Artificial annotation method for AI model with associated logic

Abstract

The invention discloses an AI model manual annotation method with associated logic, relates to the technical field of artificial intelligence and optical character recognition (OCR), and aims to solve the problems that the existing OCR recognition technology is low in accuracy and lacks an effective content verification and correction mechanism when processing complex certificates. The method comprises the following steps of: firstly, extracting a driving license, a driving license, an electronic license and text contents of various certificates through OCR (Optical Character Recognition) scanning equipment; constructing an association logic AI model for storing certificate information association logic rules; performing structured labeling processing on an OCR recognition result to realize document information classification and regularization; performing logic verification on the annotated content by utilizing an association logic AI model, marking problems and proposing a modification scheme; and finally, corrected data is fed back to the model to complete optimization iteration. According to the method, the logic contradiction can be accurately found and corrected, the identification accuracy of the AI model on multiple types of certificates is remarkably improved, the manual labeling error is reduced, and the certificate information extraction efficiency and reliability are improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence and optical character recognition (OCR) technology, specifically to a method for manually annotating AI models of association logic for eight common types of documents, including vehicle registration certificates, driver's licenses, electronic vehicle registration certificates, electronic driver's licenses, ID cards, electronic ID cards, new vehicle qualification certificates, and vehicle operating permits. This method is applicable to fields requiring accurate extraction of document information, such as government services, traffic management, financial risk control, and vehicle management, and can effectively improve the accuracy and data reliability of document OCR recognition. Background Technology

[0002] With the acceleration of digital transformation, OCR technology is being used more and more widely in the field of document information extraction. Eight types of documents, including vehicle registration certificates, driver's licenses, and ID cards, serve as core proof of personal identity and vehicle qualifications, and the rapid and accurate recognition of their information has become a crucial link in many business scenarios. However, existing OCR recognition technologies have many shortcomings in practical applications.

[0003] First, the accuracy rate of complex document recognition is low. Physical documents are prone to problems such as stains, creases, and blurry fonts, while electronic documents have differences in layout and interference from anti-counterfeiting marks, which often leads to text recognition errors and information omissions in OCR technology. In particular, the error rate is relatively high for professional fields such as vehicle parameters on new vehicle qualification certificates and qualification periods on vehicle operating permits.

[0004] Secondly, there is a lack of effective logical verification mechanisms. Traditional manual annotation only corrects single fields, ignoring the inherent logical relationships between document information. For example, the middle 8 digits of the ID card number are not matched with the annotated birth date, and the correspondence between the permitted vehicle type on a driver's license and the holder's age is not verified. This results in numerous logical inconsistencies in the annotated data, making it difficult for the trained AI model to handle such contradictory scenarios.

[0005] Finally, the quality control of annotation is inadequate. Existing methods lack a systematic quality assessment framework, the accuracy and completeness of the annotated data cannot be effectively guaranteed, and the utilization rate of the annotated data is low, making it difficult to support the continuous optimization of AI models and meet the actual needs of high-precision business scenarios.

[0006] Therefore, there is an urgent need for an AI model manual annotation method that combines correlation logic to address the shortcomings of existing technologies and improve the accuracy and reliability of OCR recognition for various types of documents. Summary of the Invention

[0007] Purpose of the invention

[0008] The purpose of this invention is to provide a method for manually annotating AI models with related logic. By using structured annotation of multiple types of documents, verification of related logic, closed-loop quality control, and iterative optimization of the model, this method solves the problems of low accuracy, prominent logical contradictions, and poor annotation quality of existing OCR recognition technologies for complex documents. It improves the recognition accuracy and data consistency of AI models for eight types of documents, reduces the cost of manual annotation, and meets the needs of various fields for high-precision extraction of document information.

[0009] (I) Technical Solution

[0010] This invention provides a method for manually labeling AI models with related logic, the specific steps of which are as follows:

[0011] 1. Multi-type Document OCR Scanning and Recognition: Employing a high-precision OCR scanning device with dynamic parameter adjustment, it performs targeted scanning of eight types of documents, including ID cards, driver's licenses, vehicle registration certificates, passports, social security cards, bank cards, business licenses, and property ownership certificates. It can be expanded to support more document types based on actual business needs. During physical document scanning, the device automatically adjusts the scanning resolution (ranging from 300 to 600 dpi), exposure, and contrast based on the degree of dirt and creases on the document surface to ensure the clarity and recognizability of the scanned image. It also has real-time image quality assessment capabilities, prompting for rescanning of blurry, overexposed, or low-contrast images. When scanning electronic documents, the device focuses on extracting anti-counterfeiting codes and core information areas, and optimizes image quality through a multi-modal image preprocessing module, effectively overcoming problems such as screen reflections, moiré patterns, and pixelation, thus improving recognition accuracy. After extracting the text content from the document, the system uses advanced natural language processing technology to intelligently standardize the formats of dates, numbers, and numbers. It supports multiple date format conversions, unification of full-width and half-width characters, and automatic correction of common spelling errors to ensure the consistency and standardization of data.

[0012] 2. Construction of the Association Logic AI Model: Based on the TensorFlow 2.0 framework, a powerful association logic AI model was built, pre-setting over 120 association logic rules for eight types of documents, and possessing good scalability to support the addition of new rules. For example, the first 6 digits of the ID card number's address code need to be matched and verified against the administrative region database; drivers with an A1 driving license must be at least 26 years old; the 10th digit of the vehicle identification number on the vehicle registration certificate represents the year of manufacture, etc. The model training process used over 100,000 document sample data, covering document images of different versions, regions, and wear conditions. Through five-fold cross-validation and adaptive learning rate algorithms, combined with a focus loss function to handle the sample imbalance problem, the model's performance and accuracy were continuously improved, ultimately achieving a field-level accuracy of 98.7% on the test set.

[0013] 3. Structured Annotation Processing: A multi-level structured format was constructed according to the JSON-LD standard, subdividing each type of document into three levels: basic identification, core information, and additional attributes. The basic identification category includes metadata such as document type and issuing country; the core information category contains key fields such as name and number; and the additional attributes category covers auxiliary information such as validity period and issuing authority. A unique code was assigned to each field (e.g., the name field on an ID card is coded as "ID-INFO-NAME"), adhering to the principles of readability and uniqueness. During the annotation process, ambiguous fields were marked "pending confirmation," and their coordinates and uncertainty scores were recorded. A collaborative review mechanism among annotators was established, ultimately forming a standardized annotated dataset, providing a solid foundation for subsequent data processing and analysis.

[0014] 4. Logic Verification and Problem Correction: A dual verification mechanism of "rule verification + data comparison" is adopted. The rule verification stage judges the consistency of information based on more than 120 preset logical rules, including format checks, logical correlation, and numerical rationality verification. The data comparison stage matches the current data with more than 500,000 historically accumulated correct data entries, using a hybrid matching algorithm based on SimHash and cosine similarity. Once a logical contradiction or data inconsistency is found, the system will automatically generate a detailed correction report, clearly indicating the problem type, suspicious fields, and recommended handling methods, and push it to the manual review stage for further correction to ensure the accuracy and reliability of the data. All correction results are fed back to the model training stage, forming a continuous optimization loop.

[0015] 5. Model Optimization and Iteration: Corrected labeled data is collected and cleaned using multiple methods including hash deduplication, manual completion, and algorithm anomaly detection. Hash deduplication combines image and text features, manual completion focuses on low-confidence samples, and the isolated forest algorithm is used to identify potential labeling anomalies for anomaly detection. Incremental training is employed to optimize the model, with weekly parameter updates and dynamic adjustments to the weights of logical rules. A hard sample mining strategy is introduced to specifically address model weaknesses, further enhancing the model's adaptability and recognition accuracy in complex scenarios, particularly its robustness under non-ideal conditions such as blurriness, tilt, and occlusion.

[0016] 6. Closed-Loop Quality Control for Annotation: A three-dimensional quality assessment system has been established, covering three dimensions: recognition accuracy, logical consistency, and annotation completeness, with 12 secondary indicators. Over 10,000 recognition results are comprehensively evaluated daily. Data that fails to meet standards automatically initiates a secondary correction process, assigning it to the appropriate correction queue based on the problem type, simultaneously optimizing annotation rules and model parameters, and generating daily quality analysis reports. By establishing a quality trend prediction model, potential quality risks are identified in advance, forming a closed-loop quality control mechanism of "assessment-correction-optimization-early warning," ensuring the high quality and stability of annotated data and consistently maintaining an overall quality compliance rate of over 99.2%.

[0017] 7. Cross-Document Association Verification: A cross-document verification rule base has been established for related documents (such as driver's license and electronic driver's license, ID card and passport, etc.), containing 28 core verification rules. By verifying the consistency of information of the same subject (such as name consistency, logical correlation of number, validity period compatibility, etc.), a knowledge graph-based association mining technology is used to further improve the reliability and integrity of data. The system supports automatically discovering cross-document association groups and generating association verification reports, ensuring the logical correlation and data consistency between various document information, and providing reliable data support for cross-document application scenarios.

[0018] (II) Beneficial Effects

[0019] 1. Significantly improved recognition accuracy: Through targeted image preprocessing and dynamic parameter adjustment, combined with correlation logic verification, the recognition accuracy of eight types of certificates has been improved by 16-20 percentage points, especially solving the recognition error problem of complex certificates and professional fields.

[0020] 2. Complete elimination of logical contradictions: The dual logical verification mechanism and cross-document association verification reduce the rate of logical contradiction omissions to 0, ensuring the consistency and reliability of document information.

[0021] 3. Significantly improved annotation efficiency: Structured annotation and automatic correction report generation reduce manual operation time, shortening manual annotation time by 34-38 seconds per sheet and reducing annotation costs.

[0022] 4. Continuously controllable annotation quality: The 3D quality assessment system and closed-loop management process ensure the accuracy and completeness of the annotation data, providing high-quality data support for AI model training.

[0023] 5. Continuously enhanced model adaptability: Incremental training and rule optimization enable the model to continuously adapt to changes in document styles and complex scenarios, thereby improving the model's generalization ability.

[0024] 6. Wide range of applications: It can cover the identification and annotation of eight types of certificates and can be extended to other types of certificates. It is suitable for certificate information extraction scenarios in multiple fields and has high application value. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating the implementation steps of OCR recognition and annotation of physical document combinations according to Embodiment 1 of the present invention.

[0026] Figure 2 This is a flowchart illustrating the implementation steps of OCR recognition and annotation for all types of certificates in Embodiment 2 of the present invention. Detailed Implementation

[0027] The specific implementation of the AI ​​model manual annotation method with associative logic of the present invention is as follows:

[0028] When performing OCR scanning and recognition of various types of documents, the corresponding scanning mode is first selected based on the type of document (physical or electronic). For physical documents, the OCR scanning device first detects the degree of stains and creases on the document surface, automatically adjusts the scanning resolution to a suitable value between 300-600 dpi, and adjusts the exposure and contrast. Then, Gaussian filtering is applied to the scanned image for noise reduction, and Hough transform tilt correction is performed. For electronic documents, the device focuses on locating the anti-counterfeiting code and core information areas, performing pixel normalization and anti-counterfeiting mark extraction. After extracting the text content, natural language processing technology is used to convert dates to YYYY-MM-DD format, and serial numbers to a combination of letters and numbers, retaining two decimal places for numerical values, and automatically correcting common spelling errors.

[0029] When constructing the association logic AI model, the basic architecture was built based on the TensorFlow 2.0 framework, and over 100,000 samples of eight types of certificates covering different degrees of damage, regional differences, and version formats were imported as the training dataset. Over 120 pre-defined association logic rules were established for the eight types of certificates, such as matching rules between ID card number encoding and birth date, and matching rules between driver's license driving license types and driving experience / age. A five-fold cross-validation method was used, dividing the training set, validation set, and test set in a 7:2:1 ratio. The initial learning rate was set to 0.001, and the learning rate was adjusted every 5 epochs based on the validation set loss value. Simultaneously, a focus loss function was used to address the sample imbalance problem, training until the F1 score on the test set reached above 0.95.

[0030] In the structured annotation processing stage, following the JSON-LD standard, three levels of annotation templates are created for each type of document: basic identifier, core information, and additional attribute. The basic identifier template includes fields such as document number and unique identifier; the core information template includes core attribute information fields for the document holder / vehicle; and the additional attribute template includes fields such as remarks and change records. During the annotation process, annotators fill in the corresponding template fields with the OCR recognition results, assign a unique field code and document type identifier to each field, mark ambiguous fields with a "pending confirmation" label, and record the image coordinates. After annotation is completed, a standardized annotation dataset is formed.

[0031] In the logic verification and issue correction phase, the rule verification module is activated first. Based on over 120 pre-defined subdivided logical rules, the system checks the format, logical coherence, and numerical reasonableness of the labeled content. Then, the data comparison module is activated, performing a hybrid matching based on SimHash and cosine similarity between the current labeled data and over 500,000 historical correctly labeled data. When both verifications determine a contradiction, the system automatically generates a correction report containing the contradiction location, conflict type, correction suggestions, and supporting evidence, and pushes it to the human review panel. Reviewers correct the contradictory data based on the report, and the correction results are fed back to the model training panel.

[0032] During the model optimization and iteration process, corrected labeled data is collected. First, a combination of "field hashing for deduplication + manual completion + algorithm anomaly detection" is used for data cleaning, ensuring a deduplication accuracy of 99.9% and an anomaly correction accuracy of no less than 98%. The cleaned data is then stored in the labeled database using encryption. Next, an incremental training approach is adopted, using the cleaned new data weekly to train the association logic AI model, dynamically updating the weights of the association logic rules in the model. A hard sample mining strategy is introduced, focusing on training complex scene samples where the model is weak, thereby improving the model's adaptability.

[0033] Regarding closed-loop management of annotation quality, a three-dimensional quality assessment system is activated daily to evaluate the accuracy, logical consistency, and completeness of the annotated data. The acceptable thresholds for these assessment indicators are set as follows: accuracy ≥ 98%, logical consistency ≥ 99%, and completeness ≥ 97%. A quality assessment report is generated. For annotated data that fails to meet the standards, a secondary correction process is automatically initiated, assigning it to the appropriate correction queue based on the problem type, while simultaneously optimizing annotation rules and model parameters. Through a quality trend prediction model, historical quality data is analyzed to identify and issue early warnings of potential quality risks.

[0034] When performing cross-document association verification, for documents with a relationship (such as driver's license and electronic driver's license, vehicle registration certificate and vehicle operating permit, etc.), the system calls upon 28 core verification rules from the cross-document verification rule base and employs knowledge graph-based relationship mining technology to verify information such as name consistency, logical correlation of numbers, and validity period compatibility for the same entity. The system automatically discovers cross-document association groups and generates an association verification report, marks inconsistent data, and after manual verification and correction, ensures the consistency of cross-document information.

[0035] Example 1: OCR recognition and annotation of physical document combinations (ID card + driver's license + vehicle registration certificate + vehicle operating permit) is attached. Figure 1 The implementation steps are as follows:

[0036] Step 1: OCR Scanning and Recognition: Using a specific brand of OCR scanning equipment, 2000 physical document samples (500 ID cards, 500 driver's licenses, 500 vehicle registration certificates, and 500 vehicle operating permits, all containing light to heavy stains and creases) were scanned. The equipment automatically adjusted parameters, and after Gaussian filtering for noise reduction and Hough transform for skew correction, core fields were extracted, such as name, date of birth, and ID number on ID cards; permitted vehicle types and validity period on driver's licenses; vehicle identification number and brand / model on vehicle registration certificates; and operating qualification period and business scope on vehicle operating permits. These were then formatted according to standard specifications.

[0037] Step 2: Building the AI ​​Model for Association Logic: The model is built based on the TensorFlow 2.0 framework, with preset association rules, such as the 7th-14th digits of the ID card number matching the marked birth date, the holder of a C1 driver's license being at least 18 years old, the 10th digit of the vehicle identification number on the vehicle registration certificate matching the vehicle's production year, and the vehicle operating permit's validity period not exceeding 5 years. The model is trained using 100,000 document samples, with 50% cross-validation and an initial learning rate of 0.001. The F1 score on the test set reaches 0.96.

[0038] Step 3: Structured Annotation Processing: Annotate according to JSON-LD format. The basic identifier class contains the document number, the core information class contains the core attributes of the certificate holder / vehicle, and the additional attribute class contains remarks. Assign a unique code to each field, such as "DRIVER-LICENSE-CORE-TYPE" for driver's license vehicle type. Mark the ambiguous fields of the 30 documents (such as the business scope of the vehicle operating permit) as "Pending Confirmation" and record the coordinates.

[0039] Step 4: Logic Verification and Problem Correction: Dual verification uncovered 120 logical contradictions, such as the inconsistency between the birth date code "19900510" and the marked date "19900610" in the ID number, and the fact that the holder of a driver's license for type A2 was 23 years old (below the legal age of 24). A correction report was generated and sent for manual review. After review, all contradictions were corrected.

[0040] Step 5: Model optimization iteration: Collect 2000 corrected data points, remove 5 duplicates, complete 8 missing information points, correct 12 abnormal data points, use incremental training to optimize the model, update the logical rule weights, and improve the model's F1 score to 0.97.

[0041] Step 6: Closed-loop quality control: Daily quality assessments are conducted. The results of the first day's assessment were: recognition accuracy 98.2%, logical consistency 99.1%, and labeling completeness 97.5%, all of which met the acceptable thresholds.

[0042] Step 7: Cross-document association verification: Verify the consistency of the holder's name on the driver's license and the vehicle operation permit. Three inconsistent data were found. After manual verification and correction, the consistency of cross-document information reached 100%.

[0043] 1. Experimental data table and analysis of beneficial effects

[0044] Experimental indicators Traditional methods Method of the present invention Increase Recognition accuracy (%) 81.5 98.2 16.7 percentage points Logical contradiction omission rate (%) 30.2 0 30.2 percentage points Manual annotation time (seconds / image) 58.6 24.3 34.3 seconds Data utilization rate (%) 73.8 99.3 25.5 percentage points Cross-document information consistency (%) 78.5 100 21.5 percentage points

[0045] Experimental data shows that the method of this invention performs excellently in the annotation of combined physical documents. The recognition accuracy is improved by 16.7 percentage points, completely resolving the problem of logical inconsistencies and omissions, improving manual annotation efficiency by 58.5%, significantly increasing the utilization rate of annotation data, and achieving 100% consistency of cross-document information, effectively meeting the needs for accurate annotation of physical documents.

[0046] Example 2: OCR recognition and annotation of all types of documents (including electronic documents + new vehicle certificate of conformity) are attached. Figure 2 The implementation steps are as follows:

[0047] Step 1: OCR Scanning and Recognition: Using professional OCR equipment, 2500 samples of all types of documents (300 electronic vehicle registration certificates, 300 electronic driver's licenses, 300 electronic ID cards, 400 new vehicle qualification certificates, 400 physical ID cards, 400 physical driver's licenses, and 300 vehicle registration certificates) were scanned. Electronic documents underwent pixel normalization and anti-counterfeiting mark extraction. Physical documents underwent standard preprocessing, extracting core fields such as the anti-counterfeiting code of electronic documents, vehicle parameters of new vehicle qualification certificates, and manufacturer information, and standardizing the format.

[0048] Step 2: Building the AI ​​Model for Association Logic: The model is built based on the TensorFlow 2.0 framework, with pre-defined rules, such as binding and matching the anti-counterfeiting code of the electronic ID card with the ID card number, uniquely corresponding the vehicle identification number on the new vehicle certificate with the engine number, and ensuring that the validity period of the electronic driver's license is consistent with the physical driver's license. It is trained with 120,000 samples and undergoes five-fold cross-validation, achieving an F1 score of 0.955 on the test set.

[0049] Step 3: Structured annotation processing: Classify and annotate according to JSON-LD format, assign a unique code to each field, such as the anti-counterfeiting code of electronic vehicle registration certificate is "E-DRIVING-LICENSE-ID-ANTI", mark the ambiguous fields of 45 documents (such as the emission standard of new vehicle certificate) as "to be confirmed" and record the coordinates.

[0050] Step 4: Logic Verification and Problem Correction: Dual verification identified 156 logical contradictions, such as mismatches between the electronic driver's license anti-counterfeiting code and the ID card number, and incorrect correspondences between the vehicle identification number and engine number on the new vehicle certificate of conformity. A correction report was generated and submitted for manual review to complete the correction of all contradictions.

[0051] Step 5: Model optimization iteration: Collect 2500 corrected data points, remove 8 duplicate data points, complete 12 missing information points, and correct 15 abnormal data points. After incremental training, the model's F1 score improved to 0.972.

[0052] Step 6: Quality closed-loop control: Daily quality assessment, the first day's recognition accuracy was 97.8%, logical consistency was 99.2%, and labeling completeness was 97.2%, meeting the qualified threshold.

[0053] Step 7: Cross-document association verification: Verify the consistency of information between the electronic driver's license and the physical driver's license, and between the electronic ID card and the physical ID card. Five inconsistent data were found. After correction, the consistency reached 100%.

[0054] 1. Experimental data table and analysis of beneficial effects

[0055] Experimental indicators Traditional methods Method of the present invention Increase Recognition accuracy (%) 78.3 97.8 19.5 percentage points Logical contradiction omission rate (%) 35.6 0 35.6 percentage points Manual annotation time (seconds / image) 65.2 26.8 38.4 seconds Data utilization rate (%) 70.5 98.8 28.3 percentage points Consistency between electronic and physical documents (%) 72.8 100 27.2 percentage points

[0056] Experimental data show that the method of this invention has significant advantages in the annotation of all types of certificates. The recognition accuracy is improved by 19.5 percentage points, the logical contradiction omission rate is reduced to 0, the manual annotation efficiency is improved by 58.9%, the utilization rate of annotation data is greatly improved, and the consistency between electronic and physical certificate information reaches 100%, effectively solving the pain points of the recognition and annotation of all types of certificates.

Claims

1. A method for manually annotating AI models with associative logic, characterized in that, Includes the following steps: Step S1: Multi-type document OCR scanning and recognition. An OCR scanning device with dynamic parameter adjustment function is used to scan documents including vehicle registration certificate, driver's license, electronic vehicle registration certificate, electronic driver's license, ID card, electronic ID card, new vehicle certificate of conformity, and vehicle operation certificate. The scanning resolution, exposure and contrast parameters are automatically adjusted according to the document type and surface condition to extract the text content of the document, covering core fields such as identity information, vehicle information, document validity period, and driving / operating qualifications. Step S2: Construction of the association logic AI model. Construct an association logic AI model based on a deep learning framework. Pre-store exclusive association logic rules for eight types of certificates, including matching logic between ID card number encoding rules and date of birth, corresponding rules between driver's license driving type and driving experience / age, association verification rules between vehicle identification number and brand / model on vehicle registration certificate, and binding rules between electronic certificate anti-counterfeiting code and certificate information, etc. Step S3: Structured annotation processing. The OCR recognition results are classified and annotated according to a preset multi-level structured format. Each type of document is divided into three levels: basic identification class, core information class, and additional attribute class. A unique field code and document type identifier are assigned to each annotation field to form a standardized annotation dataset. Step S4: Logic verification and problem correction. A dual logic verification mechanism of "rule verification + data comparison" is adopted. The AI ​​model of correlation logic is used to verify the labeled content, automatically mark logical contradictions, and generate a correction report containing the contradiction location, conflict type, correction suggestions and basis. Step S5: Model optimization and iteration. Collect the corrected labeled data. After data cleaning and processing such as deduplication, completion and anomaly correction, use incremental training to optimize the AI ​​model of association logic, update the weights of association logic rules in the model, and improve the model's adaptability to complex scenarios. Step S6: Closed-loop management of annotation quality. Establish a three-dimensional quality assessment system that includes recognition accuracy, logical consistency, and annotation completeness. Generate quality assessment reports regularly, initiate a secondary correction process for annotation data that does not meet the standards, and simultaneously optimize annotation rules and model parameters.

2. The method according to claim 1, characterized in that, In step S1, the OCR scanning device is equipped with a multimodal image preprocessing module. For physical documents, it uses "Gaussian filtering for noise reduction + Hough transform tilt correction" for processing, and for electronic documents, it uses "pixel normalization + anti-counterfeiting mark extraction" for processing to ensure the clarity and completeness of the extracted text.

3. The method according to claim 1, characterized in that, In step S2, the deep learning framework used is TensorFlow 2.0 or above. The model training dataset contains 100,000+ document samples covering eight categories with different degrees of damage, regional differences, and format versions. The training process uses an adaptive learning rate algorithm, with the initial learning rate set to 0.

001. Every 5 epochs, the learning rate is adjusted based on the validation set loss value.

4. The method according to claim 1, characterized in that, In step S3, the multi-level structured format is constructed using the JSON-LD standard. The basic identifier class includes fields such as document number and unique identification code, the core information class includes core attribute information of the certificate holder / vehicle, and the additional attribute class includes fields such as remarks and change records. During the annotation process, ambiguous fields are marked with "to be confirmed" and the image coordinates are recorded.

5. The method according to claim 1, characterized in that, In step S4, rule verification uses 120+ pre-defined subdivided logical rules. Data comparison is performed by similarity matching with 500,000+ historical correctly labeled data. Only when both verifications are determined to be contradictory will a correction report be generated and pushed to the manual review end.

6. The method according to claim 1, characterized in that, In step S5, data cleaning employs a combination of field hash deduplication, manual assistance in data completion, and algorithm anomaly detection. The deduplication accuracy reaches 99.9%, and the accuracy rate of anomaly data correction is no less than 98%. The cleaned data is stored in the annotation database in an encrypted manner.

7. The method according to claim 3, characterized in that, The model training uses a five-fold cross-validation method, with the ratio of training set, validation set, and test set being 7:2:

1. The model evaluation metrics include precision, recall, and F1 score, and the F1 score on the test set must reach 0.95 or higher.

8. The method according to claim 1, characterized in that, The qualification thresholds for the 3D quality assessment system in step S6 are: recognition accuracy ≥ 98%, logical consistency ≥ 99%, and annotation integrity ≥ 97%. The assessment cycle is once a day, and a comprehensive optimization of the annotation process is triggered when the standards are not met twice in a row.

9. The method according to claim 1, characterized in that, It also includes step S7: cross-document association verification, which establishes cross-document information verification rules for documents with related relationships, including driver's license and electronic driver's license, vehicle registration certificate and vehicle operation certificate, to verify the consistency of information of the same subject and further improve the reliability of the labeled data.

10. The method according to claim 1, characterized in that, After extracting the text content in step S1, natural language processing technology is used to standardize the format, unifying the date format as YYYY-MM-DD, the number format as a combination of "letters + numbers", and the numerical format to retain two decimal places, to ensure the uniformity of the data format.

Images