A method for generating and materializing a three-dimensional model based on perceptual voxelization
By employing a perceptual voxelization method, the problems of incomplete structure and poor visual effects in existing 3D model generation are solved. This method achieves high-density sampling, accurate color quantization, and structural repair, generating editable and printable 3D voxel models.
Patent Information
- Application Number
- CN202610911346.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-24
- Publication Date
- 2026-08-25
AI Technical Summary
Existing 3D model generation technologies, such as polygonal meshes and point clouds, have defects such as holes, self-intersections, and non-manifold surfaces, making it difficult to directly solidify them. Traditional voxelization methods are prone to losing slender structures at low resolutions, morphological repair is prone to causing erroneous adhesion, and color quantization is prone to losing key colors in small areas, reducing visual recognition.
The perceptual voxelization method is adopted, which involves multimodal 3D asset input and analysis, high-density sampling and color extraction, trilinear weighted sputtering and voxel rasterization, combined with mathematical morphological restoration and Flood Fill to identify cavities, conservative solidification processing, and color quantization based on perceptual color space to generate sparse voxel data.
It achieves structural integrity, visual recognizability, and color controllability of 3D models, supports solid assembly and AR building, ensures the integrity of slender structures, avoids erroneous adhesion and loss of visual features, and outputs editable and printable voxel models.
Smart Images
Figure CN122636863A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence 3D model generation technology, specifically a method for generating and solidifying 3D models based on perceptual voxelization. Background Technology
[0002] With the development of AI-generated content, 3D digital content creation, 3D printing, AR-assisted display, and digital physicalization technologies, users' demand for 3D model generation and low-threshold 3D creation is constantly increasing. 3D models typically exist in the form of polygonal meshes, point clouds, neural radiation fields, or 3D Gaussian splatting, and are widely used in games, film and television, virtual reality, digital twins, 3D printing, and creative design.
[0003] In existing 3D content generation technologies, common representation methods such as polygonal meshes and point clouds are difficult to directly convert into voxel structures suitable for 3D printing or solid assembly. The models often have defects such as holes, self-intersections, non-manifold surfaces, and incomplete closures, resulting in incomplete structures or the need for manual repair when outputting solidified models. During voxelization, traditional coordinate rounding or single-voxel assignment methods are prone to losing slender structures such as hair strands, sword blades, and ribbons at low resolutions. Relying solely on vertex or low-density surface sampling can cause missing areas inside triangular facets, resulting in holes, stripes, and breaks. When using morphological repair methods such as dilation and closure operations, isotropic processing can easily cause incorrect adhesion of adjacent structures such as fingers and accessories. The brute-force filling method for internal holes cannot distinguish between true closed cavities and semi-open structures such as cup rims and cuffs, which can easily destroy the original semantics of the model. In addition, traditional color quantization methods (such as K-Means) are easily dominated by large areas of primary color, causing the loss of key high-saturation colors in small areas such as eyes and highlights, reducing the visual recognition of the model.
[0004] Therefore, there is an urgent need for a method to solve the above-mentioned technical problems and to convert various types of 3D assets into 3D voxel models that are structurally complete, color-controllable, editable, exportable, printable, and can be assembled into physical objects. Summary of the Invention
[0005] The purpose of this invention is to provide a method for generating and solidifying 3D models based on perceptual voxelization, in order to solve the problems mentioned in the background art. Existing 3D models are mostly in the form of meshes and point clouds, often containing defects such as holes and self-intersections, making them difficult to directly solidify and requiring manual model repair. Traditional voxelization is prone to losing slender structures at low resolution, and insufficient sampling can easily produce holes and breaks. Morphological repair is prone to causing erroneous adhesion, and solidification can easily destroy semi-open structures. Color quantization is prone to losing key colors in small areas, reducing visual effects.
[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for generating and solidifying a 3D model based on perceptual voxelization, comprising the following steps: S1. Multimodal 3D Asset Input and Parsing: Receives multimodal 3D asset files and converts them into a standardized set of sampled data containing spatial coordinates, colors, normals, and confidence levels. S2. 3D Model Surface Sampling and Color Extraction: High-density sampling is performed on the polygonal mesh model in the 3D asset, and the color value of the sampling point is obtained by interpolating the barycenter coordinates with the UV texture coordinates. S3, 3D spatial normalization and scale constraints: Calculate the bounding box of the sampling points, translate the model to the center of the voxel space and ground the bottom, and then scale it proportionally using a uniform scaling factor to map the model to the target voxel space without deformation. S4. Voxel rasterization based on trilinear weighted sputtering: The sampling points are distributed to eight surrounding voxels by trilinear weighted sputtering. A density threshold is set, and voxels are activated according to the density field to form an initial voxel model. Key structural voxels are retained by combining local maxima to complete voxel rasterization. S5. Voxel structure repair based on mathematical morphology: Perform a three-dimensional neighborhood scan on the initial voxel model in S4, detect the broken areas in the model, perform directional bridging on the broken areas, and supplement voxels. S6. Internal cavity identification based on Flood Fill: Set external air seed points at the voxel space boundary, mark the external reachable area through three-dimensional connectivity propagation, and identify the unreachable blank area as a candidate internal cavity; S7. Conservative solidification based on distance transformation: Set a depth threshold, and fill the closed cavity that meets the conditions with solidification and inherit the surface properties through dual verification of depth threshold and escape path. S8. Color quantization based on perceptual color space: Convert voxel colors to CIELAB space, extract key colors based on frequency and saturation, and generate a finite color palette through clustering to complete color quantization. S9. Sparse Voxel Data Generation and Model Output: Extract the 3D coordinates, color index, and palette information of the activated voxels, convert the voxel model into a sparse data structure, and output the result data.
[0007] The output data includes: voxel coordinate set, color index set, color palette mapping table, model size information, layer structure information, and solidification output parameters.
[0008] Preferably, in S1, a unified 3D data parsing interface is used to receive and read 3D asset files, parse geometric, texture, color, and confidence information according to the format, filter noise, and output standardized 3D sampling data, where the 3D assets include polygon mesh models, point cloud models, and 3D Gaussian Splatting data.
[0009] Preferably, in S2, during high-density sampling, the surface area of each triangular facet is calculated, and the sampling probability is determined according to the proportion of the area of each triangular facet to the total surface area. Triangular facets with larger areas are assigned more sampling points, and triangular facets with smaller areas are assigned fewer sampling points. For each sampling point, the system calculates its centroid coordinates within its respective triangular facet.
[0010] Preferably, in S4, during weight allocation, voxels closer to the sampling point receive higher weights, while voxels farther from the sampling point receive lower weights. Each voxel unit accumulates density values, color values, and normal information from multiple sampling points.
[0011] Preferably, in S4, voxels with a density greater than the density threshold are marked as active. When retaining key structural voxels for local maxima, all non-zero density voxels are traversed and their 26-neighborhoods are checked. If the density of the current voxel is strictly greater than the density of all neighboring voxels, it is forcibly marked as active. For the already activated maximum voxel, voxels with a density higher than the second preset density threshold (i.e., the second lowest threshold) around it are also activated to form a connected region.
[0012] Preferably, in S5, for the candidate fracture region, the spatial orientation relationship and surface normal direction relationship of the voxels on both sides of the fracture are further determined: when determining the spatial orientation relationship, the spatial connection direction of the voxel groups at both ends of the fracture is calculated, and it is determined whether they belong to the extension direction of the same structure. When determining the relationship between surface normal directions, the average surface normals of voxels at both ends of the fracture are extracted, and the angle between the normals is calculated. When the average normal directions of voxels at both ends of the fracture both point to the outside of the fracture area, and the angle between the normals is less than a preset angle threshold, it is determined that the surface normal direction consistency is satisfied. When the voxels on both sides of the fracture are in the same structural extension direction, and their normal directions meet the consistency condition, a directional bridging operation is performed to supplement voxels at the fracture location. When the two sides of the fracture are close to each other, but their directions are inconsistent or they belong to different structures, bridging is not performed.
[0013] Preferably, the formulas for calculating the spatial connection direction and the angle between the normals of the voxel groups at both ends of the fracture are as follows: ; ; In the formula, Let be the direction vector of the unit space line connecting the voxel groups at both ends of the fracture. , , ( ) represents the coordinates of the three-dimensional spatial center point of the voxel set at the left end of the fracture region; , , () represents the coordinates of the three-dimensional spatial center point of the voxel group at the right end of the fracture region; The Euclidean distance between the center points of the voxel groups at both ends of the fracture is denoted as . The first voxel in the left end of the fracture region The surface normal vector corresponding to an individual element; The first voxel in the right end of the fracture region The surface normal vector corresponding to an individual element; Let be the magnitude of the sum of the left-hand normal vectors; Let be the magnitude of the sum of the normal vectors on the right side.
[0014] Preferably, in S6, blank areas accessible by external air are marked as external connected spaces, and blank areas inaccessible by external air are marked as candidate internal cavities, thus distinguishing between the external space of the model and the closed space inside the model.
[0015] Preferably, in S7, the depth threshold verification is performed: if the distance value of the cavity voxel is greater than or equal to the preset depth threshold, it is confirmed that the region is a deep closed region of the model, rather than a surface depression; the escape path verification is performed: for the cavity region, ray detection is performed along the positive and negative directions of the three coordinate axes X, Y, and Z. If there is no straight path that can directly connect to the external connected space, it is determined to be a completely closed region without an escape path.
[0016] Preferably, in S8, during color quantization, highly saturated rare colors and high-frequency primary colors are selected as key color anchors; using the key color anchors as initialization seeds for clustering centers, clustering is performed in the CIELAB space, and the clustering distance adopts the Euclidean distance of the CIELAB space. After the clustering converges, a finite number of final color centers are obtained, and the CIELAB color values of all voxels are mapped to the nearest color center.
[0017] Compared with the prior art, the beneficial effects of the present invention are: 1. In this invention, a unified parsing interface is used to convert 3D assets from different sources into a standardized set of sampling points. Spatial normalization is used to improve the stability and compatibility of converting multi-source 3D assets into voxel models. Trilinear weighted sputtering is used instead of traditional single-voxel assignment or coordinate rounding. Sampling points are distributed to the surrounding eight voxels according to distance weights, which significantly reduces sampling aliasing and ensures that slender structures such as hair, sword blades, and ribbons can still be continuously expressed in low-resolution voxel space, thus ensuring the integrity of the voxel model. High-density sampling points are allocated according to the area ratio of triangular facets. The surface color is accurately extracted by combining the centroid coordinate interpolation with UV texture coordinates. This avoids the loss of internal regions of triangular facets caused by relying solely on vertices or low-density sampling, thereby obtaining complete and continuous surface sampling data. 2. In this invention, when repairing voxel structure fractures, a dual check of spatial direction and normal consistency is combined. Oriented bridging is only performed when both conditions are met, effectively preventing adjacent but different voxels such as fingers and accessories from being incorrectly stuck together. Flood Fill distinguishes between external space and internal cavities. Combined with distance transformation and escape path check, only deep closed cavities are filled, and semi-open structures are preserved without destroying the original semantics of the model. Color quantization prioritizes the retention of highly saturated key colors in the CIELAB space to avoid the loss of visual features due to the dominance of the primary color, ensuring that the quantized model still has high recognizability and visual effect. Sparse voxel data is output, directly supporting layered assembly, block statistics, and AR building guidance, providing structural support for physicalization. Attached Figure Description
[0018] Figure 1 This is a flowchart of a method for generating and solidifying a 3D model based on perceptual voxelization according to the present invention. Detailed Implementation
[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0020] Example: Refer to Figure 1 As shown: A method for generating and solidifying 3D models based on perceptual voxelization, including the following steps: Step 1: Multimodal 3D Asset Input and Parsing: Receive the 3D asset files to be processed and convert 3D assets from different sources into a standardized set of sampled data containing spatial coordinates, color information, normal information, and confidence information. The system receives and reads 3D asset files through a unified 3D data parsing interface, and then parses the corresponding data according to their respective format specifications. The 3D assets include polygon mesh models, point cloud models, 3D Gaussian Splatting data, or other 3D data containing spatial coordinates and color information.
[0021] When parsing a polygonal mesh model, vertex coordinates, triangle patch indices, vertex normals, UV texture coordinates, and basic texture image data are extracted and parsed to establish a geometric-texture mapping relationship.
[0022] When parsing point cloud data, the point cloud file is read, and the three-dimensional spatial coordinates, color values, opacity and confidence parameters are extracted point by point to form a set of point-level attributes.
[0023] During 3D Gaussian Splatting data parsing, the Gaussian parameter file is read, and the center coordinates, opacity, and confidence of each Gaussian element are analyzed. The basic color is obtained by converting the zero-order spherical harmonic coefficient. Low-confidence noise points are filtered to reduce floating point interference, and standardized data containing spatial coordinates, color, opacity, and confidence are output.
[0024] Step 2, 3D model surface sampling and color extraction: High-density sampling is performed on the polygonal mesh model, and the color value of the sampling point is obtained by interpolating the UV texture coordinates through the centroid coordinates; Specifically, the following steps are included: 21. Calculate the area of triangular facets and allocate sampling frequency: For each triangular facet of the polygonal mesh model, calculate its surface area to obtain the area value of each facet, and count the total surface area of all triangular facets; determine the sampling probability according to the proportion of each facet area to the total surface area. The larger the area, the more sampling points are allocated, and the smaller the area, the fewer sampling points are allocated, so that the sampling points evenly cover the model surface.
[0025] 22. Generate sampling points within the triangular facets and calculate the centroid coordinates of the sampling points: Based on the sampling probability, generate high-density sampling points within each triangular facet to obtain a uniformly distributed set of surface sampling points. For each sampling point, solve for its centroid coordinates within its respective triangular facet to determine the relative position of the sampling point within the facet.
[0026] 23. Interpolate UV texture coordinates: Use the centroid coordinates to linearly interpolate the UV texture coordinates of the vertices of the triangular facet to obtain the UV coordinates corresponding to the sampling point.
[0027] 24. Extract the texture color of the sampling point: Based on the UV coordinates obtained by interpolation, read the corresponding pixel color from the texture image and use it as the color value of the sampling point.
[0028] Step 3, 3D spatial normalization and scale constraints: Calculate the bounding box of the sampling points, translate the model to the center of the voxel space and ground the bottom, and then scale it proportionally using a uniform scaling factor to map the model to the target voxel space without deformation. Traverse all sampling points and calculate the minimum and maximum values of the sampling point set in the X, Y, and Z directions to obtain the geometric bounding box of the model. Calculate the horizontal center coordinates of the model based on the bounding box extreme values and combine them with the preset horizontal center coordinates of the target voxel space. Calculate the X and Y axis translation offsets and perform an overall coordinate translation on all sampling points to make the horizontal center of the model coincide with the horizontal center of the target voxel space. In the vertical direction, using the original lowest point coordinates of the model as a reference, calculate the vertical translation on the Z axis and perform a vertical coordinate translation on all sampling points to align the overall lowest point of the model to the bottom reference plane of the target voxel space, completing the model's grounding arrangement. The spatial span of the model in the three coordinate axes is calculated separately, and the span with the largest value among the three is taken as the maximum circumscribed span of the model. The standard accommodation side length of the target voxel space is set, the global uniform scaling factor is calculated, and the three-dimensional coordinates of all sampling points after translation are scaled proportionally in sync using the uniform scaling factor. The three coordinate axes use the same scaling factor, and finally the whole model is completely mapped and contained into the target voxel space of the preset size without distortion or deformation.
[0029] Step 4: Voxel rasterization based on trilinear weighted sputtering: The sampling points are distributed to eight surrounding voxels by trilinear weighted sputtering. A density threshold is set, and voxels are activated according to the density field to form an initial voxel model. Key structural voxels are retained by combining local maxima, thus completing the voxel rasterization.
[0030] Includes the following: 41. Initialize the voxel mesh: Create three cumulative arrays corresponding to the target resolution: density, color, and normal, initially set to zero.
[0031] 42. Perform trilinear sputtering on each sampling point: Convert the sampling point coordinates into floating-point indices in voxel space, split them into integer and fractional parts, calculate the weights of the eight surrounding voxels, give higher weights to voxels closer to the sampling point and lower weights to voxels farther away from the sampling point, and add the density, color, and normal of the sampling point to the corresponding eight voxels according to their weights.
[0032] 43. Calculate the average properties of voxels: After sputtering, the total weight of each voxel is equal to the sum of the weights of all the sampling points it has obtained. If the average color and average normal are needed, divide the accumulated value by the accumulated density and renormalize the normal.
[0033] 44. Density threshold activation: Set a density threshold and mark voxels with a density greater than the threshold as active.
[0034] 45. Local maxima preserve key structural voxels: Traverse all non-zero density voxels and check their 26-neighborhood. If the current voxel density is strictly greater than the density of all neighboring voxels, then forcibly mark it as active. For an already activated maximum voxel, voxels with a density higher than the second lowest threshold can also be activated to form a connected region.
[0035] Step 5: Voxel structure repair based on mathematical morphology: Perform a 3D neighborhood scan on the initial voxel model from Step 4 to detect broken regions in the model, perform directional bridging on the broken regions, and supplement voxels.
[0036] When detecting fracture regions in the model, all voxels of the initial voxel model are traversed, and a 3×3×3 three-dimensional neighborhood window is constructed with each voxel as the center. Fracture region candidate detection is performed: if the current voxel is an inactive empty voxel, and the number of active voxels in its neighborhood window meets the low-density distribution feature of 1-3, and the active voxels are linear / symmetrically distributed, then the empty voxel is marked as a fracture candidate point; connectivity clustering is performed on all fracture candidate points, and adjacent candidate points are merged to form a fracture region candidate set, filtering out isolated noise candidates to avoid invalid calculations; For each candidate set of fracture regions, the core activation voxel groups at both ends are extracted, and a double consistency check is performed: when checking the spatial orientation relationship, the spatial connection direction of the voxel groups at both ends of the fracture is calculated to determine whether they belong to the extension direction of the same structure. The formula for calculating the spatial connection direction of the voxel groups at both ends of the fracture is: ; In the formula, The vector representing the direction of the unit space connection between the voxel groups at both ends of the fracture has a magnitude of 1 and represents the direction of structural extension. , , ( ) represents the coordinates of the three-dimensional spatial center point of the voxel set at the left end of the fracture region; , , () represents the coordinates of the three-dimensional spatial center point of the voxel group at the right end of the fracture region; The distance between the center points of the voxel sets at both ends of the fracture is the Euclidean distance.
[0037] During surface normal direction verification, the average surface normal of the voxels at both ends of the fracture is extracted, and the angle between the normals is calculated. If the angle between the normals is less than a preset threshold and the normals point to the outside of the fracture area (i.e., the fracture belongs to the crack of the same surface, rather than the opposite face of two independent surfaces), then the consistency is considered to pass. If either of the two verifications fails, it is considered to be a "near voxel of a different structure", and the fracture area is skipped directly without bridging to avoid erroneous adhesion.
[0038] The formula for calculating the included angle between normals is: ; In the formula, The first voxel in the left end of the fracture region The surface normal vector corresponding to an individual element; The first voxel in the right end of the fracture region The surface normal vector corresponding to an individual element; Let be the magnitude of the sum of the left-hand normal vectors; Let be the magnitude of the sum of the normal vectors on the right side.
[0039] For the fractured regions that pass the verification, directional bridging repair is performed. A linear interpolation sequence of bridging voxels is generated along the extension direction of the voxel groups at both ends of the fracture. For each bridging voxel, the weighted color and normal attributes of the voxels at both ends of the fracture are inherited to ensure the visual consistency between the repaired region and the original model. The bridging voxels are marked as active voxels and added to the initial voxel model. After traversing all fractured regions, the repaired voxel model is output.
[0040] Step 6: Identification of internal cavities based on Flood Fill: Set external air seed points at the boundary of the target voxel space, and mark the areas accessible by external air as external connected space through three-dimensional connectivity propagation, and mark the unreachable blank areas as candidate internal cavities; Traverse the outermost boundary voxels of the target voxel space, filter out the inactive empty voxels, and mark them all as external air seed points as the starting points for connectivity propagation.
[0041] Starting from all external air seed points, the three-dimensional 6-neighborhood connectivity rule is used to perform connectivity propagation: if the currently traversed neighboring voxel is an inactive empty voxel and has not been marked, it is marked as an external connected space and added to the propagation queue to continue spreading outward; if the neighboring voxel is an active model voxel, the propagation in that direction is stopped and no marking is done.
[0042] After completing the full-space Flood Fill propagation, a second traversal is performed on all unlabeled empty voxels in the target voxel space: all empty voxels that are not marked as external connected spaces and are not activated as model voxels are identified as candidate internal cavities; adjacent candidate cavity voxels are merged to form an independent set of internal cavity connected regions, thus completing the distinction between the external space of the model and the internal closed cavities.
[0043] Step 7: Conservative solidification based on distance transformation: Set a depth threshold, perform a three-dimensional Euclidean distance transformation on the candidate internal cavity, and fill the closed cavity that meets the conditions with solidification and inherit the surface properties through dual verification of depth threshold and escape path. The formula for calculating the distance of each candidate cavity voxel to the nearest model surface is as follows: ; in, The current cavity voxel coordinates, The calculated distance value is assigned to the corresponding cavity voxel as the coordinate of the nearest model surface voxel, forming a distance field in the cavity region that characterizes the "embedding depth" of the voxel.
[0044] A cavity is considered a true internal cavity that needs to be filled only if both of the following conditions are met simultaneously; otherwise, it is considered an internal void of a semi-open structure and is retained: (1) Depth threshold verification: Distance value of cavity voxels The depth must be greater than or equal to a preset depth threshold to ensure that the area is a deep, enclosed region of the model, rather than a surface depression. (2) Escape path verification: For the cavity area, ray detection is performed along the positive and negative directions of the three coordinate axes X, Y and Z. If there is no straight path that can directly connect to the external space, it is determined to be a completely closed area without an escape path.
[0045] Solidification is performed on the real internal cavity of the dual-condition verification: all voxels in the cavity are marked as active solid voxels to complete the solidification of the model; for each newly filled solid voxel, the nearest neighbor color inheritance rule is used to inherit the color, normal and other attributes from the nearest model surface voxel. Step 8: Color quantization based on perceptual color space: Convert voxel colors to CIELAB color space, select key colors based on color frequency and saturation and retain them first, generate a limited number of color centers through clustering, and complete voxel color quantization. Iterate through the RGB color values of all active voxels and convert them from the RGB color space to the CIELAB perceptual color space; count the frequency and saturation of each color, and select high-saturation rare colors and high-frequency primary colors as key color anchors based on color rarity and saturation; prioritize the retention of colors with low frequency but high saturation. The initialization seed is based on the key color anchor point as the cluster center. Clustering is performed in CIELAB space. The clustering distance is Euclidean distance in CIELAB space. After the clustering converges, a finite number of final color centers are obtained. The CIELAB color values of all voxels are mapped to the nearest color center to complete the color quantization. Finally, the quantized colors can be converted back to RGB space and a voxel model with a finite color palette is output.
[0046] Step 9, Sparse Voxel Data Generation and Model Output: Extract the 3D coordinates, color index, and palette information of all activated voxels; convert the voxel model into a sparse data structure that only stores the actual position and color information of the voxels, and output the result data.
[0047] The resulting data includes: voxel coordinate set, color index set, color palette mapping table, model size information, layer structure information, and solidification output parameters.
[0048] The output can be used for voxel editing, model preview, 3D printing, magnetic block assembly, AR-assisted building, or other 3D content creation scenarios.
[0049] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for generating and solidifying 3D models based on perceptual voxelization, characterized in that: Includes the following steps: S1. Multimodal 3D Asset Input and Parsing: Receives multimodal 3D asset files and converts them into a standardized set of sampled data containing spatial coordinates, colors, normals, and confidence levels. S2. 3D Model Surface Sampling and Color Extraction: High-density sampling is performed on the polygonal mesh model in the 3D asset, and the color value of the sampling point is obtained by interpolating the barycenter coordinates with the UV texture coordinates. S3, 3D spatial normalization and scale constraints: Calculate the bounding box of the sampling points, translate the model to the center of the voxel space and ground the bottom, and then scale it proportionally using a uniform scaling factor to map the model to the target voxel space without deformation. S4. Voxel rasterization based on trilinear weighted sputtering: The sampling points are distributed to eight surrounding voxels by trilinear weighted sputtering. A density threshold is set, and voxels are activated according to the density field to form an initial voxel model. Key structural voxels are retained by combining local maxima to complete voxel rasterization. S5. Voxel structure repair based on mathematical morphology: Perform a three-dimensional neighborhood scan on the initial voxel model in S4, detect the broken areas in the model, perform directional bridging on the broken areas, and supplement voxels. S6. Internal cavity identification based on Flood Fill: Set external air seed points at the voxel space boundary, mark the external reachable area through three-dimensional connectivity propagation, and identify the unreachable blank area as a candidate internal cavity; S7. Conservative solidification processing based on distance transformation: Set a depth threshold, and fill the closed cavity that meets the conditions with solidification and inherit the surface properties through dual verification of depth threshold and escape path. S8. Color quantization based on perceptual color space: Convert voxel colors to CIELAB space, extract key colors based on frequency and saturation, and generate a finite color palette through clustering to complete color quantization. S9. Sparse Voxel Data Generation and Model Output: Extract the 3D coordinates, color index, and palette information of the activated voxels, convert the voxel model into a sparse data structure, and output the result data.
2. The method for generating and solidifying 3D models based on perceptual voxelization according to claim 1, characterized in that: In S1, 3D asset files are received and read through a unified 3D data parsing interface. The geometry, texture, color, and confidence information are parsed according to the format, noise is filtered, and standardized 3D sampling data is output.
3. The method for generating and solidifying 3D models based on perceptual voxelization according to claim 1, characterized in that: In S2, during high-density sampling, the surface area of each triangular facet is calculated, and the sampling probability is determined based on the proportion of each triangular facet's area to the total surface area. Triangular facets with larger areas are assigned more sampling points, while triangular facets with smaller areas are assigned fewer sampling points. For each sampling point, the system calculates its centroid coordinates within its respective triangular facet.
4. The method for generating and solidifying 3D models based on perceptual voxelization according to claim 1, characterized in that: In S4, during weight allocation, voxels closer to the sampling point receive higher weights, while voxels farther from the sampling point receive lower weights. Each voxel unit accumulates density values, color values, and normal information from multiple sampling points.
5. The method for generating and solidifying 3D models based on perceptual voxelization according to claim 1, characterized in that: In S4, voxels with a density greater than the density threshold are marked as active. When retaining key structural voxels for local maxima, all non-zero density voxels are traversed and their 26-neighborhoods are checked. If the density of the current voxel is strictly greater than the density of all neighboring voxels, it is forcibly marked as active. For the already activated maximum voxel, voxels with a density higher than the second preset density threshold (i.e., the second lowest threshold) around it are also activated to form a connected region.
6. The method for generating and solidifying 3D models based on perceptual voxelization according to claim 1, characterized in that: In S5, for candidate fracture regions, the spatial orientation relationship and surface normal orientation relationship of voxels on both sides of the fracture are further determined: when determining the spatial orientation relationship, the spatial connection direction of the voxel groups at both ends of the fracture is calculated to determine whether they belong to the extension direction of the same structure; when determining the surface normal orientation relationship, the average surface normal of the voxels at both ends of the fracture is extracted and the normal angle is calculated. When the average normal direction of the voxels at both ends of the fracture points to the outside of the fracture region and the normal angle is less than the preset angle threshold, it is determined that the surface normal orientation consistency is satisfied. When the voxels on both sides of the fracture are in the same structural extension direction and their normal directions meet the consistency condition, a directional bridging operation is performed to supplement the voxel at the fracture location; when the two sides of the fracture are close to each other but have different directions or belong to different structures, bridging is not performed.
7. The method for generating and solidifying 3D models based on perceptual voxelization according to claim 6, characterized in that: The formulas for calculating the spatial direction of the voxel groups at both ends of the fracture and the included angle with the normal are as follows: ； ； In the formula, Let be the direction vector of the unit space line connecting the voxel groups at both ends of the fracture. , , ( ) represents the coordinates of the three-dimensional spatial center point of the voxel set at the left end of the fracture region; , , () represents the coordinates of the three-dimensional spatial center point of the voxel group at the right end of the fracture region; The Euclidean distance between the center points of the voxel groups at both ends of the fracture is denoted as . The first voxel in the left end of the fracture region The surface normal vector corresponding to an individual element; The first voxel in the right end of the fracture region The surface normal vector corresponding to an individual element; Let be the magnitude of the sum of the left-hand normal vectors; Let be the magnitude of the sum of the normal vectors on the right side.
8. The method for generating and solidifying 3D models based on perceptual voxelization according to claim 1, characterized in that: In S6, blank areas accessible by external air are marked as external connected spaces, while blank areas inaccessible by external air are marked as candidate internal cavities, thus distinguishing between the external space and the closed internal space of the model.
9. The method for generating and solidifying 3D models based on perceptual voxelization according to claim 1, characterized in that: In S7, depth threshold verification: if the distance value of the cavity voxel is greater than or equal to the preset depth threshold, it is confirmed that the region is a deep closed region of the model, rather than a surface depression; escape path verification: for the cavity region, ray detection is performed along the positive and negative directions of the three coordinate axes X, Y, and Z. If there is no straight path that can directly connect to the external connected space, it is determined to be a completely closed region with no escape path.
10. The method for generating and solidifying 3D models based on perceptual voxelization according to claim 1, characterized in that: In S8, during color quantization, highly saturated rare colors and high-frequency primary colors are selected as key color anchors. Using the key color anchors as initialization seeds for clustering centers, clustering is performed in CIELAB space. The clustering distance adopts the Euclidean distance of CIELAB space. After the clustering converges, a finite number of final color centers are obtained. The CIELAB color values of all voxels are mapped to the nearest color center.