistency and texture style consistency. The evaluation dimensions of geometric style consistency include depth distribution, contour sharpness and undulation rhythm features calculated based on concave and convex contour information. If the evaluation score is lower than the preset final evaluation threshold, the texture style is determined to be inconsistent, and style migration processing is triggered. The style migration process transfers the texture style features of the target historical building to the current model texture while keeping its geometric structure unchanged. The migration results are evaluated a second time until the style consistency score meets the standard, and the final coordinated 3D relief generation result is output.

8. The method for generating three-dimensional architectural reliefs based on artificial intelligence according to claim 3, characterized in that, The generative adversarial network model is a conditional generative adversarial network. The generator adopts a U-Net structure, and the discriminator adopts a PatchGAN structure. The generator takes random noise and architectural style labels as input and outputs an embossed feature vector.

Images