Cultural relic digital modeling and virtual reconstruction method and system based on three-dimensional Gaussian sputtering technology

Abstract

The invention relates to the technical field of digital cultural relic protection and restoration, and discloses a cultural relic digital modeling and virtual reconstruction method based on a three-dimensional Gaussian sputtering technology, and the method comprises the following steps: S1, obtaining a multi-angle image sequence of a cultural relic object, and carrying out the preprocessing of the multi-angle image sequence; s2, performing feature extraction on the preprocessed image, and constructing sparse point cloud and camera parameters; s3, based on the sparse point cloud and camera parameters, iteratively optimizing attributes of Gaussian points to generate a Gaussian point cloud model; and S4, trimming the Gaussian point cloud model, and exporting a renderable model file. According to the cultural relic digital modeling and virtual reconstruction method and system based on the three-dimensional Gaussian sputtering technology, the Gaussian point cloud model is generated through the Gaussian sputtering training step, the geometry and texture of the cultural relic are expressed through the Gaussian point set, each Gaussian point contains the attributes such as the position and the color, triangular meshes are not needed, data redundancy is reduced, the information density is improved, and the real-time performance of the cultural relic is improved. The problem that the expression efficiency of the irregular triangular mesh is low is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of digital cultural relic protection and restoration, in particular to a cultural relic digital modeling and virtual reconstruction method and system based on three-dimensional Gaussian sputtering technology. BACKGROUND

[0002] In the current field of digital restoration of cultural relics, oblique photogrammetry technology is a widely used and low-cost mainstream method. This method captures cultural relics from multiple angles, combines aerial images, ground images and other data sources, uses the structure light beam method in computer vision to construct dense point clouds, and generates textured triangular mesh models accordingly to realize three-dimensional reconstruction of the appearance structure of cultural relics.

[0003] In the practice of cultural relic protection, oblique photography technology is usually suitable for objects with regular structure or large scale such as building components and sculpture shapes. Its main workflow includes image acquisition, image registration, feature point extraction and matching, sparse point cloud reconstruction, dense point cloud generation, mesh reconstruction and texture mapping, etc. This method does not require expensive laser scanning equipment and has certain operability and popularization. It has been widely used in site mapping, exhibition display and digital filing, etc.

[0004] However, the current digital restoration technology of cultural relics still has defects:

[0005] 1. The generated mesh is an irregular triangular mesh, which is inefficient in expression. Oblique photography reconstruction results in a large number of irregular, unstructured triangular mesh bodies. Such meshes are often bloated when expressing complex surfaces, lack structure optimization, and result in large model data but low information density.

[0006] 2. The model quality depends on manual post-processing, which is time-consuming. Due to problems such as image occlusion, angle blind area and reflective material, the original modeling results often have obvious holes, cracks or misplacement, which require experienced modeling personnel to manually fill in the mesh and repair geometric errors for a long time. This method is inefficient, has poor reproducibility, and is not suitable for standardized batch modeling requirements.

[0007] 3. The UV unwrapping efficiency of the texture map is low, and the pixel utilization rate is poor. The texture map generated by oblique photography often uses automatic UV mapping, which has low UV space utilization, resulting in a large waste of pixel resources. Under the condition of limited resolution, the effective information density of the map is insufficient, and the details are blurred.

[0008] 4. The rendering efficiency is low. The generated mesh model relies on texture mapping for visualization, which often faces problems such as slow texture loading and model rendering lag, making it difficult to support real-time interactive digital display or immersive experience scenarios.

[0009] Therefore, a method and system for digital modeling and virtual reconstruction of cultural relics based on three-dimensional Gaussian sputtering technology are provided. SUMMARY

[0010] (1) Technical problems solved

[0011] In view of the deficiencies in the prior art, the present application provides a method and system for digital modeling and virtual reconstruction of cultural relics based on three-dimensional Gaussian sputtering technology, which has the advantages of expressing the geometry and texture of cultural relics through Gaussian voxels, constructing a Gaussian point automatic generation mechanism supporting multi-source input, developing a lightweight GPU real-time rendering module, forming a standardized modeling paradigm, realizing fine and efficient modeling without grids and UV maps, adapting to interactive display scenarios and reducing the modeling threshold, and solving the problems of the current cultural relic digital restoration technology, such as generating irregular triangular surface grids, low expression efficiency, model quality dependent on artificial post-repair, large workload, low UV unfolding efficiency of texture mapping, poor pixel utilization, and low rendering efficiency.

[0012] (2) Technical solutions

[0013] In order to achieve the above-mentioned purpose of expressing the geometry and texture of cultural relics through Gaussian voxels, constructing a Gaussian point automatic generation mechanism supporting multi-source input, developing a lightweight GPU real-time rendering module, forming a standardized modeling paradigm, realizing fine and efficient modeling without grids and UV maps, adapting to interactive display scenarios and reducing the modeling threshold, the present application provides the following technical solutions: a method for digital modeling and virtual reconstruction of cultural relics based on three-dimensional Gaussian sputtering technology, comprising the following steps:

[0014] S1, image acquisition and preprocessing: acquiring a multi-angle image sequence of a cultural relic object and performing preprocessing;

[0015] S2, feature extraction and sparse point cloud construction: performing feature extraction on the preprocessed image and constructing a sparse point cloud and camera parameters;

[0016] S3, Gaussian sputtering training: based on the sparse point cloud and camera parameters, the properties of Gaussian points are iteratively optimized to generate a Gaussian point cloud model;

[0017] S4, point cloud cleaning and export: pruning the Gaussian point cloud model and exporting a renderable model file.

[0018] Preferably, in the image acquisition and preprocessing step, a drone is used to carry a camera with a resolution of more than 20 million pixels for multi-angle and full-enclosure shooting, and the flight path is divided into three layers of strategies:

[0019] Long-distance encircling: for obtaining the overall structure;

[0020] Medium-distance encircling: to enhance the collection of local details;

[0021] Close-up tilt shot: complementary concave and top view information;

[0022] The camera parameter settings are resolution 3840x2160 pixels, ISO fixed at 100-200, white balance: fixed at "sunny" or "cloudy" mode.

[0023] Preferably, in the feature extraction and sparse point cloud construction step, the pre-processed image in step S1 is imported into RealityCapture software for three-dimensional modeling initialization, and the operation process is as follows:

[0024] S2.1, select "add image" function in RealityCapture to import all PNG images;

[0025] S2.2, click "image alignment", the software automatically extracts image feature points and constructs camera pose relationship network;

[0026] S2.3, set "maximum number of feature points" to 40,000 in the alignment parameter, and "image matching limit" to nearest neighbor matching within 30 images;

[0027] S2.4, the system outputs a sparse point cloud with a point number between 50,000 and 120,000 after processing, and automatically generates the camera intrinsic parameters (focal length, principal point position) and extrinsic parameters (rotation, translation matrix) of each image;

[0028] S2.5, after processing, export two key data: sparse point cloud data in Polygon File Format format, containing three-dimensional coordinates and color information; camera parameters through Reality Capture batch export tool to generate Excel or CSV format file, record the core parameters of each image, including focal length, image size, pose matrix;

[0029] S2.6, manually retake or exclude the part that fails to align the image, and ensure the integrity of the sparse point cloud without fault.

[0030] Preferably, in the Gaussian sputtering training step, first start Postshot, and import three key data:

[0031] .ply format sparse point cloud file;

[0032] CameraParameters.xlsx(or.csv) format camera parameter table;

[0033] The original image sequence participating in the training;

[0034] The parameter settings of the Gaussian sputtering training include: the input image resolution scaling factor is 1600px, the maximum number of training steps is 30000-300000, and the initial number of Gaussian points is 3000-3000000.

[0035] Preferably, the specific steps of point cloud cleaning and exporting are as follows:

[0036] S4.1, the system initializes the Gaussian point distribution according to the input image and camera parameters;

[0037] S4.2, the position, color, shape and transparency attributes of each Gaussian point are optimized by iteration, the point cloud structure is updated according to the loss function, and the geometric and visual performance of the model is enhanced;

[0038] S4.3, the number of points, PSNR peak signal-to-noise ratio and iteration progress parameters are displayed in real time during the training process, which are used to monitor the convergence of the model;

[0039] S4.4, after the training is completed, the Gaussian point data file in PLY format is exported, which can be directly used for rendering visualization or digital exhibition display;

[0040] S4.5, the shear box tool is used to frame the recovery area, and the points in the redundant area including the background, ground and non-antique structure are manually deleted, and the core antique area is reserved.

[0041] Preferably, it also includes an environment configuration step: using Anaconda to build a training environment, verifying the compatibility of CUDA version and PyTorch, installing COLMAP tool and configuring system environment variables.

[0042] The antique digital modeling and virtual reconstruction system based on three-dimensional Gaussian sputtering technology includes:

[0043] Image acquisition module: used for acquiring multi-angle image sequence of antique object;

[0044] Preprocessing module: used for preprocessing images;

[0045] Sparse point cloud construction module: used for extracting image features and constructing sparse point cloud and camera parameters;

[0046] Gaussian training module: used for generating Gaussian point cloud model based on sparse point cloud and camera parameters;

[0047] Post-processing module: used for cleaning and exporting renderable files of Gaussian point cloud model.

[0048] Preferably, the Gaussian training module comprises a parameter setting unit for configuring image resolution scaling factor, maximum training step number, initial Gaussian point number and density regulation threshold; the system further comprises a GPU real-time rendering module for supporting interactive display and VR / AR scene.

[0049] (III) Advantages

[0050] Compared with the prior art, the application provides a relic digital modeling and virtual reconstruction method and system based on three-dimensional Gaussian sputtering technology, which has the following advantages:

[0051] 1. The relic digital modeling and virtual reconstruction method and system based on three-dimensional Gaussian sputtering technology generates a Gaussian point cloud model through a Gaussian sputtering training step, expresses the geometry and texture of the relic with a Gaussian point set, each Gaussian point contains attributes such as position and color, does not require a triangular mesh, reduces data redundancy, improves information density, and solves the problem of low efficiency of irregular triangular mesh expression.

[0052] 2. In the Gaussian sputtering training, the relic digital modeling and virtual reconstruction method and system based on three-dimensional Gaussian sputtering technology optimizes the attributes such as position and color of the Gaussian points through iteration, updates the point cloud structure according to the loss function, automatically fills in the holes and repairs the misplacement, reduces the dependence on manual post-repair, improves the modeling efficiency and replicability, and solves the problem of model quality depending on manual repair.

[0053] 3. The Gaussian point cloud model of the relic digital modeling and virtual reconstruction method and system based on three-dimensional Gaussian sputtering technology directly stores color information, does not need to perform UV unfolding and texture mapping, avoids pixel waste caused by UV mapping, improves the clarity of details under limited resolution, and solves the problem of low efficiency of texture mapping UV unfolding.

[0054] 4. The relic digital modeling and virtual reconstruction method and system based on three-dimensional Gaussian sputtering technology, the GPU real-time rendering module of the system supports interactive display of the Gaussian point cloud, does not need to load texture, directly renders Gaussian point data, improves rendering speed, solves the problem of grid model rendering lag, and meets the needs of real-time interactive digital display and immersive experience. BRIEF DESCRIPTION OF DRAWINGS

[0055] Fig. 1 The relic digital modeling and virtual reconstruction method flowchart of the application;

[0056] Fig. 2 The relic digital modeling and virtual reconstruction system architecture diagram of the application. DETAILED DESCRIPTION

[0057] The technical solutions in the embodiments of the present application will be described clearly and completely below in combination with the embodiments of the present application and the accompanying drawings. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present application.

[0058] Referring to Figs. 1-2 The method for digital modeling and virtual reconstruction of cultural relics based on three-dimensional Gaussian sputtering technology comprises the following steps:

[0059] S1, image acquisition and preprocessing: acquiring a multi-angle image sequence of a cultural relic object and performing preprocessing;

[0060] S2, feature extraction and sparse point cloud construction: performing feature extraction on the preprocessed image, and constructing a sparse point cloud and camera parameters;

[0061] S3, Gaussian sputtering training: based on the sparse point cloud and camera parameters, the attributes of Gaussian points are iteratively optimized to generate a Gaussian point cloud model;

[0062] S4, point cloud cleaning and exporting: performing pruning processing on the Gaussian point cloud model, and exporting a renderable model file.

[0063] The system for digital modeling and virtual reconstruction of cultural relics based on three-dimensional Gaussian sputtering technology comprises:

[0064] An image acquisition module is configured to acquire a multi-angle image sequence of a cultural relic object;

[0065] A preprocessing module is configured to preprocess the image;

[0066] A sparse point cloud construction module is configured to extract image features and construct a sparse point cloud and camera parameters;

[0067] A Gaussian training module is configured to generate a Gaussian point cloud model based on the sparse point cloud and camera parameters;

[0068] A post-processing module is configured to clean the Gaussian point cloud model and export a renderable file.

[0069] Embodiment one:

[0070] I. Image acquisition and preprocessing step

[0071] Before Gaussian sputtering modeling, image acquisition and preprocessing work need to be completed to obtain high-quality image sequences as three-dimensional reconstruction basic data.

[0072] Image acquisition work uses a drone carrying a high-resolution camera (more than 20 million pixels) to take multiple-angle and full-enclosure shots of cultural relics objects; the flight path is divided into three strategies:

[0073] Long-distance circumnavigation: for overall structure acquisition;

[0074] Medium-distance circumnavigation: to enhance local detail collection;

[0075] Close-range pitch shooting: to supplement concave and top view information.

[0076] During flight, ensure that multiple facades (front, back, side, top) of the object are covered to avoid occluded dead angles; image overlap should be maintained between 70% and 85% to facilitate subsequent registration.

[0077] The camera should be set as follows:

[0078] Resolution: 3840 x 2160 pixels (4K);

[0079] Exposure time: dynamically adjusted according to the environment to avoid overexposure;

[0080] ISO: fixed between 100-200 to ensure low-noise images;

[0081] White balance: fixed as "sunny" or "cloudy" mode to ensure consistent light color.

[0082] If the collection method is video, use frame extraction tools to export images, with a frame rate of 5 frames per second. In a 60-second video, about 300 images can be extracted.

[0083] Output images use PNG format and are uniformly named, such as "0000.png" to "0300.png", for easy identification and processing.

[0084] After exporting the images, place them in the / input directory under the uniform path as the standard input for subsequent training programs.

[0085] II. Environment configuration and local deployment

[0086] To ensure stable operation of model training and data processing, an independent Gaussian sputtering training environment is required. The Anaconda management tool is recommended for environment isolation; the configuration steps are as follows:

[0087] Step 1: Run nvcc--version on the command line to verify the CUDA version, ensuring compatibility with the PyTorch version (use CUDA 11.7 and above);

[0088] Step 2: Download the project source code and dependency description file environment.yml;

[0089] Step three: Switch to the project root directory in the command line, and execute the commands in order:

[0090] SET DISTUTILS_USE_SDK=1

[0091] conda env create --file environment.yml

[0092] Step four: After creation, activate the environment command as conda activate gaussian_splatting;

[0093] Step five: Install the COLMAP three-dimensional reconstruction tool, and add its executable path to the system environment variable for subsequent point cloud reconstruction calls;

[0094] Step six: After the environment is deployed, it is recommended to restart the system to ensure that the dependent items are loaded correctly.

[0095] Three, image data extraction (COLMAP point cloud modeling)

[0096] After completing image preparation and environment configuration, process the image sequence through the COLMAP tool to generate sparse point cloud and camera extrinsic data, and the operation process is as follows:

[0097] Step one: Start image conversion and feature extraction using python convert.py -s <path_to_input_folder>;

[0098] Step two: COLMAP will automatically complete image de-distortion, feature point extraction and matching, and display the processing status in the terminal;

[0099] Step three: Each image will extract thousands of SIFT feature points, and output information including image focal length, principal point position, image size, etc.

[0100] Step four: After feature extraction, COLMAP enters the Structure from Motion (SfM) process to generate sparse point cloud and camera parameter file;

[0101] Step five: Data structure includes:

[0102] distorted / : original image and de-distorted image;

[0103] sparse / : sparse point cloud data (about 5-15 million points);

[0104] images / : camera pose and projection matrix;

[0105] stereo / : depth matching data.

[0106] IV. Model training

[0107] Based on the data generated by COLMAP, start the Gaussian point model training, use train.py to start the Gaussian point model construction and iterative optimization, the default training parameters are as follows:

[0108] Image resolution scaling factor -- resolution 4 (1 / 4 size of the original image);

[0109] Total number of iterations -- iterations 30000, the image quality can be appropriately increased to 100000 or more;

[0110] Point density adjustment threshold -- densify_grad_threshold 0.0002;

[0111] Encryption interval step number -- densification_interval 100;

[0112] The initial number of Gaussian points is automatically generated according to the sparse point cloud, and the training is gradually and dynamically encrypted.

[0113] V. Point cloud output and post-processing

[0114] After training, the system will automatically generate an output model file in the project directory:

[0115] The default file is stored in the / output path, which is a PLY format color Gaussian point cloud; the starting command format is:. / <SIBR directory> / bin / SIBR_gaussianViewer_app-m<training model path> where the -m parameter specifies the training model home directory; external point cloud editing tools can be used for visual inspection and clipping: delete non-antique areas (such as ground, background, ghosting); trim misaligned points, artifact points and transparent occlusion areas.

[0116] Through the above method, a high-density, high-fidelity Gaussian point cloud model with transparency and color performance is finally output, which can be directly applied to virtual exhibitions, museum displays, digital archive platforms and other cultural relic digitization scenarios.

[0117] Example two:

[0118] This example is aimed at the environmental construction scheme of Gaussian sputtering technology, to ensure stable and efficient training process.

[0119] First of all, the hardware and software version requirements are clear, and it is recommended to use NVIDIA RTX 3080Ti and above GPU, Intel i9-12900H and above CPU, 32GB memory hardware configuration.

[0120] In terms of software, CUDA11.8 needs to be compatible with PyTorch2.0.0, Visual Studio 2022 is used to compile CUDA extensions, Anaconda24.11.0 manages virtual environments, and COLMAP3.8+ version is used for three-dimensional reconstruction.

[0121] The environment building is implemented in steps, specifically:

[0122] Step 1: Verify the CUDA 11.8 version by entering "nvcc--version" in the command line;

[0123] Step 2: Switch to the project root directory, run the command "SET DISTUTILS_USE_SDK=1", then run the command "conda env create--file environment.yml" to create a virtual environment, and then use the command "conda activate gaussian_splatting" to activate the environment;

[0124] Step 3: Download COLMAP and add its executable path to the system environment variable, and restart the system after completion to ensure that the dependencies are loaded.

[0125] For version compatibility issues, if the "CUDA out of memory" error occurs, you can reduce the training image resolution or reduce the batch size; if the "environment.yml" installation fails, you need to manually install the dependencies, execute the "condainstall pytorch==2.0.0cuda-toolkit=11.8-c pytorch" and "pip install opencv-python colmap-wrapper" commands to solve the conflict.

[0126] Example 3:

[0127] This example is based on three core steps of image preparation, data conversion and model training, and describes the Gaussian sputtering reconstruction process.

[0128] I. Image sequence preparation and preprocessing

[0129] Create a " / input" folder under the working directory to store the image sequence. If using a video as the data source, import the video into Adobe Premiere Pro, perform splicing and speed adjustment, and export the images in PNG format at a rate of 5 frames per second (FPS) with a resolution of 3840x2160 to ensure quality.

[0130] Rename the files to consecutive numbers (e.g., 0000.png, 0001.png) when exporting to ensure that subsequent processing tools can recognize them. For example, a 60-second video can export 300 images with this parameter.

[0131] II. Data Conversion and Feature Extraction

[0132] After the image data extraction is complete, the next step is to use the command-line tool to perform data conversion. Run "python convert.py -s <path_to_input_folder>" to start COLMAP processing the image sequence and extract feature points from the images.

[0133] Detailed information about image processing, including the size of each image, focal length, and the number of recognized feature points, will be displayed in the terminal. For example, in the output information, you can see "Processed file 16 / 803" indicating that the 16th image has been processed, which includes the focal length, image resolution, and the number of feature points.

[0134] After feature extraction, COLMAP will continue to generate SFM (Structure from Motion) motion point cloud data. Since this process is performed by CPU, the calculation time will increase geometrically with the increase in the number of images.

[0135] After processing is complete, multiple folders including "distorted", "images", "input", "sparse", "stereo", etc. are generated, which store image de-distortion, sparse point cloud data, and stereo matching related data, respectively.

[0136] III. Gaussian Sputtering Model Training and Output

[0137] After image processing is complete, enter the training phase of the Gaussian sputtering model. Use the following command to start the training process: python train.py --resolution 4. In the "python train.py --resolution 4" command, --resolution 4 specifies that the image resolution during training is 1 / 4 of the original image, and other training parameters remain at their default settings.

[0138] During the training process, the model will be gradually optimized, and the training progress and loss function information will be displayed in real time on the terminal. After each iteration, the loss value of the training will be updated, and after 30,000 iterations, the training process will be completed.

[0139] For adjustable parameters of training, including:

[0140] --resolution;

[0141] --iterations;

[0142] --densify_grad_threshold;

[0143] --densification_interval.

[0144] Wherein, the --resolution parameter affects the resolution of image loading before training, and the values of 1, 2, 4 or 8 represent the original image and different resolution scales respectively;

[0145] --iterations controls the total number of steps of training, and 30000 iterations can usually obtain a good balance, neither too blurred image nor waste too much time;

[0146] --densify_grad_threshold determines whether to encrypt points based on position gradient, and the default value is 0.0002;

[0147] --densification_interval sets the frequency of point encryption, and the default is to encrypt once every 100 iterations.

[0148] The training process of the Gaussian sputtering model displays the training progress in real time through the command line, including the current number of points, PSNR value, training speed and other information. After training is completed, a.ply file containing Gaussian sputtering points will be generated in the Output folder under the working file directory.

[0149] In summary, the method and system for digital modeling and virtual reconstruction of cultural relics based on three-dimensional Gaussian sputtering technology generate a Gaussian point cloud model through the Gaussian sputtering training step, express the geometry and texture of cultural relics with Gaussian point sets, and each Gaussian point contains attributes such as position and color, without the need for triangular mesh, reducing data redundancy, improving information density, and solving the problem of low efficiency of irregular triangular mesh expression.

[0150] In addition, this method and system for digital modeling and virtual reconstruction of cultural relics based on three-dimensional Gaussian sputtering technology, during Gaussian sputtering training, iteratively optimizes the position, color and other attributes of Gaussian points, updates the point cloud structure according to the loss function, automatically fills in holes and repairs dislocations, reduces dependence on manual post-repair, improves modeling efficiency and reproducibility, and solves the problem of model quality relying on manual repair.

[0151] In addition, the digital modeling and virtual reconstruction method and system of cultural relics based on three-dimensional Gaussian sputtering technology directly stores color information in the Gaussian points of the Gaussian point cloud model, without the need for UV unfolding and texture mapping, avoiding pixel waste caused by UV mapping, improving detail clarity at limited resolution, and solving the problem of low efficiency of texture mapping UV unfolding.

[0152] In addition, the digital modeling and virtual reconstruction method and system of cultural relics based on three-dimensional Gaussian sputtering technology, the system's GPU real-time rendering module supports interactive display of Gaussian point clouds, without loading textures, directly rendering Gaussian point data, improving rendering speed, solving the problem of grid model rendering jams, meeting the needs of real-time interactive digital display and immersive experience, and solving the current cultural relic digital restoration technology. The grid generates irregular triangular meshes, the expression efficiency is low, the model quality depends on manual post-repair, the workload is large, the UV unfolding efficiency of texture mapping is low, the pixel utilization rate is poor, and the rendering efficiency is low.

[0153] The relevant modules involved in this system are all hardware system modules or functional modules that combine computer software programs or protocols with hardware in the existing technology. The computer software programs or protocols involved in the functional modules are themselves technologies that are well known to those skilled in the art and are not improvements to this system. The improvements to this system are the interaction or connection relationships between the modules, that is, improvements to the overall structure of the system to solve the corresponding technical problems to be solved by this system.

[0154] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for digital modeling and virtual reconstruction of cultural relics based on three-dimensional Gaussian sputtering technology, characterized in that: The following steps are involved: S1. Image acquisition and preprocessing: Acquire multi-angle image sequences of cultural relics and perform preprocessing; S2, feature extraction and sparse point cloud construction: extract features from the preprocessed image and construct sparse point cloud and camera parameters; S3, Gaussian sputtering training: Based on the sparse point cloud and camera parameters, the properties of the Gaussian points are iteratively optimized to generate a Gaussian point cloud model; S4. Point cloud cleaning and exporting: trimming the Gaussian point cloud model and exporting a renderable model file.

2. The method for digital modeling and virtual reconstruction of cultural relics based on three-dimensional Gaussian sputtering technology according to claim 1 is characterized in that: In the image acquisition and preprocessing steps, a drone equipped with a camera with a resolution of 20 megapixels or more is used to perform multi-angle and full-surround photography. The flight path is divided into three layers: Long distance circle: used to obtain the overall structure; Mid-range surround: Enhances the capture of local details; Close-up pitch shooting: Supplement concave and top perspective information; The camera parameters are set to a resolution of 3840×2160 pixels, a fixed ISO of 100-200, and a fixed white balance of "sunny" or "cloudy" mode.

3. The method for digital modeling and virtual reconstruction of cultural relics based on three-dimensional Gaussian sputtering technology according to claim 1 is characterized in that: In the feature extraction and sparse point cloud construction steps, the image pre-processed in step S1 is imported into the RealityCapture software for 3D modeling initialization. The operation process is as follows: S2.

1. Select the "Add Image" function in RealityCapture to import all PNG images; S2.

2. Click "Image Alignment" and the software will automatically extract image feature points and build a camera pose relationship network. S2.

3. Set the "Maximum number of feature points" in the alignment parameters to 40,000 and the "Image matching limit" to nearest neighbor matching within 30 images. S2.

4. After processing, the system outputs a sparse point cloud with a point count between 50,000 and 120,000, and automatically generates camera intrinsic parameters (focal length, principal point position) and extrinsic parameters (rotation and translation matrices) for each image. S2.

5. After processing, export two key data items: sparse point cloud data in Polygon File Format, containing 3D coordinates and color information; and camera parameters generated using the Reality Capture batch export tool to Excel or CSV format files, recording the core parameters of each image, including focal length, image size, and pose matrix. S2.

6. Manually retake or exclude the parts where image alignment fails to ensure that the sparse point cloud is complete and without gaps.

4. The method for digital modeling and virtual reconstruction of cultural relics based on three-dimensional Gaussian sputtering technology according to claim 1 is characterized in that: In the Gaussian sputtering training step, first start Postshot and import three key data: Sparse point cloud file in .ply format; Camera parameter table in CameraParameters.xlsx (or .csv) format; The original image sequence involved in training; The parameter settings of the Gaussian sputtering training include: the input image resolution scaling factor is 1600px, the maximum number of training steps is 30,000 to 300,000, and the initial number of Gaussian points is 3,000 to 3,000,000.

5. The method for digital modeling and virtual reconstruction of cultural relics based on three-dimensional Gaussian sputtering technology according to claim 1 is characterized in that: The specific steps for point cloud cleaning and exporting are as follows: S4.1, the system initializes the Gaussian point distribution according to the input image and camera parameters; S4.

2. Iteratively optimize the position, color, shape, and transparency attributes of each Gaussian point, update the point cloud structure according to the loss function, and enhance the geometric and visual performance of the model. S4.

3. During the training process, the number of points, PSNR peak signal-to-noise ratio, and iteration progress parameters are displayed in real time to monitor the model convergence; S4.

4. After the training is completed, the Gaussian point data file is exported in PLY format, which can be directly used for rendering visualization or digital exhibition display; S4.

5. Use the cut box tool to select the restoration area and manually delete unnecessary area points, including the background, ground, and non-cultural relic structures, to retain the core cultural relic area.

6. The method for digital modeling and virtual reconstruction of cultural relics based on three-dimensional Gaussian sputtering technology according to claim 1 is characterized in that: It also includes environment configuration steps: building the training environment using Anaconda, verifying that the CUDA version is compatible with PyTorch, installing the COLMAP tool, and configuring system environment variables.

7. The cultural relics digital modeling and virtual reconstruction system based on three-dimensional Gaussian sputtering technology is characterized by: include: Image acquisition module: used to obtain multi-angle image sequences of cultural relics; Preprocessing module: used to preprocess images; Sparse point cloud construction module: used to extract image features and construct sparse point clouds and camera parameters; Gaussian training module: used to generate Gaussian point cloud model based on sparse point cloud and camera parameters; Post-processing module: used to clean up the Gaussian point cloud model and export renderable files.

8. The method for digital modeling and virtual reconstruction of cultural relics based on three-dimensional Gaussian sputtering technology according to claim 7 is characterized in that: The Gaussian training module includes a parameter setting unit for configuring the image resolution scaling factor, the maximum number of training steps, the initial number of Gaussian points, and the density control threshold; the system also includes a GPU real-time rendering module for supporting interactive display and VR / AR scenarios.

Images