An intelligent duplicate checking method and system for rental items based on an artificial intelligence large model

Abstract

The application relates to the technical field of duplicate checking of rental objects, and discloses a rental object intelligent duplicate checking method and system based on an artificial intelligence large model, which comprises the following steps: adopting an AI large model-based intelligent identification technology to automatically identify and structurally extract rental object registration files, and determining ownership document text features; acquiring appearance image features and structural light point cloud features of a to-be-checked rental object, and constructing a multi-modal feature dataset; based on the multi-modal feature dataset, a digital fingerprint atlas is constructed; the digital fingerprint atlas is subjected to similarity detection analysis with a preset rental object historical archives library, and multi-dimensional similarity evaluation results are obtained; according to the multi-dimensional similarity evaluation results, comprehensive duplicate checking risk scores and risk tolerance parameters are determined, a duplicate checking risk grade is judged, and a duplicate checking conclusion is generated; through multi-dimensional analysis and similarity analysis, the automatic identification, intelligent comparison and risk assessment of rental object information are realized, the duplicate checking efficiency and accuracy are improved, and the business risk is reduced.

Description

Technical Field

[0001] This invention relates to the field of rental property deduplication technology, and more specifically, to an intelligent deduplication method and system for rental properties based on a large artificial intelligence model. Background Technology

[0002] Asset duplication detection refers to a risk control technology that prevents the same equipment from being mortgaged or repeatedly leased by the same person through identification and duplication detection of the assets to be leased. It is a core link in ensuring the security of assets in financial leasing business.

[0003] Existing technologies for detecting duplicate leased assets primarily rely on manual verification of ownership certificates and simple text keyword comparison, checking equipment serial numbers, purchase invoices, and other information to determine if duplicates exist. However, existing technologies are ill-equipped to handle fraudulent activities such as forged ownership certificates and cloned equipment, and cannot identify duplicate leases made by altering nameplates or forging documents. Furthermore, they depend heavily on human experience and lack intelligent risk assessment capabilities, making it difficult to identify potential risks. In addition, the single text comparison method lacks the ability to verify the physical form and microscopic features of the equipment, and cannot adapt to the varying risk tolerance levels in different business scenarios, resulting in low accuracy and high false alarm rates, making it difficult to meet the risk control needs of large-scale leasing businesses. Summary of the Invention

[0004] To address the aforementioned technical issues, the purpose of this application is to provide an intelligent deduplication method and system for leased assets based on a large-scale artificial intelligence model. This method performs comprehensive comparative analysis and similarity analysis of leased assets from multiple dimensions, enabling automatic identification, intelligent comparison, and risk assessment of leased asset information, thereby improving deduplication efficiency and accuracy and reducing business risks.

[0005] To achieve the above objectives, this invention provides an intelligent deduplication method for leased items based on a large-scale artificial intelligence model, comprising: The intelligent recognition technology based on AI large model is used to automatically identify and extract the structure of the leased property registration documents to determine the text features of the ownership documents; Obtain the appearance image features and structured light point cloud features of the leased property to be investigated, and construct a multimodal feature dataset based on the ownership document text features, appearance image features, and structured light point cloud features; Based on the multimodal feature dataset, a digital fingerprint map of the leased property to be searched is constructed. The digital fingerprint map represents the comprehensive feature distribution of the leased property to be searched in the visual geometric space and semantic description space. The digital fingerprint map is compared with a preset historical archive of leased items for similarity detection and analysis to obtain a multi-dimensional similarity evaluation result. The multi-dimensional similarity evaluation result reflects the degree of feature overlap between the leased item to be investigated and the historical archived leased items. Based on the multi-dimensional similarity assessment results, a comprehensive plagiarism risk score and risk tolerance parameters are determined. Based on the comprehensive plagiarism risk score and risk tolerance parameters, the plagiarism risk level is determined and a plagiarism conclusion is generated. The plagiarism conclusion includes a duplicate rental warning sign and risk tracing information.

[0006] Furthermore, based on the multimodal feature dataset, a digital fingerprint map of the leased property to be searched is constructed, including: Temporal alignment and coordinate system unification processing are performed on the appearance image features, the structured light point cloud features, and the ownership document text features to establish a standardized feature description framework. Geometric contour features, surface texture features, and ownership keyword features are extracted from the standardized feature description framework. The geometric contour features include edge curvature distribution and size ratio parameters. The surface texture features include local binary pattern encoding and depth difference histogram. The ownership keyword features include owner identifier and device serial number. Analyze the spatial correlation strength and semantic consistency among the geometric contour features, surface texture features, and ownership keyword features; Based on the spatial association strength and semantic consistency, a cross-modal feature mapping relationship is established, and the digital fingerprint map of the leased property is generated by fusion.

[0007] Furthermore, before analyzing the spatial correlation strength and semantic consistency among the geometric contour features, the surface texture features, and the ownership keyword features, the process also includes: Calculate the registration error between the geometric contour features and the surface texture features in the spatial coordinate system; Evaluate the semantic fit between the ownership keyword features and the geometric contour features; Based on the registration error and the semantic fit, a cross-modal consistency index is constructed; When the cross-modal consistency index falls below the confidence threshold, feature re-acquisition is triggered. When the cross-modal consistency index is higher than or equal to the confidence threshold, the spatial correlation strength and semantic consistency among the geometric contour features, the surface texture features, and the ownership keyword features are analyzed.

[0008] Furthermore, the digital fingerprint map is compared with a pre-set historical archive of leased items for similarity detection and analysis to obtain multi-dimensional similarity evaluation results, including: Calculate the feature distance between the digital fingerprint map and each archived fingerprint in the leased property's historical archive; The feature distances are sorted in descending order, and the archived fingerprints with the smallest feature distances are selected as candidate matching objects. Determine whether the minimum feature distance is lower than a preset similarity threshold; When the minimum feature distance is lower than the preset similarity threshold, it is confirmed that there is suspected duplicate rental behavior and the archived information of the candidate matching object is extracted; When the minimum feature distance is higher than or equal to the preset similarity threshold, it is determined to be a newly added leased item and the digital fingerprint map is archived and stored.

[0009] Further, calculating the feature distance between the digital fingerprint spectrum and each archived fingerprint in the leased property history database includes: The geometric contour features, surface texture features, and ownership keyword features are constructed into a multi-level feature pyramid, which includes a coarse-grained global feature layer and a fine-grained local feature layer. Based on the coarse-grained global feature layer, a fast pre-screening is performed to remove archived fingerprints whose difference from the digital fingerprint map exceeds a preset screening threshold, thereby obtaining a subset of candidate similar fingerprints. For each archived fingerprint in the candidate similar fingerprint subset, the feature discrimination index of each feature dimension in the fine-grained local feature layer is calculated, and an adaptive weight matrix is ​​dynamically generated based on the feature discrimination index. The feature discrimination index reflects the information entropy contribution of the feature dimension in distinguishing different leased items. Based on the adaptive weight matrix, the weighted subspace distance of each feature dimension is calculated, and a feature association graph model is constructed. The topological constraint relationship between the geometric contour features and the surface texture features is encoded through a graph convolutional network to obtain the structural similarity correction factor. The feature distance is calculated by fusing the weighted subspace distance and the structural similarity correction factor.

[0010] Furthermore, the method for determining the preset similarity threshold includes: Obtain the characteristic distribution of confirmed duplicate cases and false alarm cases from the leased property history archive; Analyze the statistical boundary characteristics of the characteristic distributions of duplicate cases and false alarm cases; Based on the aforementioned statistical boundary characteristics, calculate the decision boundary values ​​for recall and precision; Obtain market activity indicators for the current leasing business period; The decision boundary value is dynamically adjusted based on the market activity index to obtain the preset similarity threshold.

[0011] Furthermore, based on the multi-dimensional similarity assessment results, a comprehensive plagiarism risk score and risk tolerance parameters are determined, including: The geometric similarity index, texture matching index, and ownership association index in the multi-dimensional similarity evaluation results are evaluated. The geometric similarity index reflects the degree of physical morphological overlap, the texture matching index reflects the consistency of surface micro-features, and the ownership association index reflects the degree of ownership information conflict. A comprehensive plagiarism risk score is calculated based on the geometric similarity index, the texture matching index, and the ownership association index. Obtain the risk tolerance parameters for the current leasing business scenario.

[0012] Furthermore, based on the comprehensive plagiarism risk score and risk tolerance parameter, the plagiarism risk level is determined and a plagiarism conclusion is generated, including: Based on the comparison between the comprehensive plagiarism risk score and the risk tolerance parameter, the plagiarism risk level is classified as high risk, medium risk, or low risk. Differentiated plagiarism detection conclusions are generated based on the aforementioned plagiarism risk level.

[0013] To achieve the above objectives, the present invention also provides an intelligent deduplication system for leased items based on a large-scale artificial intelligence model, comprising: The text determination module is used to automatically identify and extract the structure of leased property registration documents using AI-based large model intelligent recognition technology, and determine the text features of ownership documents; A collection construction module is used to obtain the appearance image features and structured light point cloud features of the leased property to be investigated, and to construct a multimodal feature dataset based on the ownership document text features, appearance image features and structured light point cloud features; The graph construction module is used to construct a digital fingerprint graph of the leased property to be searched based on the multimodal feature dataset. The digital fingerprint graph represents the comprehensive feature distribution of the leased property to be searched in the visual geometric space and semantic description space. The similarity analysis module is used to perform similarity detection and analysis between the digital fingerprint map and the preset historical archive of leased items to obtain a multi-dimensional similarity evaluation result. The multi-dimensional similarity evaluation result reflects the degree of feature overlap between the leased item to be investigated and the historical archived leased items. The intelligent plagiarism detection module is used to determine a comprehensive plagiarism risk score and risk tolerance parameter based on the multi-dimensional similarity evaluation results, determine the plagiarism risk level based on the comprehensive plagiarism risk score and risk tolerance parameter, and generate a plagiarism detection conclusion. The plagiarism detection conclusion includes a duplicate rental warning sign and risk tracing information.

[0014] Furthermore, it also includes: The confidence detection module is used for: Calculate the registration error between geometric contour features and surface texture features in the spatial coordinate system; Evaluate the semantic consistency between the ownership keyword features and the geometric contour features; Based on the registration error and the semantic fit, a cross-modal consistency index is constructed; When the cross-modal consistency index falls below the confidence threshold, feature re-acquisition is triggered. When the cross-modal consistency index is higher than or equal to the confidence threshold, the spatial correlation strength and semantic consistency among the geometric contour features, the surface texture features, and the ownership keyword features are analyzed.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention discloses an intelligent deduplication method and system for leased assets based on an artificial intelligence large-scale model. The method includes: automatically identifying and structurally extracting the registration documents of the leased assets using intelligent recognition technology based on an AI large-scale model to determine the textual features of the ownership documents; acquiring the appearance image features and structured light point cloud features of the leased asset to be checked, and constructing a multimodal feature dataset; constructing a digital fingerprint map based on the multimodal feature dataset; performing similarity detection analysis between the digital fingerprint map and a preset historical archive of leased assets to obtain a multi-dimensional similarity evaluation result; determining a comprehensive deduplication risk score and risk tolerance parameters based on the multi-dimensional similarity evaluation result, judging the deduplication risk level, and generating a deduplication conclusion. Through multi-dimensional analysis and similarity analysis, the automatic identification, intelligent comparison, and risk assessment of leased asset information are achieved, improving the efficiency and accuracy of deduplication and reducing business risks. Attached Figure Description

[0016] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart illustrating an intelligent deduplication method for leased items based on a large artificial intelligence model is shown in an embodiment of the present invention. Figure 2 The diagram shows a structural schematic of an intelligent deduplication system for leased items based on a large artificial intelligence model, according to an embodiment of the present invention. Detailed Implementation

[0017] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0018] In the description of this application, it should be understood that the terms "center", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0019] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0020] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0021] The following is a description of preferred embodiments of the present invention in conjunction with the accompanying drawings.

[0022] like Figure 1 As shown, an embodiment of the present invention discloses an intelligent deduplication method for leased items based on a large artificial intelligence model, including: S110: Employ AI-based large-scale model-based intelligent recognition technology to automatically identify and extract structured information from leased property registration documents, thereby determining the textual features of ownership documents; S120: Obtain the appearance image features and structured light point cloud features of the leased property to be investigated, and construct a multimodal feature dataset based on the ownership document text features, appearance image features and structured light point cloud features; S130: Based on the multimodal feature dataset, construct a digital fingerprint map of the leased property to be searched, wherein the digital fingerprint map represents the comprehensive feature distribution of the leased property to be searched in the visual geometric space and semantic description space; S140: Perform similarity detection and analysis between the digital fingerprint map and the preset historical archive of leased items to obtain a multi-dimensional similarity evaluation result. The multi-dimensional similarity evaluation result reflects the degree of feature overlap between the leased item to be investigated and the historical archived leased items. S150: Determine the comprehensive plagiarism risk score and risk tolerance parameter based on the multi-dimensional similarity evaluation results, determine the plagiarism risk level based on the comprehensive plagiarism risk score and risk tolerance parameter, and generate a plagiarism conclusion. The plagiarism conclusion includes a duplicate rental warning sign and risk tracing information.

[0023] In this embodiment, the intelligent recognition technology supports high-precision recognition of documents in multiple formats such as PDF, JPG, and PDF with embedded images, and converts unstructured documents into structured JSON data.

[0024] In this embodiment, the appearance image features and structured light point cloud features are data describing the different dimensions of the leased property, obtained through various sensors and acquisition methods. The appearance image features are obtained by taking six-view images of the leased property with a high-resolution industrial camera, and include visual information such as color, shape, and surface patterns. The structured light point cloud features are obtained by a 3D structured light scanner and include spatial information such as the three-dimensional geometric dimensions and surface details of the leased property. The ownership document text features include text information from documents such as purchase invoices, certificates of conformity, and ownership certificates.

[0025] In some embodiments of this application, constructing a digital fingerprint map of the leased property to be searched based on the multimodal feature dataset includes: Temporal alignment and coordinate system unification processing are performed on the appearance image features, the structured light point cloud features, and the ownership document text features to establish a standardized feature description framework. Geometric contour features, surface texture features, and ownership keyword features are extracted from the standardized feature description framework. The geometric contour features include edge curvature distribution and size ratio parameters. The surface texture features include local binary pattern encoding and depth difference histogram. The ownership keyword features include owner identifier and device serial number. Analyze the spatial correlation strength and semantic consistency among the geometric contour features, surface texture features, and ownership keyword features; Based on the spatial association strength and semantic consistency, a cross-modal feature mapping relationship is established, and the digital fingerprint map of the leased property is generated by fusion.

[0026] In this embodiment, time alignment refers to unifying the acquisition timestamps of the three modalities to the same reference time (such as UTC time), ensuring that the time difference between the time of appearance image capture, point cloud scanning, and document upload is controlled within 5 seconds, avoiding data inconsistencies caused by changes in the status of the leased object. Coordinate system processing refers to transforming the image pixel coordinate system, point cloud world coordinate system, and text logical coordinate system to a standardized device coordinate system with the geometric center of the leased object as the origin, achieving a transformation accuracy of 0.1mm. The standardized feature description framework is a unified data structure containing header information (acquisition time, device ID, coordinate system parameters) and volume information (standardized feature vectors). The edge curvature distribution in geometric contour features is obtained by statistically calculating the curvature radius of edge points in point cloud data; the size ratio parameter is obtained by measuring the length-width-height ratio of the leased object. Local Binary Pattern (LBP) encoding in surface texture features is generated by analyzing the local grayscale distribution of the image; the depth difference histogram is generated by statistically analyzing the angle distribution of surface normal vectors in the point cloud. The owner identifier in ownership keyword features is the company name or individual name extracted from the text using regular expressions; the device serial number is a unique number identified from the nameplate image or document using Optical Character Recognition (OCR). Spatial association strength is obtained by calculating the average of the inverse spatial distances between keypoints in the image and keypoints in the point cloud, with a threshold set to 0.8, indicating a higher spatial location matching degree and stronger association. Semantic consistency is determined by comparing whether the device model described in the text matches the device model identified in the image; the consistency coefficient is 1 when they match and 0 when they do not. The cross-modal feature mapping relationship is a third-order tensor describing the mutual projection relationship of the three features in the latent space, learned through a multimodal autoencoder during the training phase. The fusion generation adopts a cascade fusion strategy: first, the geometric contour features (128 dimensions) and surface texture features (256 dimensions) are fused in the visual space into 384-dimensional visual features, and then fused with the ownership keyword features (128 dimensions) in the semantic space into a 512-dimensional digital fingerprint map of the leased property.

[0027] The beneficial effects of the above technical solution are: spatial and temporal consistency of multi-source data is ensured by temporal alignment and coordinate system unification; comprehensive characterization of the physical attributes of the leased object is achieved by extracting three core features: geometry, texture, and ownership; semantic association between different modalities is established by cross-modal feature mapping; and the generated digital fingerprint map has both uniqueness and robustness, and can effectively resist the loss of single-modal data or noise interference.

[0028] In some embodiments of this application, before analyzing the spatial correlation strength and semantic consistency among the geometric contour features, the surface texture features, and the ownership keyword features, the method further includes: Calculate the registration error between the geometric contour features and the surface texture features in the spatial coordinate system; Evaluate the semantic fit between the ownership keyword features and the geometric contour features; Based on the registration error and the semantic fit, a cross-modal consistency index is constructed; When the cross-modal consistency index falls below the confidence threshold, feature re-acquisition is triggered. When the cross-modal consistency index is higher than or equal to the confidence threshold, the spatial correlation strength and semantic consistency among the geometric contour features, the surface texture features, and the ownership keyword features are analyzed.

[0029] In this embodiment, registration error refers to the spatial position deviation between the image feature points projected onto the point cloud coordinate system and the corresponding point cloud feature points, calculated using the Iterative Closest Point (ICP) algorithm. Semantic consistency is calculated by comparing the device dimensions (e.g., length × width × height) declared in the ownership document with the actual dimensions obtained from the point cloud measurement, determining the relative error. A consistency score of 1 is given when the error is <5%, 0.5 when the error is 5% to 10%, and 0 when the error is >10%. The cross-modal consistency index is calculated as: 0.6 × (1 - registration error / 10 mm) + 0.4 × semantic consistency. The confidence threshold is set to 0.8. When the index is below 0.8, it indicates that there may be acquisition errors in the multimodal data (e.g., shooting with the wrong device, scanning at the wrong angle) or document forgery (e.g., forging a certificate of conformity that does not match the actual product), triggering a feature re-acquisition process (prompting for re-shooting or re-scanning).

[0030] The beneficial effects of the above technical solution are as follows: by verifying both registration error and semantic consistency, a cross-validation mechanism between multimodal data is established, which effectively identifies data inconsistencies caused by collection operation errors or malicious falsification; the cross-modal consistency index enables quantitative assessment of data quality; and the hierarchical processing of re-collection or manual verification ensures data quality while avoiding excessive reliance on manual review, thus balancing automation and accuracy.

[0031] In some embodiments of this application, the digital fingerprint map is compared with a preset historical archive of leased items for similarity detection analysis to obtain multi-dimensional similarity evaluation results, including: Calculate the feature distance between the digital fingerprint map and each archived fingerprint in the leased property's historical archive; The feature distances are sorted in descending order, and the archived fingerprints with the smallest feature distances are selected as candidate matching objects. Determine whether the minimum feature distance is lower than a preset similarity threshold; When the minimum feature distance is lower than the preset similarity threshold, it is confirmed that there is suspected duplicate rental behavior and the archived information of the candidate matching object is extracted; When the minimum feature distance is higher than or equal to the preset similarity threshold, it is determined to be a newly added leased item and the digital fingerprint map is archived and stored.

[0032] In this embodiment, feature distance is a quantitative indicator measuring the similarity between two digital fingerprint profiles. The candidate matching object is selected as the first one after sorting. Suspected duplicate leasing behavior refers to fraudulent behavior in which the same leased item (or highly similar equipment) is leased multiple times at different times or to different lessees. Archived information includes historical lease contract numbers, lessee identity information, leased item entry time, and the previous duplicate check result. If a newly added leased item is determined to be dissimilar to all equipment in the historical archive, its digital fingerprint profile is stored in the archive as a new baseline record, and a new unique equipment file number is generated.

[0033] The beneficial effects of the above technical solution are: efficient similarity retrieval and duplicate detection are achieved through descending order sorting and threshold determination, which improves the recall rate while ensuring the precision rate; through the classification and processing of suspected duplicates and new additions, the risk of duplicate rentals is prevented and the construction of the archive database is improved.

[0034] In some embodiments of this application, calculating the feature distance between the digital fingerprint spectrum and each archived fingerprint in the leased property history archive includes: The geometric contour features, surface texture features, and ownership keyword features are constructed into a multi-level feature pyramid, which includes a coarse-grained global feature layer and a fine-grained local feature layer. Based on the coarse-grained global feature layer, a fast pre-screening is performed to remove archived fingerprints whose difference from the digital fingerprint map exceeds a preset screening threshold, thereby obtaining a subset of candidate similar fingerprints. For each archived fingerprint in the candidate similar fingerprint subset, the feature discrimination index of each feature dimension in the fine-grained local feature layer is calculated, and an adaptive weight matrix is ​​dynamically generated based on the feature discrimination index. The feature discrimination index reflects the information entropy contribution of the feature dimension in distinguishing different leased items. Based on the adaptive weight matrix, the weighted subspace distance of each feature dimension is calculated, and a feature association graph model is constructed. The topological constraint relationship between the geometric contour features and the surface texture features is encoded through a graph convolutional network to obtain the structural similarity correction factor. The feature distance is calculated by fusing the weighted subspace distance and the structural similarity correction factor.

[0035] In this embodiment, the multi-level feature pyramid is a data structure that divides the digital fingerprint spectrum into two levels according to feature granularity: the coarse-grained global feature layer contains a 128-dimensional vector, obtained through principal component analysis (PCA) dimensionality reduction, summarizing the overall geometry, global texture distribution, and ownership overview of the leased property; the fine-grained local feature layer contains a 512-dimensional original vector, preserving edge curvature details, micro-texture patterns, and complete ownership description. The coarse-grained global feature layer is used for rapid pre-screening, calculating the Euclidean distance between the fingerprint to be searched and each archived fingerprint in the database. The preset screening threshold is set to 0.8 (normalized distance value), directly eliminating archived fingerprints with a distance greater than 0.8 (too large difference), retaining only a subset of candidate similar fingerprints with a distance less than or equal to 0.8. The subset of candidate similar fingerprints (e.g., 1000 records) then enters the fine-grained local feature layer for fine comparison. The feature discrimination index is obtained by calculating the information entropy of each feature dimension across all candidate pairs: for the i-th feature dimension, its value distribution across all candidate fingerprints is statistically analyzed to determine the information entropy H(i). The larger the entropy value, the stronger the ability of that dimension to distinguish different leases. The feature discrimination index is defined as D(i) = H(i) / H_max, where H_max is the maximum information entropy, and D(i) ranges from 0 to 1. An adaptive weight matrix W is dynamically generated based on the feature discrimination index. The diagonal elements of the matrix W(i,i) = D(i) / sum(D), and the off-diagonal elements are initially set to 0, so that high-discriminative features receive high weights, while low-discriminative features (such as noisy dimensions) are suppressed. Based on the adaptive weight matrix, the weighted subspace distance is calculated as follows: weighted subspace distance d_g = sqrt(sum(W_g(i)×(x_i-y_i)²)), where W_g is the weight submatrix corresponding to the i-th dimension, and x_i and y_i are the values ​​of the fingerprint to be checked and the archived fingerprint in the i-th dimension. Similarly, the texture subspace distance d_t and the weighted subspace distance d_a are calculated, where sqrt is the square root operation. A feature association graph model is constructed: geometric contour feature nodes and surface texture feature nodes are used as graph nodes, and the co-occurrence association strength of the two modal features in history is used as the edge weight (obtained through statistics from the training set), forming a bipartite graph structure. The topological constraint relationship is encoded through a graph convolutional network (GCN, containing 2 graph convolutional layers, with a hidden layer dimension of 64). The input is geometric and texture feature vectors, and the output is a structural similarity correction factor s, with the value of s ranging from 0.9 to 1.1. The feature distance calculation formula is: Feature distance = (0.4×d_g + 0.3×d_t + 0.3×d_a) / s, where (0.4×d_g + 0.3×d_t + 0.3×d_a) is the adaptively weighted comprehensive subspace distance, and the denominator s is the structural similarity correction factor.

[0036] The beneficial effects of the above technical solution are as follows: a hierarchical retrieval strategy of "coarse screening-fine comparison" is realized through a multi-level feature pyramid, which significantly improves the query efficiency of large-scale archives; dynamic importance allocation of feature dimensions is realized through feature discriminative index and adaptive weight matrix, which suppresses noise dimension interference; and a structural similarity correction mechanism is introduced by encoding topological constraints between features through graph convolutional network, which solves the defect of traditional distance calculation ignoring feature correlation, and makes feature distance more reflective of the true similarity of the leased items.

[0037] In some embodiments of this application, the method for determining the preset similarity threshold includes: Obtain the characteristic distribution of confirmed duplicate cases and false alarm cases from the leased property history archive; Analyze the statistical boundary characteristics of the characteristic distributions of duplicate cases and false alarm cases; Based on the aforementioned statistical boundary characteristics, calculate the decision boundary values ​​for recall and precision; Obtain market activity indicators for the current leasing business period; The decision boundary value is dynamically adjusted based on the market activity index to obtain the preset similarity threshold.

[0038] In this embodiment, confirmed duplicate cases refer to cases where fraudulent multiple rentals of the same item have been verified, while false positive cases refer to cases that the system initially identifies as duplicates but are later verified as involving different devices. Feature distribution is analyzed by extracting the digital fingerprints of these cases and examining their distribution density and boundaries in the feature space. Statistical boundary characteristics are analyzed using Support Vector Machine (SVM) or Logistic Regression models to find the decision hyperplane that best distinguishes between duplicate and false positive cases. Recall is defined as the proportion of correctly identified duplicate cases, and precision is defined as the proportion of truly duplicate cases among those identified as duplicates. The harmonic mean of recall and precision is calculated as the decision boundary value. Market activity indicators are obtained by statistically analyzing the month-on-month growth rate of new rental applications and new registered users in the current period (e.g., this week). A growth rate >20% indicates high activity, <5% indicates low activity, and 5% to 20% indicates normal activity. Dynamic adjustment rules: During periods of high activity, the decision boundary value will be relaxed by 5% (e.g., adjusted from 0.85 to 0.80 to reduce the requirement and minimize the interference of false alarms on business). During periods of low activity, the boundary value will be tightened by 5% (e.g., adjusted to 0.90 to increase the requirement and prevent risks). During periods of normal activity, the baseline value will remain unchanged.

[0039] The beneficial effects of the above technical solution are as follows: by analyzing the statistical distribution of historical duplicate and false alarm cases, data-driven optimization of thresholds is achieved; by balancing recall and precision calculations, the problem of missed detections or false alarms caused by traditional fixed thresholds is avoided; by introducing market activity indicators for dynamic correction, the plagiarism detection system can adapt to changes in risk control needs during peak business periods, preventing audit loopholes during busy periods and avoiding excessive interception during slow periods.

[0040] In some embodiments of this application, the comprehensive plagiarism risk score and risk tolerance parameter are determined based on the multi-dimensional similarity evaluation results, including: The geometric similarity index, texture matching index, and ownership association index in the multi-dimensional similarity evaluation results are evaluated. The geometric similarity index reflects the degree of physical morphological overlap, the texture matching index reflects the consistency of surface micro-features, and the ownership association index reflects the degree of ownership information conflict. A comprehensive plagiarism risk score is calculated based on the geometric similarity index, the texture matching index, and the ownership association index. Obtain the risk tolerance parameters for the current leasing business scenario.

[0041] In some embodiments of this application, determining the plagiarism risk level and generating a plagiarism conclusion based on the comprehensive plagiarism risk score and risk tolerance parameter includes: Based on the comparison between the comprehensive plagiarism risk score and the risk tolerance parameter, the plagiarism risk level is classified as high risk, medium risk, or low risk. Differentiated plagiarism detection conclusions are generated based on the aforementioned plagiarism risk level.

[0042] In this embodiment, the geometric similarity index is obtained by calculating the overlap rate of the 3D point clouds of the leased property to be checked and the candidate object. The index is 0.95 when the overlap rate is greater than or equal to 90%, and 0.75 when the overlap rate is less than 90%. The texture matching index is obtained by calculating the proportion of matching points of the local binary pattern (LBP) features of the image to the total number of feature points. The index is 0.9 when the matching rate is greater than or equal to 85%, and 0.7 when the matching rate is less than 85%. The ownership association index is a binary indicator. When the key fields such as the owner's name and the equipment serial number are completely consistent, the index is 1 (complete association). When the fields conflict (such as the same serial number but different owners), the index is -1 (conflict). When the fields are not completely matched but similar, the index is 0.5 (partial association). Here, "not completely matched but similar" means that when the similarity calculated by the string similarity algorithm (such as edit distance, Jaccard similarity coefficient, cosine similarity) is between 0.6 and 0.9, it is judged as "not completely matched but similar". The comprehensive plagiarism detection risk score uses a weighted summation formula: Score = 0.4 × Geometric Similarity Index + 0.3 × Texture Matching Index + 0.3 × |Ownership Association Index|, with the score ranging from 0 to 1. A score closer to 1 indicates a higher risk of duplication. The risk tolerance parameter is set according to the business type: 0.6 for high-value equipment (strict mode) and 0.8 for ordinary equipment (lenient mode). Risk level classification standards: Score ≥ Risk Tolerance Parameter + 0.2 is high risk; Risk Tolerance Parameter + 0.1 ≤ Score < Risk Tolerance Parameter + 0.2 is medium-high risk (treated as high risk); Risk Tolerance Parameter - 0.1 ≤ Score < Risk Tolerance Parameter + 0.1 is medium risk; Risk Tolerance Parameter - 0.2 < Score < Risk Tolerance Parameter - 0.1 is medium-low risk (treated as medium risk); Score ≤ Risk Tolerance Parameter - 0.2 is low risk.

[0043] The beneficial effects of the above technical solution are: through independent assessment and integrated calculation of three dimensions—geometry, texture, and ownership—it achieves refined quantification of the risk of repeated leasing; by introducing risk tolerance parameters, it adapts to the risk control needs of different business scenarios; and through risk level classification and differentiated handling suggestions, it achieves a balance between risk control and business efficiency, avoiding excessive risk control in a "one-size-fits-all" manner.

[0044] To further illustrate the technical concept of this invention, the technical solution of this invention will now be described in conjunction with specific application scenarios.

[0045] Correspondingly, such as Figure 2 As shown, this application also provides an intelligent deduplication system for leased items based on a large artificial intelligence model, including: The text determination module is used to automatically identify and extract the structure of leased property registration documents using AI-based large model intelligent recognition technology, and determine the text features of ownership documents; A collection construction module is used to obtain the appearance image features and structured light point cloud features of the leased property to be investigated, and to construct a multimodal feature dataset based on the ownership document text features, appearance image features and structured light point cloud features; The graph construction module is used to construct a digital fingerprint graph of the leased property to be searched based on the multimodal feature dataset. The digital fingerprint graph represents the comprehensive feature distribution of the leased property to be searched in the visual geometric space and semantic description space. The similarity analysis module is used to perform similarity detection and analysis between the digital fingerprint map and the preset historical archive of leased items to obtain a multi-dimensional similarity evaluation result. The multi-dimensional similarity evaluation result reflects the degree of feature overlap between the leased item to be investigated and the historical archived leased items. The intelligent plagiarism detection module is used to determine a comprehensive plagiarism risk score and risk tolerance parameter based on the multi-dimensional similarity evaluation results, determine the plagiarism risk level based on the comprehensive plagiarism risk score and risk tolerance parameter, and generate a plagiarism detection conclusion. The plagiarism detection conclusion includes a duplicate rental warning sign and risk tracing information.

[0046] In some embodiments of this application, it also includes: The confidence detection module is used for: Calculate the registration error between geometric contour features and surface texture features in the spatial coordinate system; Evaluate the semantic consistency between the ownership keyword features and the geometric contour features; Based on the registration error and the semantic fit, a cross-modal consistency index is constructed; When the cross-modal consistency index falls below the confidence threshold, feature re-acquisition is triggered. When the cross-modal consistency index is higher than or equal to the confidence threshold, the spatial correlation strength and semantic consistency among the geometric contour features, the surface texture features, and the ownership keyword features are analyzed.

[0047] In the description of the above embodiments, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0048] Although the invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the embodiments disclosed in this invention can be combined with each other in any way. The fact that not all of these combinations are described in this specification is merely for the sake of brevity and resource conservation.

[0049] It will be understood by those skilled in the art that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent deduplication of leased items based on a large-scale artificial intelligence model, characterized in that, include: The intelligent recognition technology based on AI large model is used to automatically identify and extract the structure of the leased property registration documents to determine the text features of the ownership documents; Obtain the appearance image features and structured light point cloud features of the leased property to be investigated, and construct a multimodal feature dataset based on the ownership document text features, appearance image features, and structured light point cloud features; Based on the multimodal feature dataset, a digital fingerprint map of the leased property to be searched is constructed. The digital fingerprint map represents the comprehensive feature distribution of the leased property to be searched in the visual geometric space and semantic description space. The digital fingerprint map is compared with a preset historical archive of leased items for similarity detection and analysis to obtain a multi-dimensional similarity evaluation result. The multi-dimensional similarity evaluation result reflects the degree of feature overlap between the leased item to be investigated and the historical archived leased items. Based on the multi-dimensional similarity assessment results, a comprehensive plagiarism risk score and risk tolerance parameters are determined. Based on the comprehensive plagiarism risk score and risk tolerance parameters, the plagiarism risk level is determined and a plagiarism conclusion is generated. The plagiarism conclusion includes a duplicate rental warning sign and risk tracing information.

2. The intelligent deduplication method for leased items based on a large artificial intelligence model according to claim 1, characterized in that, Based on the multimodal feature dataset, a digital fingerprint map of the leased property to be searched is constructed, including: Temporal alignment and coordinate system unification processing are performed on the appearance image features, the structured light point cloud features, and the ownership document text features to establish a standardized feature description framework. Geometric contour features, surface texture features, and ownership keyword features are extracted from the standardized feature description framework. The geometric contour features include edge curvature distribution and size ratio parameters. The surface texture features include local binary pattern encoding and depth difference histogram. The ownership keyword features include owner identifier and device serial number. Analyze the spatial correlation strength and semantic consistency among the geometric contour features, surface texture features, and ownership keyword features; Based on the spatial association strength and semantic consistency, a cross-modal feature mapping relationship is established, and the digital fingerprint map of the leased property is generated by fusion.

3. The intelligent deduplication method for leased items based on a large artificial intelligence model according to claim 2, characterized in that, Before analyzing the spatial correlation strength and semantic consistency among the geometric contour features, surface texture features, and ownership keyword features, the method further includes: Calculate the registration error between the geometric contour features and the surface texture features in the spatial coordinate system; Evaluate the semantic fit between the ownership keyword features and the geometric contour features; Based on the registration error and the semantic fit, a cross-modal consistency index is constructed; When the cross-modal consistency index falls below the confidence threshold, feature re-acquisition is triggered; When the cross-modal consistency index is higher than or equal to the confidence threshold, the spatial correlation strength and semantic consistency among the geometric contour features, the surface texture features, and the ownership keyword features are analyzed.

4. The intelligent deduplication method for leased items based on a large artificial intelligence model according to claim 3, characterized in that, The digital fingerprint map is compared with a pre-set historical archive of leased items for similarity detection and analysis to obtain multi-dimensional similarity evaluation results, including: Calculate the feature distance between the digital fingerprint map and each archived fingerprint in the historical archive of the leased property; The feature distances are sorted in descending order, and the archived fingerprints with the smallest feature distances are selected as candidate matching objects. Determine whether the minimum feature distance is lower than a preset similarity threshold; When the minimum feature distance is lower than the preset similarity threshold, it is confirmed that there is suspected duplicate rental behavior and the archived information of the candidate matching object is extracted; When the minimum feature distance is higher than or equal to the preset similarity threshold, it is determined to be a newly added leased item and the digital fingerprint map is archived and stored.

5. The intelligent deduplication method for leased items based on a large artificial intelligence model according to claim 4, characterized in that, Calculating the feature distance between the digital fingerprint spectrum and each archived fingerprint in the leased property history database includes: The geometric contour features, surface texture features, and ownership keyword features are constructed into a multi-level feature pyramid, which includes a coarse-grained global feature layer and a fine-grained local feature layer. Based on the coarse-grained global feature layer, a fast pre-screening is performed to remove archived fingerprints whose difference from the digital fingerprint map exceeds a preset screening threshold, thereby obtaining a subset of candidate similar fingerprints. For each archived fingerprint in the candidate similar fingerprint subset, the feature discrimination index of each feature dimension in the fine-grained local feature layer is calculated, and an adaptive weight matrix is ​​dynamically generated based on the feature discrimination index. The feature discrimination index reflects the information entropy contribution of the feature dimension in distinguishing different leased items. Based on the adaptive weight matrix, the weighted subspace distance of each feature dimension is calculated, and a feature association graph model is constructed. The topological constraint relationship between the geometric contour features and the surface texture features is encoded through a graph convolutional network to obtain the structural similarity correction factor. The feature distance is calculated by fusing the weighted subspace distance and the structural similarity correction factor.

6. The intelligent deduplication method for leased items based on a large artificial intelligence model according to claim 4, characterized in that, The method for determining the preset similarity threshold includes: Obtain the characteristic distribution of confirmed duplicate cases and false alarm cases from the leased property history archive; Analyze the statistical boundary characteristics of the characteristic distributions of duplicate cases and false alarm cases; Based on the aforementioned statistical boundary characteristics, calculate the decision boundary values ​​for recall and precision; Obtain market activity indicators for the current leasing business period; The decision boundary value is dynamically adjusted based on the market activity index to obtain the preset similarity threshold.

7. The intelligent deduplication method for leased items based on a large artificial intelligence model according to claim 1, characterized in that, Based on the multi-dimensional similarity assessment results, a comprehensive plagiarism risk score and risk tolerance parameters are determined, including: The geometric similarity index, texture matching index, and ownership association index in the multi-dimensional similarity evaluation results are evaluated. The geometric similarity index reflects the degree of physical morphological overlap, the texture matching index reflects the consistency of surface micro-features, and the ownership association index reflects the degree of ownership information conflict. A comprehensive plagiarism risk score is calculated based on the geometric similarity index, the texture matching index, and the ownership association index. Obtain the risk tolerance parameters for the current leasing business scenario.

8. The intelligent deduplication method for leased items based on a large artificial intelligence model according to claim 7, characterized in that, The plagiarism risk level is determined based on the comprehensive plagiarism risk score and risk tolerance parameters, and a plagiarism conclusion is generated, including: Based on the comparison between the comprehensive plagiarism risk score and the risk tolerance parameter, the plagiarism risk level is classified as high risk, medium risk, or low risk. Differentiated plagiarism detection conclusions are generated based on the aforementioned plagiarism risk level.

9. A smart deduplication system for leased assets based on an artificial intelligence large-scale model, applied to the smart deduplication method for leased assets based on an artificial intelligence large-scale model as described in any one of claims 1-8, characterized in that, include: The text determination module is used to automatically identify and extract the structure of leased property registration documents using AI-based large model intelligent recognition technology, and determine the text features of ownership documents; A collection construction module is used to obtain the appearance image features and structured light point cloud features of the leased property to be investigated, and to construct a multimodal feature dataset based on the ownership document text features, appearance image features and structured light point cloud features; The graph construction module is used to construct a digital fingerprint graph of the leased property to be searched based on the multimodal feature dataset. The digital fingerprint graph represents the comprehensive feature distribution of the leased property to be searched in the visual geometric space and semantic description space. The similarity analysis module is used to perform similarity detection and analysis between the digital fingerprint map and the preset historical archive of leased items to obtain a multi-dimensional similarity evaluation result. The multi-dimensional similarity evaluation result reflects the degree of feature overlap between the leased item to be investigated and the historical archived leased items. The intelligent plagiarism detection module is used to determine a comprehensive plagiarism risk score and risk tolerance parameter based on the multi-dimensional similarity evaluation results, determine the plagiarism risk level based on the comprehensive plagiarism risk score and risk tolerance parameter, and generate a plagiarism detection conclusion. The plagiarism detection conclusion includes a duplicate rental warning sign and risk tracing information.

10. The intelligent deduplication system for leased items based on a large artificial intelligence model according to claim 9, characterized in that, Also includes: The confidence detection module is used for: Calculate the registration error between geometric contour features and surface texture features in the spatial coordinate system; Evaluate the semantic consistency between the ownership keyword features and the geometric contour features; Based on the registration error and the semantic fit, a cross-modal consistency index is constructed; When the cross-modal consistency index falls below the confidence threshold, feature re-acquisition is triggered; When the cross-modal consistency index is higher than or equal to the confidence threshold, the spatial correlation strength and semantic consistency among the geometric contour features, the surface texture features, and the ownership keyword features are analyzed.

Images