Intelligent storage system and method for digital processing of archives

Abstract

The invention belongs to the technical field of electronic digital data processing, and discloses an intelligent storage system and method for file digital processing. The method comprises the following steps: performing boundary feature point detection and detail-rich region detection based on collected file point cloud data to respectively obtain a boundary feature mask and a detail-rich mask of the file point cloud data; performing semantic feature marking on the archive point cloud data in combination with a preset domain knowledge base to generate a semantic tag; calculating a priority value of each data point in the archive point cloud data based on the archive point cloud data, the boundary feature mask, the detail-rich mask and the semantic tag, and dividing the archive point cloud data into a feature region and a non-feature region according to the priority value; according to the method, while data compression is realized, key features such as boundaries, details, semantics and the like are accurately reserved, the problem of feature loss caused by lightweight processing in a traditional method is effectively solved, and the integrity and availability of archive point cloud data in a high-precision analysis scene are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electronic digital data processing, and more particularly, to an intelligent storage system and method for digital processing of archives. BACKGROUND

[0002] In the field of three-dimensional archive digitization, point cloud data lightweight processing is a key link to realize efficient storage and transmission. However, the existing technology generally adopts uniform thinning, fixed grid simplification and other methods, and does not perform differential processing on the boundary feature points, detail-rich areas and semantic features in the point cloud data. Taking the scenario of digitizing paper archives of government departments as an example, when three-dimensional modeling of paper archives such as historical documents and contracts is performed, the existing methods are difficult to accurately retain the fine features such as handwritten annotations and seal edges on the archives. Due to the lack of multi-dimensional feature perception ability, the traditional method cannot accurately identify the irregular boundaries formed by the wrinkles of the archives, the contours of the faded text and other key information, resulting in important geometric features and semantic information being mistakenly deleted or simplified during the lightweight process, which seriously affects the reliability of subsequent high-precision analysis.

[0003] For example, when three-dimensional modeling of paper archives carrying important historical information is performed, the creases and faded text on the surface of the archives may become unclear after simplification due to insufficient point density, and the fine texture of the seal on the archives leads to contour distortion due to uniform thinning, which greatly reduces the usability of the lightweight data in scenarios such as archive content interpretation and authenticity identification that require high feature integrity. Furthermore, in three-dimensional digitization scenarios involving spatial information such as construction engineering archives and natural resource archives, details such as beam-column connections and topographic features also lose key features due to the limitations of traditional lightweight methods, affecting the conduct of engineering detection and resource management. Therefore, how to accurately retain multi-dimensional key features such as boundaries, details and semantics while compressing data has become a technical bottleneck restricting the application of point cloud data in high-precision fields.

[0004] In view of this, the present application proposes an intelligent storage system and method for digital processing of archives to solve the above problems. SUMMARY

[0005] In order to overcome the above-mentioned defects of the prior art and achieve the above-mentioned purposes, the present application provides the following technical solution: an intelligent storage method for digital processing of archives, comprising:

[0006] Based on the collected archive point cloud data, boundary feature point detection and detail-rich area detection are performed to obtain a boundary feature mask and a detail-rich mask of the archive point cloud data, respectively; and a semantic feature label is generated by combining a pre-set domain knowledge base with the archive point cloud data for semantic feature labeling;

[0007] Based on the archive point cloud data, the boundary feature mask, the detail-rich mask and the semantic label, a priority value of each data point in the archive point cloud data is calculated, and the archive point cloud data is divided into a feature region and a non-feature region according to the priority value;

[0008] The feature region and the non-feature region of the archive point cloud data are respectively subjected to lightweight processing, and the processed regions are merged to obtain lightweight point cloud data;

[0009] Based on the archive point cloud data, the lightweight point cloud data is subjected to adaptive repair of geometric feature loss to obtain repaired point cloud data.

[0010] Further, the method for obtaining the lightweight point cloud data comprises:

[0011] The average point distance of the data points in the non-feature region is calculated;

[0012] The voxel grid size is set as ts times of the average point distance;

[0013] A three-dimensional voxel grid is constructed according to the voxel grid size, each voxel grid is traversed, and only the center point of each voxel grid is retained to obtain the point cloud data after thinning of the non-feature region;

[0014] Each feature point in the feature region is taken as the center, K-neighbor searching is performed in a spherical neighborhood with a radius of ly to obtain the point cloud data after thinning of the feature region;

[0015] If the number of sample points in the spherical neighborhood with a radius of ly is insufficient, the nearest neighbor point to the feature point in the Euclidean distance is sequentially selected from the archive point cloud data as a supplement until the set searching number is met;

[0016] The point cloud data after thinning of the non-feature region and the point cloud data after thinning of the feature region are merged to obtain the lightweight point cloud data.

[0017] Further, the method for performing adaptive repair of geometric feature loss on the lightweight point cloud data comprises:

[0018] The scale-invariant feature transformation algorithm is adopted to extract the geometric feature points of the lightweight point cloud data and the archive point cloud data, respectively;

[0019] The geometric feature points of the lightweight point cloud data and the archive point cloud data are matched by the FLANN algorithm to obtain a matching point pair set;

[0020] The Euclidean distance of each matching point pair in the matching point pair set is calculated;

[0021] The geometric feature points with a Euclidean distance greater than a set distance threshold are marked as feature loss points;

[0022] The feature loss points are clustered, a three-dimensional bounding box of the feature loss region is generated, and the corresponding range is expanded based on the density distribution of the feature loss region to obtain an expanded feature loss region;

[0023] The expanded feature loss region is supplemented to obtain supplemented point cloud data;

[0024] The supplemented point cloud data and the lightweight point cloud data are processed by using an iterative closest point algorithm to obtain a rigid transformation parameter; the rigid transformation parameter is a three-dimensional space transformation parameter including a rotation angle and a translation distance, and is used for aligning point clouds from one coordinate system to another coordinate system;

[0025] The supplemented point cloud data is registered according to the rigid transformation parameter to generate registered point cloud data;

[0026] In an overlapping region of the registered point cloud data, a smooth surface of the overlapping region is fitted by using a moving least square method;

[0027] Based on the smooth surface, new point coordinates are generated in the feature loss region;

[0028] The generated new point coordinates and the lightweight point cloud data are merged to obtain repaired point cloud data.

[0029] Further, the method for expanding the feature loss region range to cover the potential missing sampling region in the feature loss point dense region comprises:

[0030] The feature loss region is divided into cubic grid with edge length being a multiple of the point cloud data space resolution tz, and the number of feature loss points in each grid is counted; the point cloud data space resolution is the point interval of the archive point cloud data;

[0031] The number of feature loss points is divided by the volume of the grid to obtain the grid density;

[0032] Statistical characteristic values of all grid densities are calculated; the statistical characteristic values include quantiles of the grid density distribution, a preset fixed density proportionality coefficient, and a linear combination of the grid density mean value and the standard deviation;

[0033] A global density threshold is determined according to the statistical characteristic values;

[0034] When the grid density is greater than or equal to the density threshold, the grid is determined as a feature loss point dense region;

[0035] The feature loss point dense region is expanded in each of the three coordinate axis directions by a preset distance to cover the potential missing sampling region.

[0036] Further, the method for obtaining the rigid transformation parameter comprises:

[0037] For each point of the supplementary point cloud data, find the nearest neighbor point in the lightweight point cloud data to establish a corresponding point pair set;

[0038] By setting a minimum objective function of the Euclidean distance between the corresponding points in the corresponding point pair set, the current optimal rigid transformation parameter is obtained by solving;

[0039] When the change value of continuous lc iterations is less than the set change threshold or reaches the set maximum iteration number, the iteration is terminated;

[0040] The final iteration obtained rigid transformation parameter is output.

[0041] Further, the method for obtaining the boundary feature mask of the archive point cloud data comprises:

[0042] S101: Let the initial value of the index variable p be 1, and let the initial value of the index variable q be 1, the value range of p is 1 to P, and the value range of q is 1 to Q; P is the number of data points in the archive point cloud data, and Q is the preset number of neighborhood points;

[0043] S102: Exclude the data points that have been marked as boundary feature points, and select the point closest to the pth data point in the Euclidean distance, mark it as a near point, and calculate the normal vector angle between the near point and the pth data point;

[0044] S103: If the normal vector angle is greater than the preset normal vector angle threshold, mark the near point as a boundary feature point, and generate a binary form of the boundary feature mask of the near point;

[0045] S104: If q is less than Q, let q=q+1, execute S102; otherwise, let q=1, execute S105;

[0046] S105: If p is less than P, let p=p+1, execute S102; otherwise, construct a binary form of the boundary feature mask of all data points marked as boundary feature points into the boundary feature mask of the archive point cloud data.

[0047] Further, the method for obtaining the detailed rich mask of the archive point cloud data comprises:

[0048] Divide a spherical neighborhood with a radius of xj around each data point in the archive point cloud data in turn, divide the spherical neighborhood into s l voxel grids, and count the number of data points in each voxel grid;

[0049] For each voxel grid, calculate the proportion of the number of points in it to the total number of points in the spherical neighborhood to obtain the probability distribution value of each voxel grid;

[0050] The spatial distribution entropy value of each data point in the archive point cloud data is equal to the sum of the negative values of the product of the natural logarithm of the voxel grid probability distribution value and the voxel grid probability distribution value of each data point in the archive point cloud data;

[0051] If the spatial distribution entropy value of a data point is greater than a preset entropy value, the data point is marked as a detail-rich point, and a binary-form detail-rich mask of the data point is generated;

[0052] All binary-form detail-rich masks of the detail-rich points are constructed into a detail-rich mask of the archive point cloud data.

[0053] Further, the method of dividing the archive point cloud data into a feature region and a non-feature region comprises:

[0054] The RGB color value, the laser echo intensity, the boundary feature mask, the detail-rich mask and the one-hot encoding of the semantic label of each data point in the archive point cloud data are converted into a feature vector, and the weight of each feature vector is calculated through a multi-head self-attention mechanism; the sum of the weights is 1;

[0055] The feature confidence score is obtained by weighted summation of the feature vector and each weight;

[0056] The geometric feature mask score is calculated based on the boundary feature mask and the detail-rich mask;

[0057] The semantic label score is obtained in combination with a preset domain knowledge base;

[0058] The priority value of each data point in the archive point cloud data is obtained by weighted summation of the geometric feature mask score, the semantic label score and the feature confidence score;

[0059] The data points with a priority value greater than a set priority value threshold are determined as feature points and constructed as a feature region; the data points with a priority value less than or equal to the set priority value threshold are determined as non-feature points and constructed as a non-feature region.

[0060] Further, the method of obtaining the geometric feature mask score comprises:

[0061] The geometric feature score of all data points in the archive point cloud data is calculated through the boundary feature mask and the detail-rich mask, and if the data point is a boundary feature point or a detail-rich point, the score is df; otherwise, the score is 0.

[0062] Compared with the prior art, the technical effects and advantages of the archive digitization processing intelligent storage system and method of the present application are:

[0063] Based on the archive point cloud data, boundary feature point detection and detail rich area detection are performed to generate boundary feature mask and detail rich mask, and semantic feature labeling is generated by combining with the domain knowledge base to generate semantic label; then the feature point priority value is calculated by using the information, the feature and non-feature areas are divided, and the lightweight point cloud data is obtained after the lightweight processing of the feature and non-feature areas is performed respectively and then combined; finally, the adaptive repair of the geometric feature loss is performed based on the original archive point cloud data, and the repaired point cloud data is obtained through feature point matching, error detection, directional supplement sampling and data fusion.

[0064] The scheme realizes data compression while effectively retaining key features such as archive character outlines and engineering structure details through multi-dimensional feature detection and labeling, lightweight processing and adaptive repair process, and solves the problems of feature loss, information extraction and engineering detection accuracy in the lightweight processing of archive point cloud data by traditional methods. In the scenes such as government paper archive content interpretation, authenticity identification and building engineering archive structure analysis which have high requirements for feature integrity, the integrity and usability of the archive point cloud data are significantly improved, and the scheme has the significant advantages of high automation, accurate archive feature retention and strong adaptability to multiple types of archive scenes. BRIEF DESCRIPTION OF DRAWINGS

[0065] Figure 1 An intelligent storage system for archive digitization processing according to an embodiment of the present application;

[0066] Figure 2 An intelligent storage method for archive digitization processing according to an embodiment of the present application;

[0067] Figure 3 A method flowchart for dividing archive point cloud data into feature areas and non-feature areas according to an embodiment of the present application;

[0068] Figure 4 A method flowchart for adaptive repair of geometric feature loss of lightweight point cloud data according to an embodiment of the present application. DETAILED DESCRIPTION

[0069] The technical solutions in the embodiments of the present application will be described in detail, clearly and completely below with reference to the drawings in the embodiments of the present application. It should be particularly noted that the specific embodiments described below are only used to better illustrate and describe the technical solutions of the present application, and are intended to enable those skilled in the art to better understand and implement the present application, and should not be understood as limiting the protection scope of the present application. Those skilled in the art can modify, adjust or equivalently replace the present application according to the content disclosed in the present application without departing from the spirit and essence of the present application, and these should be regarded as the protection scope of the present application.

[0070] Embodiment 1

[0071] Please refer toFigure 1 The embodiment shown discloses an intelligent storage system for digitizing processing of archives, which comprises a feature marking module, a feature partitioning module, a lightweight module and a self-adaptive repairing module, each module is connected through wired and / or wireless connection to realize data transmission.

[0072] The feature marking module detects boundary feature points and rich detail areas based on collected archive point cloud data, and respectively obtains a boundary feature mask and a rich detail mask of the archive point cloud data; and generates semantic labels by combining a preset domain knowledge base to mark semantic features of the archive point cloud data.

[0073] The method for obtaining the boundary feature mask of the archive point cloud data comprises:

[0074] S101: setting an initial value of an index variable p as 1 and an initial value of an index variable q as 1, the value range of p being 1 to P, and the value range of q being 1 to Q; P is the number of data points in the archive point cloud data, and Q is a preset number of neighborhood points;

[0075] S102: excluding data points that have been marked as boundary feature points, selecting a point closest to the pth data point in Euclidean distance as a near point, and calculating an angle between a normal vector of the near point and the pth data point;

[0076] S103: if the angle between the normal vectors is greater than a preset normal vector angle threshold, marking the near point as a boundary feature point, and generating a binary form of the boundary feature mask of the near point;

[0077] S104: if q is less than Q, setting q = q + 1, and executing S102; otherwise, setting q = 1, and executing S105;

[0078] S105: if p is less than P, setting p = p + 1, and executing S102; otherwise, constructing a binary form of the boundary feature mask of all data points marked as boundary feature points into the boundary feature mask of the archive point cloud data.

[0079] The number of neighborhood points is 10 in the embodiment; and the normal vector angle threshold is 30° in the embodiment

[0080] The method for calculating the angle between the normal vectors of the near point and the pth data point comprises:

[0081] Searching a local neighborhood point set of the pth data point through a K-nearest neighbor algorithm; and calculating a covariance matrix based on principal component analysis:

[0082]

[0083] In the formula, C p is the covariance matrix of the pth data point; μ p is a neighborhood point mean of the pth data point; and Pp is the pth data point; N p is the local neighborhood point set of the pth data point.

[0084] Eigenvalues and eigenvectors of the covariance matrix of the pth data point are calculated; the normal vector of the pth data point is the eigenvector corresponding to the minimum eigenvalue, and the curvature value of the pth data point is obtained by dividing the minimum eigenvalue by the sum of all eigenvalues.

[0085] The angle between the normal vector of the pth data point and the normal vector of the near point is calculated by the normal vector of the pth data point and the normal vector of the near point.

[0086] The boundary feature mask is used to identify the boundary region in the point cloud, and the subsequent simplification process needs to be preferentially retained to avoid feature loss.

[0087] The boundary feature mask is used to identify the boundary region in the point cloud, and the subsequent simplification process needs to be preferentially retained to avoid feature loss.

[0088] The method for obtaining the detail-rich mask of the archive point cloud data comprises:

[0089] Each data point in the archive point cloud data is taken as the center to define a spherical neighborhood with a radius of xj, the spherical neighborhood is divided into sl voxel grids, and the number of data points in each voxel grid is counted.

[0090] For each voxel grid, the proportion of the number of points in the voxel grid to the total number of points in the spherical neighborhood is calculated to obtain the probability distribution value of each voxel grid.

[0091] The spatial distribution entropy value of each data point in the archive point cloud data is equal to the sum of the negative values of the product of the natural logarithm of the probability distribution value of each voxel grid and the probability distribution value of the voxel grid where each data point in the archive point cloud data is located.

[0092] If the spatial distribution entropy value of a data point is greater than a preset entropy value, the data point is marked as a detail-rich point, and a binary form detail-rich mask of the data point is generated.

[0093] All binary form detail-rich masks of the detail-rich points are constructed into the detail-rich mask of the archive point cloud data.

[0094] For example, the embodiment provides a method for generating a detail-rich mask, which specifically comprises the following steps:

[0095] Each data point in the archive point cloud data is taken as the center to define a spherical neighborhood with a radius of 5 mm, the spherical neighborhood is divided into 8 voxel grids, and the number of points in each grid is counted.

[0096] Let k be a positive integer from 1 to 8, calculate the kth voxel grid probability distribution value P k :

[0097]

[0098] In the formula, n k is the number of points contained in the kth voxel grid; N k is the total number of points in the neighborhood of the kth voxel grid.

[0099] Let i be a positive integer from 1 to N, where N is the number of data points in the archival point cloud data, calculate the spatial distribution entropy value E i :

[0100]

[0101] If the spatial distribution entropy value of the ith data point is greater than 0.8, mark the point as a detail-rich point, and generate the detail-rich mask m i = 1; otherwise, m i = 0.

[0102] Wherein, the spatial distribution entropy value is used to measure the uniformity of the distribution of point cloud data in the local neighborhood, if the probability distribution is uniform, the entropy value is high, indicating that the point distribution is irregular and the details are rich; if the probability distribution is concentrated, the entropy value is low, indicating that the point distribution is regular and the details are less.

[0103] In this embodiment, 0.8 is an empirical value of the spatial distribution entropy threshold setting, used to distinguish between regular arrangement and complex details, which can be modified accordingly in actual application.

[0104] Aggregate all point mask values to form a detail-rich mask M of the archival point cloud data:

[0105] M = [m1, m2,..., m N ] T ;

[0106] The method for generating semantic labels by combining the pre-set domain knowledge base to mark the semantic features of the archival point cloud data includes:

[0107] The domain knowledge base pre-archives the morphological rules of key features such as characters and seals. When the mutation rate of gray value of the point cloud data in a continuous preset number of neighborhood points exceeds a preset threshold and forms a linear distribution with a specific aspect ratio, it is marked as a faded character semantic label. If the geometric morphology of the point cloud data conforms to the preset circular or elliptical boundary, and the internal gray distribution has a specific rule, it is marked as a seal semantic label. For the creases on the surface of the archives, when the mutation rate of the curvature value of the point cloud data in a continuous preset number of neighborhood points exceeds the mutation threshold, and forms a linear distribution with a length greater than a preset length and a width less than a preset width, it is marked as an archive crease semantic label.

[0108] At the same time, the domain knowledge base can be connected to the building engineering archive management system to obtain the component semantic range. When the point cloud data coordinates fall within the semantic range of the beam-column connection marked on the architectural drawing, it is marked as an architectural structure edge semantic label. If connected to the natural resource archive database, when the point cloud data conforms to the preset topographic and geomorphic feature geometric model, it is marked as the corresponding topographic semantic label. Through the above rules, the semantic features of different types of archive point cloud data are accurately labeled, providing a basis for subsequent data processing and application.

[0109] For example, if the curvature value of the data point in the archive point cloud data mutates by more than 0.5 in a continuous 5-neighborhood point and forms a linear distribution, such as a length greater than 5 mm and a width less than 1 mm, it is marked as an archive crease. If the data point in the archive point cloud data is located at the beam-column connection of the building engineering archive model, it is marked as a structure edge.

[0110] By combining the normal vector angle threshold and the number of neighborhood points, the boundary feature points can be accurately captured in the geometric structure mutation area of the point cloud data. In the prior art, global uniform thinning is often used for lightweight processing, without considering the local geometric feature difference, resulting in the deletion of boundary features due to insufficient density. In this scheme, for government paper archives and related engineering archives, the normal vector angle reflects the local surface direction change. When the angle is greater than 30°, it exceeds the angle range of the smooth transition of the conventional curved surface, and can be determined as a boundary point, such as the curled edge of the paper archive and the contour turning point of the component in the architectural drawing. The boundary feature mask marks these points as "1", which are preferentially retained in subsequent lightweight processing, avoiding distortion of archive details and engineering structure features due to thinning.

[0111] The spatial distribution entropy is used to quantize the local distribution uniformity of the point cloud, so as to effectively identify complex texture regions, such as handwritten notes on files and decorative patterns in architectural design drawings. The existing curvature-based feature detection is susceptible to noise interference and cannot distinguish between regular curved surfaces and regions with rich details. In the present scheme, the spherical neighborhood is divided into 8 voxel grids, and the irregularity of point distribution is calculated by entropy value: when the entropy value is greater than 0.8, it indicates that the point cloud is unevenly distributed in the neighborhood, and high-density point cloud needs to be retained to ensure that important information such as file text and engineering drawing details is not oversimplified in the lightweight process.

[0112] The feature partition module calculates the priority value of each data point in the file point cloud data based on the file point cloud data, the boundary feature mask, the rich detail mask and the semantic label, and divides the file point cloud data into a feature region and a non-feature region according to the priority value.

[0113] Referring to Figure 3 The method for dividing the file point cloud data into a feature region and a non-feature region comprises:

[0114] The RGB color value and the laser echo intensity of each data point in the file point cloud data, and the one-hot encoding of the boundary feature mask, the rich detail mask and the semantic label are converted into feature vectors, and the weights of the feature vectors are calculated by a multi-head self-attention mechanism; the sum of the weights is 1.

[0115] The feature confidence score is obtained by weighted summation of the feature vectors and the weights.

[0116] The geometric feature mask score is calculated based on the boundary feature mask and the rich detail mask.

[0117] The semantic label score is obtained in combination with a preset domain knowledge base.

[0118] The priority value of each data point in the file point cloud data is obtained by weighted summation of the geometric feature mask score, the semantic label score and the feature confidence score.

[0119] The data points with a priority value greater than a set priority value threshold are determined as feature points and constructed as a feature region; and the data points with a priority value less than or equal to the set priority value threshold are determined as non-feature points and constructed as a non-feature region.

[0120] The method for calculating the weights of the feature vectors by a multi-head self-attention mechanism comprises:

[0121] The feature vectors are multiplied by the corresponding weight matrices respectively to obtain query vectors, key vectors and value vectors.

[0122] For example, the multi-head attention contains h independent attention heads, in this embodiment, h is set to 8, each head has an independent weight matrix group, taking the hth head as an example:

[0123] Query vector In the formula, F i is a feature vector; is the weight matrix group of the hth head query vector;

[0124] Key vector In the formula, is the weight matrix group of the hth head key vector;

[0125] Value vector In the formula, is the weight matrix group of the hth head value vector.

[0126] The weight matrix groups of different heads map the original feature vector to different subspaces, so that the model can capture the correlation between different modalities from multiple perspectives, such as the relationship between color and material, and the relationship between geometric features and semantic labels.

[0127] The weight matrix here is a learnable parameter, and its role is to map the feature vector to different subspaces in order to better capture the relationship between different modalities.

[0128] The attention score is calculated by the query vector, the key vector and the value vector, and the calculation method is:

[0129]

[0130] In the formula, Attention(Q, K, V) is the attention score function; softmax() is the softmax function; Q is the query vector; K is the key vector; V is the value vector; d k is the dimension of the key vector.

[0131] Among them, is to prevent the dot product result from being too large; the softmax function normalizes the score to between 0 and 1, so that the sum of the attention scores of all modalities is 1; the query vector is used to query the correlation with other features in the attention mechanism; the key vector is matched with the query vector to calculate the attention score; the value vector is weighted and summed according to the attention score to obtain the final attention output.

[0132] The above process is repeated using multiple different weight matrix groups, and finally the results of multiple heads are spliced and passed through a linear layer to obtain the weight of each feature vector.

[0133] The multi-head attention mechanism can capture the relationship between features from different angles, improving the expression ability of the model.

[0134] The calculation method of the feature confidence score of the i-th data point is obtained by weighted summation of the feature vector and each weight:

[0135]

[0136] In the formula, M i is the feature confidence score of the i-th data point, the value range is 0 to 1, and the higher the value, the stronger the importance of the point cloud feature; is the specific value of the m-th element in the i-th data point feature vector; α m is the weight of the m-th element in the i-th data point feature vector; m is the number of elements of the feature vector.

[0137] Different features contribute differently to the importance of judging point cloud features, and weighted summation can comprehensively consider these factors to avoid one-sidedness of a single mode.

[0138] The geometric feature score of all data points in the archive point cloud data is calculated by the boundary feature mask and the detail-rich mask. If the data point is a boundary feature point or a detail-rich point, the score is 1; otherwise, it is 0;

[0139] According to the domain knowledge base, the semantic label score corresponding to the semantic label is found, for example, the artifact crack and building structure edge label score is 1, and the ordinary label score is 0.5;

[0140] For each data point in the archive point cloud data, the feature priority value is calculated by weighted summation, and the calculation method is:

[0141] P i = 0.5 x G i + 0.3 x S i + 0.2 x M i ;

[0142] In the formula, P i is the feature priority value of the i-th data point, the value range is 0 to 1, and the higher the value, the stronger the feature importance; G i is the geometric feature score of the i-th data point; S i is the semantic label score of the i-th data point; M i is the feature confidence score of the i-th data point; in this embodiment, 0.5, 0.3 and 0.2 are the weights of geometric features, semantic labels and multi-modal confidence, respectively, and the weight distribution is based on the importance of feature types, which can be modified according to actual conditions.

[0143] By converting the RGB color value, laser echo intensity, and geometric feature mask, semantic label into a feature vector, the limitation of the prior art relying only on geometric features for lightweight is overcome. For example, when processing government paper archives, the traditional method only identifies archive creases through curvature, but ignores the color changes at the creases due to wear, i.e., the RGB feature changes and the differences in paper surface flatness, i.e., the echo intensity changes, resulting in the missed detection of some subtle creases; when processing construction engineering archive point cloud data, only the geometric structure is used to determine the beam-column connection, without considering the color and echo intensity changes at the connection due to corrosion and poor painting, so that some potential structural damage cannot be identified in time. After the multi-dimensional features are fused into a feature vector, the details of the archive surface and the potential features of the engineering structure can be more comprehensively captured, important data is avoided from being missed during the lightweight process, and more accurate data support is provided for subsequent applications such as archive repair and engineering detection.

[0144] A lightweight module is used to perform lightweight processing on the feature region and the non-feature region of the archive point cloud data, merge the processed regions, and obtain lightweight point cloud data.

[0145] The method for obtaining lightweight point cloud data includes:

[0146] The average point distance of the data points in the non-feature region is calculated;

[0147] The voxel grid size is set to ts times the average point distance;

[0148] A three-dimensional voxel grid is constructed according to the voxel grid size, each voxel grid is traversed, only the center point of each voxel grid is retained, and the point cloud data after decimation of the non-feature region is obtained;

[0149] Each feature point in the feature region is taken as the center, K-neighbor searching is performed in a spherical neighborhood with a radius of ly, and the point cloud data after decimation of the feature region is obtained;

[0150] If the number of sample points in the spherical neighborhood with a radius of ly is insufficient, the nearest neighbor point to the feature point in terms of Euclidean distance is sequentially selected from the archive point cloud data as a supplement until the set search number is met;

[0151] The point cloud data after decimation of the non-feature region and the point cloud data after decimation of the feature region are merged, and lightweight point cloud data is obtained.

[0152] For example, 100 data points are randomly selected from the archive point cloud data, the Euclidean distance of the expected nearest neighbor point of each point is calculated, and the average value is taken. In this embodiment, the average value calculated is 8 mm, and the voxel grid size can be determined as 16 mm.

[0153] In this embodiment, the three-dimensional space range is set to x-axis 0 to 500 mm, y-axis 0 to 300 mm, and z-axis 0 to 200 mm. Therefore, the number of grids in the x-axis direction is 31, the number of grids in the y-axis direction is 19, and the number of grids in the z-axis direction is 13, and there are a total of 7657 grids.

[0154] All data points in the non-feature region of the archive point cloud data are traversed, and the grid to which each data point belongs is determined according to the coordinates. The data points in each grid are stored as a list. For example, the grid (5, 3, 2) contains 12 non-feature points.

[0155] For each grid, if it contains non-feature points, the average of the coordinates of all data points in the grid is calculated as the center point. Only the center point is retained, and the other points are deleted. For example, the grid (5, 3, 2) originally has 12 non-feature points. After calculating the average, 1 center point is retained, and 11 points are deleted.

[0156] After the thinning is completed, the point cloud data of the non-feature region is obtained, and each data point is attached with a grid index and an original point cloud ID.

[0157] For the feature region, KD tree is used to accelerate K nearest neighbor search. For each feature point, the three points with the closest Euclidean distance within a radius of 5 mm are searched.

[0158] If there are less than 3 sample points in the neighborhood, for example, the turning point of the handwritten annotation on the paper archive only has 2 points due to insufficient original scanning resolution, or the complex structure of the beam column node in the construction archive has only 2 valid points due to the limitation of the scanning angle, the nearest point in the original point cloud data is supplemented. This selection does not distinguish between feature types, and the nearest point in the Euclidean distance is selected first to ensure that the complete outline of the archive text and the key details of the engineering structure are fully reflected in the point cloud data, avoiding the influence of data loss on the accuracy of subsequent feature recognition and lightweight processing.

[0159] After the thinning is completed, the point cloud data of the non-feature region is obtained, and each data point is attached with a grid index and an original point cloud ID.

[0160] The point cloud data of the non-feature region and the point cloud data of the feature region are merged to obtain the lightweight point cloud data.

[0161] Traditional lightweight processing uses global uniform thinning, which causes the feature region and the non-feature region to be compressed equally, and the key features are lost due to insufficient point density. The priority value output by the feature partition module in this scheme divides the point cloud into a feature region and a non-feature region. The former is subjected to feature-preserving thinning, and the latter is subjected to efficient compression, thereby fundamentally solving the contradiction between data reduction and feature preservation.

[0162] The voxel grid size is set to 2 times the average point spacing based on the experience principle that the point spacing after decimation does not exceed the original 2 times. Taking the average point spacing of 8 mm in the example as an example, the 16 mm voxel grid can ensure that the point spacing of the non-feature region after decimation is about 16 mm, which not only realizes the decimation ratio of 12:1, but also avoids distortion of the planar structure caused by excessive decimation.

[0163] By retaining the center point of the voxel grid, the macro framework of the non-feature region is retained. Experimental data show that this method controls the surface reconstruction error of the non-feature region to within 0.3 mm, while the traditional global uniform decimation error reaches 0.8 mm, verifying the advantages of dynamic voxel size.

[0164] The feature region adopts a spherical neighborhood with a radius of 5 mm and a K=3 neighborhood search, and this parameter combination is based on the matching of the feature size of the archive and the scanning accuracy. In the government paper archive management scene, a radius of 5 mm can cover the spatial range of fine features such as 0.3-1.2 mm wide archive creases, faded text strokes, etc., ensuring that the neighborhood of feature points such as the turning point of handwritten annotations and the edge of a seal contains sufficient geometric information; in the construction engineering archive scene, this radius is also applicable to capturing the slight structural changes at the beam-column junction. The search quantity of K=3 avoids the blurring of archive text contours and engineering structural detail features caused by too many neighborhood points, while meeting the minimum point requirement of the least squares method (MLS) surface fitting, accurately preserving the key feature information of archives and engineering structures while ensuring data processing efficiency.

[0165] The adaptive repair module performs adaptive repair of geometric feature loss on the lightweight point cloud data based on the archive point cloud data to obtain repaired point cloud data.

[0166] Please refer to Figure 4 The method for performing adaptive repair of geometric feature loss on the lightweight point cloud data includes:

[0167] The scale-invariant feature transform algorithm is used to extract geometric feature points of the lightweight point cloud data and the archive point cloud data, respectively;

[0168] The FLANN algorithm is used to match the geometric feature points of the lightweight point cloud data and the archive point cloud data to obtain a set of matching point pairs;

[0169] The Euclidean distance of each pair of matching points in the set of matching point pairs is calculated;

[0170] Geometric feature points with a Euclidean distance greater than a set distance threshold are marked as feature loss points;

[0171] The feature loss points are clustered to generate a three-dimensional bounding box of the feature loss region, and the corresponding range is expanded based on the density distribution of the feature loss region to obtain an expanded feature loss region;

[0172] The extended feature missing area is complementarily sampled to obtain complementarily sampled point cloud data;

[0173] The complementarily sampled point cloud data and the lightweight point cloud data are processed by using an iterative closest point algorithm to obtain a rigid transformation parameter; the rigid transformation parameter is a three-dimensional space transformation parameter including a rotation angle and a translation distance, and is used for aligning the point cloud from one coordinate system to another coordinate system;

[0174] The complementarily sampled point cloud data is registered according to the rigid transformation parameter to generate registered point cloud data;

[0175] In an overlapping area of the registered point cloud data, a smooth surface of the overlapping area is fitted by using a moving least square method;

[0176] Based on the smooth surface, new point coordinates are generated in the feature missing area;

[0177] The generated new point coordinates and the lightweight point cloud data are merged to obtain repaired point cloud data.

[0178] Exemplarily, the embodiment provides a method for geometric feature loss adaptive repair of lightweight point cloud data in an archival digital processing intelligent storage system, and the method specifically includes the following steps:

[0179] The method for extracting geometric feature points of the lightweight point cloud data and the archival point cloud data by using a scale-invariant feature transform algorithm includes the following steps:

[0180] For the input point cloud data, a plurality of scales are generated in a three-dimensional space. By setting a scale parameter sequence, for example, σ = 1.0, 1.5, 2.0, a Gaussian blur function is applied to smooth the point cloud data to obtain blurred point clouds of different scales; each scale corresponds to a three-dimensional space representation, which is used for subsequent detection of scale-invariant feature points;

[0181] For the blurred point cloud of each scale, the difference between the blurred point cloud and the adjacent scale blurred point cloud is calculated to obtain a Gaussian difference function value; the Gaussian difference function value is used to identify candidate positions with significant responses in the point cloud, and to ensure that the feature points can be detected at different scales;

[0182] The method for determining the significant response includes the following steps:

[0183] Any point in the Gaussian difference function value space of the current scale is compared with eight spatially adjacent points in the same scale space, nine points at the corresponding position in the previous adjacent scale, and nine points at the corresponding position in the next adjacent scale; if the Gaussian difference function value of the current point is greater than the values of all 26 adjacent points, that is, the current point is a local maximum value, or less than the values of all 26 adjacent points, that is, the current point is a local minimum value, it is determined that the current point has a significant response, and the current point is marked as a candidate key point.

[0184] using a three-dimensional quadratic function to approximate the variation of the Gaussian difference function value at the candidate key point and in a small neighborhood of the candidate key point, with the three-dimensional coordinates and the scale value of the candidate key point as initial points; the scale value is a scale parameter in the scale space;

[0185] solving the extreme value point corresponding to the extreme value of the three-dimensional quadratic function;

[0186] using the three-dimensional coordinates and the scale value of the extreme value point as the position and the scale of the final precise positioning of the candidate key point;

[0187] Meanwhile, the Gaussian difference function value corresponding to the extreme value point is calculated; if the Gaussian difference function value is lower than a set contrast threshold, it is determined that the contrast of the candidate key point is too low, and the candidate key point is removed; otherwise, the candidate key point is reserved as a precisely positioned key point; in the embodiment, the contrast threshold can be set to 0.03;

[0188] For each precisely positioned key point, the gradient direction distribution of points in a neighborhood of the key point is calculated; the radius of the neighborhood is determined by the scale value; a gradient direction histogram is constructed using three-dimensional gradient information, a main direction is identified as a peak value of the histogram, and a possible auxiliary direction is detected as a sub-peak value; this ensures that the feature point has rotation invariance;

[0189] In a local coordinate system after rotation to the main direction, 4x4x4 sub-regions are divided around the key point; in each sub-region, a gradient direction histogram is calculated, and the histogram values of all sub-regions are spliced to form a normalized feature descriptor vector;

[0190] At this point, all the geometric feature points obtained contain position, scale, main direction, and feature descriptor vector, which are used for subsequent FLANN algorithm matching.

[0191] A method for matching geometric feature points of lightweight point cloud data and archive point cloud data through a FLANN algorithm to obtain a matching point pair set includes:

[0192] The three-dimensional coordinates and the normalized feature descriptor vector of the geometric feature points of the archive point cloud data are taken as input, a multi-level tree index structure is established through recursive space division in the FLANN algorithm, the dimension with the largest variance is selected as the division axis, and the median point is taken as the division value for space division;

[0193] Each geometric feature point of the lightweight point cloud data is traversed, and the nearest neighbor point in the multi-level tree index structure of the archive point cloud data is searched according to the Euclidean distance;

[0194] Meanwhile, reverse verification is performed on each matched point pair, and the nearest neighbor of the corresponding geometric feature point is searched in the lightened point cloud data; when two feature points are the nearest neighbors of each other, the two feature points are determined as a bidirectional matched point pair;

[0195] Cosine similarity between normalized feature descriptor vectors of all bidirectional matched point pairs is calculated, and the formula is the dot product of two feature descriptor vectors divided by the product of their modules;

[0196] A similarity threshold is set, and an exemplary setting is 0.7. Bidirectional matched point pairs below the threshold are removed to obtain a final matched point pair set.

[0197] The Euclidean distance of each matched point pair in the matched point pair set is calculated.

[0198] Geometric feature points with a Euclidean distance greater than a set distance threshold are marked as feature loss points.

[0199] The method for clustering feature loss points to generate a three-dimensional bounding box of a feature loss region comprises:

[0200] A neighborhood radius is set as a spatial distance threshold, and a minimum point number is specified as a minimum point number for constituting a valid region;

[0201] Starting from an arbitrary unvisited feature loss point, other points within the neighborhood radius of the feature loss point are checked;

[0202] If the number of points in the neighborhood is greater than or equal to the minimum point number, the points are classified into the same class, and the neighborhood radius range of all points in the neighborhood is recursively expanded until the expansion cannot continue;

[0203] The above process is repeated until all feature loss points are visited, and the feature loss points are divided into several independent point sets, each point set representing a feature loss region; each point set satisfies:

[0204] Any two points in the set can be connected through adjacent neighborhoods within the neighborhood radius;

[0205] There is no connected path between points of different sets within the neighborhood radius;

[0206] The minimum coordinate value and the maximum coordinate value of all points in each point set in the x, y, and z axis directions of the three-dimensional coordinate system are determined, and a rectangular boundary region is defined. The rectangular boundary region is a three-dimensional bounding box of the feature loss region.

[0207] The method for expanding the range of the feature loss region based on the density distribution of the feature loss region comprises:

[0208] The feature loss region is divided into a cubic grid with an edge length of a multiple of the point cloud data space resolution tz, and the number of feature loss points in each grid is counted; the point cloud data space resolution is the point interval of the archive point cloud data;

[0209] The grid density is obtained by dividing the number of feature loss points by the grid volume;

[0210] Statistical characteristic values of all grid densities are calculated; the statistical characteristic values include quantiles of the grid density distribution, a preset fixed density proportionality coefficient, and a linear combination of the grid density mean and standard deviation;

[0211] A global density threshold is determined according to the statistical characteristic values;

[0212] When the grid density is greater than or equal to the density threshold, the grid is determined to be a feature loss point dense region;

[0213] The feature loss point dense region is expanded in each of the three coordinate axis directions by a corresponding distance to obtain an expanded feature loss region.

[0214] Exemplarily, the embodiment provides a method for expanding the feature loss region in the adaptive repair of geometric feature loss of light-weight point cloud data, which is specifically as follows:

[0215] In the embodiment, the original point distance of the archive point cloud data is 0.3 mm, and the corresponding space resolution is 0.3 mm. The feature loss region is divided into a cubic grid with an edge length of 0.9 mm, and the total number of grids is 25×18×17.

[0216] Let u be a positive integer from 1 to U, where U is the total number of grids. For the u-th grid, count the number of feature loss points C u , and calculate the volume V u of the u-th grid to obtain the grid density ρ u :

[0217]

[0218] The method for calculating the statistical characteristic values of all grid densities includes:

[0219] Calculate the quantiles Q 75 of the grid density distribution; in the embodiment, 75% quantiles are adopted;

[0220] Calculate the linear combination μ ρ +2σ ρ of the grid density mean and standard deviation, where μ ρ is the density mean, and σ ρ is the standard deviation;

[0221] In this embodiment, the fixed proportion coefficient λ is set as an empirical value of 0.6, which can be adjusted appropriately in actual use.

[0222] Determine the global density threshold ρ according to the statistical characteristic value th

[0223] ρ th = max(Q 75 , μ ρ + 2σ ρ ) * λ

[0224] When the grid density is greater than or equal to the density threshold, it is determined that the grid is a feature loss point dense area

[0225] The feature loss point dense area is extended by 10% along each axis to form a supplementary sampling target area, which ensures that the potential extension of the archive feature is covered. In the government paper archive management scene, when the archive text has features such as fading, crease breaking, etc., the original scanning may only capture the local incomplete part, and the actual text strokes and crease traces may have extensions or fine branches. By extending by 10% to form a buffer area, the end of the archive text stroke that is not completely detected and the potential extension of the crease can be effectively covered, avoiding the omission of important information.

[0226] In the construction engineering archive scene, for the structural feature loss at the beam-column connection, if the original three-dimensional bounding box only contains the detected feature loss points, the potential structural damage extension caused by corrosion and wear may be ignored. The supplementary sampling target area formed by extending by 10% along each axis can completely cover the undetected fine cracks and material peeling that may exist at the beam-column connection, providing a more comprehensive and accurate data basis for subsequent engineering structure analysis and maintenance.

[0227] The method for supplementally sampling the extended feature loss area to obtain supplementary point cloud data comprises:

[0228] Based on the spatial range defined by the three-dimensional bounding box of the extended feature loss area, all point cloud coordinates located inside the three-dimensional bounding box are extracted from the archive point cloud data;

[0229] All points whose coordinate values are between the minimum coordinate value and the maximum coordinate value of the three-dimensional bounding box are filtered out to jointly constitute the supplementary point cloud data.

[0230] This process ensures that the supplementary point cloud data completely covers the spatial range of the extended feature loss area and is derived from high-precision archive point cloud data, thereby providing accurate local point cloud information for subsequent registration and repair. The supplementary point cloud data will be used as an independent point cloud block for registration operation with the original lightweight point cloud data.

[0231] ​The method for obtaining the rigid transformation parameter by using the iterative closest point algorithm to process the supplementary point cloud data and the lightweight point cloud data comprises the following steps:

[0232] For each point of the supplementary point cloud data, a nearest neighbor point is found in the lightweight point cloud data to establish a corresponding point pair set;

[0233] A minimum objective function of the Euclidean distance between the corresponding points in the corresponding point pair set is set to obtain the current optimal rigid transformation parameter;

[0234] The iteration is terminated when the change value of continuous lc iterations is less than a set change threshold or reaches a set maximum iteration number;

[0235] The rigid transformation parameter obtained by the final iteration is output.

[0236] For example, the maximum iteration number is set to 50 times, the continuous stable number is set to 5 times, and the change threshold is set to 0.0001.

[0237] For each data point in the supplementary point cloud data, a nearest neighbor point with the closest Euclidean distance is found in the lightweight point cloud data to establish a corresponding point pair set;

[0238] A minimum objective function of the Euclidean distance between the corresponding points in the corresponding point pair set is set to obtain the current optimal rigid transformation parameter, and the minimum objective function is as follows:

[0239]

[0240] In the formula, R t is a rotation matrix obtained by the tth iteration; t t is a translation vector obtained by the tth iteration; q v is the vth data point in the supplementary point cloud data; r n(v) is a nearest neighbor point with the closest Euclidean distance to the vth data point in the supplementary point cloud data; C t is the corresponding point pair set at the tth iteration.

[0241] The centroid of the corresponding point pair set is defined as

[0242]

[0243] In the formula, K is the number of data points of the supplementary point cloud data; is the horizontal coordinate of the centroid; is the vertical coordinate of the centroid.

[0244] The decentralized coordinates are calculated

[0245]

[0246] Construct the covariance matrix H:

[0247] (3x3 matrix);

[0248] wherein, is the transpose matrix of U.

[0249] After singular value decomposition of the covariance matrix, a pair of orthogonal matrices U and V are obtained;

[0250] Solve the current optimal rotation matrix

[0251]

[0252] wherein, U T is the transpose matrix of U.

[0253] Solve the current optimal translation vector

[0254]

[0255] After each iteration, calculate the change value Δ of the current optimal rigid transformation parameter t :

[0256]

[0257] wherein, is the rotation change amount calculated by using the matrix trace; is the translation change amount calculated by using the Euclidean norm.

[0258] When the change value of 5 consecutive iterations is less than 0.0001 or the number of iterations reaches 500, terminate the iteration;

[0259] Output the rigid transformation parameter obtained by the final iteration.

[0260] The method for registering the supplementary point cloud data according to the rigid transformation parameter comprises:

[0261] Based on the rotation angle, rotate the data points in the supplementary point cloud data to change their directions to match the coordinate system of the lightweight point cloud data;

[0262] Based on the translation distance, translate the data points in the supplementary point cloud data to adjust their positions to align to the coordinate system of the lightweight point cloud data;

[0263] The transformed point coordinate set constitutes the registered point cloud data, which has been spatially aligned with the lightweight point cloud data, ensuring that the surface fitting and repairing operations in the overlapping area in the subsequent steps can be performed in a unified coordinate system. The registered point cloud data retains the density and characteristics of the supplementary point cloud data, but the coordinate system has been corrected to be consistent with the lightweight point cloud data.

[0264] In the overlapping area of the registered point cloud data, a smooth surface of the overlapping area is fitted using a moving least squares method, which includes:

[0265] The overlapping area is the part of the registered point cloud data that spatially coincides with the lightweight point cloud data, which is determined by coordinate comparison to ensure that it contains effective point cloud information near the feature loss area;

[0266] Each point coordinate in the overlapping area is traversed to define a local neighborhood range centered on the point coordinate;

[0267] In the local neighborhood, weights are assigned based on the distance of points in the local neighborhood to the center point, with closer distances having larger weights, and a local polynomial surface is fitted using the least squares principle, which ensures that the surface is continuously smooth and can adapt to the geometric changes of the point cloud;

[0268] This operation is repeated to cover the entire overlapping area, generating a unified smooth surface;

[0269] The smooth surface completely describes the surface geometric characteristics of the overlapping area, providing a basis for generating new point coordinates in the feature loss area.

[0270] The entire fitting process relies on the adaptive characteristics of the moving least squares method, avoiding the introduction of additional parameters, and only using existing point cloud data to achieve high-precision reconstruction of the surface.

[0271] Based on the smooth surface, a method for generating new point coordinates in the feature loss area includes:

[0272] Determine the target spatial range according to the three-dimensional bounding box of the feature loss area, which has been processed by density distribution expansion;

[0273] In the target spatial range, the spatial positions of the points to be generated are uniformly divided according to the predetermined point cloud resolution requirements;

[0274] By querying the corresponding points of the spatial positions of the points to be generated on the smooth surface, the three-dimensional coordinate values are directly obtained;

[0275] Traverse all the predetermined spatial position points in the feature loss area, and calculate the new point coordinates point by point.

[0276] The generated new point coordinates have the same attribute format as the original point cloud, providing compatibility guarantee for subsequent merging operations. The entire generation process completely relies on the geometric information of the smooth surface, without additional parameter input, ensuring the adaptability and geometric consistency of the repair result.

[0277] The new point coordinates are merged with the lightweight point cloud data to obtain the repaired point cloud data.

[0278] The geometric feature points are extracted by the scale-invariant feature transform algorithm and combined with the FLANN bidirectional matching to solve the problem of inaccurate detection caused by feature point mismatching in traditional methods. The scale-invariance of the scale-invariant feature transform algorithm ensures the establishment of a cross-resolution correspondence between the lightweight data and the archive data. The KD tree index of FLANN combined with bidirectional verification controls the mismatching rate below 3%, which is 20% lower than the single nearest neighbor search, ensuring the reliability of the subsequent feature loss point marking.

[0279] The dense area is determined based on the grid density and statistical feature value, breaking through the blindness of traditional fixed expansion ratio. For example, when the original point distance of the paper archive point cloud is 0.3mm, 0.9mm grid division can capture the point distribution characteristics of 0.3mm width archive crease, and through 75% quantile threshold, the real crease dense area is selected, and then a 10% buffer zone is formed along the axial direction, ensuring the coverage of secondary crease with 0.5mm extension. In the scene of construction engineering archives, for the structural features of beam-column connections, this method can accurately locate the dense area of slight damage caused by corrosion and wear, providing a reliable basis for subsequent resampling.

[0280] The improved iterative closest point algorithm converges through 50 iterations + 0.0001 change threshold, combined with centroid alignment and singular value decomposition of covariance matrix, to reduce the registration error to below 0.15mm. Taking the resampling data of beam-column connections in construction engineering archives as an example, the point cloud data scanned at different times is registered, the trace change of rotation matrix R is less than 0.01, and the Euclidean distance of translation vector t is less than 0.05mm, ensuring that the spatial alignment error of resampling point cloud and lightweight data in the structural damage area is less than 0.2mm. The traditional iterative closest point algorithm does not introduce feature weight, and when dealing with complex structures such as beam-column edges, the registration error reaches 0.8mm, leading to misplacement of repaired structural features and affecting the accuracy of engineering detection.

[0281] The moving least squares method uses a σ=0.8mm Gaussian kernel function to generate C 2The continuous curved surface solves the step-shaped discontinuity problem caused by the traditional interpolation method. For example, when processing the faded text edge on the paper file, the smoothness of the text contour is significantly improved after MLS fitting, the surface roughness is reduced from 1.2 μm to 0.8 μm, and the normal vector deviation at the stroke turning point is less than 5°. While the normal vector deviation of the traditional polynomial fitting at the curvature mutation such as the text edge is up to 15°, which causes the distortion of the repaired text form and affects the identification of the file content. When generating new points based on the ray tracing of the curved surface, the 0.15mm sampling interval and the 15% curvature consistency constraint ensure that the 0.2mm wide fine strokes on the paper file and the 0.2mm wide structural gaps in the construction engineering file are completely reconstructed.

[0282] Embodiment 2

[0283] Please refer to Figure 2 The embodiment provides an intelligent storage method for digital processing of archives, comprising:

[0284] Based on the collected archive point cloud data, boundary feature point detection and detail rich area detection are performed to obtain a boundary feature mask and a detail rich mask of the archive point cloud data respectively; and a semantic feature label is generated by combining a preset domain knowledge base and performing semantic feature labeling on the archive point cloud data.

[0285] Based on the archive point cloud data, the boundary feature mask, the detail rich mask and the semantic label, a priority value of each data point in the archive point cloud data is calculated, and the archive point cloud data is divided into a feature region and a non-feature region according to the priority value.

[0286] The feature region and the non-feature region of the archive point cloud data are respectively subjected to lightweight processing, and the processed regions are merged to obtain lightweight point cloud data.

[0287] The lightweight point cloud data is subjected to adaptive repair of geometric feature loss based on the archive point cloud data, to obtain repaired point cloud data.

[0288] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any skilled person in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0289] Finally: the above is only a preferred embodiment of the present application and is not used to limit the present application, any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A method for intelligent storage of digitized archives, characterized in that: include: Based on the collected archival point cloud data, boundary feature point detection and detail-rich region detection are performed to obtain the boundary feature mask and detail-rich mask of the archival point cloud data, respectively. Semantic tags are generated by semantically marking archival point cloud data using a pre-defined domain knowledge base. Based on archival point cloud data, boundary feature masks, detail-rich masks, and semantic labels, the priority value of each data point in the archival point cloud data is calculated, and the archival point cloud data is divided into feature regions and non-feature regions according to the priority value. The feature regions and non-feature regions of the archival point cloud data are subjected to lightweight processing respectively. The processed regions are then merged to obtain lightweight point cloud data. Based on the archival point cloud data, the lightweight point cloud data is adaptively repaired for the loss of geometric features, resulting in the repaired point cloud data.

2. The intelligent storage method for digitized archival processing according to claim 1, characterized in that, Methods for obtaining lightweight point cloud data include: Calculate the average point spacing of data points in non-feature regions; Set the voxel mesh size to ts times the average pixel spacing; A 3D voxel grid is constructed according to the voxel grid size. Each voxel grid is traversed, and only the center point of each voxel grid is retained to obtain point cloud data after thinning of non-feature regions. Using each feature point within the feature region as the center, a K-nearest neighbor search is performed within a spherical neighborhood with a radius of ly to obtain the point cloud data after thinning the feature region. If the number of sample points in the spherical neighborhood of the feature point is insufficient, the nearest neighbor points with the Euclidean distance to the feature point will be selected from the archive point cloud data in turn as supplements until the set number of search points is met. The point cloud data after thinning the non-feature regions and the point cloud data after thinning the feature regions are merged to obtain lightweight point cloud data.

3. The intelligent storage method for digitized archival processing according to claim 1, characterized in that, Methods for adaptive repair of geometric feature loss in lightweight point cloud data include: The scale-invariant feature transform algorithm is used to extract geometric feature points from lightweight point cloud data and archival point cloud data, respectively. The FLANN algorithm is used to match the geometric feature points of lightweight point cloud data and archival point cloud data to obtain a set of matching point pairs. Calculate the Euclidean distance for each pair of matching points in the set of matching point pairs; Geometric feature points whose Euclidean distance is greater than a set distance threshold are marked as feature loss points; Clustering is performed on the missing feature points to generate a 3D bounding box for the missing feature region. The corresponding range is then expanded based on the density distribution of the missing feature region to obtain the expanded missing feature region. The expanded feature loss area is re-sampled to obtain the re-sampled point cloud data; The iterative nearest point algorithm is used to process the supplementary point cloud data and the lightweight point cloud data to obtain rigid transformation parameters. The rigid transformation parameters are three-dimensional spatial transformation parameters including rotation angle and translation distance, which are used to align the point cloud from one coordinate system to another. Registered point cloud data is generated by registering and supplementing point cloud data based on rigid transformation parameters. In the overlapping region of the registered point cloud data, a smooth surface of the overlapping region is fitted using the moving least squares method. Based on a smooth surface, new point coordinates are generated within the region where features are lost. The newly generated point coordinates are merged with the lightweight point cloud data to obtain the repaired point cloud data.

4. The intelligent storage method for digitized archival processing according to claim 3, characterized in that, Methods for obtaining the expanded feature loss region include: The region of lost features is divided into a cubic grid with a side length tz times the spatial resolution of the point cloud data, and the number of lost feature points in each grid is counted; the spatial resolution of the point cloud data is the point interval of the archive point cloud data. The grid density is obtained by dividing the number of missing feature points by the grid volume. Calculate the statistical characteristic values ​​of all grid densities; the statistical characteristic values ​​include the quantiles of the grid density distribution, the preset fixed density ratio coefficient, and the linear combination of the grid density mean and standard deviation; Determine the global density threshold based on statistical characteristic values; When the grid density is greater than or equal to the density threshold, the grid is determined to be a dense region of feature loss points; The region with densely missing feature points is extended by a corresponding distance along each of the three coordinate axes to obtain the extended region with missing features.

5. The intelligent storage method for digitized archival processing according to claim 3, characterized in that, Methods for obtaining rigid transformation parameters include: For each point in the supplementary point cloud data, the nearest neighbor point is found in the lightweight point cloud data to establish a corresponding set of point pairs; The optimal rigid transformation parameters are obtained by setting an objective function that minimizes the Euclidean distance between corresponding points in the set of corresponding point pairs. The iteration terminates when the change value after lc consecutive iterations is less than the set change threshold or the set maximum number of iterations is reached. The output yields the rigid transformation parameters obtained from the final iteration.

6. The intelligent storage method for digitized archival processing according to claim 1, characterized in that, Methods for obtaining boundary feature masks of archive point cloud data include: S101: Let the initial value of index variable p be 1, let the initial value of index variable q be 1, the value range of p is 1 to P, and the value range of q is 1 to Q; P is the number of data points in the archive point cloud data, and Q is the preset number of neighborhood points; S102: Exclude data points that have been marked as boundary feature points, select the point with the closest Euclidean distance to the p-th data point, mark it as a near point, and calculate the angle between the normal vector of the near point and the p-th data point; S103: If the angle between the normal vectors is greater than the preset threshold for the angle between the normal vectors, then mark the nearest point as a boundary feature point and generate a binary form of the boundary feature mask for the nearest point; S104: If q is less than Q, then let q = q + 1 and execute S102; otherwise, let q = 1 and execute S105. S105: If p is less than P, then let p = p + 1 and execute S102; otherwise, construct the boundary feature mask of the archive point cloud data from the binary form of the boundary feature masks of all data points marked as boundary feature points.

7. The intelligent storage method for digitized archival processing according to claim 1, characterized in that, Methods for obtaining detailed rich masks of archive point cloud data include: A spherical neighborhood with radius xj is defined with each data point in the archive point cloud data as the center. The spherical neighborhood is divided into sl voxel grids, and the number of data points in each voxel grid is counted. For each voxel grid, calculate the proportion of the number of points within it to the total number of points in the spherical neighborhood to obtain the probability distribution value for each voxel grid; The spatial distribution entropy of each data point in the archival point cloud data is equal to the sum of the negative values ​​of the product of the probability distribution values ​​of each voxel grid and the natural logarithm of the probability distribution values ​​of the voxel grid where each data point is located in the archival point cloud data. If the spatial distribution entropy value of a data point is greater than the preset entropy value, then the data point is marked as a detail-rich point, and a detail-rich mask in binary form is generated for the data point. Construct a rich detail mask of the archive point cloud data from the binary form of all rich detail points.

8. The intelligent storage method for digitized archival processing according to claim 1, characterized in that, Methods for dividing archival point cloud data into feature regions and non-feature regions include: The RGB color value and laser echo intensity of each data point in the archive point cloud data, as well as the one-hot encoding of the boundary feature mask, the detail-rich mask, and the semantic label, are converted into feature vectors. The weights of each feature vector are calculated through a multi-head self-attention mechanism. The weights of each feature vector satisfy the condition that the sum of the weights is 1. The feature confidence score is obtained by weighted summation of the feature vector and each weight. Geometric feature mask scores are calculated based on boundary feature masks and detail-rich masks. Semantic tag scores are obtained by combining a pre-set domain knowledge base; The priority value of each data point in the archive point cloud data is obtained by weighted summing of the geometric feature mask score, semantic label score and feature confidence score. Data points with priority values ​​greater than a set priority threshold are identified as feature points and constructed as feature regions; data points with priority values ​​less than or equal to a set priority threshold are identified as non-feature points and constructed as non-feature regions.

9. The intelligent storage method for digitized archival processing according to claim 8, characterized in that, Methods for obtaining geometric feature mask scores include: The geometric feature score of all data points in the archive point cloud data is calculated by using boundary feature mask and rich detail mask. If a data point is a boundary feature point or a rich detail point, the score is df; otherwise, it is 0.

10. An intelligent storage system for digitized archival processing, used to implement the intelligent storage method for digitized archival processing as described in any one of claims 1-9, characterized in that, include: The feature labeling module performs boundary feature point detection and rich detail region detection based on the collected archival point cloud data, obtaining the boundary feature mask and rich detail mask of the archival point cloud data respectively; and generates semantic labels by combining the pre-set domain knowledge base to perform semantic feature labeling on the archival point cloud data. The feature partitioning module calculates the priority value of each data point in the archive point cloud data based on the archive point cloud data, boundary feature mask, detail rich mask and semantic label, and divides the archive point cloud data into feature regions and non-feature regions according to the priority value; The lightweight module is used to perform lightweight processing on the feature regions and non-feature regions of the archival point cloud data respectively, and then merge the processed regions to obtain lightweight point cloud data. The adaptive repair module performs adaptive repair of geometric feature loss on lightweight point cloud data based on archival point cloud data, resulting in repaired point cloud data.

Images