Insurance document data management method and system and intelligent terminal

Abstract

The invention relates to the technical field of data management, in particular to an insurance document data management method and system and an intelligent terminal. The method comprises the following steps: acquiring insurance document data uploaded by a user, and marking the insurance document data to obtain checked insurance document data; constructing an insurance document labeling agent according to the checked insurance document data; acquiring insurance document data uploaded by a user in real time, and classifying the insurance document data to obtain normally-labeled document data and abnormally-labeled document data; evaluating the abnormal labeled certificate data to obtain high-deviation labeled certificate data and low-deviation labeled certificate data; reconstructing the high-deviation labeled certificate data to obtain credible high-deviation labeled certificate data; correcting the low-deviation labeled certificate data and the credible high-deviation labeled certificate data to obtain deviation correction labeled certificate data; and packaging the normal labeling certificate data and the deviation correction labeling certificate data. According to the invention, the efficiency, accuracy and automation level of insurance document management are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data management, and in particular to a data management method, system and intelligent terminal for insurance certificates. Background Art

[0002] Insurance documents serve as core credentials for protecting customer rights, documenting insurance relationships, and facilitating subsequent claims processing and verification. Digital management and intelligent processing of their information have become critical components of insurance information system development. These documents include, but are not limited to, insurance contracts, application forms, policy approvals, claim vouchers, payment records, risk assessment materials, and other diverse document data from multiple types, sources, and formats. These documents are typically archived in paper or PDF scans. With the growth of business volume, traditional management methods face challenges such as high data storage pressure, low retrieval efficiency, and severe information silos, making them unable to meet the needs of insurance institutions for efficient compliance management and intelligent business operations.

[0003] Traditional insurance certificate management methods often use file directory or database field indexing to classify and store certificates and annotate information. For example, a structured database is established by manually entering key fields (such as policy number, ID number, insurance type name, etc.), supplemented by a directory storage method to save certificate images. However, this type of method generally has the following technical defects: (1) The data structure relies on manual input, which may lead to field missing, information ambiguity, and human error, seriously affecting the accuracy and availability of subsequent data; (2) The format and content of different types of insurance certificates vary significantly, and there is a lack of a unified data abstraction mechanism, which makes it impossible to achieve cross-insurance and cross-product certificate content association and analysis. Summary of the Invention

[0004] Based on this, it is necessary for the present invention to provide a data management method, system and intelligent terminal for insurance certificates to solve at least one of the above technical problems.

[0005] To achieve the above objectives, a data management method for insurance certificates includes the following steps:

[0006] Step S1: Obtain insurance certificate data uploaded by the user, and classify and verify the certificate data based on the insurance certificate data uploaded by the user to obtain verified insurance certificate data;

[0007] Step S2: Constructing a quotation automatic labeling model and an electronic insurance policy automatic labeling model based on the verified insurance certificate data; combining the quotation automatic labeling model and the electronic insurance policy automatic labeling model to obtain an insurance certificate labeling agent;

[0008] Step S3: Obtain the real-time user-uploaded insurance certificate data, and use the insurance certificate annotation agent to classify the real-time user-uploaded insurance certificate data into abnormal certificate annotations to obtain normal annotation certificate data and abnormal annotation certificate data;

[0009] Step S4: Perform signature stroke shape deviation evaluation on the abnormally annotated document data to obtain high-deviation annotated document data and low-deviation annotated document data; perform signature behavior atomic reconstruction on the high-deviation annotated document data to obtain credible high-deviation annotated document data;

[0010] Step S5: Perform insurance rule reasoning correction on the low-deviation annotated document data and the credible high-deviation annotated document data to obtain deviation-corrected annotated document data;

[0011] Step S6: Encapsulate the normal marked document data and the deviation-corrected marked document data, and upload the encapsulation results to the insurance document management platform for archiving.

[0012] Optionally, step S1 specifically includes:

[0013] Step S11: Obtain the quotation file data and electronic policy file data submitted by the user through the insurance certificate upload interface, and construct a document original index table based on the upload timestamp and device identification information, thereby generating the user-uploaded insurance certificate data;

[0014] Step S12: Perform image quality detection and layout structure analysis on the insurance certificate data uploaded by the user, eliminate low-quality files and images, and generate a standardized insurance certificate image set;

[0015] Step S13: performing document image classification on the standardized insurance certificate image set, extracting quotation image features and electronic policy image features respectively, and performing similarity comparison with a preset insurance certificate template library to generate classified and annotated insurance certificate data;

[0016] Step S14: extracting the field position distribution based on the classified and annotated insurance certificate data and performing a position offset comparison with the standard field templates in the insurance certificate template library; performing corrections and position adjustments based on the offset comparison results to generate structure-verified insurance certificate data;

[0017] Step S15: extracting field content semantic vectors from the insurance certificate data based on the structure, and performing cluster comparison with the preset industry standard field semantic set to complete field-level label assignment, thereby generating semantically annotated insurance certificate data;

[0018] Step S16: The semantically annotated insurance certificate data is bound to the original document index table to obtain the verified insurance certificate data.

[0019] Optionally, step S12 is specifically as follows:

[0020] Step S121: Read the file images of the insurance certificate data uploaded by the user one by one, extract the image clarity features, illumination balance index and noise distribution parameters, and construct an image quality assessment tensor;

[0021] Step S122: Normalize the image quality assessment tensor, set the clarity threshold to 0.65, the illumination balance threshold to 0.7, and the noise tolerance upper limit to 0.3, mark images that do not meet any threshold as redundant images and remove them, and generate an image retention discrimination list;

[0022] Step S123: performing regional structure analysis on the images in the image retention discrimination list, extracting the regional coordinates of the title area, field area, seal area, and signature area in the image, thereby generating a layout structure partition map;

[0023] Step S124: mapping the coordinates of each region in the layout structure partition diagram to a preset A4 standard coordinate system, reordering the relative positions of the regions, adjusting the image rotation angle and scale distortion, and generating a structure normalized image set;

[0024] Step S125: performing color channel alignment and boundary extension and padding operations on the images in the structural normalized image set to generate a standardized insurance document image set.

[0025] Optionally, step S14 is specifically as follows:

[0026] Step S141: extracting coordinates of field areas in the classified and annotated insurance certificate data, extracting field position information using a bounding box positioning method, and generating a field position distribution matrix;

[0027] Step S142: Retrieving a standard field template corresponding to the classified and annotated insurance certificate data from the insurance certificate template library, parsing the field structure diagram of the standard field template, and thereby constructing a standard field position matrix;

[0028] Step S143: performing position offset calculation on the field position distribution matrix and the standard field position matrix, and marking the field position with a field offset greater than 10px in the position offset calculation result as an offset abnormality, thereby obtaining an offset abnormality field;

[0029] Step S144: Based on the structural semantic level of the field structure diagram, sequentially correct the abnormally offset fields, perform field bounding box repositioning, and obtain a position-corrected field matrix;

[0030] Step S145: Compare the field coordinates of the position correction field matrix with the alignment reference points of the standard field template, and iterate steps S143 and S144 until the field offset converges within the maximum tolerable error range of <5px, thereby generating a structurally consistent field position matrix;

[0031] Step S146: Bind the structure-consistent field position matrix to the corresponding fields in the classified and annotated insurance certificate data to generate structure-verified insurance certificate data.

[0032] Optionally, step S2 is specifically:

[0033] Step S21: Based on the verified insurance document data, extract the quotation data subset and the electronic policy data subset according to the document type label, perform field-level region slicing and semantic label alignment operations respectively, and construct the quotation training set and the electronic policy training set;

[0034] Step S22: Based on the quotation training set, a multi-channel cross-attention structure model is performed to obtain a quotation automatic labeling model; based on the electronic insurance policy training set, a bidirectional symmetric multi-scale perception network model is performed to obtain an electronic insurance policy automatic labeling model;

[0035] Step S23: performing structural model splicing on the automatic quotation annotation model and the automatic electronic insurance policy annotation model, thereby obtaining a structural splicing intermediate model;

[0036] Step S24: Perform dynamic path optimization training on the structure splicing intermediate model to obtain an insurance document labeling agent;

[0037] Step S25: Connect the insurance certificate annotation agent to the insurance certificate management platform, register the input structure, identify the service interface and callback output specifications to execute the agent deployment task.

[0038] Optionally, the document abnormality marking classification in step S3 is specifically:

[0039] Perform structured parsing and input tensor reconstruction on real-time user-uploaded insurance certificate data to obtain the certificate structured tensor;

[0040] Input the document structured tensor into the insurance document annotation agent, perform document type guided path selection and field annotation, and thus obtain the document initial annotation graph;

[0041] Based on the initial document annotation graph, a label similarity edge graph between fields is constructed, and the label similarity edge graph is corrected for field label space consistency to generate a document field annotation graph;

[0042] Based on the label vector of each field in the document field annotation graph, the Mahalanobis distance between the standard semantic vectors of the corresponding fields in the quotation training set and the electronic insurance policy training set is calculated to construct an anomaly score vector.

[0043] A field deviation density map is constructed based on the anomaly score vector, and the field label vector is divided into a deviation normal set and a deviation abnormal set according to the field deviation density map, thereby obtaining a label deviation cluster map;

[0044] Perform semantic structure integrity analysis on the label deviation cluster graph, and divide the field label vector into a normal label set and an abnormal label set based on the integrity analysis results, thereby obtaining a structural deviation cluster graph;

[0045] According to the structural deviation clustering diagram, the abnormal data of the certificate field annotation diagram is collected and separated, and the certificate field annotation diagrams that are divided into both the deviation normal set and the normal annotation set are separated into the normal annotation certificate data; the certificate field annotation diagrams that are divided into both the deviation abnormal set and the abnormal annotation set are separated into the abnormal annotation certificate data.

[0046] Optionally, the signature stroke shape deviation assessment in step S4 is specifically as follows:

[0047] Perform image enhancement and handwriting separation on the signature area in the abnormally marked document data, remove background interference and seal area overlap noise, extract the signature area image, and generate a signature structure image set;

[0048] Extract curvature points based on the signature structure image set, extract the stroke start and end points, turning points and curvature peak points, and construct a stroke sequence encoding map for each signature in the signature structure image set to generate a signature stroke sequence map;

[0049] Perform curvature function fitting and dynamic trajectory restoration on each stroke in the signature stroke sequence map to generate a signature shape parameter matrix;

[0050] Perform structural similarity matching on each parameter group in the signature shape parameter matrix and the corresponding signature structure in the verified insurance document data, calculate the residual distribution of the stroke shape of each signature group, and generate a structural deviation residual map;

[0051] Clustering is performed based on the residual mean, maximum offset, and structural deformation index in the structural deviation residual map, and a deviation level label set is output;

[0052] The signature deviation significance score is calculated by combining the deviation level label set and the distribution characteristics of the corresponding signature stroke sequence map. The abnormal annotated document data are divided into high-deviation annotated document data and low-deviation annotated document data according to the signature deviation significance score.

[0053] Optionally, the atomic reconstruction of the signature behavior in step S4 is specifically as follows:

[0054] Perform frame-by-frame trajectory reconstruction on the signature area image in the high-deviation annotated document data to generate a handwriting dynamic reconstruction trajectory set;

[0055] Based on the handwriting dynamic reconstruction trajectory set, segmented trajectory analysis is performed to extract the micro-action units of each signing behavior and generate a set of atomic fragments of signing behavior;

[0056] Constructing a behavior probability transition graph based on the signature behavior atomic fragment set;

[0057] Screening a signature area sample set based on verified insurance document data, and constructing a behavioral benchmark map based on the signature area sample set;

[0058] Using the behavioral benchmark graph, we perform graph alignment and behavioral topology similarity analysis on the behavioral probability transition graph, calculate the dynamic consistency score of each stroke path, and eliminate trajectories with low dynamic consistency scores to obtain valid trajectory data.

[0059] Based on the valid trajectory data, the atomic behavior sequence is reconstructed according to the dynamic consistency score distribution to obtain the behavior completion and reconstruction fragment set;

[0060] The behavior completion reconstruction fragment set is merged with the signature behavior atomic fragment set, and conditional consistency playback is performed to generate a signature behavior reconstruction trajectory graph;

[0061] The signature behavior reconstruction trajectory diagram is redrawn and format mapped as a static image, and mapped to the corresponding area in the high-deviation annotated document data to generate reliable high-deviation annotated document data.

[0062] This invention converts unstructured insurance document data into a structured digital form through real-time structured parsing and input tensor reconstruction of user-uploaded insurance document data, significantly improving data storage and retrieval efficiency. By mapping document data into a structured document tensor, it facilitates subsequent intelligent analysis and automatic annotation, avoiding errors associated with manual input and ensuring data accuracy and consistency. Furthermore, by implementing document type-guided path selection and field annotation, it enables automated document annotation, reducing manual intervention and improving annotation efficiency and accuracy. This process not only enhances document data processing capabilities but also lays a solid foundation for cross-insurance and cross-product document content correlation and analysis, avoiding the structural inconsistencies caused by format differences in traditional management methods. By constructing a label similarity edge graph between fields and performing label space consistency correction, it is possible to analyze and optimize annotation accuracy at a semantic level, addressing the inability to understand semantic relationships in traditional management and improving intelligent recognition capabilities. This label similarity-based correction reduces annotation inconsistencies caused by human error and further ensures the accuracy of subsequent decisions. During anomaly detection and data aggregation, clustering based on Mahalanobis distance and density maps effectively identifies and eliminates invalid data, ensuring document data quality. By constructing a structurally consistent field position matrix, the system automatically corrects and aligns document fields, effectively avoiding recognition errors caused by data offset and ensuring data consistency and reliability across different document formats. By constructing a signature behavior reconstruction trajectory graph, the system intelligently restores signature behavior, eliminates anomalous trajectories during the signing process, and reconstructs trusted data, further improving the usability and verification capabilities of signature data. During this process, parameter settings such as trajectory point matching thresholds and dynamic consistency score calculations effectively ensure the accuracy and reliability of trajectory reconstruction, avoid the adverse effects of low-quality trajectories, and ensure the integrity and consistency of signature behavior. By modeling a multi-channel cross-attention structure and applying a bidirectional multi-scale perception network, specialized models can be trained for different types of insurance documents (such as quotations and electronic policies), resulting in higher accuracy in automatic annotation, especially when processing complex document formats. Furthermore, the construction and alignment of behavioral probability transition graphs enables effective alignment and comparison of the temporal structures of different signature data, ensuring the accuracy of signature reconstruction. Overall, through the implementation of this invention, insurance document data management no longer relies on manual input, enabling efficient and accurate automatic annotation and intelligent processing. This addresses issues such as missing data, format discrepancies, and information silos inherent in traditional methods, providing a solid data foundation for advanced applications such as intelligent claims review, risk warning, and customer profiling. Furthermore, the implementation of dynamic security policies and privacy protection ensures data security during storage and transmission, improving overall performance and adaptability.

[0063] Optionally, this specification also provides an insurance certificate data management system for executing the above-mentioned insurance certificate data management method, the insurance certificate data management system comprising:

[0064] The verification and marking module is used to obtain the insurance certificate data uploaded by the user, and classify and verify the certificate data based on the insurance certificate data uploaded by the user to obtain the verified insurance certificate data;

[0065] The agent generation module is used to build an automatic quotation labeling model and an automatic electronic insurance policy labeling model based on the verified insurance certificate data; the automatic quotation labeling model and the automatic electronic insurance policy labeling model are combined to obtain an insurance certificate labeling agent;

[0066] The certificate anomaly labeling module is used to obtain real-time user-uploaded insurance certificate data and use the insurance certificate labeling agent to classify the certificate anomaly labels on the real-time user-uploaded insurance certificate data to obtain normal labeling certificate data and abnormal labeling certificate data;

[0067] The signature behavior reconstruction module is used to evaluate the signature stroke shape deviation of abnormally annotated document data to obtain high-deviation annotated document data and low-deviation annotated document data; the signature behavior atomic reconstruction of the high-deviation annotated document data is performed to obtain credible high-deviation annotated document data;

[0068] The certificate correction module is used to perform insurance rule reasoning correction on the low-deviation annotated certificate data and the trusted high-deviation annotated certificate data to obtain the deviation-corrected annotated certificate data;

[0069] The data encapsulation module is used to encapsulate the normal marked document data and the deviation corrected marked document data, and upload the encapsulation results to the insurance document management platform for archiving.

[0070] The data management system for insurance certificates of the present invention is capable of implementing any of the data management methods for insurance certificates of the present invention, and is used as a medium for combining operations and signal transmission between various modules to complete the data management method for insurance certificates. The internal modules of the system cooperate with each other, thereby improving the efficiency, accuracy and automation level of insurance certificate management.

[0071] Optionally, this specification also provides an intelligent terminal having a computer program stored therein, which implements the above-mentioned insurance certificate data management method when executed. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments thereof made with reference to the following drawings:

[0073] Figure 1A schematic flow chart of the steps of the data management method for insurance certificates of the present invention;

[0074] Figure 2 Detailed step flow diagram of step S1 in the present invention;

[0075] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0076] The following is a clear and complete description of the technical method of the present invention in conjunction with the accompanying drawings. It is obvious that the embodiments described are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts are within the scope of protection of the present invention.

[0077] In addition, the accompanying drawings are merely schematic illustrations of the present invention and are not necessarily drawn to scale. Identical reference numerals in the figures denote identical or similar parts, and thus repetitive descriptions thereof will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities that do not necessarily correspond to physically or logically separate entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor and / or microcontroller approaches.

[0078] It should be understood that although the terms "first," "second," and the like may be used herein to describe various elements, these elements should not be limited by these terms. These terms are used solely to distinguish one element from another. For example, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element, without departing from the scope of the exemplary embodiments. The term "and / or" as used herein includes any and all combinations of one or more of the listed associated items.

[0079] To achieve this, please refer to Figures 1 to 2 The present invention provides a data management method for insurance certificates, the method comprising the following steps:

[0080] Step S1: Obtain insurance certificate data uploaded by the user, and classify and verify the certificate data based on the insurance certificate data uploaded by the user to obtain verified insurance certificate data;

[0081] In this embodiment, users upload scanned insurance document images through the front-end interface and input the images into an image classification model built on the EfficientNet-B4 architecture. This architecture uses an initial convolution kernel size of 7×7 and extracts image features layer by layer using MBConv residual units. A spatial attention mechanism is added before the fully connected layers to enhance local image information. The model output dimension is (1×3), corresponding to the classification confidence level for the quotation, electronic policy, and other types of documents. To ensure subsequent annotation accuracy, images with a confidence level less than 0.9 are sent to a manual review channel. For classified images, an OCR annotation process combining template matching and text localization is invoked. A pre-built field location information template (for example, the policy number field is set to the image-relative coordinate interval [0.3W:0.5W, 0.1H:0.2H]) is compared against the OCR output field by field. A list of fields and their corresponding content is output and saved in a JSON structure named "verified_doc_i.json" for subsequent processing.

[0082] Step S2: Constructing a quotation automatic labeling model and an electronic insurance policy automatic labeling model based on the verified insurance certificate data; combining the quotation automatic labeling model and the electronic insurance policy automatic labeling model to obtain an insurance certificate labeling agent;

[0083] In this embodiment, cluster analysis is performed on verified image data. Quotations and electronic policies are used as two separate training sets, each serving as input for constructing two independent field labeling models. The quotation labeling model utilizes a Bi-LSTM-CRF architecture, with a maximum text sequence length of 128 characters. The embedding layer uses pre-trained vectors with a word vector dimension of 100, and the output is a sequence of field labels. The electronic policy labeling model is constructed as a three-channel parallel image-text fusion architecture: the first channel encodes the character sequence extracted by OCR, the second channel contains image convolutional features (extracted using ResNet18), and the third channel encodes the spatial position of the field template (mapping 2D coordinates to vector space). The outputs of these three channels are concatenated and fed into a 128-dimensional Transformer architecture for weighted feature fusion, ultimately outputting the field and its position interval within the image. After both models are constructed, the outputs are encapsulated and concatenated using a unified intermediate representation structure (key-value pairs + position matrix) to generate an insurance document labeling agent. This agent automatically selects the appropriate path for labeling through internal confidence branching, supporting automatic format recognition and automatic switching of labeling strategies.

[0084] Step S3: Obtain the real-time user-uploaded insurance certificate data, and use the insurance certificate annotation agent to classify the real-time user-uploaded insurance certificate data into abnormal certificate annotations to obtain normal annotation certificate data and abnormal annotation certificate data;

[0085] In this embodiment, the user upload behavior is obtained in real time, and the newly uploaded image is input into the insurance document annotation agent processing flow in real time. The image is first extracted through the deep residual network to obtain the main feature tensor of the image. (, combined with OCR text encoding and field position matrix , and are input into the annotation agent together. The output structure is the set of annotated fields {F_k} and their confidence distribution P_k. Set the confidence threshold θ=0.75. If the P_k of a certain field is lower than the threshold or the number of missing annotated fields is greater than the set upper limit N=3, it is judged as abnormal annotated document data. This type of data will be classified and saved in the abnormal annotated data area, and marked with the type of field missing or position offset. The rest of the data determined to be normal will directly enter the subsequent data packaging process. During this process, the processing time of each annotation and the field matching accuracy index will also be recorded for subsequent model fine-tuning and dynamic verification mechanism training.

[0086] Step S4: Perform signature stroke shape deviation evaluation on the abnormally annotated document data to obtain high-deviation annotated document data and low-deviation annotated document data; perform signature behavior atomic reconstruction on the high-deviation annotated document data to obtain credible high-deviation annotated document data;

[0087] In this embodiment, for abnormally annotated document data, the signature region localization model is first invoked. This model is based on a U-Net image semantic segmentation architecture. The input image size is uniformly 256×256, and the output is a single-channel signature region mask (size 64×64). The segmented signature image is input into the stroke vectorization extractor, which extracts signature morphological features based on a modified ResNet34 architecture and outputs a stroke morphology vector V_i of dimension (1×128). Subsequently, the original signature template vector V_ref is loaded and the Euclidean distance calculation is performed with V_i: . Set the high deviation threshold δ_h=15 and the low deviation threshold δ_l=8. If D>δ_h, it is classified as a high deviation document. If D∈[δ_l,δ_h], it is classified as a low deviation document. For high deviation documents, call the graph-structured signature behavior reconstruction process to abstract the signature trajectory into a directed graph G=(V,E), where V represents the stroke node (with timestamp and pressure parameters) and E is the stroke connection sequence. Using the historical signing behavior trajectory database, a set of alternative signature behavior trajectories is generated through graph similarity matching (based on shortest path matching and node rotation alignment), and reconstructed into a set of trusted stroke images to replace the original deviation area to ensure the compliance of the signature structure.

[0088] Step S5: Perform insurance rule reasoning correction on the low-deviation annotated document data and the credible high-deviation annotated document data to obtain deviation-corrected annotated document data;

[0089] In this embodiment, low-deviation annotated document data and trusted high-deviation annotated document data are uniformly input into the insurance rule inference engine. The engine consists of five rule sets, corresponding to: field temporal integrity, policyholder age requirement, signature time logical constraint, insurance company issuance location constraint, and insurance type field matching rules. Taking field temporal integrity as an example, the order of field appearance is abstracted into a field path sequence S = [f_1, f_2, ..., f_n] and compared with the rule graph R (represented as an adjacency matrix ) to perform path consistency check. If the order of two fields does not match (for example, the signature time is earlier than the issuance time), the field repairer is automatically called to map the original field content into a vector , and uses nearest neighbor logical matching to generate reasonable replacement values. The corrected results are repackaged into field-value-position triples and the correction instructions are recorded. Finally, the deviation-corrected and annotated document data is output.

[0090] Step S6: Encapsulate the normal marked document data and the deviation-corrected marked document data, and upload the encapsulation results to the insurance document management platform for archiving.

[0091] In this embodiment, the data of documents that have been determined to be normally marked and those marked with deviation corrections are uniformly encapsulated. The encapsulation format is a standard JSON structure, containing field names, field values, and field location coordinates (expressed in image-relative pixel format). Information items such as the data's unique ID, document processing timestamp, and processing strategy flags (normal / low_deviation / high_deviation_rebuild) are also included. After encapsulation, each data copy is accompanied by a SHA-256 digest value for integrity verification. The data is then submitted via the insurance document management platform's RESTful API, using the path " / api / upload_verified_doc." The submission process supports TLS 1.3 encryption and an asynchronous receipt mechanism. Upon successful upload, the platform returns a status code of 200 and a file ID, which is then bound to the user ID and stored in a local index table for subsequent service calls and verification.

[0092] It is particularly important that step S5 is specifically:

[0093] Obtain the insurance industry rule set, which includes the insurance validity period range, policyholder age requirement range, and insurance type field matching rules;

[0094] In this embodiment, insurance industry rules are provided by the insurance company's business department and stored in a configuration database in the form of a structured table. The rules include validity period ranges (e.g., the policy effective date must be before the issuance date and no later than the upload date, with a maximum time limit of T = 5 years), policyholder age ranges (e.g., auto insurance requires an age range of 18 to 70), and a mapping table for insurance type fields (e.g., "motor vehicle loss insurance" should be associated with fields such as "license plate number," "engine number," and "vehicle type"). This rule table is read as three structures: R_date, R_age, and R_type. R_date is a two-dimensional interval matrix R_date = [[start_min, start_max], [end_min, end_max]], R_age is a pair of age upper and lower limits (18, 70), and R_type is a key-value mapping matrix (e.g., "auto insurance" → [field numbers 101, 102, 103]). All rules are encapsulated in a unified data structure as JSON configuration units, which serve as the input basis for subsequent inference logic transformations.

[0095] Performing reasoning logic transformation on insurance industry rules to obtain the insurance rule reasoning logic structure, and constructing an insurance rule reasoning model based on the insurance rule reasoning logic structure;

[0096] In this embodiment, the read industry rules are converted through structural mapping and logical instruction tables to construct the insurance rule reasoning logic structure. Taking the insurance type field matching rule as an example, each insurance type is regarded as a central node, and the required field is its adjacent node, and a rule dependency graph R_graph=(V,E) is generated, where the node V is the field identifier and the edge E represents the field dependency relationship. If the insurance type node v_0 is associated with the field v_1, then the edge e_01 is established. For the validity period and age rules, they are converted into a rule template set R_template, each template is in the form of: [field name, operator, comparison value], for example ["insured person's age", "≥", 18]. The logical structure consists of three parts: rule matching subgraph, field constraint matrix (listed as fields, conditional symbols, thresholds), and reasoning triggering logic chains (expressing logical rule flows with IF-THEN). Based on this, an insurance rule reasoning model is constructed, using a graph convolutional network (GCN) structure, with the input being a field graph structure feature vector set. , the output is the correction recommendation result , used to guide whether the field needs to be adjusted. The number of graph convolution layers is set to 2, the output dimension of each layer is 64, the activation function is ReLU, and the learning rate is set to 0.001.

[0097] Input the low-deviation annotated document data and the credible high-deviation annotated document data into the insurance rule inference model, perform field-by-field correction on the input data, and obtain preliminary corrected document data;

[0098] In this example, the pre-processed low-deviation annotated document data and the trusted high-deviation annotated document data are input into the insurance rule inference model. Before input, all field contents are encoded into vector form. Numeric fields (such as amount and age) are directly normalized into real numbers in the interval [0, 1]. Text fields (such as insurance type name) are embedded using word vectors with a dimension of 100. The data structure is uniformly converted into a field vector set. , where m is the number of fields and d is the vector dimension. The inference model takes X_doc as input, makes inferences on each field one by one, and outputs the corresponding correction label , where Y_hat_i is 0, indicating the original field is trustworthy, and 1, indicating that correction is required. For fields requiring correction, the most likely replacement value is selected from a candidate set based on the field dependency graph and contextual field values. The corrected document structure is then output, forming the preliminary corrected document data. The field recommendation engine is built based on historical structured insurance document data, the insurance industry rule map, and the model inference confidence matrix. Its core structure includes a field co-occurrence probability matrix, a field embedding vector library, and a multi-factor scoring mechanism. The engine first uses the co-occurrence matrix to identify the dependency relationship between the target field and its context. It then uses the embedding vector to calculate the semantic similarity between candidate field values and the context, and then ranks them based on historical frequency and rule matching. Finally, outlier fields are automatically replaced based on confidence thresholds or sent to manual review, achieving precise correction and recommendation of insurance document fields. This process automatically records the corrected field name, original value, recommended value, and field confidence for subsequent analysis and retrospective review.

[0099] The preliminary correction document data is screened for correction data, abnormal data is eliminated, and deviation-corrected and labeled document data is obtained.

[0100] In this embodiment, the document data set after preliminary correction is screened to eliminate potential abnormal samples. The abnormality judgment mechanism includes two dimensions: field confidence threshold and field consistency check. The confidence judgment rule is: if the field correction confidence is lower than the set threshold θ=0.6, or the number of corrected fields in the entire document exceeds α=30% of the total number of fields, then the document is marked as high-risk abnormal data. The field consistency check is performed by establishing a logical consistency table. For example, if the insurance policy is "home property insurance" and the field contains "license plate number", it is considered a type conflict. For the screened samples, they are packaged and entered into the manual review channel; the remaining data is stored in the compliance data pool as deviation correction annotated document data. All screening operations are recorded and processed in the log format of field anomaly list + screening label + processing timestamp, and are archived and managed as operation audit data.

[0101] Optionally, step S1 specifically includes:

[0102] Step S11: Obtain the quotation file data and electronic policy file data submitted by the user through the insurance certificate upload interface, and construct a document original index table based on the upload timestamp and device identification information, thereby generating the user-uploaded insurance certificate data;

[0103] In this embodiment, a data access gateway configured on the insurance service platform retrieves user-uploaded insurance document files, including PDF or image-based documents such as quotations and electronic policies. Each upload automatically records basic information such as the upload timestamp (in the format YYYY-MM-DDHH:MM:SS), the upload terminal device identifier (IMEI code or device serial number), and the file name and size. This information is combined to construct a document raw index table with the structure [doc_id, user_id, device_id, upload_time, file_name, file_size], where doc_id is generated using a hashing method to ensure uniqueness. This ultimately creates user-uploaded insurance document data with tracking properties for subsequent structural analysis and annotation.

[0104] Step S12: Perform image quality detection and layout structure analysis on the insurance certificate data uploaded by the user, eliminate low-quality files and images, and generate a standardized insurance certificate image set;

[0105] In this embodiment, received file data first enters an image quality inspection process. Blurred, incomplete, or low-quality images are eliminated using a resolution threshold (set to 300 dpi), an image blur score (using a Laplacian operator score greater than 150 to qualify), and file integrity verification (e.g., ensuring the number of PDF pages matches the template). Qualified images then enter a layout structure analysis phase, where a visual region segmentation mechanism is used to identify key content blocks and standardize their position, size, and text coverage density. These standardized images are stored in JPEG format and uniformly resized to A4 size at 300 dpi, forming a standardized insurance document image set that provides high-quality input for subsequent recognition tasks.

[0106] Step S13: performing document image classification on the standardized insurance certificate image set, extracting quotation image features and electronic policy image features respectively, and performing similarity comparison with a preset insurance certificate template library to generate classified and annotated insurance certificate data;

[0107] In this example, an image classification model based on the EfficientNet-B4 architecture was used to perform document classification on a standardized image set. This architecture uses an initial convolution kernel size of 7×7, extracts image features layer by layer using MBConv residual units, and adds a spatial attention mechanism before the fully connected layers to enhance local image information. The model outputs a dimension of (1×3), corresponding to the classification confidence levels for quotations, electronic insurance policies, and other types of documents. The image classification model removes images with confidence levels less than 0.9, extracts structural and font distribution features from the remaining images, and generates 128-dimensional feature vectors for each of the two image categories. These vectors are then compared with the average vectors of known document categories in a template library using cosine similarity. Those with a similarity greater than 0.85 are considered to match the document type, forming a preliminary classification label. The comparison results are then combined with the template path record to complete the classification and labeling process, outputting classified and labeled insurance document data with the structure [doc_id, doc_type, template_id, similarity_score].

[0108] Step S14: extracting the field position distribution based on the classified and annotated insurance certificate data and performing a position offset comparison with the standard field templates in the insurance certificate template library; performing corrections and position adjustments based on the offset comparison results to generate structure-verified insurance certificate data;

[0109] In this embodiment, based on the classification results, the relative position coordinates of the standard fields are identified from the image (in pixels, with the upper left corner as the origin). A field position matrix is constructed, formatted as [field_name, x, y, width, height]. This matrix is then compared with the standard field matrix in the template library for each field offset calculation. Fields with offsets exceeding 5 pixels trigger a position correction mechanism, fine-tuning the field position through an affine transformation. A comparison record of the field coordinates before and after the adjustment is retained and updated to structure the insurance certificate data. This data is used to guide subsequent field content extraction and semantic analysis.

[0110] Step S15: extracting field content semantic vectors from the insurance certificate data based on the structure, and performing cluster comparison with the preset industry standard field semantic set to complete field-level label assignment, thereby generating semantically annotated insurance certificate data;

[0111] In this example, the text content of each field in the insurance certificate data is extracted and verified. A text vector generation component is used to convert the field content into a 300-dimensional word vector representation. This vector is then compared to the standard word vectors in a pre-set industry-standard field semantics set. All matching pairs with a similarity greater than 0.75 are then entered into a semantic clustering module for classification. A soft classification mechanism is used to associate the field content with the most likely standard label, and the field-level labeling results are output, generating semantically annotated insurance certificate data with the structure [field_name, field_value, semantic_tag, confidence_score].

[0112] Step S16: The semantically annotated insurance certificate data is bound to the original document index table to upload the certificate and obtain the verified insurance certificate data.

[0113] In this embodiment, the generated semantically annotated insurance document data is bound to the original index table using a primary key join. The upload source is matched using the doc_id field to construct the final verified insurance document data. This data retains field location and content tag information, along with the original upload time and user device identifier. It is ultimately stored in a unified data integration platform with a structure such as [doc_id, field_name, field_value, semantic_tag, upload_time, device_id]. This creates a high-quality, traceable insurance document data asset, providing fundamental data support for subsequent model training and business intelligence processing.

[0114] Optionally, step S12 is specifically as follows:

[0115] Step S121: Read the file images of the insurance certificate data uploaded by the user one by one, extract the image clarity features, illumination balance index and noise distribution parameters, and construct an image quality assessment tensor;

[0116] In this embodiment, during the image quality assessment process, image files are read one by one from the insurance document data uploaded by the user, and three key image quality parameters are extracted using a feature sampling module. First, clarity is quantified by the variance of the image's grayscale gradients, and its Laplace gradient statistics are calculated after the image is resized to 480×680. Illumination balance is assessed by determining the presence of severe exposure or shadowing using the entropy of the image's histogram distribution. Finally, the noise distribution parameter uses the root mean square noise index to assess the random distribution of high-frequency components in the image, quantifying the degree of image interference. These three metrics are then unified into an image quality assessment tensor Q of the form (Qn×3), where Qn is the number of images and the three columns correspond to clarity, illumination, and noise metrics, respectively, for subsequent quality screening.

[0117] Step S122: Normalize the image quality assessment tensor, set the clarity threshold to 0.65, the illumination balance threshold to 0.7, and the noise tolerance upper limit to 0.3, mark images that do not meet any threshold as redundant images and remove them, and generate an image retention discrimination list;

[0118] In this embodiment, in the constructed image quality assessment tensor Q, the quality indicators of the three dimensions are linearly normalized, and each indicator is adjusted to the numerical range of [0,1]. The minimum acceptable standard values for clarity, illumination balance, and noise level are set as 0.65, 0.7, and 0.3, respectively. If an image does not meet the corresponding threshold in any indicator, the image is marked as redundant. A set of logical discriminant vectors L (of the form Qn×1) is generated, in which the element value is 0 or 1, indicating whether it is retained. Finally, the low-quality images are filtered out from the initial set through this discriminant list, and the results after elimination will enter the structural analysis process to ensure that subsequent structural analysis tasks are based only on valid image data.

[0119] Step S123: performing regional structure analysis on the images in the image retention discrimination list, extracting the regional coordinates of the title area, field area, seal area, and signature area in the image, thereby generating a layout structure partition map;

[0120] In this embodiment, after obtaining the image retention discrimination list, a structural region decomposition operation is performed on each valid image. In the specific steps, the title area is first located based on the image binary edge projection method, and then the field area is identified by vertical density slicing and character area detection, and the seal area and signature area are identified by the shape template matching technology of the bottom area of the image. The coordinate information of all areas is organized into a structural coordinate matrix S, which is in the form of (4×4), and records the upper left and lower right pixel coordinates of each area in the image, as well as the physical height and width of the corresponding area. Through this coordinate matrix, a preliminary layout structure partition map is constructed and used for subsequent layout reconstruction operations.

[0121] Step S124: mapping the coordinates of each region in the layout structure partition diagram to a preset A4 standard coordinate system, reordering the relative positions of the regions, adjusting the image rotation angle and scale distortion, and generating a structure normalized image set;

[0122] In this embodiment, the image structure is mapped to the standard pixel coordinate system corresponding to A4 paper (size 794×1123) based on the preliminarily generated structural coordinate matrix S. During the mapping process, the image is first affine transformed to correct the rotation angle. The horizontal scaling ratio and vertical alignment offset are calculated based on the vertical spacing of the field area, thereby correcting the proportional distortion and regional displacement caused by the shooting angle or image cropping. Finally, a set of structurally normalized images is generated, and the key structural areas in all images are uniformly aligned to the standard coordinate template, providing a consistent reference for subsequent field recognition and position offset analysis.

[0123] Step S125: performing color channel alignment and boundary extension and padding operations on the images in the structural normalized image set to generate a standardized insurance document image set.

[0124] In this embodiment, after completing the structural normalization, in order to ensure the visual consistency and boundary integrity of the image in subsequent processing, the three color channels of each image are aligned for channel response. The specific method is to normalize the brightness of the three RGB channels separately, and use the inter-channel cross-correlation matrix analysis to determine the alignment offset vector. Subsequently, based on the edge information of the layout structure area, the boundary filling area of the image is expanded by 10 to 20 pixels up, down, left and right to ensure that each area will not be lost due to image cropping. Finally, a standardized insurance certificate image set is output, with the image format unified as PNG, the size unified as 794×1123, and the pixel density set to 300dpi to support subsequent classification recognition and structural analysis operations.

[0125] Optionally, step S14 is specifically as follows:

[0126] Step S141: extracting coordinates of field areas in the classified and annotated insurance certificate data, extracting field position information using a bounding box positioning method, and generating a field position distribution matrix;

[0127] In this embodiment, based on the classified and labeled insurance document data, a coordinate extraction operation is performed on the field area in each document image. In the processing flow, the bounding box area of each field in the image is first located based on the labeled field category label. The position coordinates of each field in the image are recorded using a four-tuple in the format of [x_min, y_min, x_max, y_max]. The number, category, and location record of each field are combined into a field location information set. The location information of all fields is integrated into a field location distribution matrix F with dimensions of (n×4), where n is the total number of fields identified in the current document. Each row represents the coordinates of the four corners of the bounding box of a field, which is used for subsequent structural matching analysis.

[0128] Step S142: Retrieving a standard field template corresponding to the classified and annotated insurance certificate data from the insurance certificate template library, parsing the field structure diagram of the standard field template, and thereby constructing a standard field position matrix;

[0129] In this embodiment, for each classified and labeled insurance certificate data corresponding to the template category, a matching template is retrieved from the insurance certificate template library provided by the insurance business department, and the field structure diagram under the template is parsed. The structure diagram consists of a field logical hierarchy, standard coordinates, and alignment reference points. All field positions are predefined according to the standard A4 coordinate system, and the record format is also in the matrix form of [x_min, y_min, x_max, y_max]. The bounding boxes of all standard fields are extracted and a standard field position matrix T is constructed with a dimension of (m×4), where m is the number of fields in the template. This standard matrix will serve as a reference for subsequent position offset calculations to ensure that position offset detection has a unified standard.

[0130] Step S143: performing position offset calculation on the field position distribution matrix and the standard field position matrix, and marking the field position with a field offset greater than 10px in the position offset calculation result as an offset abnormality, thereby obtaining an offset abnormality field;

[0131] In this embodiment, the field position distribution matrix F and the standard field position matrix T are analyzed for field-level offset differences. For each pair of matching fields, the Euclidean distance of the four corner coordinates of the corresponding bounding box is calculated, and this is used as a position offset metric. If the average distance between the corner points of the bounding box of any field is greater than 10 pixels, the field is regarded as an offset abnormal field, and its field number and corresponding offset are recorded. All abnormal fields are organized into an offset abnormal field set E, and the data structure is a triple array of [k, dx, dy], where k is the field number, dx and dy are the horizontal and vertical offset values, respectively, for subsequent structural correction.

[0132] Step S144: Based on the structural semantic level of the field structure diagram, sequentially correct the abnormally offset fields, perform field bounding box repositioning, and obtain a position-corrected field matrix;

[0133] In this embodiment, after identifying the abnormally offset field, a sequential correction operation is performed based on the field hierarchy relationship defined in the standard field structure diagram. By traversing the abnormal field set E and combining the logical order of the fields (for example, "insured person's name" should be before "date of birth") and the layout structure of the area where it is located, its bounding box position is re-estimated. When repositioning, it is prioritized to maintain the logical alignment of the field within the horizontal line, and adjust its vertical offset according to the position constraints of the adjacent fields to form a position correction field matrix C with a dimension of (k×4), which represents the new coordinates of the corrected field. This matrix will be used in the iterative structural consistency optimization process.

[0134] Step S145: Compare the field coordinates of the position correction field matrix with the alignment reference points of the standard field template, and iterate steps S143 and S144 until the field offset converges within the maximum tolerable error range of <5px, thereby generating a structurally consistent field position matrix;

[0135] In this embodiment, to ensure the ultimate accuracy of field positions, the position-corrected field matrix C is re-aligned with the standard field position matrix T. Specifically, the residual error between the corrected field position and the standard field position is calculated based on the field alignment reference point provided by the template (e.g., left-aligned baseline or central axis of symmetry). If the average residual for a field is still greater than 5 pixels, the position repositioning process in the previous step is returned to, and corrections are iterated until all offsets converge to an error tolerance of <5 pixels. This iterative process is repeated up to three times; if the number of iterations exceeds the limit, the image will be marked as a layout anomaly and enter the manual review process.

[0136] Step S146: Bind the structure-consistent field position matrix to the corresponding fields in the classified and annotated insurance certificate data to generate structure-verified insurance certificate data.

[0137] In this embodiment, after completing the above iterations, a structurally consistent field position matrix S' is obtained. Its format is consistent with the initial position matrix, and all fields are within the position range defined by the template. This matrix is bound to the classified and annotated insurance document data, and each field information structure is updated to [field number, field category, corresponding image doc_id, field coordinates, offset residual], thereby generating a structurally verified insurance document data record. This structured data is then output as a standard JSON format file for subsequent semantic recognition and rule inference.

[0138] Optionally, step S2 is specifically:

[0139] Step S21: Based on the verified insurance document data, extract the quotation data subset and the electronic policy data subset according to the document type label, perform field-level region slicing and semantic label alignment operations respectively, and construct the quotation training set and the electronic policy training set;

[0140] In this embodiment, on the insurance document data that has completed structural verification, based on the doc_type tag field corresponding to each document, the data that has been labeled and has a status of "verified" is selected as two subsets of "quotation" and "electronic insurance policy", and field-level training sample sets are established for each. For each image, combined with the bounding box coordinates recorded in the structurally consistent field position matrix, each field area in the image is accurately sliced, the size of each slice image is uniformly adjusted to 112×32 pixels, and the field semantic labels provided by the standard template are used for annotation. All slice images and labels form training samples in the format of [field_img, field_label]. After the training set is constructed, two training data sets with the structure of {img_id, field_crop, field_name} are generated, named QuoteDataset and PolicyDataset respectively.

[0141] Step S22: Based on the quotation training set, a multi-channel cross-attention structure model is performed to obtain a quotation automatic labeling model; based on the electronic insurance policy training set, a bidirectional symmetric multi-scale perception network model is performed to obtain an electronic insurance policy automatic labeling model;

[0142] In this embodiment, the model architecture employed in the automatic quotation field annotation task employs a multi-channel cross-attention architecture, primarily aiming to improve recognition stability and generalization when field image information is incomplete or ambiguous. The model receives three input channels: 1) a color image channel: the original RGB image, uniformly sized [3, 112, 32]; 3 represents the number of channels; 112 represents the height, indicating the vertical dimension of the feature map; and 32 represents the width, indicating the horizontal dimension of the feature map. 2) a grayscale texture map channel: the original image is converted to grayscale and then a local binary pattern (LBP) is used to generate a texture feature map of uniform size. 3) a character edge enhancement map channel: a character boundary map is obtained through Sobel edge detection combined with a dilation operation. Preliminary feature extraction is performed on each of these three channels through three branched convolutional blocks. Each convolutional block consists of two concatenated convolutional layers (Conv2D) followed by batch normalization (BatchNorm) and ReLU activation, resulting in an output feature dimension of [64, 56, 16]. The feature maps from the three branches are then fused in a "cross-attention fusion architecture," which uses a channel-attention mechanism to dynamically adjust the weights of feature maps from different sources. The core of this architecture is a squeeze-and-excitation (SE) module: Each channel feature map undergoes global average pooling to generate a channel statistic vector, which is then passed through two fully connected layers to generate channel weights, ultimately weighting the feature responses of different channels. The fused feature maps are of size [128, 56, 16] and enter the subsequent backbone architecture: a compression path consisting of three groups of residual structures (each containing three convolutional layers), each followed by a max pooling operation. This transforms the feature maps layer by layer from [128, 28, 8] → [256, 14, 4] → [512, 7, 2]. After the Flatten operation, the feature map is flattened to [1,7168] and then connected to a two-layer fully connected structure (FC→ReLU→FC) to generate a semantic vector [1,256]. Finally, a softmax operation is performed to output field categories. The number of categories is set to 12, corresponding to the field types in the quotation, such as "Insurance Company," "Owner's Name," and "Quoted Amount." The electronic insurance policy annotation model uses a bidirectional symmetric multi-scale perceptual network architecture, which takes a single-channel normalized image [1,112,32] as input. The network consists of three parallel convolutional paths with convolution kernels of 5×5, 3×3, and 1×1, respectively. Each path outputs a feature map [32,112,32] after convolution. This is then compressed to [64,56,16] through two stages of stacked convolution. A path-symmetric mechanism is used, where the outputs of each set of convolution kernels are concatenated forward and backward to generate a fused feature map [128,56,16].In the post-processing stage, the model incorporates a positional information guidance layer: the normalized coordinates [x / W, y / H] of each pixel in the image are concatenated to the feature map channels, generating an expanded feature map of [130, 56, 16] to enhance the spatial location information of the fields. This feature map then enters the semantic aggregation module, which first compresses the dimensions using 3×3 convolutions, followed by average pooling and flattening to generate a high-level semantic vector of [1, 1024]. Finally, the output is fed into a two-layer fully connected structure, which outputs [1, 15], corresponding to 15 common fields found in electronic insurance policies, such as "Policyholder," "Electronic Signature," and "Effective Date." This structure is trained with a batch size of 24, using cross-class weighting to enhance training for less common fields. After training, the two models are saved as structured files named quote_classifier.pth and policy_classifier.pth, respectively, containing the network definition, weight parameters, and output mapping tables. After model training, these files serve as the building blocks for subsequent structure concatenation and the construction of an insurance document annotation agent.

[0143] Step S23: performing structural model splicing on the automatic quotation annotation model and the automatic electronic insurance policy annotation model, thereby obtaining a structural splicing intermediate model;

[0144] In this embodiment, after completing the training of two independent models, a structural splicing operation is performed, the main structure of the quotation model is used as the feature generation path, the symmetrical branch of the electronic insurance policy model is used as the supplementary path, and a fusion layer is introduced to unify the output semantic representation and the path selector based on document type conditional gating. A shared attention module is introduced in the splicing structure to integrate the spatial dimensions of feature vectors from different sources, and finally generate a unified semantic expression layer with an output feature dimension of (1×320). The splicing model is named "structural splicing intermediate model", and the original branch weights are retained during construction for subsequent dynamic optimization. The entire structural splicing graph is expressed in the Tensor structure definition as: [batch_size,3,112,32]→feature_branch_1→fusion_layer→output_layer.

[0145] Step S24: Perform dynamic path optimization training on the structure splicing intermediate model to obtain an insurance document labeling agent;

[0146] In this embodiment, after obtaining the structural splicing intermediate model, in order to further improve the model's ability to recognize uncertain structures, dynamic path optimization training is performed on the model structure. The training strategy is to automatically switch the main path activation status in each batch of samples according to the type of sample, that is, the quotation sample activates the main path 1, the electronic insurance policy sample activates the main path 2, and the cross sample activates the fusion path. The model output is evaluated for branch accuracy through two parallel softmax output heads, and the loss function is integrated in a dynamic weighted manner to fine-tune the path selection weight. During the training process, a dynamic scheduler is used to decay the initial learning rate 1e-4 by 0.8 after every five rounds until the model converges on the validation set. Finally, the training results in an insurance document labeling intelligent agent with multi-path autonomous decision-making capabilities, whose structure has both classification accuracy and structural adaptability.

[0147] Step S25: Connect the insurance certificate annotation agent to the insurance certificate management platform, register the input structure, identify the service interface and callback output specifications to execute the agent deployment task.

[0148] In this embodiment, after completing the training and verification of the insurance document labeling agent, the agent is connected to the insurance document management platform as a serviceable component. During deployment, a unified field image input structure with the input structure {doc_id, field_crop_image} is registered for the agent, and a standard REST interface is defined to support asynchronous recognition service calls. The output structure is registered in the format {doc_id, field_label, confidence_score, position}, and a callback address mechanism is used to support the centralized return of batch recognition results. By introducing service identification, access tokens, and model version control mechanisms, the deployed agent is guaranteed to operate stably in real-world environments, and supports subsequent online fine-tuning and hot update mechanisms to meet the ever-changing insurance document field standards.

[0149] Optionally, the document abnormality marking classification in step S3 is specifically:

[0150] Perform structured parsing and input tensor reconstruction on real-time user-uploaded insurance certificate data to obtain the certificate structured tensor;

[0151] In this example, real-time uploaded insurance document image data is received and uniformly decoded using the platform's registered document input structure. Field region coordinates and candidate field content are extracted. The image coordinates, region labels, and text candidate content are combined to construct a document structured tensor of size [N, 128, 64], where N represents the number of fields, 128 is the embedding dimension, and 64 is the sub-region feature dimension within the field image region. This tensor serves as the primary input for the subsequent labeling agent recognition task.

[0152] Input the document structured tensor into the insurance document annotation agent, perform document type guided path selection and field annotation, and thus obtain the document initial annotation graph;

[0153] In this example, the constructed document structured tensor is fed into the insurance document annotation agent. This agent integrates a path selector based on document type conditional gating. Based on the path configuration, it calls the field annotation branch network in the corresponding structure splicing model to perform the field-by-field recognition task. The output is an initial annotation graph with the structure [N, 1, 256], where a 256-dimensional vector represents the label embedding representation of each field.

[0154] Based on the initial document annotation graph, a label similarity edge graph between fields is constructed, and the label similarity edge graph is corrected for field label space consistency to generate a document field annotation graph;

[0155] In this embodiment, based on the output initial annotation graph, an edge connection graph between field labels is further constructed in the field embedding space. This graph generates an adjacency weight matrix A∈[N,N] by calculating the Euclidean distance and semantic direction angle between any pair of fields. Edges with a weight greater than 0.85 are considered highly similar and used to build a label similarity edge graph. Subsequently, a local mean calibration strategy is used to perform consistency correction on the edge relationships in the similarity graph, resulting in a more structurally stable document field annotation graph.

[0156] Based on the label vector of each field in the document field annotation graph, the Mahalanobis distance between the standard semantic vectors of the corresponding fields in the quotation training set and the electronic insurance policy training set is calculated to construct an anomaly score vector.

[0157] In this embodiment, the core label semantic vector is extracted from the embedding vector of each field in the document field annotation graph. , comparing the standard semantic center vectors in the quotation training set and the electronic insurance policy training set , using Mahalanobis distance Calculate their degree of difference. Each field generates an anomaly score, and the process forms a sequence of anomaly score vectors with a dimension of [N].

[0158] A field deviation density map is constructed based on the anomaly score vector, and the field label vector is divided into a deviation normal set and a deviation abnormal set according to the field deviation density map, thereby obtaining a label deviation cluster map;

[0159] In this example, this anomaly score vector sequence is used to generate a field deviation density map. The anomaly score vectors are clustered according to a normal distribution, and the density extremes are identified and used to divide the score segments accordingly. A deviation threshold θ = 2.5σ is set to divide the field label vectors into a normal deviation set and an abnormal deviation set. The corresponding output is a label deviation cluster map, which is a binary label tensor with a structure of [N, 1].

[0160] Perform semantic structure integrity analysis on the label deviation clustering graph, and divide the field label vector into a normal label set and an abnormal label set based on the integrity analysis results, thereby obtaining a structural deviation clustering graph;

[0161] In this embodiment, based on the label deviation clustering graph, the semantic structural integrity between field labels is further analyzed. A structural depth-first aggregation method is used to assess the semantic adjacency between fields and the continuity of label boundaries. If any field fails to meet the edge weight connectivity requirements within its structural block, it is classified as an abnormal label set; otherwise, it is included in the normal label set. This process generates the final structural deviation clustering graph and outputs the [N, 1] structural flags as a tensor.

[0162] According to the structural deviation clustering diagram, the abnormal data of the certificate field annotation diagram is collected and separated, and the certificate field annotation diagrams that are divided into both the deviation normal set and the normal annotation set are separated into the normal annotation certificate data; the certificate field annotation diagrams that are divided into both the deviation abnormal set and the abnormal annotation set are separated into the abnormal annotation certificate data.

[0163] In this embodiment, the structural deviation clustering diagram is used to aggregate abnormal annotations in the document field annotation diagram. During the aggregation, if there is a field in the document field annotation diagram that belongs to both the deviation normal set and the normal annotation set in the structural deviation clustering diagram, it is marked as a normal annotation field, and the corresponding document field annotation diagram is separated into normal annotation document data; if there is a field in the document field annotation diagram that appears in both the deviation abnormal set and the abnormal annotation set in the structural deviation clustering diagram, it is marked as an abnormal annotation field, and the corresponding document field annotation diagram is separated into abnormal annotation document data. This process achieves reliable classification of insurance document fields, and outputs normal annotation document data and abnormal annotation document data respectively for subsequent annotation verification or manual review processes.

[0164] Optionally, the signature stroke shape deviation assessment in step S4 is specifically as follows:

[0165] Perform image enhancement and handwriting separation on the signature area in the abnormally marked document data, remove background interference and seal area overlap noise, extract the signature area image, and generate a signature structure image set;

[0166] In this example, image enhancement is performed on the signature region of insurance document images identified as abnormal. This enhancement process utilizes a YUV luminance channel enhancement strategy, combined with CLAHE local histogram equalization technology, to enhance handwriting line contrast. Subsequently, a saturation threshold segmentation method in the HSV color space is used to remove color interference from the background image and high-saturation red overlapping areas caused by seal overlays. The processed result retains only the handwriting information of the signature portion, which is then normalized to 128×64 pixels. This generates a set of signature images with a unified structure, forming a signature structure image tensor set with the structure [M, 1, 128, 64], where M is the number of signature samples.

[0167] Extract curvature points based on the signature structure image set, extract the stroke start and end points, turning points and curvature peak points, and construct a stroke sequence encoding map for each signature in the signature structure image set to generate a signature stroke sequence map;

[0168] In this embodiment, a structural analysis is performed on each image in the signature structure image set, and stroke contours are separated using Canny edge enhancement and connected domain separation techniques. Subsequently, key geometric points are extracted based on the curvature trend of the stroke trajectory, focusing on turning points, starting and ending points, and curvature peaks, as stroke node sequences. Each group of nodes is sequentially reconstructed and connected according to the writing path, forming a stroke chain structure. Finally, each signature stroke structure is abstracted into a sequence encoding graph, represented as a fixed-length sequence tensor with the structure [M, N_i, 3], where N_i is the number of curvature points in the i-th signature character and 3 is the (x, y, κ) triple.

[0169] Perform curvature function fitting and dynamic trajectory restoration on each stroke in the signature stroke sequence map to generate a signature shape parameter matrix;

[0170] In this embodiment, dynamic trajectory recovery is performed segment by segment on the constructed stroke sequence atlas. Cubic Bezier splines are used to curve fit the curvature sequence, and the velocity and acceleration changes of each trajectory segment are reversely deduced to reconstruct its spatial dynamic trajectory morphology. The shape parameters of each curve segment are extracted, including the starting and ending position vectors, average curvature, maximum deformation radius, trajectory energy density, and other indicators. These form a shape parameter matrix for each signature group. The format is [M, P], where P = 8, representing the 8-dimensional shape feature parameters.

[0171] Perform structural similarity matching on each parameter group in the signature shape parameter matrix and the corresponding signature structure in the verified insurance document data, calculate the residual distribution of the stroke shape of each signature group, and generate a structural deviation residual map;

[0172] In this embodiment, the generated signature shape parameter matrix is structurally matched against the standard structure of homologous signatures (same customer number or same name) in verified insurance document data, with vector-level comparison performed using Euclidean distance and direction cosine analysis. A corresponding residual vector is calculated for each stroke segment, and its positional offset and shape tension changes are statistically analyzed. All residual data is summarized as a structural deviation residual map with the structure [M, P], where each row records the stroke-level difference distribution of the current signature relative to the reference structure.

[0173] Clustering is performed based on the residual mean, maximum offset, and structural deformation index in the structural deviation residual map, and a deviation level label set is output;

[0174] In this embodiment, based on the structural deviation residual map generated in the previous order, the parameters of the stroke shape residual information of each signature are extracted, specifically including: ① residual mean, that is, the average Euclidean offset length of all strokes in the signature, the numerical range of which usually falls between 2 and 18 pixels; ② maximum offset, that is, the maximum Euclidean distance between the most severely offset stroke in the signature and its standard template, which directly reflects the degree of local structural distortion; ③ structural deformation index, which is expressed by constructing the covariance matrix of the stroke residual vector and calculating the variance ratio of its principal component eigenvalues. This index measures whether the overall shape of the signature has undergone concentrated nonlinear expansion and contraction in a certain dimension. The above three parameters will together constitute the structural deviation feature vector of each signature, and its tensor structure is in the form of ,in is the number of signature samples, and the three columns correspond to the residual mean, maximum offset, and structural deformation index. In order to avoid the scale effect caused by different dimensions, the matrix is first normalized by column, so that the mean of each column of feature data is 0 and the standard deviation is 1, generating a normalized feature tensor Next, the Gaussian kernel function is introduced to construct the similarity weight matrix between samples in the three-dimensional feature space for the normalized feature tensor. Each element of the matrix ,in It is a scale parameter set empirically, and its value is generally 0.6. Subsequently, the similarity matrix is used to construct the degree matrix and Laplace matrix of the graph structure, and the eigenvalue decomposition of the normalized Laplace matrix is performed to extract the eigenvectors corresponding to the first three smallest non-zero eigenvalues to form the feature projection tensor , projecting the original high-dimensional deviation structure into a lower-dimensional space with stronger clustering discriminability. Finally, the samples in the projected tensor are spatially partitioned. The number of clusters is preset to 5, representing five levels of structural deviation (from low to extremely high). The cluster centers are initialized by the Euclidean distance uniform distribution. After the partitioning is completed, a deviation level label is assigned to each signature. , corresponding to low, medium-low, medium, medium-high, and high deviation levels. The final generated structure is The deviation level label set is used for subsequent signature significance evaluation and high-risk document identification.

[0175] The signature deviation significance score is calculated by combining the deviation level label set and the distribution characteristics of the corresponding signature stroke sequence map. The abnormal annotated document data are divided into high-deviation annotated document data and low-deviation annotated document data according to the signature deviation significance score.

[0176] In this example, the deviation level label for each signature data piece is combined with features such as spatial compactness, node overlap, and stroke balance within the stroke sequence graph to generate a comprehensive significance score. This scoring criterion uses a linear combination, with a weight of 0.6 for the residual principal component contribution and a weight of 0.4 for the structural clustering score. The significance score interval is [0, 1], with a threshold of 0.5 used to divide samples into high-deviation and low-deviation document data. This provides a precise classification basis for subsequent anomaly repair and annotation verification processes.

[0177] Optionally, the atomic reconstruction of the signature behavior in step S4 is specifically as follows:

[0178] Perform frame-by-frame trajectory reconstruction on the signature area image in the high-deviation annotated document data to generate a handwriting dynamic reconstruction trajectory set;

[0179] In this embodiment, when processing signature area images from high-deviation annotated document data, the image is first decomposed at the frame level, and the handwriting trajectory is reconstructed in time series at a sampling rate of 8 frames per second. For each frame, the dynamic stroke path is extracted through edge detection and corner tracking, and a three-dimensional handwriting trajectory tensor is constructed. The tensor size is set to [n×m×3], where n is the number of handwriting points, m is the number of trajectory segments, and the third dimension records the (x, y, t) terms. By fitting the trajectory trends segment by segment, the original signature motion path can be restored, forming a dynamic handwriting trajectory set, which provides a dynamic behavioral foundation for subsequent analysis.

[0180] Based on the handwriting dynamic reconstruction trajectory set, segmented trajectory analysis is performed to extract the micro-action units of each signing behavior and generate a set of atomic fragments of signing behavior;

[0181] In this embodiment, after obtaining the handwriting dynamic reconstruction trajectory set, micro-motion recognition processing is performed on each trajectory. Specifically, the path bending angle, speed change rate and pause time are used as the segmentation basis to divide the trajectory into several atomic behavior units. Each unit consists of a stroke segment with a timestamp and pressure curve value. The specific method is: first, calculate the angle formed by any three consecutive points in the trajectory. If the bending angle exceeds the set threshold (such as 45°), it is marked as a potential turning point; secondly, analyze the trajectory speed change rate. If the speed of a certain section changes suddenly (such as the speed ratio of adjacent segments is greater than 2), it is regarded as a stroke rhythm change node; at the same time, the time interval of each point in the trajectory is counted. If the pause time of a point exceeds the set threshold (such as 200ms), it is identified as a pause action. The above three types of feature points are combined as the demarcation basis to segment the trajectory sequence, and finally divided into multiple atomic behavior units with independent action intentions. Each unit represents the natural action process of a continuous writing. Each atomic unit is encoded as a five-tuple: (start_x, start_y, end_x, end_y, delta_t), forming a set of atomic fragments of the signature behavior. This set of fragments can be used to model the temporal regularity and action feature patterns of the signature behavior, and serves as structural input for subsequent behavior graph construction.

[0182] Constructing a behavior probability transition graph based on the signature behavior atomic fragment set;

[0183] In this embodiment, based on the set of atomic behavior segments, a directed probability transition graph G = [V, E, P] is constructed, where V represents the set of all atomic segment nodes, E represents the temporal connections between nodes, and P represents the conditional probability distribution of transitioning from the current segment to the next. The connections between nodes are determined based on the matching relationship between the start and end points of the trajectory and the time interval. The P value is calculated by statistically normalizing the co-occurrence frequency of atomic segments in a large number of highly reliable signature behaviors. This behavior probability transition graph is used to describe the dynamic occurrence path of each type of signature behavior, providing support for temporal structure modeling.

[0184] Screening a signature area sample set based on verified insurance document data, and constructing a behavioral benchmark map based on the signature area sample set;

[0185] In this embodiment, in order to obtain a behavioral benchmark reference, 1,000 samples with clear signatures and stable structures were extracted from the verified insurance certificate data, each signature was atomically fragmented, and its corresponding behavioral graph was constructed. After aggregation, a behavioral benchmark map was constructed using the largest common subgraph structure. This map uses the steady-state pattern of segment connections as the standard behavioral path, and defines a segment concentration threshold of 0.85 to screen for high-frequency structural paths. This map serves as a reference standard for subsequent behavioral consistency assessments, ensuring the robustness of trajectory evaluation.

[0186] Using the behavioral benchmark graph, we perform graph alignment and behavioral topology similarity analysis on the behavioral probability transition graph, calculate the dynamic consistency score of each stroke path, and eliminate trajectories with low dynamic consistency scores to obtain valid trajectory data.

[0187] In this example, graph matching analysis is performed using a behavioral baseline graph and a behavioral probability transition graph. Substructure matching and node embedding reconstruction are used during the matching process to measure the topological overlap rate and the difference in node transition probability distribution between the two graphs. Finally, a dynamic consistency score is calculated for each trajectory path, ranging from 0 to 1. Trajectory paths with a score below 0.6 are considered abnormal and do not conform to the baseline behavior and are removed. The remaining trajectories are aggregated into a valid trajectory dataset, ensuring a reasonable structure, high temporal consistency, and clear motion patterns.

[0188] Based on the valid trajectory data, the atomic behavior sequence is reconstructed according to the dynamic consistency score distribution to obtain the behavior completion and reconstruction fragment set;

[0189] In this example, valid trajectory data is sorted by consistency score to construct a sorted sequence. Trajectory paths with a score above 0.75 are selected to construct a reconstructed sequence index. From this index, a trajectory set covering all high-frequency atomic segments is selected. By combining the segment connection order within the path, missing or omitted low-frequency segments are supplemented through trajectory interpolation and time node restoration to generate a complete behavioral path. This reconstructed behavioral path is then reassembled into a standardized set of segments, known as the behavioral completion reconstructed segment set, which provides the data foundation for dynamic signature reproduction.

[0190] The behavior completion reconstruction fragment set is merged with the signature behavior atomic fragment set, and conditional consistency playback is performed to generate a signature behavior reconstruction trajectory graph;

[0191] In this embodiment, the behavioral completion and reconstruction segments are sequentially fused with the original signature behavior segments, and conditional playback is performed based on the sequence of segments and consistency of the action pattern. During playback, the time intervals and trajectory curvature continuity are controlled. A complete trajectory stream is generated through the trajectory simulation engine, and a signature behavior reconstruction trajectory graph is constructed. The trajectory graph nodes contain the action start and end coordinates, time nodes, pressure values, and direction vectors, and edges represent the relationship between consecutive actions. This graph provides structural support for the dynamic reproduction of signatures.

[0192] The signature behavior reconstruction trajectory diagram is redrawn and format mapped as a static image, and mapped to the corresponding area in the high-deviation annotated document data to generate reliable high-deviation annotated document data.

[0193] In this example, the reconstructed signature trajectory graph is redrawn at the image level, with all trajectories drawn in time sequence on a blank canvas using vector graphics. A 4px line width is used for drawing, with grayscale values controlling simulated pressure. Angle smoothing is performed between strokes to ensure naturalness. The redrawn image is scaled and converted to a PNG format, ultimately overlaying the original signature area within the high-deviation annotated document data. This result creates reliable high-deviation annotated document data, effectively compensating for recognition errors caused by structural deviations in the initial annotation.

[0194] What is particularly important is that the sequence of reconstructed atomic behaviors is:

[0195] The dynamic consistency scores of all stroke paths in the behavior probability transition graph are extracted into vector form, sorted from high to low according to the scores, and divided into consistency intervals to obtain a dynamic consistency level map;

[0196] In this embodiment, for each signature trajectory path in the behavior probability migration map, the corresponding dynamic consistency score is extracted. This score is calculated based on the dynamic feature mean, speed change matching and trajectory coverage of similar paths in the behavior benchmark map, and the value range is [0,1]. The consistency scores of all trajectory paths are organized into a one-dimensional vector of length N. ,in Indicates the The vectors are then sorted from high to low according to their values and divided into three levels based on the set threshold range (e.g., high consistency: 0.8–1.0, medium consistency: 0.5–0.8, low consistency: <0.5), thus constructing a dynamic consistency level map.

[0197] The trajectory stability test is performed on the trajectories in the low consistency area of the dynamic consistency level map. If the path changes drastically or the deviation is large, the trajectory is eliminated and the low consistency screening trajectory data is obtained. The segmented entropy increase evaluation is performed on the trajectories in the medium consistency area. If the information increment is within the preset range, the medium consistency screening trajectory data is obtained.

[0198] In this embodiment, a stability detection operation is performed on the trajectory in the low consistency section. Specifically, the total direction change of the continuous path and the drift rate of the stroke center are calculated for each trajectory segment. If the total change exceeds 30 degrees / frame, or the drift rate exceeds 1.5 pixels / frame, the trajectory path is considered unstable and is eliminated to form a low consistency screening trajectory data set. For trajectories in the medium consistency area, the entropy increase of each trajectory segment is calculated using a segmented information entropy evaluation method. If the information entropy increment is within the set range (such as 0.5-1.2), it means that although the trajectory is not optimal, it still retains the real behavior characteristics and is acceptable, forming a medium consistency screening trajectory data set.

[0199] The low-consistency filtered trajectory data and the medium-consistency filtered trajectory data are aligned at the frame level to establish trajectory point correspondences. Based on the trajectory point correspondences, curvature refitting and local smoothing completion are performed to form continuous stroke sub-segments.

[0200] In this embodiment, the low consistency and medium consistency trajectory sets are merged and aligned at the frame level on the time axis. The method is to interpolate and match the trajectory points based on the timestamp sequence of each trajectory segment to construct a corresponding trajectory point sequence matrix with consistent length and frame number. ,in is the number of frames, is the trajectory dimension (usually 2D or 3D coordinates). After establishing the correspondence, local curvature refitting is performed within each stroke subsegment. Cubic spline interpolation is used to smooth sharp transitions in the trajectory and fill in gaps in short time periods, thus generating continuous and smooth stroke subsegments.

[0201] Combine the continuous stroke sub-segments according to the state transition relationship of the behavior probability transition diagram to generate an atomic behavior sequence;

[0202] In this embodiment, the behavior probability transition graph is a directed graph structure. Indicates that Represents a collection of behavioral state nodes (such as starting, lifting, turning, and retracing). represents the set of state transition edges, Status Towards After completing the reconstruction of the continuous stroke segments, each segment is first mapped to a certain behavior state node, such as the starting segment corresponding to , the turning section corresponds to , the end segment corresponds to Then, for any two adjacent trajectory segments and , respectively mark their status 、 , and query Does it exist in If exists and (If the threshold is set to 0.15), the path combination is retained; otherwise, it is judged as an unconventional behavior sequence and a supplementary segment is inserted or disassembled. Through this mechanism, all trajectory sub-segments are spliced segment by segment to form an atomic behavior sequence that follows the historical behavior trajectory transfer pattern. .

[0203] Perform sample migration and interpolation on the same type of behavior segments in the atomic behavior sequence to form a set of behavior completion and reconstruction fragments.

[0204] In this embodiment, after the atomic behavior sequence is generated, the standard signature behavior sample library is called for the part of the sequence with incomplete structure or low trajectory density. , each of which contains a complete trajectory matrix ( is the standard stroke length), state path label sequence , and trajectory pressure and velocity distribution characteristics For a certain incomplete stroke subsequence in the current atomic behavior sequence First, according to the state label and spatial direction characteristics, the sample set with completely consistent or partially overlapping state sequences in the sample library is retrieved. Then the trajectory data in these samples are normalized (for example, all coordinates are normalized to the range of [0,1]) and aligned to After the overlapping window is established, the closest sample is selected according to the minimum Euclidean distance criterion. , and then perform affine transformation and shape registration on the spatial coordinates of the sample trajectory segment to make its starting point, end point and total duration consistent with The interpolation process adheres to the principle of structural continuity, i.e., the starting direction of the interpolated trajectory must differ by no more than 20 degrees from the ending direction of the previous trajectory, and the ending direction must align with the starting point of the subsequent trajectory. This ultimately results in a structurally complete, dynamically natural set of behavioral completion and reconstruction fragments.

[0205] Optionally, this specification also provides an insurance certificate data management system for executing the above-mentioned insurance certificate data management method, the insurance certificate data management system comprising:

[0206] The verification and marking module is used to obtain the insurance certificate data uploaded by the user, and classify and verify the certificate data based on the insurance certificate data uploaded by the user to obtain the verified insurance certificate data;

[0207] The agent generation module is used to build an automatic quotation labeling model and an automatic electronic insurance policy labeling model based on the verified insurance certificate data; the automatic quotation labeling model and the automatic electronic insurance policy labeling model are combined to obtain an insurance certificate labeling agent;

[0208] The certificate anomaly labeling module is used to obtain real-time user-uploaded insurance certificate data and use the insurance certificate labeling agent to classify the certificate anomaly labels on the real-time user-uploaded insurance certificate data to obtain normal labeling certificate data and abnormal labeling certificate data;

[0209] The signature behavior reconstruction module is used to evaluate the signature stroke shape deviation of abnormally annotated document data to obtain high-deviation annotated document data and low-deviation annotated document data; the signature behavior atomic reconstruction of the high-deviation annotated document data is performed to obtain credible high-deviation annotated document data;

[0210] The certificate correction module is used to perform insurance rule reasoning correction on the low-deviation annotated certificate data and the trusted high-deviation annotated certificate data to obtain the deviation-corrected annotated certificate data;

[0211] The data encapsulation module is used to encapsulate the normal marked document data and the deviation corrected marked document data, and upload the encapsulation results to the insurance document management platform for archiving.

[0212] Optionally, this specification also provides an intelligent terminal having a computer program stored therein, which implements the above-mentioned insurance certificate data management method when executed.

[0213] The present invention is therefore intended to be illustrative and non-restrictive in all respects, with the scope of the invention being defined by the appended claims rather than the foregoing description, and all changes that come within the meaning and range of equivalents of the application documents are intended to be embraced therein.

[0214] The foregoing description is intended only to provide specific embodiments of the present invention, which will enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is to be construed in the widest possible manner consistent with the principles and novel features disclosed herein.

Claims

1. A data management method for insurance certificates, characterized in that: The following steps are involved: Step S1: Obtain insurance certificate data uploaded by the user, and classify and verify the certificate data based on the insurance certificate data uploaded by the user to obtain verified insurance certificate data; Step S2: Constructing a quotation automatic labeling model and an electronic insurance policy automatic labeling model based on the verified insurance certificate data; combining the quotation automatic labeling model and the electronic insurance policy automatic labeling model to obtain an insurance certificate labeling agent; Step S3: Obtain the real-time user-uploaded insurance certificate data, and use the insurance certificate annotation agent to classify the real-time user-uploaded insurance certificate data into abnormal certificate annotations to obtain normal annotation certificate data and abnormal annotation certificate data; Step S4: performing signature stroke shape deviation evaluation on the abnormally marked document data to obtain high-deviation marked document data and low-deviation marked document data; The high-deviation annotated document data is atomically reconstructed by signature behavior to obtain credible high-deviation annotated document data; Step S5: Perform insurance rule reasoning correction on the low-deviation annotated document data and the credible high-deviation annotated document data to obtain deviation-corrected annotated document data; Step S6: Encapsulate the normal marked document data and the deviation-corrected marked document data, and upload the encapsulation results to the insurance document management platform for archiving.

2. The data management method for insurance certificates according to claim 1, characterized in that: Step S1 is specifically as follows: Step S11: Obtain the quotation file data and electronic policy file data submitted by the user through the insurance certificate upload interface, and construct a document original index table based on the upload timestamp and device identification information, thereby generating the user-uploaded insurance certificate data; Step S12: Perform image quality detection and layout structure analysis on the insurance certificate data uploaded by the user, eliminate low-quality files and images, and generate a standardized insurance certificate image set; Step S13: performing document image classification on the standardized insurance certificate image set, extracting quotation image features and electronic policy image features respectively, and performing similarity comparison with a preset insurance certificate template library to generate classified and annotated insurance certificate data; Step S14: extracting the field position distribution based on the classified and annotated insurance certificate data and performing a position offset comparison with the standard field templates in the insurance certificate template library; performing corrections and position adjustments based on the offset comparison results to generate structure-verified insurance certificate data; Step S15: extracting field content semantic vectors from the insurance certificate data based on the structure, and performing cluster comparison with the preset industry standard field semantic set to complete field-level label assignment, thereby generating semantically annotated insurance certificate data; Step S16: The semantically annotated insurance certificate data is bound to the original document index table to obtain the verified insurance certificate data.

3. The data management method for insurance certificates according to claim 2, characterized in that: Step S12 is specifically as follows: Step S121: Read the file images of the insurance certificate data uploaded by the user one by one, extract the image clarity features, illumination balance index and noise distribution parameters, and construct an image quality assessment tensor; Step S122: Normalize the image quality assessment tensor, set the clarity threshold to 0.65, the illumination balance threshold to 0.7, and the noise tolerance upper limit to 0.3, mark images that do not meet any threshold as redundant images and remove them, and generate an image retention discrimination list; Step S123: performing regional structure analysis on the images in the image retention discrimination list, extracting the regional coordinates of the title area, field area, seal area, and signature area in the image, thereby generating a layout structure partition map; Step S124: mapping the coordinates of each region in the layout structure partition diagram to a preset A4 standard coordinate system, reordering the relative positions of the regions, adjusting the image rotation angle and scale distortion, and generating a structure normalized image set; Step S125: performing color channel alignment and boundary extension and padding operations on the images in the structural normalized image set to generate a standardized insurance document image set.

4. The data management method for insurance certificates according to claim 2, characterized in that: Step S14 is specifically as follows: Step S141: extracting coordinates of field areas in the classified and annotated insurance certificate data, extracting field position information using a bounding box positioning method, and generating a field position distribution matrix; Step S142: Retrieving a standard field template corresponding to the classified and annotated insurance certificate data from the insurance certificate template library, parsing the field structure diagram of the standard field template, and thereby constructing a standard field position matrix; Step S143: performing position offset calculation on the field position distribution matrix and the standard field position matrix, and marking the field position with a field offset greater than 10px in the position offset calculation result as an offset abnormality, thereby obtaining an offset abnormality field; Step S144: Based on the structural semantic level of the field structure diagram, sequentially correct the abnormally offset fields, perform field bounding box repositioning, and obtain a position-corrected field matrix; Step S145: Compare the field coordinates of the position correction field matrix with the alignment reference points of the standard field template, and iterate steps S143 and S144 until the field offset converges within the maximum tolerable error range of <5px, thereby generating a structurally consistent field position matrix; Step S146: Bind the structure-consistent field position matrix to the corresponding fields in the classified and annotated insurance certificate data to generate structure-verified insurance certificate data.

5. The data management method for insurance certificates according to claim 1, characterized in that: Step S2 is specifically as follows: Step S21: Based on the verified insurance document data, extract the quotation data subset and the electronic policy data subset according to the document type label, perform field-level region slicing and semantic label alignment operations respectively, and construct the quotation training set and the electronic policy training set; Step S22: Based on the quotation training set, a multi-channel cross-attention structure model is performed to obtain a quotation automatic labeling model; based on the electronic insurance policy training set, a bidirectional symmetric multi-scale perception network model is performed to obtain an electronic insurance policy automatic labeling model; Step S23: performing structural model splicing on the automatic quotation annotation model and the automatic electronic insurance policy annotation model, thereby obtaining a structural splicing intermediate model; Step S24: Perform dynamic path optimization training on the structure splicing intermediate model to obtain an insurance document labeling agent; Step S25: Connect the insurance certificate annotation agent to the insurance certificate management platform, register the input structure, identify the service interface and callback output specifications to execute the agent deployment task.

6. The data management method for insurance certificates according to claim 5, characterized in that: The specific classification of abnormal document marking in step S3 is: Perform structured parsing and input tensor reconstruction on real-time user-uploaded insurance certificate data to obtain the certificate structured tensor; Input the document structured tensor into the insurance document annotation agent, perform document type guided path selection and field annotation, and thus obtain the document initial annotation graph; Based on the initial document annotation graph, a label similarity edge graph between fields is constructed, and the label similarity edge graph is corrected for field label space consistency to generate a document field annotation graph; Based on the label vector of each field in the document field annotation graph, the Mahalanobis distance between the standard semantic vectors of the corresponding fields in the quotation training set and the electronic insurance policy training set is calculated to construct an anomaly score vector. A field deviation density map is constructed based on the anomaly score vector, and the field label vector is divided into a deviation normal set and a deviation abnormal set according to the field deviation density map, thereby obtaining a label deviation cluster map; Perform semantic structure integrity analysis on the label deviation clustering graph, and divide the field label vector into a normal label set and an abnormal label set based on the integrity analysis results, thereby obtaining a structural deviation clustering graph; According to the structural deviation clustering diagram, the abnormal data of the certificate field annotation diagram is collected and separated, and the certificate field annotation diagrams that are divided into both the deviation normal set and the normal annotation set are separated into the normal annotation certificate data; the certificate field annotation diagrams that are divided into both the deviation abnormal set and the abnormal annotation set are separated into the abnormal annotation certificate data.

7. The data management method for insurance certificates according to claim 6, characterized in that: The signature stroke shape deviation assessment in step S4 is specifically as follows: Perform image enhancement and handwriting separation on the signature area in the abnormally marked document data, remove background interference and seal area overlap noise, extract the signature area image, and generate a signature structure image set; Extract curvature points based on the signature structure image set, extract the stroke start and end points, turning points and curvature peak points, and construct a stroke sequence encoding map for each signature in the signature structure image set to generate a signature stroke sequence map; Perform curvature function fitting and dynamic trajectory restoration on each stroke in the signature stroke sequence map to generate a signature shape parameter matrix; Perform structural similarity matching on each parameter group in the signature shape parameter matrix and the corresponding signature structure in the verified insurance document data, calculate the residual distribution of the stroke shape of each signature group, and generate a structural deviation residual map; Clustering is performed based on the residual mean, maximum offset, and structural deformation index in the structural deviation residual map, and a deviation level label set is output; The signature deviation significance score is calculated by combining the deviation level label set and the distribution characteristics of the corresponding signature stroke sequence map. The abnormal annotated document data are divided into high-deviation annotated document data and low-deviation annotated document data according to the signature deviation significance score.

8. The method for managing insurance certificate data according to claim 1, wherein: The atomic reconstruction of the signature behavior in step S4 is specifically as follows: Perform frame-by-frame trajectory reconstruction on the signature area image in the high-deviation annotated document data to generate a handwriting dynamic reconstruction trajectory set; Based on the handwriting dynamic reconstruction trajectory set, segmented trajectory analysis is performed to extract the micro-action units of each signing behavior and generate a set of atomic fragments of signing behavior; Constructing a behavior probability transition graph based on the signature behavior atomic fragment set; Screening a signature area sample set based on verified insurance document data, and constructing a behavioral benchmark map based on the signature area sample set; Using the behavioral benchmark graph, we perform graph alignment and behavioral topology similarity analysis on the behavioral probability transition graph, calculate the dynamic consistency score of each stroke path, and eliminate trajectories with low dynamic consistency scores to obtain valid trajectory data. Based on the valid trajectory data, the atomic behavior sequence is reconstructed according to the dynamic consistency score distribution to obtain the behavior completion and reconstruction fragment set; The behavior completion reconstruction fragment set is merged with the signature behavior atomic fragment set, and conditional consistency playback is performed to generate a signature behavior reconstruction trajectory graph; The signature behavior reconstruction trajectory diagram is redrawn and format mapped as a static image, and mapped to the corresponding area in the high-deviation annotated document data to generate reliable high-deviation annotated document data.

9. A data management system for insurance certificates, characterized in that: For executing the data management method of insurance documents according to claim 1, the data management system of insurance documents comprises: The verification and marking module is used to obtain the insurance certificate data uploaded by the user, and classify and verify the certificate data based on the insurance certificate data uploaded by the user to obtain the verified insurance certificate data; The agent generation module is used to build an automatic quotation labeling model and an automatic electronic insurance policy labeling model based on the verified insurance certificate data; the automatic quotation labeling model and the automatic electronic insurance policy labeling model are combined to obtain an insurance certificate labeling agent; The certificate anomaly labeling module is used to obtain real-time user-uploaded insurance certificate data and use the insurance certificate labeling agent to classify the certificate anomaly labels on the real-time user-uploaded insurance certificate data to obtain normal labeling certificate data and abnormal labeling certificate data; The signature behavior reconstruction module is used to evaluate the signature stroke shape deviation of abnormally annotated document data to obtain high-deviation annotated document data and low-deviation annotated document data; the signature behavior atomic reconstruction of the high-deviation annotated document data is performed to obtain credible high-deviation annotated document data; The certificate correction module is used to perform insurance rule reasoning correction on the low-deviation annotated certificate data and the trusted high-deviation annotated certificate data to obtain the deviation-corrected annotated certificate data; The data encapsulation module is used to encapsulate the normal marked document data and the deviation corrected marked document data, and upload the encapsulation results to the insurance document management platform for archiving.

10. An intelligent terminal, characterized in that: A computer program is stored therein, which, when executed, implements the data management method for insurance certificates according to any one of claims 1 to 8.

Images