Three-dimensional model acquisition and restoration system and method

Abstract

The invention relates to the technical field of three-dimensional model processing, in particular to a three-dimensional model collecting and repairing system and method. The system scans an ancient cultural relic through a laser scanning device, obtains three-dimensional point cloud data and texture information of the ancient cultural relic, and constructs an original three-dimensional damage model. On this basis, the system obtains the same type of cultural relic images similar to the target ancient cultural relic through an image style clustering algorithm, and forms a style label library. Thirdly, performing style feature extraction on the image in the style label library by adopting a deep learning network to obtain a style feature vector for condition generation; the system further generates a model based on the three-dimensional diffusion obtained by training, takes the style feature vector as condition input, and generates a complete three-dimensional repair model consistent with the target style through a gradual denoising process. And finally, outputting a repaired three-dimensional model result by the system, and providing a reference for subsequent digital display, three-dimensional printing or material object repair.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional model processing technology, and in particular to a three-dimensional model acquisition and repair system and method. Background Technology

[0002] Many ancient artifacts have suffered varying degrees of damage due to natural erosion, environmental changes, or human activities. This damage not only affects the appearance and structure of the artifacts but may also lead to the loss of their historical and cultural value. Bronze artifacts, in particular, are highly susceptible to cracking, corrosion, and breakage during long-term preservation due to their unique materials and craftsmanship. Traditional artifact restoration methods primarily rely on manual restoration. While this method can restore the appearance of artifacts, the process is complex and time-consuming, and it is difficult to maintain the original style and historical traces of the artifacts, often resulting in restoration effects that fall short of ideal standards.

[0003] Currently, the field of cultural relic restoration has begun to explore digital restoration technologies, among which 3D scanning and 3D printing technologies are widely used in the digital modeling and restoration of ancient artifacts. However, existing 3D restoration methods still have many problems. For example, the models obtained by 3D scanning often only contain physical morphological information, making it difficult to restore details such as the texture, color, and style of the artifacts. In addition, existing digital restoration methods usually lack in-depth learning of the historical style of the artifacts, often resulting in stylistic distortion during the restoration process, and failing to accurately restore the original appearance of the artifacts. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides a three-dimensional model acquisition and restoration system and method, aiming to improve the problem that existing restoration methods are difficult to achieve in terms of style consistency and detail restoration.

[0005] In a first aspect, the present invention provides the following technical solution: a three-dimensional model acquisition and repair system, comprising: The 3D data acquisition module is used to scan ancient artifacts using laser scanning equipment, obtain their 3D point cloud data and texture information, and construct the original 3D damage model. The image style clustering module is used to acquire multiple sets of images of similar ancient artifacts related to the target ancient artifact, and to classify the style of the image set through a clustering algorithm to form a style tag library; The style feature extraction module is used to learn style vector embeddings of images in the style tag library based on a deep learning network, and obtain style feature vectors for conditional generation. The 3D restoration model generation module is used to generate a complete 3D restoration model with the same style as the input by taking the style feature vector as the conditional input and through a step-by-step denoising process on the basis of the original 3D damaged model. The Repair Suggestion Output Module is used to output the 3D repair model results for subsequent digital display, 3D printing, or physical repair reference.

[0006] Furthermore, the steps for constructing the original three-dimensional damage model are as follows: The target ancient artifact is scanned from all angles using laser scanning equipment to obtain dense point cloud data; The acquired point cloud data is preprocessed, including noise filtering, coordinate system unification, and data format standardization. Extract texture information from point cloud data and establish the correspondence between color attributes and spatial coordinates; The point cloud data is converted into a 3D mesh model by a triangulation algorithm, and the coordinates and geometric features of the damaged areas in the model are identified and recorded to complete the construction of the original 3D damaged model.

[0007] Furthermore, the steps to create a style tag library are as follows: Collect image data of reference ancient artifacts of the same era, type, and craftsmanship as the target ancient artifact, and preprocess the images, including size standardization, illumination equalization, and noise removal. Visual features of an image are extracted, and hierarchical clustering algorithm is used to classify the image features. The visual features include: texture features, geometric features, and color features. A unique style label is assigned to each cluster category to establish a mapping relationship between style categories and image samples. By verifying the effectiveness of the clustering results, the clustering parameters are adjusted and the style label library is improved.

[0008] Furthermore, the steps to obtain the style feature vector are as follows: A deep learning network model is trained using image data from a style tag library, and the network parameters are optimized using the backpropagation algorithm to minimize the style classification loss function. The trained network model is forward-propagated to each image in the style tag library, and the feature representations of the intermediate layers of the network are extracted as style feature vectors. The extracted style feature vectors are dimensionality reduced and standardized to form a fixed-dimensional conditional input vector.

[0009] Furthermore, the training process for the 3D diffusion generation model includes: Collect multiple complete 3D models of similar ancient artifacts and preprocess them in point cloud or mesh format to unify their size and coordinates; The ancient artifact images corresponding to the training samples are input into the style feature extraction network to obtain style vectors for conditional modeling; A progressive noise-adding process is applied to the training samples, and the reverse denoising process of the diffusion model is guided by style conditions. By minimizing the error between the original model and the reconstruction result, the parameters of the diffusion network are optimized to complete the training of the 3D diffusion generation model.

[0010] Furthermore, the steps for generating the 3D repair model are as follows: The original 3D damaged model to be repaired is input into the trained 3D diffusion generation model, and the style feature vector extracted from the target style image is input as conditional information. The diffusion generation model performs a stepwise reverse denoising process, predicting and completing the geometric structure and texture information of the damaged area under style conditions, and outputting a 3D restoration model with consistent style and complete structure that matches the input.

[0011] Furthermore, the output 3D repair model is in OBJ, PLY, or STL format, supporting manual editing, size adjustment, and material modification.

[0012] Secondly, this invention provides the following technical solution: a method for acquiring and repairing a three-dimensional model, comprising: Ancient artifacts are scanned using laser scanning equipment to obtain their three-dimensional point cloud data and texture information, and an original three-dimensional damage model is constructed. Acquire multiple sets of images of similar ancient artifacts related to the target artifact, and classify the image set by style using a clustering algorithm to form a style tag library; Style vector embedding learning is performed on images in the style tag library based on deep learning network to obtain style feature vectors for conditional generation. Based on the trained 3D diffusion generation model, the style feature vector is used as a conditional input. On the basis of the original 3D damaged model, a complete 3D restoration model with the same input style is generated through a step-by-step denoising process. Output the 3D restoration model results for subsequent digital display, 3D printing, or physical restoration reference.

[0013] The present invention has the following beneficial effects: 1. This invention collects image data of cultural relics of the same type, era, and craftsmanship as the target ancient artifact, and uses a hierarchical clustering algorithm to classify the texture, geometry, color, and other features of the images, thereby forming a precise style tag library. This style clustering process effectively groups images of cultural relics with similar styles into one category, ensuring that the restoration process can accurately follow the historical style of the cultural relics.

[0014] 2. In the process of 3D cultural relic restoration, this invention extracts style feature vectors from images through a deep learning network and inputs them as conditional information into a 3D diffusion generation model. This process utilizes the inverse denoising mechanism of the diffusion model to gradually repair the geometric structure and texture of the damaged area through multiple iterations, ensuring that the restored cultural relic is highly consistent with the original in form and style. Attached Figure Description

[0015] Figure 1 This is a structural diagram of a three-dimensional model acquisition and repair system proposed in this invention; Figure 2 This is a flowchart of a three-dimensional model acquisition and repair method proposed in this invention. Detailed Implementation

[0016] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Example 1 In a first embodiment of the present invention, a three-dimensional model acquisition and repair system is provided, such as... Figure 1 As shown, it includes: The 3D data acquisition module is used to scan ancient artifacts using laser scanning equipment, obtain their 3D point cloud data and texture information, and construct the original 3D damage model. Furthermore, the steps for constructing the original three-dimensional damage model are as follows: The target ancient artifact is scanned from all angles using laser scanning equipment to obtain dense point cloud data; The acquired point cloud data is preprocessed, including noise filtering, coordinate system unification, and data format standardization. Extract texture information from point cloud data and establish the correspondence between color attributes and spatial coordinates; The point cloud data is converted into a 3D mesh model by a triangulation algorithm, and the coordinates and geometric features of the damaged areas in the model are identified and recorded to complete the construction of the original 3D damaged model.

[0018] Specifically, a high-precision 3D laser scanning device (such as an area array laser scanner or a portable structured light scanner) is used to perform a 360-degree omnidirectional scan of the target ancient artifact. During the scanning process, multiple observation angles and positions are set to avoid occlusion areas, ensuring the comprehensiveness and continuity of the collected point cloud data. The output is high-density 3D point cloud data, including spatial 3D coordinates (x, y, z) and optional reflectivity or texture mapping reference information.

[0019] The acquired point cloud data generally suffers from noise, discrete points, and coordinate system inconsistencies. Therefore, the following processing is required: noise filtering, using methods such as radius filtering, statistical filtering, or voxel filtering to remove outliers and scanning errors; coordinate system unification, standardizing the spatial reference coordinate system of the point cloud data and performing attitude alignment or registration (e.g., using the ICP algorithm); and data format conversion, unifying the point cloud into a standard format (e.g., .ply, .pcd, .xyz) for subsequent processing.

[0020] Laser scanning equipment acquires texture data, requiring the modeling and processing of the point cloud texture. Specific steps include: mapping each point cloud data point to its acquired image using pixel coordinates to obtain the corresponding color attributes (R, G, B); establishing a one-to-one mapping relationship between texture and geometric data, storing it as an additional attribute of the point cloud for texture restoration in subsequent 3D reconstruction and restoration.

[0021] Triangular mesh reconstruction algorithms (such as Poisson reconstruction, Ball Pivoting Algorithm, and Delaunay triangulation) are used to convert point clouds into continuous 3D mesh models. After mesh generation, damaged areas are automatically or semi-automatically labeled, including using geometric discontinuity analysis (such as boundary length anomalies and local curvature anomalies) to identify typical damaged structures such as cracks and holes; combining historical repair data or manual assistance, the coordinate range and morphological characteristics of missing areas are marked; and the damaged areas are saved as mask areas for subsequent model repair.

[0022] The above construction method can preserve the original structure and style information of cultural relics to the greatest extent, forming a digital twin basic model with high fidelity, which provides accurate support for subsequent three-dimensional restoration based on style conditions.

[0023] The image style clustering module is used to acquire multiple sets of images of similar ancient artifacts related to the target ancient artifact, and to classify the style of the image set through a clustering algorithm to form a style tag library; Furthermore, the steps to create a style tag library are as follows: Collect image data of reference ancient artifacts of the same era, type, and craftsmanship as the target ancient artifact, and preprocess the images, including size standardization, illumination equalization, and noise removal. Visual features of an image are extracted, and hierarchical clustering algorithm is used to classify the image features. The visual features include: texture features, geometric features, and color features. A unique style label is assigned to each cluster category to establish a mapping relationship between style categories and image samples. By verifying the effectiveness of the clustering results, the clustering parameters are adjusted and the style label library is improved.

[0024] Specifically, image data similar to the target ancient artifacts in terms of era, material, shape, use, or pattern are collected through museum digital collection platforms, archaeological databases, or open cultural relic image resource libraries to form an image sample set. To improve clustering quality, the collected images should have high resolution and discernible details. The image sample set undergoes standardization processing, including size standardization, unifying the image resolution to a fixed size (e.g., 224×224 or 512×512) to adapt to subsequent network processing; illumination equalization, using methods such as histogram equalization and adaptive contrast enhancement (e.g., CLAHE) to improve the usable information of the images; noise removal, applying algorithms such as mean filtering, Gaussian blur, or nonlocal mean denoising to reduce the interference of background noise on feature extraction; and format conversion, unifying the image data format to .jpg or .png and performing normalization processing.

[0025] After preprocessing, deep neural networks (such as ResNet, VGG, and Inception) are used to perform forward propagation on the image samples, extracting multi-scale feature maps from the intermediate convolutional layers and performing global average pooling to obtain fixed-dimensional feature vectors. To comprehensively express the stylistic features of cultural relics, the following three types of visual features can be extracted: texture features, such as Gabor filter response, gray-level co-occurrence matrix statistics, and deep texture embedding vectors; geometric features, obtained through image edge detection (such as Canny) or contour extraction (such as SIFT and HOG) to obtain structural contour distribution; and color features, extracted through color histograms and dominant hue analysis to extract color distribution vectors. These features are concatenated or projected into the same embedding space to form the joint style feature vector of the image. Hierarchical clustering algorithms are used to perform cluster analysis on the samples, automatically identifying subsets of images with similar style features.

[0026] After clustering, a unique style label ID is automatically assigned to each image class, and a mapping table between labels and image samples is constructed. To verify the accuracy of clustering and the representativeness of style labels, a silhouette coefficient assessment can be performed to measure intra-cluster consistency and inter-cluster differences. If there are obvious errors or class imbalances in the clustering results, the clustering threshold, number of layers, or feature weights can be adjusted for re-clustering. In the final style label library, each category represents a style type of ancient artifact, which can serve as the conditional input for subsequent style feature extraction and generation models.

[0027] The style tag library constructed using the above method can accurately reflect the stylistic differences in texture, geometry, and color among different ancient artifact images, providing clear and representative prior information for style feature vector extraction and 3D restoration models.

[0028] The style feature extraction module is used to learn style vector embeddings of images in the style tag library based on a deep learning network, and obtain style feature vectors for conditional generation. Furthermore, the steps to obtain the style feature vector are as follows: A deep learning network model is trained using image data from a style tag library, and the network parameters are optimized using the backpropagation algorithm to minimize the style classification loss function. The trained network model is forward-propagated to each image in the style tag library, and the feature representations of the intermediate layers of the network are extracted as style feature vectors. The extracted style feature vectors are dimensionality reduced and standardized to form a fixed-dimensional conditional input vector.

[0029] Specifically, a classification structure based on a convolutional neural network (CNN) is used to train image samples from the style tag library. The neural network model used can be a mainstream architecture such as ResNet-50, Inception-v3, or VGG16, or a Transformer-based image encoding network. Image samples labeled with style categories in the style tag library are divided into training and validation sets. Data augmentation (including rotation, scaling, cropping, and color perturbation) is performed on each image to improve model robustness. The network training process includes standardizing the input image size to, for example, 224×224 pixels; using the cross-entropy loss function as the optimization objective; selecting Adam or SGD as the optimization algorithm; and setting the learning rate to [value missing]. ; During training, the network weight parameters are automatically adjusted using the backpropagation algorithm to minimize the classification error between the predicted and true labels. After training, the model that performs best on the validation set is selected as the subsequent feature extraction network.

[0030] Using a trained deep learning network, perform a forward propagation on each image in the style tag library. Extract the activation output of a specific intermediate layer (such as the penultimate fully connected layer or a convolutional layer) as the style feature representation of that image. If using ResNet-50, the output of the avgpool layer (typically 2048 dimensions) can be used; if using VGG16, the output of the fc2 layer (typically 4096 dimensions) can be extracted; if using a Transformer architecture, the CLS token vector can be used as the global feature. The resulting raw high-dimensional feature vector is the preliminary style description of the image.

[0031] To improve computational efficiency and enhance the discriminative power of style features, a dimensionality reduction algorithm is used to compress the original style vectors. First, principal component analysis, t-SNE, UMAP, or other dimensionality reduction methods are applied to all original style vectors, reducing the feature dimension to, for example, 128 or 256 dimensions. Then, the dimensionality-reduced vectors are standardized (e.g., zero-mean unit variance normalization, or L2 normalization) to enhance the consistency of model generation under different style conditions. The final fixed-dimensional, normalized vector is the style feature vector described in this invention, which can be used as the conditional input for the subsequent 3D diffusion generation model.

[0032] The style feature vectors extracted in this step can efficiently and accurately represent the style characteristics of the target ancient artifact, achieving style transfer and completion while maintaining geometric consistency, effectively supporting the personalized restoration process of the subsequent 3D model.

[0033] The 3D restoration model generation module is used to generate a complete 3D restoration model with the same style as the input by taking the style feature vector as the conditional input and through a step-by-step denoising process on the basis of the original 3D damaged model. Furthermore, the training process for the 3D diffusion generation model includes: Collect multiple complete 3D models of similar ancient artifacts and preprocess them in point cloud or mesh format to unify their size and coordinates; The ancient artifact images corresponding to the training samples are input into the style feature extraction network to obtain style vectors for conditional modeling; A progressive noise-adding process is applied to the training samples, and the reverse denoising process of the diffusion model is guided by style conditions. By minimizing the error between the original model and the reconstruction result, the parameters of the diffusion network are optimized to complete the training of the 3D diffusion generation model.

[0034] Specifically, multiple complete and representative 3D model data of similar ancient artifacts are collected. The collected 3D models can come from public databases (such as digital archives of cultural heritage), museum-built databases, or on-site scan data, and file formats include .OBJ, .PLY, or .STL, etc. The collected 3D models undergo unified preprocessing, including scaling all models to a specified scale range (e.g., within a unit cube); standardizing the model pose using PCA principal orientation or manual keypoint alignment; and converting the mesh model into point cloud or voxel representation according to the training model requirements, and unifying it to a fixed resolution (e.g., 64×64×64 voxels, or 1024-point cloud).

[0035] The ancient artifact image corresponding to the above 3D model is input into a trained style feature extraction network (such as ResNet-50 or Transformer). The style feature vector of the image is obtained through forward propagation, and this vector will be used as the conditional input of the diffusion model.

[0036] Diffusion models or their variants are used as the underlying architecture for the 3D generative models. Each 3D model data is treated as a high-dimensional data tensor (e.g., voxels are represented as...). During the training phase, progressive Gaussian noise is added to construct a forward diffusion sequence. Where T is the total number of diffusion steps (e.g., T=1000), and noise is added in each step according to the formula: ; in, Indicates the first Noisy samples of the step, Noise scheduling parameters are typically expressed using linear, cosine, or learned attenuation. Indicates standard Gaussian noise. Represents a multidimensional Gaussian distribution. This represents the unit covariance matrix. This process progressively maps the 3D model to a pure noise space, forming training sample pairs. ,in This represents the style feature vector.

[0037] Construct a conditional diffusion neural network (such as 3D U-Net, PointNet++, or Voxel UNet) to learn the process of recovering the original model from noise given style feature vectors. The goal is to learn a function. To make it approximate real noise The training loss function is defined as: ; in, This is represented as the noise estimate output by the diffusion model. This represents the model parameters to be learned. Training, by minimizing this loss function, enables the network to effectively recover the 3D model from any diffusion step.

[0038] The 3D diffusion generation model trained through the above steps can accurately generate a complete 3D model that conforms to a specified artistic style under the guidance of the input style feature condition vector, achieving high-fidelity reconstruction from a structurally damaged model to a stylized complete model.

[0039] Furthermore, the steps for generating the 3D repair model are as follows: The original 3D damaged model to be repaired is input into the trained 3D diffusion generation model, and the style feature vector extracted from the target style image is input as conditional information. The diffusion generation model performs a stepwise reverse denoising process, predicting and completing the geometric structure and texture information of the damaged area under style conditions, and outputting a 3D restoration model with consistent style and complete structure that matches the input.

[0040] Specifically, the original 3D damaged model (point cloud or triangular mesh format) of the ancient artifact to be restored is input into the system, and operations such as unified coordinate normalization and point density adjustment are performed to make it conform to the input standard dimensions of the diffusion model. Simultaneously, the corresponding target style image is input into a pre-trained style feature extraction network to obtain its style vector. .

[0041] Using the trained 3D diffusion model, the damaged model is progressively denoised and reconstructed following the backsampling process of the diffusion model. This process is based on the conditional diffusion mechanism, and at each step... The current generated result is Predicting noise through models The approximate denoising result of the previous stage is calculated. The denoising process employs the following sampling strategy: ; in, and The noise scheduling parameters represent the diffusion process. This represents a conditional diffusion model, receiving the current state. Style conditions and time step , Indicates random noise. This represents the noise standard deviation at the current time step. This process iterates until... The final denoised model is output.

[0042] This method combines a 3D model of the damaged cultural relic to be restored with an image of the target style. A trained 3D diffusion generation model is then used to progressively denoise and complete the geometric structure and texture information of the damaged area. The resulting restoration model maintains the original structure of the cultural relic while conforming to the visual characteristics of the target style, ensuring a restoration effect that is both accurate and natural.

[0043] The Repair Suggestion Output Module is used to output the 3D repair model results for subsequent digital display, 3D printing, or physical repair reference.

[0044] Furthermore, the output 3D repair model is in OBJ, PLY, or STL format, supporting manual editing, size adjustment, and material modification.

[0045] Example 2: Bronze artifacts commonly exhibit varying degrees of damage and cracks due to long-term preservation and environmental erosion. Traditional manual restoration methods have limitations, such as inconsistent restoration results, time-consuming processes, and a tendency to distort the style. To address these issues, this invention provides a three-dimensional model acquisition and restoration method, such as... Figure 2 As shown. The specific implementation process of this method is as follows: A high-precision laser scanning device was used to perform a 360-degree scan of the bronze artifact. During the scan, the laser continuously emitted laser beams and measured the time it took for the laser to reflect back, thereby obtaining the three-dimensional coordinates of the bronze artifact's surface. After each scan, the device generated a point cloud dataset consisting of millions of points. The acquired point cloud data was preprocessed. First, a denoising algorithm was used to remove noise caused by scanning errors or external interference. Then, the coordinates of the point cloud data were unified to ensure that all scanned data were stitched together under the same coordinate system. Finally, the data format was standardized so that it could be used in subsequent steps. Texture information was extracted from the point cloud data, and the correspondence between color attributes and spatial coordinates was established. Using a triangulation algorithm, the processed point cloud data was transformed into a three-dimensional mesh model. By marking the coordinates and geometric features of the damaged areas, the original three-dimensional damaged model was constructed, providing accurate basic data for subsequent restoration.

[0046] Images of other artifacts from the same period, type, and craftsmanship as the target bronze artifact were collected. These images may come from museums, archaeological research sites, and other historical sites. The collected image data underwent preprocessing, including size normalization, illumination equalization, and noise removal, to ensure consistency in style analysis. Deep learning algorithms (such as convolutional neural networks, CNNs) were used to extract visual features from the images, including texture, color, and geometric features. Then, a hierarchical clustering algorithm was used to classify the image features, grouping images with similar styles into the same category. Each category corresponds to a style label, and a style label library is gradually formed. A unique style label is assigned to each cluster category, and the clustering results are validated with expert feedback, adjusting clustering parameters and refining the style label library. Ultimately, the style label library contains all image data and corresponding labels that are stylistically similar to the target bronze artifact.

[0047] A deep neural network is trained using image data from a style tag library. The network parameters are optimized using backpropagation to minimize the style classification loss function, allowing the network to gradually learn feature representations that distinguish different styles. The trained deep learning model is applied to each image in the style tag library, and the feature representations of the intermediate layers of the network are obtained through forward propagation; these feature representations are the style feature vectors. These vectors contain the style information of the image and are used for subsequent 3D restoration model generation. The extracted style feature vectors are then dimensionality-reduced (e.g., using PCA or t-SNE) and standardized to form fixed-dimensional style feature input vectors. These feature vectors will serve as conditional inputs for the subsequent 3D diffusion generation model.

[0048] The original 3D damaged model of the bronze artifact to be restored is input into a trained 3D diffusion generation model, along with style feature vectors extracted from the target style image as conditional information. This information guides the generation of the restoration model. Based on the principle of diffusion models, the 3D diffusion generation model applies progressive noise and gradually restores the geometric structure and texture information of the damaged area during the reverse denoising process. Through multiple iterations, the model gradually approximates the restored 3D model. In each denoising step, the model adjusts the artifact's geometry and texture according to the input style feature vectors, ensuring that the restored artifact is consistent with the target style, preserving the artifact's historical appearance while repairing its damaged parts.

[0049] The generated, repaired 3D model is converted into a standard file format (such as OBJ, PLY, STL) for easy digital display, 3D printing, or physical restoration. This module allows users to adjust the size and modify the materials of the repaired model to suit different restoration needs.

[0050] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A three-dimensional model acquisition and repair system, characterized in that, include: The 3D data acquisition module is used to scan ancient artifacts using laser scanning equipment, obtain their 3D point cloud data and texture information, and construct the original 3D damage model. The image style clustering module is used to acquire multiple sets of images of similar ancient artifacts related to the target ancient artifact, and to classify the style of the image set through a clustering algorithm to form a style tag library; The style feature extraction module is used to learn style vector embeddings of images in the style tag library based on a deep learning network to obtain style feature vectors for conditional generation. The 3D restoration model generation module is used to generate a complete 3D restoration model with the same style as the input by taking the style feature vector as the conditional input and through a step-by-step denoising process on the basis of the original 3D damaged model. The Repair Suggestion Output Module is used to output the 3D repair model results for subsequent digital display, 3D printing, or physical repair reference.

2. The three-dimensional model acquisition and repair system according to claim 1, characterized in that, The steps to construct the original 3D damaged model are as follows: The target ancient artifact is scanned from all angles using laser scanning equipment to obtain dense point cloud data; The acquired point cloud data is preprocessed, including noise filtering, coordinate system unification, and data format standardization. Extract texture information from point cloud data and establish the correspondence between color attributes and spatial coordinates; The point cloud data is converted into a 3D mesh model by a triangulation algorithm, and the coordinates and geometric features of the damaged areas in the model are identified and recorded to complete the construction of the original 3D damaged model.

3. The three-dimensional model acquisition and repair system according to claim 1, characterized in that, The steps to create a style tag library are as follows: Collect image data of reference ancient artifacts of the same era, type, and craftsmanship as the target ancient artifact, and preprocess the images, including size standardization, illumination equalization, and noise removal. Visual features of an image are extracted, and hierarchical clustering algorithm is used to classify the image features. The visual features include: texture features, geometric features, and color features. A unique style label is assigned to each cluster category to establish a mapping relationship between style categories and image samples. By verifying the effectiveness of the clustering results, the clustering parameters are adjusted and the style label library is improved.

4. The three-dimensional model acquisition and repair system according to claim 1, characterized in that, The steps to obtain style feature vectors are as follows: A deep learning network model is trained using image data from a style tag library, and the network parameters are optimized using the backpropagation algorithm to minimize the style classification loss function. The trained network model is forward-propagated to each image in the style tag library, and the feature representations of the intermediate layers of the network are extracted as style feature vectors. The extracted style feature vectors are dimensionality reduced and standardized to form a fixed-dimensional conditional input vector.

5. A three-dimensional model acquisition and repair system according to claim 1, characterized in that, The training process of the 3D diffusion generation model includes: Collect multiple complete 3D models of similar ancient artifacts and preprocess them in point cloud or mesh format to unify their size and coordinates; The ancient artifact images corresponding to the training samples are input into the style feature extraction network to obtain style vectors for conditional modeling; A progressive noise-adding process is applied to the training samples, and the reverse denoising process of the diffusion model is guided by style conditions. By minimizing the error between the original model and the reconstruction result, the parameters of the diffusion network are optimized to complete the training of the 3D diffusion generation model.

6. A three-dimensional model acquisition and repair system according to claim 1, characterized in that, The steps to generate a 3D repair model are as follows: The original 3D damaged model to be repaired is input into the trained 3D diffusion generation model, and the style feature vector extracted from the target style image is input as conditional information. The diffusion generation model performs a stepwise reverse denoising process, predicting and completing the geometric structure and texture information of the damaged area under style conditions, and outputting a 3D restoration model with consistent style and complete structure that matches the input.

7. A three-dimensional model acquisition and repair system according to claim 1, characterized in that, The output 3D repair model is in OBJ, PLY, or STL format, and supports manual editing, size adjustment, and material modification.

8. A method for acquiring and repairing three-dimensional models, characterized in that, include: Ancient artifacts are scanned using laser scanning equipment to obtain their three-dimensional point cloud data and texture information, and an original three-dimensional damage model is constructed. Acquire multiple sets of images of similar ancient artifacts related to the target artifact, and classify the image set by style using a clustering algorithm to form a style tag library; Style vector embedding learning is performed on images in the style tag library based on deep learning network to obtain style feature vectors for conditional generation. Based on the trained 3D diffusion generation model, the style feature vector is used as a conditional input. On the basis of the original 3D damaged model, a complete 3D restoration model with the same input style is generated through a step-by-step denoising process. Output the 3D restoration model results for subsequent digital display, 3D printing, or physical restoration reference.

Images