Three-dimensional model digital collection high-precision reconstruction and credible asset identification method

Abstract

The invention belongs to the crossing field of digital image processing, three-dimensional model reconstruction and block chain technologies, and particularly relates to a three-dimensional model digital collection high-precision reconstruction and credible asset identification method, and the method comprises the specific steps: carrying out the scanning through employing corresponding equipment according to the characteristics of a cultural relic, and obtaining point cloud data; two-channel CNN feature extraction is divided into a geometric channel and a texture channel, the geometric channel extracts space geometric shape information of the point cloud, and the texture channel extracts texture information of the surface of the model; based on a self-adaptive attention weight mechanism, a division reconstruction strategy is adopted, texture information and form information are reconstructed, and a triangular surface model is generated; main feature points are screened out from the reconstructed model, coordinates and normal vectors of the main feature points are extracted to generate feature hash, the feature hash is stored by adopting a mixed chain storage architecture, and an intelligent contract for transaction constraint is set.

Description

TECHNICAL FIELD

[0001] The application belongs to the cross field of digital image processing, three-dimensional model reconstruction and blockchain technology, and particularly relates to a three-dimensional model digital collection high-precision reconstruction and credible asset identification method, which is suitable for cultural heritage digitization protection, digital collection assetization circulation and other application scenarios. BACKGROUND

[0002] In the field of three-dimensional digitization of cultural heritage and circulation of digital collections, high-precision model reconstruction and credible identification are core technical requirements. The existing technology has the following defects:

[0003] 1. Insufficient reconstruction precision and scene adaptability: traditional reconstruction technology uses a fixed weight algorithm, with an error > 0.3 mm; weak light in grottoes, bronze reflectivity and other environments easily produce point cloud noise, further amplifying distortion. For example, the patent technology "Three-dimensional reconstruction method of cultural relics based on monocular vision", the reconstruction error is 0.3 mm, which cannot meet the high-precision requirement.

[0004] 2. Poor identification stability: the existing "file overall hash" method fails to identify after model simplification. For example, the patent technology "Digital collection hash identification method", the identification consistency is only 60% after model simplification, which cannot support cross-scene circulation.

[0005] 3. Insufficient universality: existing solutions are mostly targeted at a single type of cultural relics, without covering scenes such as murals and damaged cultural relics, and key parameters lack experimental basis. SUMMARY

[0006] Therefore, the application provides a three-dimensional model digital collection high-precision reconstruction and credible asset identification method, which solves the precision and credibility problems of complex digital collections such as grotto Buddha statues through the collaborative design of "feature hierarchical reconstruction + stable feature point identification", and achieves the technical goals of high-precision reconstruction, stable identification and cross-scene adaptability.

[0007] The technical scheme of the application is as follows:

[0008] A three-dimensional model digital collection high-precision reconstruction and credible asset identification method, the specific process is as follows:

[0009] Data acquisition and environment adaptation preprocessing: according to the characteristics of cultural relics, corresponding equipment is used to scan and obtain point cloud data, and environment adaptation processing is performed during the scanning process; the point cloud data is subjected to adaptive filtering and normalization processing, and when multiple devices are used, the point cloud data obtained by the multiple devices needs to be fused before adaptive filtering;

[0010] Differential feature-driven model reconstruction: divide the dual-channel CNN feature extraction into a geometric channel and a texture channel, the geometric channel extracts the spatial geometric shape information of the point cloud, and the texture channel extracts the texture information of the model surface; based on the adaptive attention weight mechanism, a division reconstruction strategy is adopted to reconstruct the texture information and the shape information, and generate a triangular surface model;

[0011] Hybrid chain trust identification: select the main feature points from the reconstructed model, extract the coordinates and normal vectors of the main feature points to generate a feature hash, store the feature hash using a hybrid chain storage architecture, and set an intelligent contract for transaction constraints.

[0012] Optionally, according to the characteristics of cultural relics, the corresponding equipment is used for scanning, specifically: when the height of cultural relics is greater than 3m and the surface curvature change is greater than 0.5mm -1 , the area ratio is greater than 30%, a laser scanner and a structured light scanner are combined for scanning; when the height of cultural relics is less than 3m or the area ratio is less than 30%, a structured light scanner is used alone for scanning; through a multi-device cooperative acquisition strategy, point cloud data of cultural relics is obtained.

[0013] When the cultural relics are murals, the three-dimensional structure of the wall where the murals are located also needs to be scanned synchronously to preserve the texture spatial position information.

[0014] Optionally, the environmental adaptation processing of the present application includes weak light processing and reflection / shading processing

[0015] Weak light processing: set a white light light supplement device, and process the point cloud data by combining the Retinex algorithm;

[0016] Reflection / shading processing: use a polarized light filter to collect the reflection area of bronze wares; fuse point clouds from more than 3 perspectives in the shading area, and complete the missing part by 4th order polynomial interpolation, and mark the "interpolation supplement" data.

[0017] Optionally, when multiple devices are used in the present application, the point cloud data obtained is fused, and the specific process is as follows:

[0018] Calculate the coordinate conversion matrix between devices through a chessboard calibration plate, input the primary converted multi-source point cloud into the ICP algorithm for secondary calibration using the conversion matrix, optimize the conversion matrix parameters by minimizing the Euclidean distance error of corresponding points of the point cloud, and finally realize multi-source point cloud fusion.

[0019] Optionally, the specific process of the adaptive filtering of the present application is as follows: according to the degree of change of the geometric shape of the surface of cultural relics, it is divided into high curvature area, intermediate area and flat area, and different proportions of feature points are retained for different areas.

[0020] Optionally, the geometric channel of the present application adopts a 3x3x3 convolution kernel to output a 64-channel curvature gradient feature and performs a batch normalization operation, and the texture channel adopts a 5x5x5 convolution kernel to output a 32-channel texture entropy feature and is subjected to a ReLU activation process.

[0021] Optionally, the present application adopts a split reconstruction strategy based on an adaptive attention weight mechanism to reconstruct the model for texture information and morphological information.

[0022] Firstly, a pre-trained classifier is used to identify the type of the current artifact, and basic weights are allocated for the core area and secondary area of different types of artifacts;

[0023] Secondly, the basic weights are locally adjusted by calculating the feature response intensity of the core area. If the feature response intensity is greater than a preset threshold, the basic weight is increased. If the feature response intensity is less than the threshold, the basic weight is maintained.

[0024] Finally, the reconstruction level is divided according to the optimized weight values, different order polynomials are used for different levels, and the model is reconstructed for texture information and morphological information.

[0025] Optionally, the present application selects main feature points from the reconstructed model and extracts the coordinates and normal vectors of the main feature points to generate a feature hash, the specific process being:

[0026] The point cloud coordinate points that can uniquely represent the three-dimensional morphology and texture core features of the artifact are regarded as main feature points. If the number of extracted main feature points is less than the set number, a "point cloud + image" dual carrier adaptation method is used to supplement. The "point cloud + image" dual carrier adaptation method is as follows:

[0027] ① Image matching: call the historical document auxiliary module, input the historical image before the artifact is damaged, and perform feature matching between the historical image and the point cloud texture mapping image of the current damaged artifact by SIFT algorithm. Image feature points with a matching degree of ≥80% are regarded as effective references.

[0028] ② Virtual feature point generation: based on the matched image feature points, combined with the point cloud geometric trend of the damaged area, virtual feature points are generated. The proportion of virtual feature points is ≤30%, and the Euclidean distance from the existing main feature points is ≤0.5mm.

[0029] ③ Metadata labeling: when virtual feature points are stored on the chain, they need to be labeled with the 'virtually generated' attribute and the basis for generation to ensure the transparency of asset identification.

[0030] Optionally, the present application stores the feature hash in a hybrid chain storage architecture and sets a smart contract for transaction constraints, the specific process being:

[0031] Consortium blockchains store feature hashes and the creator's public key; public blockchains store extended metadata and consortium blockchain hash pointers.

[0032] Cross-chain synchronization: Synchronization is triggered at set times and uses the HTLC protocol to ensure the atomicity of data transmission between consortium blockchains and public blockchains;

[0033] Smart contract verification: Built-in feature point comparison algorithm: When the vertex change rate is >5% or the main feature point deviation is >0.01mm, the transaction is rejected and a hash update is triggered.

[0034] Optionally, the present invention also includes adaptive post-processing and verification steps, including: terminal adaptation simplification and API verification interface;

[0035] Terminal adaptation simplification: Based on the GPU memory size, an edge collapse algorithm is used to simplify the regions outside the non-simplified areas of the model.

[0036] API verification interface: It adopts a three-level permission system, that is, cultural relics protection institutions can view all data, partner institutions can view hash and verification results, and ordinary users can only view public metadata.

[0037] Beneficial effects

[0038] First, the reconstruction accuracy is significantly improved: compared with patent CN114283251A, the reconstruction error of dual-channel CNN is reduced from 0.3mm to 0.08mm (73% reduction compared with the existing technology), and the signal-to-noise ratio in low light environment is improved to more than 15dB, which meets the accuracy requirements of cultural relic digitization.

[0039] Second, enhanced stability of the identifier: Compared with the file hashing method of patent CN115083672B, the hash identifier based on the main feature point of this invention maintains 100% consistency after 5 simplifications (the existing technology only achieves 60%), solving the problem of disconnect between on-chain and off-chain.

[0040] Third, scene adaptability optimization: It supports multiple types of cultural relics such as grotto Buddha statues, bronzes, and murals, and damaged cultural relics can be identified by virtual feature points;

[0041] Fourth, practical application verification: In a digital project of Buddhist statues in a Northern Wei grotto, cultural relics experts improved the model detail matching score from 6.2 points (traditional method) to 9.5 points (out of 10). The texture restoration error of the Tang Dynasty mural project was 0.06mm. Both meet the needs of digital collection circulation and support flexible display on multiple terminals.

[0042] Fifth, production efficiency optimization: Regional calculation reduces processing time from 12 hours to 9.6 hours, meeting the needs of mass production. Attached Figure Description

[0043] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a flowchart of the technical solution of the present invention;

[0045] Figure 2 This is a schematic diagram of the dual-channel CNN structure of the present invention;

[0046] Figure 3 The stability verification diagram of feature points (coordinate deviation of 128 feature points after 5 simplifications, all <0.01mm). Detailed Implementation

[0047] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0048] It should be noted that, in the absence of conflict, the following embodiments and features can be combined with each other; and, based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0049] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0050] like Figure 1 As shown in the embodiment of this application, a method for high-precision reconstruction and reliable asset identification of a three-dimensional model digital collection is described, and the specific process is as follows:

[0051] Step 1, Data Acquisition and Environment Adaptation Preprocessing: Point cloud data is obtained by scanning with appropriate equipment based on the characteristics of the cultural relic, and environment adaptation processing is performed during the scanning process. Adaptive filtering and normalization are then applied to the point cloud data. When multiple devices are used, the point cloud data acquired by multiple devices needs to be fused before adaptive filtering. The specific process of this step is as follows:

[0052] 1. Multi-device collaborative data acquisition strategy

[0053] Equipment combination rules: The cultural relic is greater than 3m in height and the surface curvature change is greater than 0.5mm. -1 When the area occupied by the artifact is greater than 30%, a combination of Leica BLK360 laser scanner (0.6mm accuracy within 50m) and Artec Eva structured light scanner (0.1mm accuracy within 30-150cm) is used for data acquisition; when the height of the artifact is less than 3m or the area occupied by the artifact is less than 30%, Artec Eva scanner is used alone; point cloud data of the artifact is obtained through a multi-device collaborative acquisition strategy.

[0054] Mural adaptation: Simultaneously scan the three-dimensional structure of the wall where the mural is located to preserve the spatial location information of the texture.

[0055] 2. Environmental adaptation methods

[0056] Low-light processing: A 550nm white light supplementary illumination device (illuminance 500-1000 lux) combined with the Retinex algorithm was used to improve the signal-to-noise ratio of the point cloud data to over 15dB. Through testing in 10 low-light environments (illuminance < 100 lux), the average signal-to-noise ratio of the point cloud before processing was 8.2dB, and after processing, it reached an average of 16.5dB, meeting the signal quality requirements for subsequent feature extraction. The selection of 550nm white light supplementary illumination was based on comparative testing of three wavelengths: 450nm, 550nm, and 650nm. 550nm achieved the optimal balance between artifact preservation and data quality.

[0057] Reflection / Occlusion Processing: Reflective areas of bronze artifacts were captured using a polarizing filter (65% noise reduction); occluded areas were fused with point clouds from more than 3 viewpoints, and missing parts were completed using 4th-order polynomial interpolation (85% completion accuracy), with "interpolation supplementation" data marked.

[0058] 3. Point cloud fusion and filtering

[0059] Fusion Calibration: For multi-source point cloud data collected by multiple devices (such as laser scanners and structured light scanners), a coordinate transformation matrix between devices is calculated using a 1m×1m checkerboard calibration board (accuracy ±0.01mm). This matrix is ​​used to unify point clouds collected by different devices into the same coordinate system; the matrix dimension is 4×4, including translation and rotation parameters. First, the transformation matrix is ​​initially calculated based on the calibration board data. Then, the multi-source point cloud data after the initial transformation is input into the ICP (Iterative Closest Point) algorithm for secondary calibration. By minimizing the Euclidean distance error between corresponding points in the point cloud (target error ≤0.05mm), the transformation matrix parameters are optimized, ultimately achieving a multi-source point cloud fusion error ≤0.05mm. The initially calculated transformation matrix serves as the initial input parameter for the ICP algorithm, which achieves higher precision calibration by iteratively updating the matrix parameters.

[0060] Adaptive filtering: First, define the region curvature attribute—high curvature regions refer to areas where the geometric shape of the artifact surface changes drastically (such as the folds of clothing on a Buddha statue in a grotto or the edges of inscriptions on bronze artifacts). Then, calculate the curvature value of the point cloud within the local region (curvature > 0.5mm). -1 Judgment: A flat area (low curvature area) refers to an area on the surface of an artifact where the geometric shape changes gently (such as the back plane of a Buddha statue in a grotto or the plane of a wall where a mural is located). The criterion for judgment is a curvature < 0.1mm. -1 For points in the intermediate region, the proportion of their local neighborhood (e.g., k-nearest neighbors) belonging to high-curvature regions (>0.5) and low-curvature regions (<0.1) is calculated. If the proportion of high-curvature points in the neighborhood exceeds a threshold (e.g., 50%), the point is considered closer to a high-curvature region and processed using a high-curvature strategy. If the proportion of low-curvature points is higher, a low-curvature strategy is applied. If the proportions are close (e.g., 40%–60%), a weighted average filter is used. Using a 1mm×1mm×1mm cube as the local region unit, the point density is calculated using the K-nearest neighbor method (taking the nearest 20 points). High-curvature regions (point density > 50 points / mm) are considered. 2 Retain 90% of the dots (to ensure no loss of detail), and maintain a density of <10 dots / mm in flat areas. 2 60% of the points are retained (to reduce redundant data); the noise ratio of the original point cloud with 150 million points is reduced to 6.7% after processing.

[0061] The coordinates are normalized to the interval [-1, 1].

[0062] Step 2: Differentiated Feature-Driven Reconstruction: The dual-channel CNN feature extraction is divided into geometric channels and texture channels. The geometric channel extracts the spatial geometric morphology information of the point cloud, and the texture channel extracts the texture information of the model surface. Based on the adaptive attention weight mechanism, a reconstruction strategy is adopted to realize the reconstruction of texture information and morphological information.

[0063] 1. Dual-channel CNN feature extraction

[0064] (1) Channel segmentation logic: Based on the core information dimensions of the 3D model of the cultural relic, the CNN feature extraction is divided into geometric channels and texture channels—the geometric channel focuses on the spatial geometric morphology information of the point cloud, and the texture channel focuses on the texture information of the model surface, realizing the comprehensive extraction of 'geometric + texture' dual-dimensional features, such as Figure 2 As shown;

[0065] (2) Geometric channel: A 3×3×3 convolution kernel (stride 1) is used to perform convolution operation on the voxelized point cloud data to output 64 channels of curvature gradient features (referring to the rate of curvature change between adjacent voxels, used to characterize the geometric details of the surface of cultural relics such as the concavity, folds, etc., such as the facial contour of Buddha statues and the turning point of bronze vessel shape).

[0066] (3) BatchNormalization processing: The 64-channel feature map output by the geometric channel is subjected to batch normalization, specifically by calculating the mean (μ) and variance (σ) of each feature channel in the batch samples (batchsize=32). 2 According to the formula (where ε = 1e-5, to avoid denominator 0) Normalize the feature values ​​so that the feature distribution satisfies mean 0 and variance 1; update the mean and variance every 100 samples. This operation can accelerate the convergence of CNN model training and avoid the gradient vanishing problem.

[0067] (4) Feature relationship: The texture entropy feature and the curvature gradient feature of the geometric channel are complementary. The curvature gradient feature describes the "spatial morphology", while the texture entropy feature describes the "surface texture". The combination of the two can fully characterize the three-dimensional morphology and surface information of the cultural relic, avoiding the loss of details caused by a single feature.

[0068] (5) ReLU activation: ReLU (Rectified Linear Unit) activation is performed on the 32-channel feature map output by the texture channel. The function expression is f(x) = max(0,x), which means retaining pixels with positive feature values ​​and suppressing negative feature values ​​(considered as noise). This operation can enhance the feature response of texture edges (such as the edges of inscription strokes and the boundaries of mural color blocks) and improve the detail accuracy of subsequent model reconstruction.

[0069] Geometric channels: Employs a 3×3×3 convolution kernel (stride 1) to output 64 channels of curvature gradient features; processed by BatchNormalization (mean 0, variance 1, ε = 1e-5), updating the mean (μ) and variance (σ) every 100 samples. 2 );

[0070] Texture Channels: A 5×5×5 convolution kernel (stride 2) is used to perform convolution operations on the model surface texture image (generated from point cloud texture mapping), outputting 32-channel texture entropy features (referring to the information complexity of the texture region, used to characterize texture details such as the distribution of color blocks in murals and the texture of bronze inscriptions);

[0071] Convolution kernel optimization verification: Through comparative experiments of four sizes, 2×2×2, 3×3×3, 5×5×5, and 7×7×7, the combination of 3×3×3 and 5×5×5 showed the best performance in terms of reconstruction accuracy (0.08mm), detail retention rate (95%), and computational efficiency (1.8 seconds / region).

[0072] Channel count verification: The 64+32 channel combination achieves the optimal balance between facial error (0.08mm) and processing time (9.6 hours).

[0073] 2. Adaptive attention weight fusion

[0074] Dynamic allocation rules: Based on the ResNet18 classifier (the pre-trained dataset contains 10 types of cultural relics samples, including grotto Buddha statues, bronzes, and murals, with a classification accuracy of 97%), the type of cultural relic being processed is identified. First, basic weights (initial weights for the core and secondary areas of the cultural relic) are allocated: grotto Buddha statue face (core area) basic weight 0.6, clothing folds (secondary area) basic weight 0.3; bronze inscription (core area) basic weight 0.6, vessel shape (secondary area) basic weight 0.3; mural texture color blocks (core area) basic weight 0.6, wall structure (secondary area) basic weight 0.3.

[0075] Weight Iterative Optimization: Local adjustment of basic weights—by calculating the feature response intensity of the core region (such as the curvature gradient value of facial feature points, the texture entropy value of the inscription region), if the feature response intensity > a preset threshold (such as curvature gradient > 1.2mm), the weights are adjusted accordingly. -1 If the texture entropy is greater than 1.0, the base weight is increased to 0.8-0.9 (considered as a 'high-priority core region'); if the feature response intensity is less than the threshold, the base weight is maintained (0.3-0.6).

[0076] Weight range verification: 10 sets of tests showed that the optimal range for core feature weights is 0.5-0.9. Weights < 0.5 will result in loss of details (such as blurred facial features of a Buddha statue), while weights > 0.9 will amplify noise (such as color interference in texture channels).

[0077] 3. Layered Poisson Reconstruction: Based on the optimized weight values, the reconstruction layers are divided as follows: High-priority core areas with weights > 0.8 (such as key facial features of Buddha statues and main inscriptions on bronze artifacts) are reconstructed using an 8th-order polynomial (precision priority, 20% higher accuracy than a 4th-order polynomial); ordinary areas with weights between 0.3 and 0.8 (such as drapery of Buddha statues and mural walls) are reconstructed using a 4th-order polynomial (efficiency priority); non-core areas with weights < 0.3 (such as the background of cultural relics) are reconstructed using a 2nd-order polynomial (simplification priority); the error in high-curvature areas is controlled to be ≤ 0.08 mm through curvature constraints.

[0078] The selection criteria for 8th-order polynomials versus 4th-order polynomials: 8th-order polynomials improve accuracy by 20% in high-weight regions (mural texture color blocks) compared to 4th-order polynomials, but increase computation time by 35%. Therefore, they are only used in regions with weights > 0.8 (verified by 10 sets of mural tests).

[0079] Multi-resolution optimization: Gradually increase from 1 / 16 resolution (37500 triangles), randomly sample 200 points in each region to calculate error; apply a relaxation factor of 0.1-0.15 to core features (tested to be the optimal range for accuracy and efficiency);

[0080] The optimal number of triangles was determined: a model with 2.5 million triangles (storage of about 500MB) has a loading time of ≤5 seconds on VR terminals, and the core feature deviation is ≤0.02mm after simplification to 500,000 triangles on mobile terminals.

[0081] Step 3: Hybrid Chain Trustworthiness Identification: Select the main feature points from the reconstructed point cloud model, extract the coordinates and normal vectors of the main feature points to generate feature hashes, store the feature hashes using a hybrid chain storage architecture, and set up smart contracts for transaction constraints.

[0082] 1. Feature point selection and hash generation

[0083] Definition of principal feature point: A principal feature point refers to a point cloud coordinate point that can uniquely represent the core features of the three-dimensional shape and texture of an artifact. It must meet the following conditions: The principal feature point of a three-dimensional artifact (such as a Buddha statue or bronze ware) must be located at a curvature > 0.5 mm. *1 The geometric key regions (such as the brow ridge of a Buddha statue, the intersection of bronze inscriptions), the main feature points of the mural must be located in the texture key regions with texture entropy > 0.8 (such as the boundary of mural color blocks, the outline points of patterns), and the deviation of all main feature points after 5 simplifications of the model must be ≤ 0.01mm. Figure 3 As shown; the curvature is >0.5mm *1 The selection criteria were as follows: tests on 10 sets of three-dimensional cultural relics showed that this threshold could cover 98% of the core feature points, which is less than 0.4mm. *1 Redundant points will be introduced; the proportion of virtual feature points is ≤30% and the Euclidean distance from the main feature point is ≤0.5mm; the test basis for the proportion of virtual feature points ≤30% is: when it exceeds 30%, the consistency of the 10 groups of samples is <60%, which cannot meet the circulation requirements.

[0084] Main feature point selection: 128 main feature points are distributed in a core:minor ratio of 3:2 (77 core main feature points and 51 minor main feature points). They are automatically extracted by the SIFT3D feature detector of the PCL (PointCloudLibrary) point cloud processing library. After extraction, false detection points are manually verified and removed (verification pass rate ≥95%).

[0085] Adaptation for Damaged Cultural Relics: For damaged cultural relics (such as Buddha statues missing faces or broken bronzes), if the number of automatically extracted main feature points is less than 50, a dual-carrier adaptation method of "point cloud + image" is adopted.

[0086] ① Image matching: Call the historical document auxiliary module, input historical images of cultural relics before they were damaged (such as photos of cultural relics in archaeological excavation reports), and use the SIFT algorithm to perform feature matching between the historical images and the point cloud texture mapping map of the current damaged cultural relics. Image feature points with a matching degree of ≥80% are considered as valid references.

[0087] ② Virtual feature point generation: Based on the matched image feature points, combined with the point cloud geometric trend of the damaged area (such as by fitting the curvature gradient of adjacent intact areas), virtual feature points are generated; the proportion of virtual feature points is ≤30% (when it exceeds 30%, the consistency of the 10 groups of sample identifiers is <60%, which cannot meet the circulation requirements), and the Euclidean distance with the existing main feature points is ≤0.5mm (to ensure spatial correlation).

[0088] ③ Metadata annotation: Virtual feature points must be annotated with the 'virtual generation' attribute and generation basis (such as historical image number, fitting algorithm parameters) when stored on the blockchain to ensure the transparency of asset identification.

[0089] Hash generation: Extract feature point coordinates (keeping 6 decimal places) and normal vector (keeping 4 decimal places) to generate a 256-bit SHA-3 hash; comparative tests show that SHA-3 has approximately 20% better collision resistance than SHA-256;

[0090] Hash update mechanism: The hash is automatically updated when the deviation of the main feature point is >0.01mm; after the virtual feature point is repaired to correspond to the physical object (deviation <0.01mm), it is updated synchronously and the historical version is retained.

[0091] 2. Hybrid Chain Storage Architecture

[0092] Consortium blockchain (Hyperledger Fabric) stores feature hashes and the creator's public key (written only by authorized institutions); public blockchain (Polygon) stores extended metadata and consortium blockchain hash pointers (metadata capacity ≤ 10KB); chain selection criteria: Hyperledger Fabric (block generation 10 seconds) is superior to FISCOBCOS (15 seconds), and Polygon (gas fee 0.001 ETH / transaction) has a cost that is 90% lower than Ethereum;

[0093] Cross-chain synchronization: Synchronization is triggered at 00:00 every hour (carrying block height and signature), using the HTLC (Hash Time Locked Contract) protocol to ensure the atomicity of data transmission between the consortium chain and the public chain; the synchronization success rate has been tested to reach 99.9%; two consecutive failures trigger SMS and email alerts (authorized nodes can manually synchronize within 24 hours).

[0094] 3. Smart Contract Verification

[0095] Built-in feature point comparison algorithm: When the vertex change rate is greater than 5% or the main feature point deviation is greater than 0.01mm (virtual point is greater than 0.03mm), the transaction is rejected and a hash update is triggered.

[0096] Step 4: Adaptive Post-processing and Validation

[0097] 1. Simplified terminal adaptation

[0098] Simplified tiering rules: The number of triangles is controlled according to GPU memory tiers (ultra-lightweight: 50,000-100,000; low-performance: 300,000-500,000; medium-performance: 1,200,000-1,500,000; high-performance: 2,000,000-2,500,000). Specifically, the number of triangles is tiered according to terminal performance: ultra-lightweight terminals with <1GB GPU memory retain 50,000-100,000; low-performance terminals with 1-2GB GPU memory retain 300,000-500,000; medium-performance terminals with 2-4GB GPU memory retain 1,200,000-1,500,000; and high-performance terminals with >4GB GPU memory retain 2,000,000-2,500,000. The loading time for the 50,000-100,000 triangles has been verified to be 2.3-2.8 seconds through 10 sets of tests, meeting the smoothness threshold of ≤3 seconds. Ultra-lightweight terminals retain a minimum of 50,000 triangles (≥60 main feature points).

[0099] Simplified algorithm selection: The edge collapse algorithm (main feature point retention rate of 98%) is better than vertex clustering (85%) and triangle folding (90%); the specific shrinkage threshold of the edge collapse algorithm: the shrinkage threshold is set to 0.01mm. After 10 sets of tests, it is verified that this threshold can make the main feature point retention rate reach 98%. Below 0.005mm, the amount of computation will increase, and above 0.02mm, the feature will be lost.

[0100] No simplified area settings: 3mm for the face of Buddha statues in grottoes, 5mm for bronze artifacts (tested to cover 98% of main feature points);

[0101] Ultra-high performance adaptation: For terminals with GPU memory > 16GB, the core area is not simplified + non-core area is simplified by 1.2 times, and the core area deviation is ≤ 0.005mm.

[0102] 2. API verification interface

[0103] Interface specification: HTTP POST request (parameters include model ID and on-chain hash), return code 200 (verification passed), 400 (parameter error), 500 (data error);

[0104] Access Control: A three-tiered access control system (cultural heritage protection institutions can access all data, partner institutions can access hashes and verification results, and ordinary users can only access public metadata); Specific implementation of API interface access control: The public key uses the RSA2048 encryption algorithm. During verification, the public key hash stored in the consortium blockchain must be compared. Only after a successful match can the interface be called. This method has been tested and can resist 99% of illegal call attempts.

[0105] Error handling: When the return code is 500, it will automatically retry 3 times (with an interval of 10 seconds). If it fails, it will be logged (stored in the consortium blockchain's auxiliary database) and an alarm will be triggered.

[0106] Based on the simplified model of terminal performance, the feature point deviation was verified to be ≤0.02mm through the API interface.

[0107] Example of a Buddhist statue from a Northern Wei Dynasty grotto:

[0108] S1, Data Acquisition and Preprocessing

[0109] Device collaboration: Leica BLK360 scans the entire area (150 million points), Artec Eva scans the face (20 areas) and clothing folds (30 areas); a supplementary lighting device is activated in low-light areas, and the signal-to-noise ratio is improved to 18dB after processing with the Retinex algorithm;

[0110] Fusion filtering: After ICP calibration, the fusion error is 0.04mm. Adaptive filtering retains 90% of the points in the high curvature region, and the coordinates are normalized to [-1,1].

[0111] S2, Feature Extraction and Reconstruction

[0112] Feature extraction: Point cloud was converted into a 64×64×64 voxel mesh, with a voxel size of 0.5mm. 3 (i.e., each voxel has a side length of 0.5mm), and a 64×64×64 grid covers a 32mm×32mm×32mm area, which is suitable for facial detail scanning range; input is a dual-channel CNN, and output is a 128-channel feature map (facial weight > 0.8);

[0113] Feature extraction was implemented using the open-source library TensorFlow 2.0, and the convolutional layers were initialized using a He normal distribution. The TensorFlow 2.0 training parameters were: batchsize = 32, initial learning rate = 0.001, Adam optimizer was used, training was iterated for 5000 times, and training was stopped when the accuracy on the validation set reached 98%.

[0114] Reconstruction and optimization: The face uses an 8th-order polynomial, and the clothing folds use a 4th-order polynomial. The initial model with 6 million triangles is generated through 100 iterations.

[0115] Iteration termination condition: When the average error of 200 sampling points in the facial region stabilizes at 0.06mm (≤0.08mm threshold) for 3 consecutive iterations, the iteration stops;

[0116] After multi-resolution optimization, a model with 2.5 million triangle faces is generated;

[0117] Redundant triangular facet removal criteria: curvature < 0.1mm *1 And the area is greater than 1mm 2 The triangular faces were automatically removed, with a total of 3.5 million (58%) removed. Among the 2.5 million triangular faces that were retained, the core feature regions accounted for 60%.

[0118] S3, Trusted Identifier

[0119] Feature point selection: 128 main feature points (40 for the face, 50 for clothing folds, and 38 for relief), with a maximum deviation of 0.008mm after 5 simplifications;

[0120] Hash and Storage: Generate hash 0x7f3a...9d2b, write the core identifier to HyperledgerFabric, write metadata to Polygon, and synchronize block hashes hourly.

[0121] S4, Post-processing and Verification

[0122] Adaptive simplification: When the terminal is detected as a PC (medium performance), it is simplified to 1.5 million triangles, while retaining 3mm triangles around facial feature points;

[0123] Verification result: The API comparison showed that the deviation of all 128 feature points was less than 0.01 mm, and the result was "Verification passed".

[0124] Mural Example:

[0125] Taking the "Creation of Digital Collections of Tang Dynasty Murals" as an example, the specific implementation steps are as follows:

[0126] Data acquisition: Using an ArtecEva scanner with 550nm illumination (800 lux) and a scanning resolution of 0.1mm, 120 million point cloud data were acquired;

[0127] Data processing: Point cloud data was preprocessed using CloudCompare 2.12.4, and fusion calibration and adaptive filtering were performed. After processing, the noise ratio of point cloud was reduced to 5.2%.

[0128] Model Reconstruction: The model reconstruction adopted MeshLab 2022.02, and feature extraction was based on the TensorFlow 2.10 framework. A dual-channel CNN was used to extract texture color block features (32 channels). The convolutional layers were initialized using a He normal distribution. The training iterations were 5000 times (batch size = 32), and the training stopped when the accuracy on the validation set reached 98%. The attention weights were allocated as follows: 0.6 for texture color blocks and 0.3 for wall structures. A model with 2 million triangles was generated, and the texture restoration error was 0.06mm. Compared with the traditional single-channel CNN, the dual-channel CNN improved the integrity of texture color block restoration by 40% and reduced the processing time by 25%.

[0129] Identifier generation: 128 main feature points (texture entropy > 0.8) are selected, SHA-3 hashes are generated, and written to the Hyperledger Fabric consortium blockchain;

[0130] Terminal adaptation and verification: The mid-performance terminal (3GB video memory) is simplified to 1.3 million triangles, and the edge collapse algorithm is used with a shrinkage threshold of 0.01mm; the API interface verification returned a deviation of 0.009mm, and the identification consistency was 100%.

[0131] Examples of bronze artifacts:

[0132] S1, Data Acquisition and Preprocessing

[0133] A polarizing filter was used to reduce surface reflection of bronze artifacts, resulting in a 65% reduction in point cloud noise; the average fusion error after ICP calibration was 0.03 mm.

[0134] S2, Feature Extraction and Reconstruction

[0135] Feature extraction: Dual-channel CNN geometric channel focuses on inscription edges (curvature > 0.3mm) -1 Attention weight allocation is 0.6 for inscription and 0.3 for shape; inscription area error is 0.05mm.

[0136] S3, Trusted Identifier

[0137] Feature point selection: Based on the PCL library, after preprocessing with voxel_grid filtering, the SIFT feature detector was used to extract the inscription edge points; among the 128 feature points, there were 40 stroke intersection points (31%), 50 stroke endpoint points (39%), and 38 edge turning points (29%), which met the rule of "core feature points accounting for ≥60%"; after 5 simplifications, the feature point deviation was <0.01mm;

[0138] Hash generation: Extract feature point coordinates and normal vectors to generate SHA-3 hashes, write the core identifier to the consortium blockchain, and write the metadata to the public blockchain.

[0139] Post-processing and verification

[0140] Adaptive simplification: When the terminal is detected as a mobile phone (high-end), it is simplified to 600,000 triangular faces, and the 5mm triangular face around the inscription feature point is retained;

[0141] Verification result: The API call for comparison returned "Verification passed";

[0142] API interface output format: JSON format, including "Verification Result" (pass / fail), "Obstacle Point Coordinates" (if failed), "On-chain Hash" and "Local Hash" fields, with a response time of ≤500ms.

[0143] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for high-precision reconstruction and reliable asset identification of 3D model digital collectibles, characterized in that, The specific process is as follows: Data acquisition and environmental adaptation preprocessing: Point cloud data is obtained by scanning with corresponding equipment according to the characteristics of cultural relics, and environmental adaptation processing is performed during the scanning process; adaptive filtering and normalization processing are performed on the point cloud data. When multiple devices are used, the point cloud data acquired by multiple devices need to be fused before adaptive filtering. Differentiated feature-driven model reconstruction: The dual-channel CNN feature extraction is divided into geometric channels and texture channels. The geometric channel extracts the spatial geometric morphology information of the point cloud, and the texture channel extracts the texture information of the model surface. Based on the adaptive attention weight mechanism, a reconstruction strategy is adopted to reconstruct the texture information and morphology information to generate a triangular face model. Hybrid Chain Trustworthiness Identification: Select the main feature points from the reconstructed model, extract the coordinates and normal vectors of the main feature points to generate feature hashes, store the feature hashes using a hybrid chain storage architecture, and set up smart contracts for transaction constraints.

2. The method for high-precision reconstruction and reliable asset identification of three-dimensional model digital collectibles according to claim 1, characterized in that, The scanning process is based on the characteristics of the cultural relic, specifically: the cultural relic's height is greater than 3m and its surface curvature change is greater than 0.5mm. -1 When the area occupied by the artifact is greater than 30%, a combination of laser scanner and structured light scanner is used for scanning; when the height of the artifact is less than 3m or the area occupied by the artifact is less than 30%, a structured light scanner is used alone for scanning; point cloud data of the artifact is obtained through a multi-device collaborative acquisition strategy. When the cultural relic is a mural, it is also necessary to simultaneously scan the three-dimensional structure of the wall where the mural is located to preserve the spatial location information of the texture.

3. The method for high-precision reconstruction and reliable asset identification of three-dimensional model digital collections according to claim 2, characterized in that, The environmental adaptation process includes low-light processing and reflection / obstruction processing. Low light processing: Set up a white light supplementary lighting device and process the point cloud data in conjunction with the Retinex algorithm; Reflection / Occlusion Processing: Reflective areas of bronze artifacts were captured using a polarizing filter; occluded areas were fused with point clouds from more than three viewpoints, and missing parts were filled in using fourth-order polynomial interpolation, with the data labeled "interpolation supplementation".

4. The method for high-precision reconstruction and reliable asset identification of three-dimensional model digital collectibles according to claim 3, characterized in that, When multiple devices are used, the point cloud data acquired from multiple sources are fused. The specific process is as follows: The coordinate transformation matrix between devices is calculated using a checkerboard calibration board. The transformation matrix is ​​then used to input the initially transformed multi-source point cloud into the ICP algorithm for secondary calibration. By minimizing the Euclidean distance error between corresponding points in the point cloud, the transformation matrix parameters are optimized, ultimately achieving multi-source point cloud fusion.

5. The method for high-precision reconstruction and reliable asset identification of three-dimensional model digital collectibles according to claim 4, characterized in that, The specific process of the adaptive filtering is as follows: based on the degree of drastic change in the geometric shape of the artifact surface, it is divided into high curvature region, intermediate region and flat region, and different proportions of feature points are retained for different regions.

6. The method for high-precision reconstruction and reliable asset identification of three-dimensional model digital collectibles according to claim 1, characterized in that, The geometric channel uses a 3×3×3 convolution kernel to output 64 channels of curvature gradient features and performs batch normalization. The texture channel uses a 5×5×5 convolution kernel to output 32 channels of texture entropy features and performs ReLU activation.

7. The method for high-precision reconstruction and reliable asset identification of three-dimensional model digital collectibles according to claim 1, characterized in that, Based on the adaptive attention weight mechanism, a segmentation and reconstruction strategy is adopted to reconstruct the model based on texture information and morphological information. First, a pre-trained classifier is used to identify the type of cultural relic being processed, and basic weights are assigned to the core and secondary areas of different types of cultural relics. Secondly, the basic weights are locally adjusted—by calculating the feature response intensity of the core region, if the feature response intensity is greater than the preset threshold, the basic weights are increased. If the feature response intensity is less than the threshold, then the basic weights are maintained; Finally, the reconstruction levels are divided according to the optimized weight values, and different order polynomials are used for different levels to reconstruct the model based on texture and morphological information.

8. The method for high-precision reconstruction and reliable asset identification of three-dimensional model digital collections according to claim 1, characterized in that, The process of selecting principal feature points from the reconstructed model and extracting their coordinates and normal vectors to generate feature hashes is as follows: Point cloud coordinates that uniquely represent the core features of the three-dimensional shape and texture of the cultural relic are considered as principal feature points. A set number of principal feature points are extracted from the point cloud model. If the number of extracted principal feature points is less than the set number, a dual-carrier adaptation method of 'point cloud + image' is used to supplement them. The dual-carrier adaptation method of 'point cloud + image' is as follows: ① Image matching: Call the historical document auxiliary module, input the historical image of the cultural relic before it was damaged, and use the SIFT algorithm to perform feature matching between the historical image and the point cloud texture map of the current damaged cultural relic. Image feature points with a matching degree of ≥80% are considered as valid references. ② Virtual feature point generation: Based on the matched image feature points and combined with the point cloud geometric trend of the damaged area, virtual feature points are generated; the proportion of virtual feature points is ≤30%, and the Euclidean distance between them and the existing main feature points is ≤0.5mm; ③ Metadata labeling: Virtual feature points must be labeled with the 'virtual generation' attribute and the basis for generation when stored on the blockchain to ensure the transparency of asset identification.

9. The method for high-precision reconstruction and reliable asset identification of three-dimensional model digital collectibles according to claim 8, characterized in that, The process involves storing the feature hash using a hybrid chain storage architecture and setting up a smart contract for transaction constraints. Consortium blockchains store feature hashes and the creator's public key; public blockchains store extended metadata and consortium blockchain hash pointers. Cross-chain synchronization: Synchronization is triggered at set times and uses the HTLC protocol to ensure the atomicity of data transmission between consortium blockchains and public blockchains; Smart contract verification: Built-in feature point comparison algorithm: When the vertex change rate is >5% or the main feature point deviation is >0.01mm, the transaction is rejected and a hash update is triggered.

10. The method for high-precision reconstruction and reliable asset identification of three-dimensional model digital collectibles according to claim 1, characterized in that, It also includes adaptive post-processing and verification steps, including: simplified terminal adaptation and API verification interface; Terminal adaptation simplification: Based on the GPU memory size, an edge collapse algorithm is used to simplify the regions outside the non-simplified areas of the model. API verification interface: It adopts a three-level permission system, that is, cultural relics protection institutions can view all data, partner institutions can view hash and verification results, and ordinary users can only view public metadata.

Images