A 3DGS cultural relic digital reconstruction method and system based on a blockchain

Abstract

This invention discloses a blockchain-based 3DGS digital reconstruction method and system for cultural relics. It collects RGB image data and depth sensing data of the cultural relics, eliminates the influence of different shooting conditions through illumination separation processing technology, and establishes a standardized image dataset. Based on image analysis, it identifies suitable surface areas for reconstruction, uses kernel density estimation to determine the distribution of feature points, and generates an initial 3D representation through Gaussian ellipsoid fitting. It performs gradient calculation and feature extraction on the depth data, combining it with the Gaussian representation to form a geometric constraint mechanism. For sparsely represented regions, it performs adaptive encryption based on visual importance, and achieves layered rendering effects through opacity parameter adjustment. It analyzes the rendering characteristics and transformation relationships of different perspectives, determines key observation points through stability analysis, and constructs a smooth multi-view display sequence. It integrates multi-view rendering information to generate a volumetric representation, and completes the ownership verification of the high-quality 3D digital model through digital signature.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and cultural relic conservation technology, and in particular to a blockchain-based 3DGS digital reconstruction method and system for cultural relics. Background Technology

[0002] As an important carrier of human civilization, the digital preservation and display of cultural relics has become a crucial development direction in the field of cultural heritage. Traditional methods for digitizing cultural relics mainly rely on laser scanning or photogrammetry. While laser scanning offers high precision, the equipment is expensive and inefficient. Photogrammetry is low-cost but lacks precision in reconstructing complex surfaces. Both methods fall short in terms of restoring surface details, expressing material properties, and providing interactive displays, making it difficult to meet the needs of modern cultural heritage applications.

[0003] In recent years, 3D Gaussian scattering (3DGS), as an emerging 3D representation method, has demonstrated advantages in fields such as virtual reality by achieving high-quality scene rendering through differentiable Gaussian ellipsoids. However, existing 3DGS technology is mainly geared towards natural scene design and has significant limitations in the application of cultural relic digitization: the quality of multi-view data acquisition in the complex environment of museums is unstable, affecting the reconstruction effect; existing methods lack the ability to differentiate the importance of different areas on the surface of cultural relics (such as inscriptions, patterns, and flat areas); the smoothness and stability of perspective transitions during rendering need to be improved; and the generated digital results lack effective technical protection measures, posing intellectual property risks. These problems limit the in-depth application of 3DGS technology in the field of cultural relic preservation. Therefore, a method is urgently needed to solve at least one of the above problems. Summary of the Invention

[0004] This invention discloses a blockchain-based 3DGS digital reconstruction method and system for cultural relics. It aims to obtain standardized input data through multi-angle data acquisition and illumination separation processing, construct an initial Gaussian representation through Gaussian kernel density estimation and ellipsoidal fitting, achieve geometric constraints through gradient analysis and feature coupling of depth information, and optimize the representation quality through perception-driven adaptive subdivision and multi-level rendering strategies. Ultimately, it forms a high-fidelity, multi-view, and interactive 3D model of the cultural relic and achieves blockchain-based ownership verification, providing a complete and high-quality digital solution for applications such as digital preservation, virtual display, and cultural dissemination of cultural relics.

[0005] The first aspect of this invention proposes a blockchain-based 3DGS digital reconstruction method for cultural relics, comprising the following steps:

[0006] Collect RGB image data and depth perception data of cultural relics, and perform illumination separation processing on the RGB image data to form a standardized image library;

[0007] Based on the standardized image library, Gaussian kernel density estimation and ellipsoid fitting are performed to generate an initial Gaussian point set. The initial Gaussian point set is then intelligently density-allocated to determine the Gaussian topological domain.

[0008] Gradient parsing and geometric constraint decomposition are performed on the depth sensing data to obtain structural guidance information. The structural guidance information is coupled with the Gaussian topological domain to generate a depth feature field. A geometric correction space is constructed based on the depth feature field.

[0009] Density analysis is performed on the Gaussian topological domain to identify sparse distribution regions. A perception-driven subdivision operation is performed on the sparse regions to generate a dense Gaussian field. Opacity allocation processing is performed on the dense Gaussian field to form a transparency control sequence. A multi-level rendering strategy is constructed based on the fusion of the transparency control sequence and the geometric correction space.

[0010] The multi-level rendering strategy is projected onto the view sensitivity space to generate key view intervals. The importance of the key view intervals is evaluated to identify view inflection point positions. The stability of the view inflection point positions is tested to generate view anchor points. A multi-view rendering sequence is constructed based on the view anchor points.

[0011] A volumetric rendering matrix is ​​generated based on the multi-view rendering sequence. The volumetric rendering matrix is ​​then optimized to generate a 3D model of the cultural relic. Finally, the 3D model of the cultural relic is hash-verified and digitally signed to complete the ownership confirmation.

[0012] A second aspect of this invention proposes a blockchain-based 3DGS digital reconstruction system for cultural relics, comprising:

[0013] The data processing module is used to collect RGB image data and depth perception data of cultural relics, and to perform illumination separation processing on the RGB image data to form a standardized image library;

[0014] The intelligent modeling module is used to generate an initial Gaussian point set by performing Gaussian kernel density estimation and ellipsoid fitting based on the standardized image library, and to determine the Gaussian topological domain by performing intelligent density allocation on the initial Gaussian point set.

[0015] The feature fusion module is used to perform gradient parsing and geometric constraint decomposition on the depth sensing data to obtain structural guidance information, and to perform feature coupling between the structural guidance information and the Gaussian topological domain to generate a depth feature field, and to construct a geometric correction space based on the depth feature field.

[0016] The perceptual optimization module is used to perform density analysis on the Gaussian topological domain to identify sparse distribution regions, perform perceptual-driven subdivision operation on the sparse regions to generate a dense Gaussian field, perform opacity allocation processing on the dense Gaussian field to form an opacity control sequence, and construct a multi-level rendering strategy based on the fusion of the opacity control sequence and the geometric correction space.

[0017] The viewpoint anchoring module is used to project the multi-level rendering strategy into the viewpoint sensitivity space to generate key viewpoint intervals, evaluate the importance of the key viewpoint intervals to identify viewpoint inflection point positions, perform stability tests on the viewpoint inflection point positions to generate viewpoint anchor points, and construct a multi-view rendering sequence based on the viewpoint anchor points.

[0018] The quality reconstruction module is used to generate a volume rendering matrix based on the multi-view rendering sequence, optimize the quality of the volume rendering matrix to generate a 3D model of the cultural relic, and perform hash verification and digital signature on the 3D model of the cultural relic to complete the confirmation of ownership.

[0019] The beneficial effects of this invention are reflected in the following points: First, by combining RGB image data acquisition with depth perception data and processing it using illumination separation technology to form a standardized image library, the inconsistency in texture caused by changes in illumination conditions in traditional methods is eliminated, ensuring the reliability of the input data. Simultaneously, based on the standardized image library, Gaussian kernel density estimation automatically determines the seed region of the point cloud, avoiding the subjectivity and inefficiency of manual settings, making the generation of the initial Gaussian point set more scientific and reasonable. Second, by extracting structural guidance information through depth gradient analysis and coupling it with the Gaussian topological domain to establish a geometric correction space, a deep fusion of depth geometric constraints and Gaussian representation is achieved, effectively solving the problem of insufficient geometric accuracy in traditional 3DGS methods. For the perception-driven subdivision mechanism targeting sparsely distributed regions, the Gaussian density is adaptively adjusted according to visual importance, ensuring accurate representation of important details while avoiding waste of computational resources. Finally, through perspective sensitivity spatial projection analysis, key perspective intervals and perspective inflection points are automatically identified, and perspective anchor points generated through stability testing ensure the smoothness of perspective switching. Based on the volume rendering matrix constructed from the multi-view rendering sequence, the effective fusion of multi-view information is achieved through pixel correlation analysis and quality propagation chain mechanism. Combined with blockchain technology, this not only ensures the high fidelity of the final 3D model, but also provides a complete copyright confirmation and intellectual property protection mechanism.

[0020] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating a blockchain-based 3DGS digital reconstruction method for cultural relics according to the present invention.

[0022] Figure 2 This is a structural block diagram of a blockchain-based 3DGS digital reconstruction system for cultural relics according to the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0024] It should be noted that all directional indicators (such as up, down, left, right, front, back, etc.) in the embodiments of this application are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicator will also change accordingly.

[0025] It should also be noted that when a component is described as "fixed to" or "set on" another component, it can be directly on the other component or there may be an intervening component present. When a component is described as "connected to" another component, it can be directly connected to the other component or there may be an intervening component present.

[0026] Furthermore, the use of terms such as "first" and "second" in this application is for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. Additionally, the technical solutions of the various embodiments can be combined with each other, but only on the basis of being achievable by those skilled in the art. When the combination of technical solutions is contradictory or impossible to implement, such a combination of technical solutions should be considered non-existent and not within the scope of protection claimed in this application.

[0027] The technical solutions of the embodiments of this application will be described below.

[0028] like Figure 1 As shown, this embodiment of the invention provides a blockchain-based 3DGS digital reconstruction method for cultural relics, including the following steps S110-S160:

[0029] Step S110: Collect RGB image data and depth perception data of cultural relics, and perform illumination separation processing on the RGB image data to form a standardized image library.

[0030] Specifically, RGB image data and depth sensing data of the cultural relic were collected. A circular camera array containing 24 high-resolution RGB cameras was deployed around the relic, evenly distributed on a horizontal circle with adjacent cameras spaced 15 degrees apart to ensure complete coverage of the relic's surface. Each RGB camera used a 24-megapixel CMOS sensor, equipped with a fixed-focus lens with a focal length of 50mm and an aperture of f / 8 to ensure sufficient depth of field. The camera array was controlled by a synchronous triggering system to ensure that all cameras were exposed at the same time, eliminating errors caused by minor movements of the relic. Three layers of shooting height were set in the vertical direction, corresponding to the upper, middle, and lower parts of the relic, respectively. Each layer independently completed a 360-degree panoramic shot, acquiring a total of 72 RGB images from different perspectives. Depth sensing data was collected using a structured light scanner. The scanner projected a coded stripe pattern onto the surface of the relic, and the depth value of each surface point was calculated using the principle of triangulation. The scanning resolution was set to 0.1 mm to ensure the capture of fine geometric features of the relic's surface. The depth data and RGB images were precisely registered in space, with each RGB pixel corresponding to a depth value, forming RGBD four-channel data. The collection process was carried out in a standard environment with constant temperature and humidity to avoid the impact of environmental changes on the cultural relics and equipment.

[0031] A standardized image library is formed by performing illumination separation processing on RGB image data. The illumination conditions of each RGB image are analyzed, and the image is decomposed into two components: a reflectance image and an illumination image, following the image formation model I(x,y)=R(x,y)·L(x,y), where I is the observed image intensity, R is the surface reflectance, and L is the illumination intensity. An optimization-based intrinsic image decomposition algorithm is employed to effectively separate the reflectance and illumination components through optimized calculations. The algorithm simultaneously considers data fidelity and image smoothness constraints to ensure the accuracy and visual continuity of the decomposition results. Color correction is performed on the decomposed reflectance images using a standard color chart as a reference. The color transformation matrix is ​​calculated using the least squares method to eliminate the influence of different light source color temperatures. The reflectance images from all viewpoints are normalized to a uniform brightness range, and histogram matching technology is used to select the intermediate viewpoint image as a reference template. A multi-scale image pyramid is constructed, with each scale retaining a different level of detail information. The pyramid has 5 layers, and the resolution difference between adjacent layers is a factor of 2. The images for each pyramid level were quality-assessed, with metrics such as sharpness, contrast, and signal-to-noise ratio calculated. Images that did not meet quality standards were removed. The resulting standardized image library contained multi-angle visual information of the artifacts under uniform lighting conditions, eliminating interference from lighting variations in the original photographs.

[0032] Step S120: Based on the standardized image library, Gaussian kernel density estimation and ellipsoid fitting are performed to generate an initial Gaussian point set. The initial Gaussian point set is then intelligently density-allocated to determine the Gaussian topological domain.

[0033] Specifically, an initial Gaussian point set is generated by Gaussian kernel density estimation and ellipsoid fitting based on a standardized image library. Feature points from multiple perspectives are extracted from the standardized image library, and a scale-invariant feature transformation algorithm is used to detect key points in each image. Each key point contains location coordinates, scale, orientation, and high-dimensional descriptor information. Feature points are matched between images from different perspectives using a nearest neighbor distance ratio test method, comparing the distance ratio between the nearest and second nearest neighbors to determine the reliability of the matching. A sparse 3D point cloud is constructed based on reliable feature matching. Triangulation is used to recover the 3D spatial coordinates of the 2D matching points from multiple perspectives. The accuracy of the triangulation results is verified using depth-sensing data to ensure the geometric consistency of the reconstructed point cloud. The local geometric properties of the 3D point cloud are analyzed, calculating the normal vector direction, principal curvature value, and neighborhood point density of each point. Based on these properties, point clusters with similar geometric features are identified. Point clusters are grouped and merged according to surface continuity and geometric similarity to form different surface segments, each segment representing a semantic unit of the artifact surface. The reconstructability of each surface fragment is evaluated using a scoring function S = w1·G + w2·T + w3·(1-Q), where S is the reconstructability score, G is the geometric complexity index, T is the texture richness index, Q is the occlusion rate, and w1, w2, and w3 are the corresponding weighting coefficients. Surface fragments with scores exceeding a set threshold are selected as reconstructable surface patterns, which contain the main geometric structures and texture features of the artifact surface. Three-dimensional points in the reconstructable surface patterns are used as sample points for density estimation. A Gaussian kernel function is placed at each sample point to perform density diffusion. The kernel function adopts a multidimensional Gaussian distribution to calculate the density contribution at any location in three-dimensional space. The bandwidth parameter for kernel density estimation is determined using an adaptive strategy. The estimation accuracy is adjusted according to the reconstructability score S of the surface fragment; a smaller bandwidth is used in high-score areas to preserve geometric details, while a larger bandwidth is used in low-score areas to ensure the continuity of the density field. The kernel density value ρ(x) = (1 / n)Σ at each location in the computation space is calculated. i K(xx i ), where ρ(x) is the density value at position x, n is the total number of sample points, K is the Gaussian kernel function, and x i Let be the location of the i-th sample point. A density threshold is set to divide the space into high-density and low-density regions. Within the high-density region, a mean-shift algorithm is used to find local density maxima as seed points. These seed points are then used to define the influence range, forming independent point cloud seed regions. For each point cloud seed region, the centroid positions of all 3D points within the region are calculated. Based on these centroids, a covariance matrix Σ=(1 / m)Σ is constructed. j (p j -μ)(p j -μ) T Where m is the total number of points in the region, p j Let be the three-dimensional coordinates of the j-th point, and μ be the average position of all points, with superscript .T The transpose operation is used to perform eigenvalue decomposition on the covariance matrix, yielding three eigenvalues ​​and corresponding eigenvectors. Based on the eigenvalue decomposition results, 3DGS ellipsoid parameters are constructed. The ellipsoid center is set as the centroid, the principal axis direction is determined by the eigenvectors, and the semi-axis length is proportional to the square root of the eigenvalue. Spherical harmonic coefficients and opacity parameters are initialized for each ellipsoid. Through the system's ellipsoid fitting process, each point cloud seed region is converted into a parameterized 3DGS ellipsoid representation. All ellipsoids together constitute the initial Gaussian point set on the artifact's surface.

[0034] In some embodiments, the step of intelligently distributing the initial Gaussian point set to determine the Gaussian topological domain includes: constructing an ellipsoidal coverage analysis graph based on the initial Gaussian point set; locating sparsely covered nodes in the ellipsoidal coverage analysis graph; using the sparsely covered nodes to induce Gaussian compensation growth to generate compensated Gaussian points; and forming a Gaussian topological domain based on the compensated Gaussian points.

[0035] An ellipsoidal coverage analysis map is constructed based on an initial Gaussian point set. Each ellipsoidal element in the initial Gaussian point set is traversed, and the 3D spatial coverage is precisely calculated based on its spatial location, rotation orientation, and scaling parameters. The analysis space is discretized into a regular voxel grid, with the grid resolution adaptively set according to the artifact size and detail requirements, ensuring both surface detail capture and minimal computational burden. The coverage contribution of all ellipsoids is evaluated for each voxel location, and the influence intensity of the ellipsoid on the voxel center is calculated using a Gaussian decay function, with the influence decreasing further away from the ellipsoid center. The total coverage intensity is obtained by summing the contributions of all ellipsoids to each voxel, and the comprehensive coverage effect of each ellipsoid is calculated using a Gaussian distribution overlay. The coverage intensity values ​​are mapped to an intuitive color representation using a heatmap color scheme: dark red represents fully covered areas, blue represents under-covered areas, and intermediate shades represent transitional states. The coverage analysis map is rendered in 3D space, using volumetric rendering technology to display the spatial distribution of coverage intensity, supporting interactive viewing and analysis. The coverage analysis map allows for intuitive identification of the coverage quality of the initial Gaussian point set, revealing potential coverage holes and overlapping areas.

[0036] Locate sparse coverage nodes in the ellipsoidal coverage analysis map. Scan all voxel units in the coverage analysis map to identify voxels with coverage intensity below a preset minimum threshold; these low-coverage voxels correspond to locations where the artifact surface reconstruction is insufficient. Perform 3D connectivity analysis on the identified low-coverage voxels, clustering spatially adjacent voxels into connected regions, each forming an independent coverage cavity. Calculate the geometric characteristic parameters of each coverage cavity, including cavity volume, surface area, maximum inscribed sphere radius, and shape compactness, to comprehensively evaluate the cavity's scale and morphological characteristics. Intelligently select sparse node locations within each coverage cavity, taking into account factors such as the distance to the nearest ellipsoid, cavity centrality, and surface visibility. Verify the effectiveness of the sparse nodes through back projection, projecting the node coordinates onto the original multi-view image to confirm that the location corresponds to the actual artifact surface rather than the background or occluded area. Calculate the filling priority for each verified sparse node, determining the processing order by comprehensively considering cavity volume, distance, and visibility indicators. Establish a priority queue for sparse nodes; high-priority nodes will undergo Gaussian compensation processing first to ensure coverage quality in critical areas.

[0037] Gaussian compensated ellipsoids are generated by stimulating Gaussian growth using sparse nodes. Starting with the highest priority sparse node, a new compensated Gaussian ellipsoid is initialized at the node location. The initial parameters are determined by analyzing the statistical characteristics of surrounding existing ellipsoids. The initial scale of the new ellipsoid is set to an adaptive multiple of the average scale of the neighboring ellipsoids to avoid coverage problems caused by excessively large or small scales. For grooves or fine texture areas on the artifact surface, the initial ellipsoid may not completely cover them. The compensated growth algorithm automatically adjusts the shape and position of the ellipsoid to ensure that these detailed features are accurately represented. An objective function for growth optimization is defined, comprehensively considering coverage improvement, overlap penalty, and smoothing constraints. The coverage term encourages filling holes, the overlap penalty term prevents excessive overlap, and the smoothing term maintains the continuous transition between adjacent ellipsoids. The objective function is optimized using gradient descent, calculating the gradient of each term with respect to the ellipsoid parameters, updating the parameters along the negative gradient direction, and the step size is determined by line search. Multiple growth termination conditions are set, including coverage improvement increment falling below a threshold, reaching the maximum number of iterations, and parameter convergence. Growth of the current ellipsoid stops when any of these conditions are met. Post-processing optimization is performed on all generated compensated Gaussian points to identify and merge highly similar ellipsoids, and remove redundant ellipsoids that contribute little, ensuring the simplicity of the Gaussian representation.

[0038] A Gaussian topological domain is formed based on compensated Gaussian points. The initial set of Gaussian points is integrated with the newly generated compensated Gaussian points to form a complete and unified set of Gaussian points, containing all the ellipsoidal elements necessary to fully cover the artifact's surface. The spatial relationships between the integrated Gaussian points are analyzed, calculating the Euclidean distance between any two ellipsoid centers and the minimum distance between ellipsoidal boundaries, establishing adjacency relationships between ellipsoids based on distance thresholds. Ellipsoidal parameters are extracted from the Gaussian topological domain, systematically analyzing the geometric features, spatial location, and neighborhood relationships of each ellipsoid, forming a Gaussian constraint template containing complete information such as ellipsoid center coordinates, principal axis directions, scaling parameters, and spherical harmonic coefficients. This Gaussian constraint template provides a standardized parameter framework for subsequent geometric constraints and optimization processes. Graph theory methods are used to analyze the adjacency graph, identifying connected components, each corresponding to an independent surface region of the artifact. Key characteristic parameters of the topological domain are calculated, including ellipsoidal density distribution and average connectivity, comprehensively describing the organizational characteristics of the Gaussian representation within the domain. The construction of the Gaussian topological domain completes the transformation from discrete ellipsoids to structured organization.

[0039] Step S130: Gradient analysis and geometric constraint decomposition are performed on the depth sensing data to obtain structural guidance information. The structural guidance information is coupled with the Gaussian topological domain to generate a depth feature field. A geometric correction space is constructed based on the depth feature field.

[0040] Specifically, gradient analysis and geometric constraint decomposition are performed on the depth-sensing data to obtain structural guidance information. Depth image sequences are extracted from the depth data; each depth image records the distance information from the artifact surface to the sensor from a specific viewpoint. The depth images are preprocessed using a bilateral filter to remove measurement noise while maintaining edge sharpness; the filter applies Gaussian weights to both the spatial and depth domains. The first-order gradient of the depth image is then calculated. ,in The depth gradient vector and The gradient vectors, representing the partial derivatives of depth D in the x and y directions respectively, reflect the local tilt of the surface. Further calculation of the second derivative of depth detects geometric feature inflection points; this operator is sensitive to surface unevenness. Gradient information is analyzed at multiple scales to construct a depth gradient pyramid. Each scale captures geometric changes of different granularities; the coarse scale reflects the overall shape, while the fine scale reveals surface details. Statistical analysis of gradient magnitude and direction generates a gradient direction histogram, identifying dominant surface orientations and geometric texture patterns. Multi-scale, multi-order gradient information is integrated to form a depth feature spectrum, which comprehensively encodes the geometric change characteristics of the artifact's surface, including key information such as flat areas, edge contours, and curvature changes. The gradient distribution pattern in the depth feature spectrum is analyzed, and principal component analysis is used to extract the main geometric change directions, which correspond to the artifact's structural axes. Based on the distribution of gradient magnitude, the surface region is classified into planar areas, curved areas, and feature areas. The gradient in planar areas is close to zero, the gradient in curved areas changes continuously, and the gradient in feature areas exhibits abrupt changes. Within the feature area, geometric primitives are further identified, including edge lines, corners, ridges, and valleys. These primitives constitute the skeletal structure of the artifact. For example, in bronze artifacts, the strokes of inscriptions form clear ridges and valleys, the outline of the object forms continuous edge lines, and the turning points of decorative patterns produce obvious corners. These geometric primitives collectively constitute the structural skeleton of the artifact. The identified geometric primitives are parametrically represented: edge lines are fitted with polynomial curves, corners are recorded with their positions and the angles between adjacent edges, and ridges and valleys are described using geodesics. Topological relationships between geometric primitives are established, recording the connection methods, intersection patterns, and hierarchical structures to form a structural relationship diagram. Constraint rules are extracted from the structural relationship diagram, including coplanar constraints, perpendicular constraints, and symmetry constraints. These rules reflect the design intent and manufacturing process of the artifact. The geometric primitives and constraint rules are integrated to form structural guidance information, guiding the subsequent shape optimization process.

[0041] In some embodiments, the step of generating a depth feature field by feature coupling the structural guidance information with the Gaussian topological domain includes: performing depth gradient analysis on the structural guidance information to form a geometric constraint template; extracting ellipsoidal parameters from the Gaussian topological domain to form a Gaussian constraint template; performing interactive mapping between the geometric constraint template and the Gaussian constraint template to generate a constraint fusion template; and performing depth guidance reconstruction based on the constraint fusion template to determine the depth feature field.

[0042] A geometric constraint template is generated by performing deep gradient analysis on the structural guidance information. Identified geometric primitives, including key features such as edge segments, corners, ridges, and valleys, are extracted from the structural guidance information. Each primitive carries local geometric information about the artifact's surface, recording its spatial location and type attributes. The distribution pattern of the depth gradient is analyzed within each primitive's location and its neighborhood. Depth values ​​are sampled using local windows, with the window size adaptively adjusted according to the primitive type. Slender windows are used along the edge direction for edges, while square windows cover multiple directions for corners. The gradient feature vector of the geometric primitive is calculated, containing statistics on gradient magnitude and direction information. Analysis of these features determines whether the primitive's geometric properties are sharp or smooth. Geometric primitives are finely classified based on gradient features: sharp edges exhibit high gradient magnitude and consistent gradient direction, smooth transition edges show a gradual change in gradient direction, and gradient directions diverge at corners. A dedicated constraint strength function is designed for each type of geometric primitive: a fast decay function is used for sharp edges, and a slow decay function is used for smooth features, ensuring that the constraint influence range matches the primitive characteristics. A radial basis function network is used to spatially interpolate the constraints of discrete primitives. The basis function centers are set at the primitive locations, and a continuous constraint field is generated through interpolation. The constraint functions of all geometric primitives are then weighted and superimposed to form a unified constraint field C(x) = Σ. i W i ·S i (||xx i ||), where C(x) is the constraint strength at spatial location x, W i S represents the importance weight of the i-th primitive. i Let ||xx| be the constraint strength function of this primitive. i || represents the distance from position x to primitive position x i The distance is then considered. The generated constraint field is normalized and smoothed to eliminate abrupt changes between different primitives, forming a continuously varying geometric constraint template.

[0043] Ellipsoidal parameters are extracted from the Gaussian topological domain to form a Gaussian constraint template. The system traverses all ellipsoidal elements in the Gaussian topological domain, extracting complete parameter information from each ellipsoid to construct the basic data for the Gaussian constraint template. The complete parameter vector for each ellipsoid is extracted, including the three-dimensional center coordinates, quaternions describing rotation, scaling factors for the three principal axes, opacity, and spherical harmonic coefficient vectors. Spatial analysis is performed on the extracted ellipsoidal parameters to calculate the parameter differences between adjacent ellipsoids. This difference information is recorded in the Gaussian constraint template, identifying regions of smooth parameter changes and abrupt boundaries. Constraint influence weights are defined based on the geometric properties of the ellipsoids, and these weights are integrated into the constraint network of the Gaussian constraint template. Large-scale ellipsoids receive stronger global constraints, and high-opacity ellipsoids receive higher weights. Adjacency constraint relationships are established between ellipsoids, and the network structure of the Gaussian constraint template is continuously updated. When the influence ranges of two ellipsoids overlap, constraint connections are created to ensure a smooth transition of parameters between adjacent ellipsoids. The constraint information is organized through the network structure and optimized to form a globally consistent Gaussian constraint template.

[0044] A constraint fusion template is generated by interactively mapping geometric constraint templates and Gaussian constraint templates. A bidirectional mapping mechanism is designed to facilitate information exchange between the two constraint templates, establishing a correspondence between geometric spatial coordinates and Gaussian parameter space to ensure correct transmission and effect of constraints. The compatibility of the two constraints at each spatial location is evaluated. When the direction of the geometric constraint is consistent with the direction of the Gaussian principal axis, the constraints are considered compatible and can mutually reinforce each other. When there is a large deviation, coordination is required to find a balance point. The two constraints are fused through weighted combination and interaction terms. The weights of the linear combination are determined based on data reliability, and the coefficients of the interaction terms control the strength of the synergistic effect of the two constraints. The local consistency of the fused constraints is calculated. The degree of agreement between the two constraints is evaluated by comparing the constraint gradient directions. Regions with high consistency show good fusion results, while conflicting regions require special handling. Conflict detection is performed on the fusion results. Minor conflicts are resolved through weighted averaging, while severe conflicts are prioritized based on physical rationality and reconstruction goals. The generated constraint fusion template integrates deep geometric information and Gaussian representation characteristics, providing a unified constraint framework for subsequent feature field construction.

[0045] Depth feature field determination is achieved through depth-guided reconstruction based on a constraint fusion template. The comprehensive constraint strength in the constraint fusion template is converted into a feature representation. Strongly constrained regions correspond to low-feature attraction ellipsoid clustering, while weakly constrained regions have smoother features, allowing for free ellipsoid configuration. A feature function is constructed, where the spatial location and ellipsoid parameter vector jointly determine the function value. The function value represents the energy required to configure an ellipsoid with specific parameters at that location. The distribution of this feature function in 3D space forms a continuous depth feature field, and ellipsoid configuration is optimized by minimizing the total feature. The depth feature field is discretized using numerical methods, dividing the 3D space into grid cells. Interpolation is performed within each grid cell to approximate the feature distribution. The partial derivatives of the depth feature field with respect to the ellipsoid parameters are calculated. These derivatives constitute the gradient information for optimization, where the position derivative indicates the direction the ellipsoid should move, the rotation derivative guides attitude adjustment, and the scale derivative controls shape changes. Considering the mutual influence between multiple ellipsoids, a change in one ellipsoid parameter will affect surrounding ellipsoids through the depth feature field. A joint optimization framework needs to be established to simultaneously update the parameters of all related ellipsoids.

[0046] A geometric correction space is constructed based on the depth feature field. The spatial distribution characteristics of the depth feature field are analyzed to identify local minima, which correspond to geometrically stable ellipsoidal configurations. The gradient vector field of the feature field is calculated, with the gradient pointing in the direction of fastest feature growth and the negative gradient pointing in the optimization direction of feature reduction. A correction vector field is constructed along the negative gradient direction, with each spatial location associated with a correction vector indicating how the ellipsoid at that location should move to reduce the overall feature. The correction vector field is regularized to limit the maximum magnitude of a single correction, preventing overly drastic parameter updates that could lead to system instability. A hierarchical correction strategy is established: large-scale correction adjusts the overall configuration, medium-scale correction optimizes local regions, and small-scale correction handles detailed features. Adaptive correction weights are designed, increasing the correction intensity in regions with clear geometric features to ensure accuracy and decreasing the correction intensity in ambiguous regions to maintain robustness. All correction information is integrated to construct a complete geometric correction space, which provides optimized target positions and shape parameters for each ellipsoid, guiding the evolution of the Gaussian representation towards a more accurate geometric form.

[0047] Step S140: Perform density analysis on the Gaussian topological domain to identify sparse distribution regions, perform perception-driven subdivision operation on the sparse distribution regions to generate dense Gaussian fields, perform opacity allocation processing on the dense Gaussian fields to form a transparency control sequence, and construct a multi-level rendering strategy based on the fusion of the transparency control sequence and the geometric correction space.

[0048] Specifically, density analysis is performed on the Gaussian topological domain to identify sparsely distributed regions. All ellipsoidal elements within the Gaussian topological domain are traversed, and the spatial coverage volume of each ellipsoid and the distance between adjacent ellipsoids are calculated to establish a density metric for the ellipsoidal distribution. The analysis space is divided into a regular cubic grid, with the grid size determined based on the average scale of the ellipsoids. The number of ellipsoids and coverage within each grid cell are statistically analyzed. A local density function ρ(x) = Σ is defined. i α i ·exp(-||x-μ i ||² / 2σ i ²), where ρ(x) is the density value at position x, and α i Let μ be the opacity of the i-th ellipsoid. i For the center of the ellipsoid, σ i To influence the radius, the spatial density distribution is calculated by summing the contributions of all ellipsoids. A density threshold ρ_min is set to distinguish between sparse and dense regions. When the average density of a region is below the threshold, it is marked as a sparse distribution region. These regions typically correspond to flat parts or undersampled areas on the surface of the artifact. For example, in the surface reconstruction of a bronze ding (a type of ancient Chinese cooking vessel), the flat sidewalls of the vessel may show sparse ellipsoidal coverage, while areas with complex decorative patterns are usually densely distributed with ellipsoids. Density analysis can identify weak areas that need reinforcement. Connectivity analysis is performed on the identified sparse regions to merge spatially adjacent sparse grids into complete sparse distribution regions, recording the boundaries, volume, and geometric features of each region. The impact of sparse regions on the overall reconstruction quality is evaluated, and processing priorities are determined based on the region's location and visibility on the artifact's surface. Sparse problems in the front and detailed areas require priority resolution.

[0049] In some embodiments, performing a perception-driven subdivision operation on the sparsely distributed region to generate a dense Gaussian field includes: establishing a visual importance scanning grid within the sparsely distributed region; identifying perception priority and detail sensitivity through the visual importance scanning grid; weighting the detail sensitivity based on the perception priority to generate a perception-driven factor; and performing an adaptive subdivision labeling dense Gaussian field according to the perception-driven factor.

[0050] A visual importance scanning grid is established within sparsely distributed regions. Based on the processing priorities determined in the first section, 3D scanning grids are constructed preferentially in high-priority sparse regions. The grid resolution is determined according to the human eye's resolving power at typical viewing distances, ensuring the capture of even the smallest visually perceptible details. An initial visual importance value is assigned to each grid node, calculated based on the node's position relative to the preset observation point's angle and distance; nodes facing the viewing direction receive higher initial weights. Geometric and textural features surrounding the grid nodes are analyzed, including normal vector change rate, color gradient, and depth discontinuity, features closely related to human visual attention allocation. A saliency detection algorithm is used to calculate the visual saliency of each node, combining spatial contrast, color uniqueness, and geometric prominence to generate a comprehensive saliency score. Importance assessment criteria are adjusted according to the type of artifact and display requirements; decorative patterns receive increased texture weight, while structural components emphasize geometric features, ensuring the scanning grid reflects actual aesthetic value. Visual importance values ​​are mapped to the attributes of grid nodes, forming a spatially distributed importance field. High-importance regions receive more computational resources and more refined representations in subsequent processing.

[0051] For example, the step of identifying perception priority and detail sensitivity through the visual importance scanning grid includes: locating high-attention grid nodes in the visual importance scanning grid; determining node weight values ​​based on the high-attention grid nodes; performing a diffusion scan to surrounding grids based on the node weight values ​​to generate a diffusion influence map; and converging the diffusion influence map to the perception focal point to form perception priority and detail sensitivity.

[0052] High-attention grid nodes are located within a visual importance scanning grid. The entire 3D grid space is scanned, and the local importance index of each node is calculated, including a comprehensive score based on geometric curvature, texture complexity, and semantic label weights. Non-maximum suppression (NMS) is used to filter importance peaks within the local neighborhood; when a node's importance value is greater than all its neighbors, it is marked as a candidate high-attention node. Hierarchical clustering is performed on candidate nodes; nodes with spatial distances less than a threshold and similar importance are merged into a single attention center to avoid overly dense focus distributions. The effectiveness of high-attention nodes is validated by projecting them onto multiple viewpoints to check their visibility, eliminating nodes that are occluded in most viewpoints to ensure the selected nodes have actual display value. An attribute profile is created for each confirmed high-attention node, recording its spatial coordinates, importance score, feature type, and influence radius, forming an anchor point set for high-attention nodes.

[0053] Node weights are determined based on high-attention grid nodes. The multidimensional attributes of each high-attention node are analyzed, comprehensively considering its geometric importance, visual saliency, and cultural value, to design a multi-factor weight calculation scheme. Geometric weights are calculated based on the curvature variation and normal vector dispersion at the node; areas with dramatic curvature variations indicate rich geometric features and require higher representation accuracy. Visual weights consider the differences in human eye sensitivity to different spatial frequencies; nodes corresponding to mid-frequency details receive higher weights, consistent with the characteristics of the human visual system. Cultural weights are determined based on the type of artifact and the semantic information of the node's location; special areas such as decorative patterns, inscriptions, and restoration marks receive additional weight bonuses. The analytic hierarchy process (AHP) is used to integrate multiple weight factors, constructing a judgment matrix to compare the relative importance of each factor, and the final comprehensive weight value is calculated using the eigenvector method.

[0054] A diffusion influence map is generated by a diffusion scan of the surrounding mesh based on node weight values. Starting from each high-weight node, an anisotropic diffusion model is used to propagate influence to the surrounding mesh, with the diffusion speed and direction adjusted according to local geometric features. During diffusion, influence propagates faster along the surface tangent and slower along the normal direction; this directionality ensures effective propagation of influence on geometrically continuous surfaces. A diffusion propagation algorithm is employed, where each scanning step distributes a certain proportion of influence from the current node to neighboring nodes, with the weights calculated based on the Euclidean distance and geometric similarity between nodes. Influence gradually propagates and superimposes within the mesh; the influence of multiple source nodes interacts, generating a reinforcing effect in the intersection region, forming a complex influence distribution pattern. Boundary conditions for diffusion are set, with reflection boundaries at geometric and material boundaries to weaken diffusion intensity, prevent influence from crossing unrelated areas, and maintain the independence of different surface features. The diffusion process employs a time-stepping method, updating the influence values ​​of all mesh nodes in each iteration and monitoring the convergence state of the global influence distribution. Through iterative calculation, the influence distribution gradually converges to a stable state, forming a diffusion influence map. The value at each location in the map reflects its overall attention and spatial connectivity strength.

[0055] The diffusion impact map is converged to perception focal points to establish perception priority and detail sensitivity. The topological structure of the diffusion impact map is analyzed, the gradient distribution of the influence field is calculated, and local extrema with zero gradients are identified as candidate focal areas. A watershed algorithm is used to identify convergence areas of influence, treating the influence field as three-dimensional terrain; areas where water flows naturally converge form perception focal points. For each focal point, the size of its attraction basin and average influence intensity are calculated. The size of the attraction basin is measured by the watershed area, and the influence intensity is represented by statistical measures of the influence value within the basin. A hierarchical relationship is established between focal points, classifying them according to the size of the attraction basin and the influence intensity. Primary focal points control large-scale perception allocation, while secondary focal points are responsible for enhancing local details, forming a tree-like perception control structure. The spatial distribution of perception priorities is determined based on the distribution and intensity of the focal points. A Gaussian kernel function is used to establish a priority decay pattern around the focal points, with higher priority closer to the focal point. The rate of change of local influence is analyzed to determine detail sensitivity. The magnitude of the influence gradient is calculated to assess the intensity of local changes; areas with dramatic changes indicate rich detail and require fine representation, while areas with gradual changes can be represented coarsely. Perception priority and detail sensitivity are encoded as continuous spatial functions with the function range normalized to the [0,1] interval.

[0056] Perceptual driving factors are generated by weighting and enhancing detail sensitivity based on perceptual priority. A non-linear combination of perceptual priority and detail sensitivity is used, mapped using a sigmoid function to ensure that detail requirements in high-priority regions are sufficiently enhanced. A more aggressive enhancement strategy is applied to high-priority regions, where even small changes in detail sensitivity can trigger significant weight increases, ensuring that critical details are not overlooked. The direction and rate of change of spatial gradient recognition enhancement effects are calculated, and transition bands are inserted in boundary regions with large gradients to avoid abrupt changes between different enhancement levels. Multi-scale perceptual information is integrated: the coarse scale provides the global importance distribution, while the fine scale captures local detail requirements, forming multi-resolution driving factors through weighted fusion. The generated perceptual driving factors are normalized and smoothed to ensure that the numerical range is suitable for subsequent calculations and that the spatial distribution is continuous without abrupt changes, ultimately forming a complete driving field to guide subdivision operations.

[0057] Adaptive subdivision labeling of a dense Gaussian field is performed based on perceptual driving factors. The required ellipsoid density target for each region is determined based on the magnitude of the perceptual driving factor; regions with higher driving factors require a denser ellipsoid distribution to fully represent details. Existing ellipsoids are subdivided into candidate labels. When the average driving factor of the region covered by an ellipsoid exceeds a threshold, the ellipsoid is labeled as needing subdivision, with the subdivision level determined based on the driving factor. A hierarchical subdivision strategy is implemented, allowing ellipsoids in high-driving-factor regions to be recursively subdivided multiple times. Each subdivision generates sub-ellipsoids that are further evaluated for subdivision until the density requirement is met. The continuity of ellipsoid parameters is maintained during subdivision; the spherical harmonic coefficients of sub-ellipsoids are obtained through interpolation of the parent ellipsoid, and small perturbations are added to the rotational attitude to increase representational flexibility. Topology optimization is performed on the subdivided ellipsoid group, removing redundant ellipsoids, merging overly close ellipsoids, and adjusting the overlap between ellipsoids to ensure efficient spatial coverage. After adaptive subdivision, the originally sparse regions are densed to varying degrees according to perceptual importance, forming a dense Gaussian field that satisfies visual requirements while maintaining computational efficiency.

[0058] A transparency control sequence is formed by applying opacity allocation to a dense Gaussian field. The spatial overlap relationship of ellipsoids in the dense Gaussian field is analyzed, and the intersection volume and overlap rate of each pair of overlapping ellipsoids are calculated to establish an occlusion relationship diagram between ellipsoids. Ellipses are sorted by depth according to viewpoint position and viewing direction to determine the rendering order from front to back; this order directly affects the transparency blending result. An opacity allocation strategy is designed, adjusting the base opacity of the ellipsoids in conjunction with the perceptual driving factor. Ellipses in high-driving-factor regions receive higher base opacity to enhance visual performance, surface ellipsoids receive higher opacity to ensure surface solidity, and internal ellipsoids are assigned lower opacity to avoid excessive occlusion. The effective opacity of each ellipsoid is calculated as α_eff = α_base·(1-Σ). j The opacity is defined as `overlap_j`, where `α_eff` is the effective opacity, `α_base` is the base opacity, and `overlap_j` is the overlap rate with the j-th ellipsoid, ensuring that the cumulative opacity does not exceed 1. A spatial gradient mechanism for opacity is established, increasing opacity near geometric edges to enhance contour sharpness and appropriately decreasing opacity in smooth areas to achieve a soft transition. Opacity values ​​are organized into a temporal control sequence, supporting dynamic adjustment to adapt to different viewing conditions and rendering needs. The sequence records the identifier, depth value, and opacity parameter of each ellipsoid. The resulting opacity control sequence not only optimizes the rendering effect but also provides a flexible opacity management mechanism, supporting interactive visual adjustments.

[0059] In some embodiments, the step of constructing a multi-layered rendering strategy based on the fusion of the transparency control sequence and the geometric correction space includes: performing hierarchical analysis on the transparency control sequence to identify transparency transition boundaries; using the transparency transition boundaries to locate geometric change sensitive areas in the geometric correction space; mapping the geometric change sensitive areas to render intensity to form an intensity distribution map; and constructing a multi-layered rendering strategy based on the intensity distribution map.

[0060] Hierarchical analysis is used to identify transparency transition boundaries in the transparency control sequence. All ellipsoidal transparency values ​​in the sequence are traversed, and the transparency difference between adjacent ellipsoids is calculated. Points exceeding a set threshold are marked as potential transition points. Spatial clustering analysis is performed on these marked transition points to connect consecutive points, forming transparency transition boundaries. These boundaries typically correspond to locations of geometric or material changes. For example, in bronze artifacts, smooth transitions from the body to the handle and sharp boundaries between inscriptions and the surface form transparency transition boundaries. Identifying these boundaries allows for accurate capture of structural feature changes in the artifact. Spatial continuity constraints are established for these boundaries, and broken boundary segments are repaired to ensure each boundary forms a complete, continuous curve or surface in three-dimensional space. Geometric descriptive parameters of the boundaries are calculated, including spatial coordinates, local tangent direction, and neighborhood coverage.

[0061] Geometric change-sensitive regions are located in the geometric correction space using transparency jump boundaries. The transparency jump boundaries are mapped to the geometric correction space, and a correspondence between transparency features and geometric features is established through coordinate transformation. Combining the occlusion relationship graph between ellipsoids, a neighborhood search is performed along the jump boundary in the geometric correction space to analyze the differences in geometric properties on both sides of the boundary, including changes in curvature, normal vector, and depth values. When changes in geometric properties are synchronized with transparency jumps, the region is marked as a geometric change-sensitive region; these regions require precise rendering control to maintain visual realism. The extent of the sensitive regions is expanded to include transition zones, and morphological dilation operations are used to generate buffers on both sides of the boundary to ensure smooth transitions during rendering. A geometric complexity index is calculated for each sensitive region, and a spatial relationship graph between sensitive regions is established, taking into account surface roughness, feature density, and topological changes.

[0062] An intensity distribution map is generated by mapping the rendering intensity of geometrically sensitive areas. A rendering intensity function is designed for each geometrically sensitive area, with a base intensity value determined based on the area's geometric complexity and visual importance. Nonlinear mapping is used to convert geometric features into rendering parameters; areas with high curvature have increased specular intensity to highlight shape features, while areas with dramatic depth changes have enhanced shadow effects to create a sense of depth. Gradual intensity transitions are designed at the boundaries of sensitive areas, and spline interpolation is used to generate a smooth intensity distribution inside and outside the sensitive areas, avoiding abrupt changes in rendering effects. The influence of lighting conditions on rendering intensity is considered; rendering parameters for each area are adjusted according to preset light source positions and intensities to ensure overall lighting harmony. The rendering intensities of all sensitive areas are integrated into a unified spatial map, and weighted overlay is used to handle overlapping areas.

[0063] A multi-layered rendering strategy is constructed based on intensity distribution maps. The rendering space is divided into multiple layers according to the intensity distribution maps, each corresponding to different rendering quality and computational complexity requirements. A base layer rendering strategy is designed, using low-order spherical harmonic expansion and a simplified lighting model to provide a fast preview effect, suitable for long-distance observation or performance-constrained scenarios. A detail layer rendering strategy is constructed, using full spherical harmonic coefficients and complex lighting calculations in high-intensity areas to accurately represent the material and texture features of the surface. An enhancement layer rendering strategy is set, applying special rendering techniques, such as subsurface scattering and ambient occlusion, in geometrically sensitive areas to enhance visual realism. An automatic switching mechanism between layers is established, dynamically selecting the appropriate rendering layer based on viewing distance, viewpoint, and system performance to ensure a smooth interactive experience. For example, when users browse cultural relics at a distance, the base layer rendering is automatically used to provide a smooth experience; when zooming in to observe the details of the inscription, it automatically switches to the enhancement layer rendering to ensure the clarity and sharpness of the characters. The completed multi-layered rendering strategy achieves an optimized balance of quality, performance, and resources, providing a flexible rendering solution for different application scenarios.

[0064] Step S150: Project the multi-layer rendering strategy onto the view sensitivity space to generate key view intervals, evaluate the importance of the key view intervals to identify view inflection point positions, perform stability tests on view inflection point positions to generate view anchor points, and construct a multi-view rendering sequence based on view anchor points.

[0065] Specifically, the multi-layered rendering strategy is projected onto the viewpoint sensitivity space to generate key viewpoint intervals. A three-dimensional viewpoint sensitivity space is constructed, establishing a spherical coordinate system with the observation point as the origin. The azimuth angle φ ranges from [0° to 360°], the pitch angle θ ranges from [-90° to 90°], and the distance r is determined to be within a reasonable range based on the size of the artifact. The parameters of each layer in the multi-layered rendering strategy are mapped to the viewpoint space, and the activation degree and contribution weight of each rendering layer under different viewpoints are analyzed. The basic layer dominates at distant viewpoints, while the detail layer is activated at specific angles at close range. The rendering complexity index C(φ,θ,r) = Σ is calculated for each viewpoint position. i W i ·L i Where C is the rendering complexity, φ is the azimuth angle, θ is the pitch angle, r is the distance, and W is the distance to the target value. i Let L be the activation weight of the i-th layer. i To assess the computational complexity of this layer, the overall viewpoint sensitivity distribution is evaluated using surface integrals. Gradient analysis is used to identify regions of rapid change in rendering complexity. When the gradient exceeds a threshold, it indicates that the viewpoint region is sensitive to the rendering strategy, and small changes in viewpoint can lead to significant differences in visual effects. Cluster analysis is performed on sensitive regions, grouping spatially consecutive high-sensitivity points into key viewpoint intervals. Each interval represents a specific type of observation condition, such as close-up frontal observation or medium-distance side observation. For example, for bronze ding artifacts, the close-up frontal observation interval highlights the details of the decorative patterns on the vessel, the medium-distance side observation interval showcases the overall shape and proportions, and the top-down angle interval emphasizes the inscription information inside the vessel. Rendering strategies are optimized for different viewpoint intervals. Feature parameters are calculated for each key viewpoint interval, including the center viewpoint, angle range, and peak sensitivity.

[0066] The importance of key viewpoint intervals is assessed to identify viewpoint inflection points. Within each key viewpoint interval, a refined analysis is performed to calculate the impact of viewpoint changes on the rendering effect, and a rendering quality difference metric is used to assess the visual differences between adjacent viewpoints. An importance function is defined as I(v) = Q(v)·F(v)·U(v), where I is the importance value of viewpoint v, Q(v) is the rendering quality score of viewpoint v, F(v) is the predicted access frequency for that viewpoint, and U(v) is the user preference weight. This weight is adjusted in conjunction with the rendering complexity metric C(φ,θ,r) from the first section, with viewpoint intervals of higher complexity showing increased importance. Continuous sampling is performed along the boundaries of the viewpoint intervals, calculating the first and second derivatives of the importance function, and identifying locations where the derivative sign changes. These locations correspond to local extrema or inflection points of importance. Viewpoint inflection points typically occur at the critical points where geometric features transition from visible to invisible, or at the boundaries of different detail levels. Rendering strategies at these locations require special optimization to ensure visual continuity. For example, when the viewing angle shifts from the side to the front, the decorations on the side walls of an object are gradually obscured while the patterns on the front begin to appear. This transition point is the viewing angle inflection point, which requires special optimization to ensure the continuity of feature display. Identified inflection points are categorized into geometric occlusion inflection points, detail level inflection points, and lighting change inflection points. Different types of inflection points require different processing strategies. The relationships between inflection points are established; similar inflection points may belong to different sides of the same geometric feature, requiring coordinated processing to ensure the consistency of the overall effect.

[0067] In some embodiments, the step of generating a view anchor point by performing a stability test on the view inflection point position includes: constructing a rendering response matrix based on the view inflection point position; performing a stability oscillation scan in the rendering response matrix to identify oscillation suppression regions; performing response smoothing processing on the oscillation suppression regions to generate a stable rendering domain; and generating a view anchor point by fixing the view based on the stable rendering domain.

[0068] A rendering response matrix is ​​constructed based on the viewpoint inflection point location. In stability testing, the test accuracy is adjusted according to the viewpoint importance function I(v), with stricter stability standards applied to high-importance inflection points. A local coordinate system is defined at each viewpoint inflection point, and a spherical neighborhood is established centered on the inflection point. The radius is adaptively determined based on viewpoint sensitivity, with smaller radii used in highly sensitive areas to ensure accuracy. System sampling is performed within the neighborhood, generating sampling grids along the azimuth and pitch directions. Each sampling point represents a small viewpoint perturbation, recording the perturbation vector and corresponding rendering parameter changes. A rendering response matrix is ​​constructed, using partial derivatives to represent the influence of viewpoint component changes on rendering parameters, and calculating the partial derivative relationship using the finite difference method. Rendering parameters include key indicators such as activation weights of each layer, ellipsoidal visibility, and illumination intensity. Viewpoint components include viewing angle and distance, forming a complete input-output mapping relationship. The numerical stability of the matrix is ​​evaluated using the condition number; a larger condition number indicates a more sensitive system to input perturbations. By constructing the rendering response matrix, the local linear relationship between viewpoint changes and rendering effects is quantitatively characterized.

[0069] Stability oscillation scanning is performed on the rendered response matrix to identify oscillation suppression regions. Eigenvalue analysis is performed on the rendered response matrix to calculate all eigenvalues ​​and their corresponding eigenvectors. The presence of complex eigenvalues ​​indicates that the system may exhibit oscillatory behavior. An oscillation detection index is designed; when conjugate complex eigenvalue pairs exist and the imaginary part is large, the system is prone to periodic oscillations in the corresponding characteristic direction, affecting the stability of the rendering. Dynamic simulations are performed along different viewpoint perturbation directions, and the system response is observed by applying sinusoidal perturbation signals. The oscillatory components and their frequency characteristics in the response are identified through spectral analysis. Oscillation suppression regions are searched in the parameter space of the response matrix. These regions have all real eigenvalues ​​with small absolute values, and the system response is stable and oscillatory. Numerical optimization methods are used to adjust local rendering parameters. By adding damping terms or modifying interpolation functions, originally unstable regions are transformed into oscillation suppression regions. Boundary models of the oscillation suppression regions are established, and classification methods are used to predict the stability category of the regions based on the characteristics of the response matrix. The identified oscillation suppression regions provide a reliable candidate range for selecting stable viewpoint anchor points.

[0070] A stable rendering domain is generated by smoothing the response in the oscillation suppression region. Spatial filtering techniques are applied within the oscillation suppression region, using a Gaussian filter to smooth the rendering response function. The filter's standard deviation is adaptively adjusted based on the rate of change of the local response. An interpolation scheme for the response function is designed, using cubic spline interpolation between discrete sampling points to generate a continuous response surface, ensuring continuity and avoiding abrupt changes in derivatives. Curvature analysis is performed on the interpolated response surface to identify and locally correct areas of excessive curvature. Energy minimization is used to reduce the bending energy of the surface, improving smoothness. A temporal smoothing constraint is introduced to ensure that the temporal derivatives of the rendering parameters are bounded when the viewpoint changes continuously, preventing sudden jumps from affecting the viewing experience. A multi-resolution smoothing strategy is established to ensure overall trend smoothness at the global scale while preserving necessary detail variations at the local scale, balancing smoothness and expressiveness. After response smoothing, the originally discrete oscillation suppression regions are merged to form a continuous, stable rendering domain.

[0071] Anchor points are generated based on a stable rendering domain with a fixed viewpoint. An optimization algorithm is applied within the stable rendering domain to search for the optimal anchor position. The objective function comprehensively considers rendering quality, stability metrics, and coverage to find the globally optimal solution. Anchor point selection criteria are defined, including the minimum condition number criterion, the maximum coverage criterion, and the uniform distribution criterion, with the weights of each criterion determined according to application requirements. A greedy algorithm is used to progressively select anchor points, choosing new anchor points that maximize overall coverage at each step until the preset coverage requirement or anchor point number limit is met. Local fine-tuning is performed on the selected anchor points to optimize viewpoint parameters while maintaining stability, making key features more prominent and the overall composition more aesthetically pleasing. An influence radius and transition region are set for each anchor point, and a gradient function is defined from the anchor point to its influence boundary to ensure a smooth transition between adjacent anchor points. Metadata records for the anchor points are established, including precise viewpoint parameters, expected rendering effects, and applicable observation scenarios.

[0072] In some embodiments, constructing a multi-view rendering sequence based on the view anchor points includes: performing spatial trajectory planning based on the view anchor points to generate an observation path network; performing connectivity analysis on the observation path network to identify key path nodes; performing view quality assessment at the key path nodes to generate a view score matrix; and arranging the view score matrix in a temporal sequence to generate a multi-view rendering sequence.

[0073] A spatial trajectory planning system is used to generate an observation path network based on viewpoint anchor points. During path planning, anchor points within stable rendering domains are prioritized to ensure overall sequence stability and viewing comfort. All viewpoint anchor points are treated as key nodes in the path planning, and their distribution is visualized in 3D view space to analyze spatial relationships and accessibility between nodes. Spline curves are used to connect adjacent anchor points, with control points set according to the smoothness requirements of viewpoint transitions to ensure continuous curvature of the path and that it does not exceed the rotation speed limit comfortable for the human eye. Various types of observation paths are designed, including a wraparound path showcasing a 360-degree panoramic view of the artifact, a spiral path displaying details at different scales from far to near, and a focus path highlighting specific artifact features. Obstacle avoidance constraints are considered in path planning to ensure that the view is not obstructed by the artifact itself and to avoid entering viewpoint dead zones with poor rendering quality. A hierarchical path network structure is constructed, with main paths connecting major anchor points to form a basic observation framework, and branch paths providing options for in-depth observation of specific details. Geometric properties are calculated for each path, including path length, maximum curvature, and average height variation; these properties are used to evaluate path complexity and viewing comfort. The completed observation path network provides a rich selection of viewpoint traversals to meet different observation needs and interaction modes.

[0074] Connectivity analysis is performed on the observation path network to identify critical path nodes. Graph theory is used to represent the path network as a directed graph, where nodes represent anchor points and path intersections, edges represent feasible viewpoint transition paths, and edge weights reflect the cost of the transition. Network connectivity metrics are calculated, including strongly connected components, cut vertices, and bridge edges. Disruption of these critical structures leads to network fragmentation and requires focused protection and optimization. Centrality analysis is applied to identify critical path nodes. Nodes with high betweenness centrality are located on multiple shortest paths and are hubs for viewpoint transitions; nodes with high degree centrality connect multiple paths, providing rich observation options. The importance of path intersections is assessed, as viewpoint selection at intersections affects the observation flow and requires clear navigation guidance and preview information. Bottleneck nodes in the network are identified; these nodes have concentrated rendering loads and may become performance bottlenecks, requiring special optimization such as pre-rendering or multi-level caching. Dependencies between nodes are established; some nodes must be visited after specific nodes to obtain the best observation results, and this temporal dependency needs to be considered in path planning.

[0075] A viewpoint quality assessment is performed at critical path nodes to generate a viewpoint scoring matrix. A comprehensive rendering quality assessment is conducted at each critical path node, using multiple evaluation metrics to quantify the quality of the viewpoint, including geometric integrity, texture sharpness, and lighting realism. Geometric integrity is assessed by calculating the proportion of visible surfaces and the degree of self-occlusion, ensuring that the main structure of the artifact is fully displayed from this viewpoint and avoiding the occlusion of important features. Texture sharpness is calculated based on the projected pixel density and sampling rate, considering the combined effects of viewpoint, distance, and surface normals to ensure effective rendering of texture details. Lighting assessment analyzes the distribution of light and shadow and highlight effects under this viewpoint, avoiding overexposed or underexposed areas and ensuring accurate representation of material properties. A viewpoint scoring matrix is ​​constructed, and normalization is used to make scores across different dimensions comparable. The Analytic Hierarchy Process (AHP) is applied to determine the weights of each evaluation dimension, adjusting the weight allocation according to the type of artifact and the purpose of the display, and calculating the overall quality score. Statistical analysis of the scoring matrix identifies high-scoring nodes as key display areas, while low-scoring nodes need improvement or should be weakened in the sequence.

[0076] A multi-view rendering sequence is generated by arranging the viewpoint rating matrix in a temporal sequence. Based on the viewpoint rating matrix and path network structure, a dynamic programming algorithm is used to solve for the optimal viewpoint access sequence, aiming to maximize the accumulated viewpoint quality score. Constraints for sequence optimization are defined, including time limits, path smoothness requirements, and mandatory inclusion of important features. The optimal solution is searched within these constraints. A progressive display strategy is designed: the sequence begins with a distant view to establish an overall impression, gradually approaches to display details, and ends with a landmark view, forming a complete narrative structure. For example, the artifact display sequence establishes an overall impression of a bronze tripod from a distance, gradually approaches to reveal decorative details, and finally focuses on important inscriptions. The entire process is like a guided tour route, providing a structured viewing experience. Visual fatigue factors are considered in the temporal arrangement, with concise transitional views arranged after complex details are displayed, maintaining the observer's attention through rhythmic changes. Duration is allocated to each segment in the sequence, with more time allocated to important nodes and transitional segments passed quickly, keeping the total duration within the user's attention span. Metadata descriptions of the sequence are generated, including viewpoint parameters, rendering configurations, and expected effect descriptions for each time point, supporting sequence editing and customization. The final multi-view rendering sequence is optimized and arranged to provide a high-quality, rhythmic, and information-rich experience for displaying cultural relics.

[0077] Step S160: Generate a volume rendering matrix based on the multi-view rendering sequence, optimize the quality of the volume rendering matrix to generate a 3D model of the cultural relic, and perform hash verification and digital signature on the 3D model of the cultural relic to complete the confirmation of ownership.

[0078] Specifically, a volumetric rendering matrix is ​​generated based on a multi-view rendering sequence. Rendering results at each time point are extracted from the multi-view rendering sequence, including color images, depth images, and ellipsoidal attribute data. These data record the surface information of the artifact from different angles. The three-dimensional space is discretized into a regular voxel mesh. The mesh resolution is determined according to the size of the artifact and the detail requirements, with a typical value of 512×512×512 to ensure the capture of millimeter-level geometric details. The projection contribution of each voxel under various viewpoints is calculated. A ray casting algorithm is used to trace the ray path from the viewpoint to the voxel, accumulating color and opacity information along the way. A volumetric rendering matrix V is constructed, where V(x,y,z) represents the rendering attribute vector at spatial location (x,y,z), containing multi-dimensional information such as color value, opacity, and normal vector. A multi-view fusion algorithm is applied to integrate the observation results from different viewpoints. Multiple observation values ​​of the same voxel are weighted and averaged, with weights determined according to viewpoint quality and observation angle. Inconsistencies between viewpoints are addressed. When the observation results of different viewpoints at the same location differ significantly, median filtering or robust estimation methods are used to eliminate outliers. By gradually refining the volume rendering matrix over time, an accurate 3D representation is formed.

[0079] In some embodiments, the step of optimizing the volume rendering matrix to generate a 3D model of the cultural relic includes: performing pixel correlation analysis based on the volume rendering matrix to identify quality propagation nodes; injecting attribute correction signals into the quality propagation nodes to generate a quality propagation chain; performing diffusion enhancement on the quality propagation chain to form a global optimization field; and performing rendering reconstruction based on the global optimization field to generate a 3D model of the cultural relic.

[0080] Pixel correlation analysis is performed based on the volumetric rendering matrix to identify quality propagation nodes. Each voxel in the volumetric rendering matrix is ​​traversed, and its attribute similarity with neighboring voxels is calculated. Vector cosine distance is used to measure the consistency of color and normal vectors. A pixel correlation function R(i,j) = exp(-||V_i-V_j||² / 2σ²) is defined, where R(i,j) represents the correlation strength between voxels i and j, V_i and V_j are the corresponding attribute vectors, and σ controls the sensitivity of the correlation. A voxel correlation graph is constructed, connecting voxel pairs with correlation higher than a threshold with edges to form a network structure. The network connectivity reflects the spatial distribution of rendering quality. Spectral clustering is used to analyze the correlation graph, calculating the eigenvectors of the graph Laplacian matrix to identify voxel groups with high cohesion and good quality consistency within these groups. The highest-quality voxel in each group is selected as the seed node. The quality score comprehensively considers factors such as rendering sharpness, multi-view consistency, and geometric rationality, adjusting the scoring criteria based on the weight information of multi-view fusion. The contribution of high-weighted views receives a larger quality influence coefficient. Expanding the influence range of seed nodes involves incorporating directly adjacent and highly correlated voxels into the quality propagation node set, forming the starting point for quality optimization. For example, in the 3D reconstruction of bronze artifacts, areas with clear surface textures and consistent multi-view observation are identified as quality propagation nodes. The rendering attributes of these high-quality areas will propagate to surrounding blurry or incomplete areas, gradually improving the overall model quality.

[0081] A quality propagation chain is generated by injecting attribute correction signals into the quality propagation nodes. Specific rendering attributes are extracted from the propagation nodes, combining RGB color values ​​(e.g., R=180, G=120, B=85), opacity values ​​(e.g., α=0.85), and normal vector coordinates (e.g., nx=0.6, ny=0.3, nz=0.7) to form the attribute correction signal. For example, the color values ​​and opacity parameters of clearly textured areas on a bronze artifact's surface are extracted and prepared for propagation to surrounding blurry areas. The extracted values ​​are directly written into the node's attribute vector, replacing the original blurry colors with accurate RGB values ​​and correcting over-transparency or over-opaqueness with standard opacity values. Adjacent nodes receive these correction values ​​and update their parameters through numerical interpolation. Closer nodes receive the full correction values, while farther nodes receive attenuated values. For example, a node one unit away from the propagation source receives 80% of the correction intensity, and a node two units away receives 50% of the correction intensity. Nodes calculate the average value when receiving multiple correction signals. For example, if a node receives RGB values ​​from three sources (180, 120, 85), (175, 125, 90), and (185, 115, 80), it will eventually update to the average value (180, 120, 85). Continuous numerical propagation forms a propagation chain in space, making the colors of each node in the chain more accurate, the transparency more reasonable, and the surface normal vectors more correct. By correcting the surrounding blurred and missing areas with accurate data from high-quality areas, the complete reproduction of the surface details of the cultural relic is achieved.

[0082] A global optimized field is formed by diffusion enhancement of the mass propagation chain. The effects of all mass propagation chains are superimposed to form a continuous mass field in the volume space, where the field strength represents the degree of quality improvement at that location. An iterative diffusion algorithm is used to model the spatial propagation of mass, achieving spatial diffusion of the mass field through multiple iterations. The diffusion coefficient adjusts anisotropy based on local structural features. Special processing is applied at the intersections of propagation chains; the interaction of mass signals from multiple chains produces an enhancement effect through nonlinear combination, avoiding oversaturation caused by simple superposition. Boundary conditions for diffusion are set to maintain the mass gradient on the artifact surface and interpolate missing mass values ​​in void regions. A multi-mesh method accelerates diffusion calculations, rapidly propagating global information on a coarse mesh and accurately refining local details on a fine mesh, balancing computational efficiency and accuracy. The diffusion results are regularized to limit the maximum gradient of the mass field, avoiding unnatural abrupt changes and maintaining a smooth spatial transition. After diffusion enhancement, the entire volume space is endowed with optimized mass properties, forming a global optimized field that guides the final rendering. For example, the optimized field has a high field strength in areas with rich decorative patterns, guiding the use of fine rendering parameters in that area, while the field strength is lower in flat surface areas, adopting a simplified rendering strategy, thus realizing intelligent allocation of computing resources.

[0083] A 3D model of the cultural relic is generated through rendering reconstruction based on a global optimization field. The rendering attributes of each voxel are recalculated according to the distribution of the global optimization field, using field strength as a confidence weight. Attributes in high-field-strength regions remain unchanged, while those in low-field-strength regions are corrected. A Markov random field model is applied for attribute optimization, using the global optimization field as a data term and incorporating smoothness constraints as a regularization term. The optimal attribute configuration is obtained by minimizing energy. An improved ray casting algorithm is used for volumetric rendering, integrating the optimized voxel attributes along the viewing direction and adaptively adjusting the sampling step size to balance quality and efficiency. Early ray termination is applied during rendering, stopping ray tracing when the accumulated opacity reaches a threshold, reducing unnecessary computation while maintaining rendering quality. A multi-resolution rendering strategy is implemented, using different voxel sampling rates based on viewing distance and importance; high resolution is used for near and important areas, while low resolution is used for distant areas. Post-processing enhancements are applied to the rendering results, including ambient occlusion, subsurface scattering, and tone mapping effects, to improve visual realism. The generated 3D model of the cultural relic integrates multi-view information and quality optimization results, accurately restoring the geometric shape and surface characteristics of the relic, supporting high-quality interactive display and analysis applications.

[0084] The process involves hash verification and digital signature to establish ownership of the 3D cultural relic model. A SHA-256 hash value is calculated on the generated 3D model data to create a unique digital fingerprint, ensuring the integrity and immutability of the model data. Key attribute information of the model is extracted, including the total number of ellipsoidal parameters, rendering quality level, reconstruction timestamp, algorithm version identifier, and other technical features, to construct a standardized model metadata structure. The model hash value is encrypted and signed using an elliptic curve digital signature algorithm, generating a legally valid digital certificate. The signing key is managed by the reconstruction institution's PKI certificate system to ensure the authority and credibility of the signature. The model hash, digital signature, and metadata information are organized into a complete ownership data package, and the validity of the signature and the consistency of the model data are verified through a smart contract. A mapping relationship between model identifiers and ownership records is established to support subsequent copyright inquiries, usage authorization, and infringement detection, providing a complete intellectual property protection mechanism for the results of digital reconstruction of cultural relics.

[0085] To implement the above-described method embodiments, a blockchain-based 3DGS digital reconstruction method for cultural relics is proposed to achieve the corresponding functionalities and technical effects. See also... Figure 2 , Figure 2 This diagram illustrates a structural block diagram of a blockchain-based 3DGS digital reconstruction system 200 for cultural relics, as provided in an embodiment of this application. For ease of explanation, only the parts relevant to this embodiment are shown. The blockchain-based 3DGS digital reconstruction system 200 for cultural relics provided in this embodiment includes:

[0086] Data processing module 201 is used to collect RGB image data and depth perception data of cultural relics, and to perform illumination separation processing on the RGB image data to form a standardized image library;

[0087] The intelligent modeling module 202 is used to generate an initial Gaussian point set by performing Gaussian kernel density estimation and ellipsoid fitting based on the standardized image library, and to determine the Gaussian topological domain by performing intelligent density allocation on the initial Gaussian point set.

[0088] The feature fusion module 203 is used to perform gradient parsing and geometric constraint decomposition on the depth sensing data to obtain structural guidance information, perform feature coupling between the structural guidance information and the Gaussian topological domain to generate a depth feature field, and construct a geometric correction space based on the depth feature field.

[0089] The perception optimization module 204 is used to perform density analysis on the Gaussian topological domain to identify sparse distribution regions, perform perception-driven subdivision operation on the sparse regions to generate dense Gaussian fields, perform opacity allocation processing on the dense Gaussian fields to form a transparency control sequence, and construct a multi-level rendering strategy based on the fusion of the transparency control sequence and the geometric correction space.

[0090] The viewpoint anchoring module 205 is used to project the multi-level rendering strategy into the viewpoint sensitivity space to generate key viewpoint intervals, evaluate the importance of the key viewpoint intervals to identify viewpoint inflection point positions, perform stability tests on the viewpoint inflection point positions to generate viewpoint anchor points, and construct a multi-view rendering sequence based on the viewpoint anchor points.

[0091] The quality reconstruction module 206 is used to generate a volume rendering matrix based on the multi-view rendering sequence, optimize the quality of the volume rendering matrix to generate a three-dimensional model of the cultural relic, and perform hash verification and digital signature on the three-dimensional model of the cultural relic to complete the confirmation of ownership.

[0092] The aforementioned blockchain-based 3DGS digital reconstruction system 200 can implement one of the blockchain-based 3DGS digital reconstruction methods for cultural relics described in the above-described method embodiments. The options in the above method embodiments are also applicable to this embodiment and will not be detailed here. The remaining content of this application's embodiments can be referred to the content of the above method embodiments, and will not be repeated in this embodiment.

[0093] The above description is only a part or preferred embodiment of this application. Neither the text nor the drawings should limit the scope of protection of this application. All equivalent structural transformations made using the content of this application's specification and drawings under the overall concept of this application, or direct / indirect applications in other related technical fields, are included within the scope of protection of this application.

Claims

1. A blockchain-based 3DGS digital reconstruction method for cultural relics, characterized in that, include: Collect RGB image data and depth perception data of cultural relics, and perform illumination separation processing on the RGB image data to form a standardized image library; Based on the standardized image library, Gaussian kernel density estimation and ellipsoid fitting are performed to generate an initial Gaussian point set. The initial Gaussian point set is then intelligently density-allocated to determine the Gaussian topological domain. Gradient parsing and geometric constraint decomposition are performed on the depth sensing data to obtain structural guidance information. The structural guidance information is coupled with the Gaussian topological domain to generate a depth feature field. A geometric correction space is constructed based on the depth feature field. Density analysis is performed on the Gaussian topological domain to identify sparse distribution regions. A perception-driven subdivision operation is performed on the sparse regions to generate a dense Gaussian field. Opacity allocation processing is performed on the dense Gaussian field to form a transparency control sequence. A multi-level rendering strategy is constructed based on the fusion of the transparency control sequence and the geometric correction space. Projecting the multi-layered rendering strategy onto the view sensitivity space to generate key view intervals includes: constructing a three-dimensional view sensitivity space, mapping the parameters of each layer in the multi-layered rendering strategy to the view space, and calculating the rendering complexity index C(φ,θ,r)=Σ for each view position. i W i ·L i Where C is the rendering complexity, φ is the azimuth angle, θ is the pitch angle, r is the distance, and W is the distance to the target value. i Let L be the activation weight of the i-th layer. i To reduce the computational complexity of this layer, gradient analysis is used to identify sensitive regions where rendering complexity changes rapidly, and cluster analysis is performed to group spatially continuous high-sensitivity points into key viewpoint intervals. The importance of these key viewpoint intervals is assessed to identify viewpoint inflection point locations, including: defining an importance function I(v) = Q(v)·F(v)·U(v), where I is the importance value of viewpoint v, Q(v) is the rendering quality score of viewpoint v, F(v) is the predicted access frequency for that viewpoint, and U(v) is the user preference weight. Continuous sampling is performed along the boundary of the viewpoint interval, and the first and second derivatives of the importance function are calculated. The location where the derivative sign changes is identified as the viewpoint inflection point location. The process of generating viewpoint anchor points through stability testing includes: constructing a rendering response matrix based on the viewpoint inflection point positions; performing stability oscillation scanning within the rendering response matrix to identify oscillation suppression regions; performing response smoothing processing on the oscillation suppression regions to generate a stable rendering domain; fixing the viewpoint based on the stable rendering domain to generate viewpoint anchor points; and constructing a multi-view rendering sequence based on the viewpoint anchor points, including: performing spatial trajectory planning based on the viewpoint anchor points to generate an observation path network; performing connectivity analysis on the observation path network to identify critical path nodes; performing viewpoint quality assessment at the critical path nodes to generate a viewpoint scoring matrix; and arranging the viewpoint scoring matrix temporally to generate a multi-view rendering sequence. A volumetric rendering matrix is ​​generated based on the multi-view rendering sequence. The volumetric rendering matrix is ​​then optimized to generate a 3D model of the cultural relic. Finally, the 3D model of the cultural relic is hash-verified and digitally signed to complete the ownership confirmation.

2. The method according to claim 1, characterized in that, The step of intelligently density-assigning the initial Gaussian point set to determine the Gaussian topological domain includes: An ellipsoidal coverage analysis diagram is constructed based on the initial Gaussian point set; Locate the sparse coverage nodes in the ellipsoidal coverage analysis diagram; Gaussian compensation points are generated by exciting Gaussian compensation growth using the sparse nodes covered; A Gaussian topological domain is formed based on the compensated Gaussian points.

3. The method according to claim 1, characterized in that, The step of generating a depth feature field by feature coupling the structural guidance information with the Gaussian topological domain includes: Deep gradient analysis is performed on the structural guidance information to form a geometric constraint template; The Gaussian topological domain is subjected to ellipsoidal parameter extraction to form a Gaussian constraint template; The geometric constraint template and the Gaussian constraint template are interactively mapped to generate a constraint fusion template. The depth feature field is determined by depth-guided reconstruction based on the constraint fusion template.

4. The method according to claim 1, characterized in that, The step of performing a perception-driven subdivision operation on the sparsely distributed region to generate a dense Gaussian field includes: Establish a visual importance scanning grid within the sparsely distributed region; Perceptual priority and detail sensitivity are identified through the visual importance scanning grid; Based on the perception priority, the detail sensitivity is weighted and enhanced to generate a perception driving factor; An adaptive subdivision labeling dense Gaussian field is performed based on the perception driving factor.

5. The method according to claim 1, characterized in that, The multi-layered rendering strategy constructed based on the fusion of the transparency control sequence and the geometric correction space includes: Hierarchical analysis is performed on the transparency control sequence to identify transparency transition boundaries; The transparency jump boundary is used to locate the geometric change sensitive area in the geometric correction space; The intensity distribution map is generated by rendering the geometrically sensitive area; A multi-level rendering strategy is constructed based on the intensity distribution map.

6. The method according to claim 1, characterized in that, The process of optimizing the volume rendering matrix to generate a 3D model of the cultural relic includes: Pixel correlation analysis is performed based on the volume rendering matrix to identify quality propagation nodes; A quality propagation chain is generated by injecting attribute correction signals into the quality propagation node. The mass propagation chain is diffused and enhanced to form a global optimization field; The three-dimensional model of the cultural relic is generated by rendering and reconstruction based on the global optimization field.

7. The method according to claim 4, characterized in that, The identification of perceived priority and detail sensitivity through the visual importance scanning grid includes: Locate high-attention grid nodes within the visual importance scanning grid; The node weight values ​​are determined based on the high-attention grid nodes; Based on the node weight values, a diffusion effect map is generated by performing a diffusion scan to the surrounding grid. The diffusion effect map is converged to the perception focal point to form perception priority and detail sensitivity.

8. A blockchain-based 3DGS digital reconstruction system for cultural relics, characterized in that, include: The data processing module is used to collect RGB image data and depth perception data of cultural relics, and to perform illumination separation processing on the RGB image data to form a standardized image library; The intelligent modeling module is used to generate an initial Gaussian point set by performing Gaussian kernel density estimation and ellipsoid fitting based on the standardized image library, and to determine the Gaussian topological domain by performing intelligent density allocation on the initial Gaussian point set. The feature fusion module is used to perform gradient parsing and geometric constraint decomposition on the depth sensing data to obtain structural guidance information, and to perform feature coupling between the structural guidance information and the Gaussian topological domain to generate a depth feature field, and to construct a geometric correction space based on the depth feature field. The perceptual optimization module is used to perform density analysis on the Gaussian topological domain to identify sparse distribution regions, perform perceptual-driven subdivision operation on the sparse regions to generate a dense Gaussian field, perform opacity allocation processing on the dense Gaussian field to form an opacity control sequence, and construct a multi-level rendering strategy based on the fusion of the opacity control sequence and the geometric correction space. The viewpoint anchoring module is used to project the multi-layered rendering strategy onto the viewpoint sensitivity space to generate key viewpoint intervals. This includes: constructing a three-dimensional viewpoint sensitivity space; mapping the parameters of each layer in the multi-layered rendering strategy to the viewpoint space; and calculating the rendering complexity index C(φ,θ,r)=Σ for each viewpoint position. i W i ·L i Where C is the rendering complexity, φ is the azimuth angle, θ is the pitch angle, r is the distance, and W is the distance to the target value. i Let L be the activation weight of the i-th layer. i To reduce the computational complexity of this layer, gradient analysis is used to identify sensitive regions where rendering complexity changes rapidly, and cluster analysis is performed to group spatially continuous high-sensitivity points into key viewpoint intervals. The importance of these key viewpoint intervals is assessed to identify viewpoint inflection point locations, including: defining an importance function I(v) = Q(v)·F(v)·U(v), where I is the importance value of viewpoint v, Q(v) is the rendering quality score of viewpoint v, F(v) is the predicted access frequency for that viewpoint, and U(v) is the user preference weight. Continuous sampling is performed along the boundary of the viewpoint interval, and the first and second derivatives of the importance function are calculated. The location where the derivative sign changes is identified as the viewpoint inflection point location. The process of generating viewpoint anchor points through stability testing includes: constructing a rendering response matrix based on the viewpoint inflection point positions; performing stability oscillation scanning within the rendering response matrix to identify oscillation suppression regions; performing response smoothing processing on the oscillation suppression regions to generate a stable rendering domain; fixing the viewpoint based on the stable rendering domain to generate viewpoint anchor points; and constructing a multi-view rendering sequence based on the viewpoint anchor points, including: performing spatial trajectory planning based on the viewpoint anchor points to generate an observation path network; performing connectivity analysis on the observation path network to identify critical path nodes; performing viewpoint quality assessment at the critical path nodes to generate a viewpoint scoring matrix; and arranging the viewpoint scoring matrix temporally to generate a multi-view rendering sequence. The quality reconstruction module is used to generate a volume rendering matrix based on the multi-view rendering sequence, optimize the quality of the volume rendering matrix to generate a 3D model of the cultural relic, and perform hash verification and digital signature on the 3D model of the cultural relic to complete the confirmation of ownership.

Images