An intelligent customs declaration system and method for international trade

Abstract

The application discloses an international trade intelligent customs declaration system and method, and relates to the technical field of international trade.The application comprises a data acquisition module for acquiring single certificate image data and associated form data of a batch to be declared; a multi-modal analysis and verification module for loading a trained multi-modal perception model; and the following steps: S1: acquiring single certificate image data and associated form data of a batch to be declared.The application has the advantages that single certificate information extraction and compliance logic verification are simultaneously completed in single semantic understanding, the recognition result directly carries a quantifiable and reliable verification conclusion, eliminating the information gap of the traditional separate architecture; at the same time, the maintenance work of creating fixed templates according to countries and single certificate types is eliminated, the application has self-adaptive capacity for format changes and unexpected abnormalities, and the model can be continuously optimized from the correction feedback of the customs officer, so that compliance conflicts can be effectively intercepted in the declaration preparation stage, and the single rejection rate and customs clearance time are reduced.

Description

Technical Field

[0001] This invention relates to the field of international trade technology, and in particular to an intelligent customs declaration system and method for international trade. Background Technology

[0002] With the continuous growth of international trade, the customs declaration process for import and export of goods involves a variety of business documents such as sales contracts, commercial invoices, waybills, certificates of origin, and inspection and quarantine certificates. Customs declarants need to extract key fields such as the country of trade, currency of transaction, commodity code, and mode of transport from these documents and organize the declaration data in accordance with the standardized format of the international trade "single window". In this process, errors in field filling or logical inconsistencies between fields are one of the main reasons for customs rejection.

[0003] Existing technologies have attempted to incorporate OCR (Optical Character Recognition) and RPA (Robotic Process Automation) technologies into customs declaration. For example, Chinese invention patent CN119624346A discloses an intelligent customs declaration method that identifies business fields in documents by creating OCR templates grouped by trade countries, and then uses an RPA robot to input the data into a "single window." This solution alleviates the burden of manual data entry to some extent, but still has the following drawbacks: When the document format is slightly adjusted or a non-standard format appears, template matching fails, recognition is interrupted, and the system has poor robustness. The OCR stage only outputs text strings without carrying business semantics. The verification rules are independent of the recognition process and cannot be traced back to verify the original image. Once an error occurs in OCR, subsequent verification is performed based on the erroneous data. The risk of order cancellation is not reduced, but rather made more difficult to trace because the system hides the process.

[0004] Therefore, it is necessary to design an intelligent customs declaration system and method for international trade. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent customs declaration system and method for international trade, which solves the problems mentioned in the background section.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: An intelligent customs declaration system for international trade includes: The data acquisition module is used to acquire the document image data and associated form data of the batch to be declared; The multimodal parsing and verification module is used to load a trained multimodal perception model. The multimodal perception model is configured to: perform global semantic parsing on the document image data, identify and extract key-value pair information of each business field; and, based on preset trade compliance domain knowledge, perform cross-field logical consistency verification on the values ​​of the extracted multiple business fields, and generate verification results carrying confidence labels. The declaration data generation module is used to merge the key-value pair information with the form data based on the verification result to generate structured declaration data.

[0007] Furthermore, the multimodal perception model is a large vision-language model, and the multimodal parsing and verification module is also used to inject the preset trade compliance domain knowledge into the inference context of the multimodal perception model through a retrieval-enhanced generation architecture.

[0008] Furthermore, the cross-field logical consistency verification performed by the multimodal perception model includes: identifying the business logic relationship between two or more field values ​​on the same document or related documents, and determining whether the business logic relationship conforms to the association rules stored in the trade compliance domain knowledge; the association rules include at least one of the following: the customary combination rules of trade country and currency, the mapping rules of commodity code and regulatory conditions, and the matching rules of transportation mode and port qualification.

[0009] Furthermore, the multimodal parsing and verification module is also configured to: when the confidence level of a specific field in the verification result is lower than a preset threshold, based on the attention weight or position coordinates output by the multimodal perception model, backtrack the local area in the document image data corresponding to the specific field for secondary recognition, and update the verification result.

[0010] Furthermore, the multimodal perception model generates verification results carrying confidence labels, including: For any validation item, let the combination of field values ​​involved be {v1,v2,…,vn}, the recognition confidence of each field be {c1,c2,…,cn}, and the expected logical relationship be Rk. Then the confidence score of the validation item is calculated according to the following formula: Conf(k)=f(c1,c2,…,cn;Rk)·g(v1,v2,…,vn;Rk); In the formula, f(·) is the identification confidence factor, which represents the joint reliability of the identification confidence of each field under the expected logical relationship Rk. Its calculation method can assign different weights to the confidence of each field according to different logical relationships; g(·) is the logical conformity factor, which represents the degree of matching between the actual value combination of the field and the expected logical relationship Rk. The confidence level label is determined based on the comparison between the confidence level score and a preset threshold.

[0011] Furthermore, it also includes: The result review interaction module is used to perform differentiated rendering of the verification items on the visualization interface based on the comparison results of the confidence score and the preset threshold, and to receive user confirmation or correction instructions for at least the low confidence field. The model evolution module is used to collect user feedback for correction after processing by the result review and interaction module, forming online training samples, and to perform small-sample iterative fine-tuning of the multimodal perception model using these online training samples; wherein the loss function used for fine-tuning is: L(θ)=L_task(θ)+λ·L_align(θ); In the formula, θ is the parameter of the multimodal perception model, L_task(θ) is the original task loss, L_align(θ) is the alignment loss, which is used to constrain the difference between the model prediction result and the user correction result, and λ is the balancing hyperparameter.

[0012] An intelligent customs declaration method for international trade, using the aforementioned intelligent customs declaration system for international trade, includes the following steps: S1: Obtain the document image data and associated form data for the batch to be declared; S2: Use a trained multimodal perception model to perform global semantic parsing on the document image data, identify and extract key-value pair information of each business field; S3: Using the multimodal perception model based on preset trade compliance domain knowledge, perform cross-field logical consistency verification on the values ​​of multiple extracted business fields, and generate verification results carrying confidence labels; S4: Based on the verification result, the key-value pair information is merged with the form data to generate structured declaration data. Furthermore, the multimodal perception model is a large vision-language model, and the preset trade compliance domain knowledge is injected into the reasoning context of the multimodal perception model in a retrieval-enhanced generation manner.

[0013] Furthermore, the generation of the verification result carrying the confidence label includes: For any validation item, let the combination of field values ​​involved be {v1,v2,…,vn}, the recognition confidence of each field be {c1,c2,…,cn}, and the expected logical relationship be Rk. Then, the confidence score of the validation item is calculated according to the formula Conf(k)=f(c1,c2,…,cn;Rk)·g(v1,v2,…,vn;Rk), where f(·) is the recognition confidence factor, which can assign different weights to the confidence of each field according to different logical relationships; g(·) is the logical compliance factor. Based on the comparison between the confidence score and the preset threshold, a corresponding confidence label is assigned to the verification item.

[0014] Furthermore, it also includes: On the visual interface, differentiated rendering is performed based on the confidence score of the verification item; Receive confirmation or correction instructions from users for at least low-confidence fields, and update the structured declaration data based on the corrected data; The corrected data of the user is used as training samples, and the multimodal perception model is fine-tuned by small sample iteration using the loss function L(θ)=L_task(θ)+λ·L_align(θ), where L_task(θ) is the original task loss, L_align(θ) is the alignment loss, and λ is the balancing hyperparameter.

[0015] Compared with existing technologies, the advantages of this invention are: 1: By performing global semantic parsing on document images through a multimodal parsing and verification module and simultaneously completing cross-field logical consistency verification, combined with confidence score calculation and hierarchical label output, the design achieves the effect of directly carrying verification conclusions in the recognition results and making the reliability quantifiable. It has the advantages of eliminating information gaps in traditional split architecture, distinguishing the source of uncertainty in recognition from contradictions in the data itself, and enabling customs brokers to prioritize handling suspicious items, thereby improving review efficiency.

[0016] 2: By adopting a visual-language large model for template-free semantic information extraction and injecting trade compliance knowledge into the verification and reasoning process in real time through retrieval-enhanced generation, the design achieves the effect of natural adaptation to changes in document format and dynamic compliance judgment of any combination of fields based on rule semantics. It has the advantages of saving a lot of template creation and maintenance work, reducing system operation and maintenance costs, and being able to deal with anomalies of unpreset field combinations.

[0017] 3: The result review interaction module collects the customs broker's correction records for low-confidence items, and the model evolution module continuously updates the design of the multimodal perception model through small-sample iterative fine-tuning. At the same time, in conjunction with the processing sequence of completing compliance risk review before declaration submission, the system achieves the effect of gradually improving adaptability as the system is used over time and intercepting hidden conflicts such as currency mismatch during the data preparation stage. This has the benefits of positive accumulation of declaration pass rate and reduction of customs rejection rate and shortening customs clearance time.

[0018] In summary, this invention simultaneously extracts document information and verifies compliance logic in a single semantic understanding process, enabling the recognition results to directly carry quantifiable and reliable verification conclusions, thus eliminating information gaps in traditional separate architectures. It also eliminates the maintenance work of creating fixed templates based on country and document type, has adaptive capabilities to format changes and unpreset anomalies, and can continuously optimize the model based on feedback from customs brokers. This effectively intercepts compliance conflicts during the declaration preparation stage, reducing rejection rates and customs clearance time. Attached Figure Description

[0019] Figure 1 This is a structural block diagram of an intelligent customs declaration system for international trade proposed in this invention; Figure 2 This is a schematic diagram of the main process of an intelligent customs declaration method for international trade proposed in this invention. Detailed Implementation

[0020] Reference Figures 1 to 2 An intelligent customs declaration system for international trade includes a data acquisition module, a multimodal parsing and verification module, a declaration data generation module, a result review and interaction module, and a model evolution module, which are used to realize closed-loop processing from document input to compliant declaration data output.

[0021] The data acquisition module is used to acquire the image data of the documents and related form data for the batch to be declared, specifically including: Front-end data collection unit: Provides a human-computer interaction interface for customs brokers to fill in form data including company name, unified credit code, competent customs, port of entry and exit, commodity information, means of transport information, country of origin and destination, consignor and consignee information, country of trade, trade method, transaction amount, and currency of transaction, and upload scanned copies or photos of business documents; System Interface Unit: Provides an application programming interface to receive form data structures and document encoding streams pushed by the Enterprise Resource Planning system according to agreed specifications; The data acquisition module assigns a unique task number to each batch of received application materials, stores the form data in the original form database, stores the document image data in the original document database, and sends a processing readiness notification to the multimodal parsing and verification module.

[0022] The multimodal parsing and verification module is used to load the trained multimodal perception model, which is specifically a large vision-language model that can perform global semantic parsing and cross-field logical verification on document images. The module has a knowledge retrieval unit that injects pre-defined trade compliance knowledge into the inference context of the multimodal perception model through a retrieval-enhanced generation architecture.

[0023] The knowledge base in the field of trade compliance stores at least one of the following related rules: the usual combination rules of trade country and currency, the mapping rules of commodity code and regulatory conditions, and the matching rules of transportation mode and port qualifications.

[0024] The workflow of the multimodal parsing and verification module includes: In the first stage, global semantic parsing: the multimodal perception model receives the entire document image, extracts image features through a visual encoder, and generates structured key-value pairs through a language decoder. During this process, the model not only transcribes the text but also maps it to standard field names and values ​​based on contextual semantics. For example, for "NetWeight: 15,000KGS" on an invoice, the output key-value pair is {"Field Name": "Net Weight", "Value": "15000", "Unit": "Kilogram"}; for "CountryofOrigin:Germany" on a certificate of origin, it is mapped to {"Field Name": "Country of Origin", "Value": "DE", "Original Statement": "Germany"}. Each extracted field value vi is accompanied by a recognition confidence score ci, which is derived from the visual region sharpness, text segmentation certainty, and decoding probability distribution, and takes a value between 0 and 1.

[0025] The second stage is cross-field logical consistency verification: the model combines the values ​​of multiple fields involved in the current declaration task and performs consistency inference based on the injected association rules; Taking certificates of origin and commercial invoices as examples, the model identifies the country of origin as "DE" (Germany) and the currency of transaction as "VND" (Vietnamese Dong). According to the retrieved rule "when the country of origin is an EU member state, the currency of transaction is usually EUR", it is determined that the combination {country of origin = DE, currency of transaction = VND} is contrary to the usual rule.

[0026] The third stage involves confidence score calculation and label generation: For any validation item k, let its associated field value combination be {v1,v2,…,vn}, the confidence score of each field be {c1,c2,…,cn}, and the expected logical relationship be Rk. Then, the confidence score Conf(k) of the validation item is calculated according to the following formula: Conf(k)=f(c1,c2,…,cn;Rk)·g(v1,v2,…,vn;Rk); In the formula, f(·) is the identification confidence factor, which represents the joint reliability of the identification confidence of each field under the logical relation Rk.

[0027] The confidence factor f(·) is calculated as follows: In the preset field weight configuration table, the confidence weight wi corresponding to each field is queried according to the type of expected logical relationship Rk, and then the weight is calculated by weighting according to the following formula: f=Σ(wi·ci) / Σwi; The field weight configuration table stores the correspondence between the association rule types and the weights of each field. For example, for the "trading country and currency customary combination" validation item, the weight of the country of origin field is set to 0.6, and the weight of the transaction currency field is set to 0.4. For the "commodity code and regulatory condition mapping" verification item, the weight of the commodity code field is set to 0.7, and the weight of the regulatory condition field is set to 0.3. In some other embodiments, to enhance the sensitivity to low confidence levels of a single field, a calculation form with an exponential factor can be used: f=[Σ(wi·ci^p) / Σwi]^(1 / p), where p is an adjustment factor greater than 1, so that the field with the lowest confidence level has a greater impact on the overall result.

[0028] The logical compliance factor g(·) is calculated by inputting the actual combination of field values ​​into the preset logical consistency quantification scoring rules to obtain the logical compliance factor.

[0029] The scoring rules are stored in a trade compliance knowledge base in the form of structured data tables, specifically including three types of mappings: The first category consists of standard combination templates and their corresponding ratings. These templates store a list of commonly used currencies for each trading country. For example, if the country of origin is Germany, the list of commonly used currencies is {EUR, USD}. Any currency combination with Germany within this list is rated 1.0. Acceptable alternative currencies not on the list (such as CHF) are rated 0.6 to 0.8. Rare currencies not on the list and with which there is no trade (such as VND) are rated 0.1 to 0.3. This list is compiled from historical international trade data and customs regulations, and is stored in a quadruple format: {Country of Origin Code, List of Commonly Used Currencies, List of Acceptable Currencies, List of Rare Currencies}.

[0030] The second category is a blacklist of contradictory combinations and their corresponding scores. The blacklist stores field combinations explicitly deemed non-compliant by customs, such as: {Transportation mode = International road transport, Import port type = Non-land designated port for complete vehicles} rated 0; {Commodity HS code = Commodity subject to import license management, License number = Null value} rated 0. Combinations in the blacklist directly lock the g-value to 0, blocking subsequent submissions.

[0031] The third category is the tolerance scoring function for continuous variables. For the validation of continuous numerical fields such as invoice amount and contract amount, g(·) is calculated based on the relative deviation between the two, δ=|v1-v2| / max(v1,v2): when δ≤5%, g takes 1.0; when 5%<δ≤10%, g takes 0.6 to 0.8; when δ>10%, g takes less than 0.2.

[0032] Taking the validation item "Country of Origin = DE, Currency of Transaction = VND" as an example: The model identifies the confidence level of the Country of Origin field as 0.98 and the confidence level of the Currency field as 0.95; the field weight configuration table shows that the weights of the two fields under this logical relationship type are 0.6 and 0.4 respectively, so f = (0.6 × 0.98 + 0.4 × 0.95) / (0.6 + 0.4) = 0.968; the logical consistency quantification scoring rules show that the {DE, VND} combination is a rare currency combination, with a corresponding g score of 0.2; so Conf(k) = 0.968 × 0.2 = 0.1936, which is lower than the second threshold of 0.40, and is marked as low confidence.

[0033] After calculating the confidence score, it is compared with a preset threshold: a score greater than or equal to the first threshold (e.g., 0.85) is marked as "high confidence," a score between the first and second thresholds (e.g., 0.40) is marked as "medium confidence," and a score below the second threshold is marked as "low confidence." For example, if the g value of {country of origin = DE, currency = VND} is 0.2, and f is min(0.98, 0.95) = 0.95, then Conf(k) = 0.19, which is lower than the second threshold and is marked as low confidence, with the anomaly noted as "a rare combination of country of origin and currency of transaction."

[0034] The fourth stage is the secondary recognition mechanism: when the confidence score of a specific field is lower than the second threshold, the multimodal parsing and verification module determines the local region of the field in the original document image based on the attention weight or position coordinates output by the multimodal perception model during the parsing process; and crops out the local high-resolution image of the region from the original image. The data is then re-entered into the multimodal perception model for focused recognition. The secondary recognition result is compared with the original result. If they are inconsistent, the field value is updated with the secondary result, and the confidence score of the associated verification item is recalculated. For example, the original invoice amount is identified as "15000". Because it does not match the contract amount "10500", a secondary recognition is triggered. After cropping a portion of the image of the amount area, the recognition result is "10500". After the amount is updated and matched, the g value becomes 1.0, and the score of the relevant verification item automatically rises to the high confidence interval.

[0035] The declaration data generation module, based on the complete field set with confidence labels output by the multimodal parsing and verification module, merges key-value pair information with the original form data according to the international trade "single window" standard data format to generate structured declaration data. High-confidence fields are used directly, while medium- and low-confidence fields retain their current optimal values ​​and are embedded with an "audit flag" attribute for subsequent differentiation and processing. After assembly, the data is stored in the standard declaration database.

[0036] The results review interaction module provides a visual interface, which differentiates the rendering of the verification results based on the confidence score: high confidence items are displayed in a collapsed green background, medium confidence items are displayed in a yellow background with identification screenshots and reasoning explanations, and low confidence items are displayed in a red background with an error message. Customs brokers can click on low confidence fields to bring up partial screenshots of documents and editing boxes for confirmation or direct correction. After the correction instruction is submitted, the correction value is written back to the declaration data generation module to update the structured declaration data, and the complete correction record is saved.

[0037] The model evolution module collects a certain number (e.g., 50) or accumulated user feedback at a fixed period (e.g., weekly) to form online training samples, which are then used for small-sample iterative fine-tuning of the multimodal perception model. The loss function used for fine-tuning is: L(θ)=L_task(θ)+λ·L_align(θ); In the formula, θ represents the parameters of the multimodal perception model; L_task(θ) is the original task loss, which uses the cross-entropy loss of the vision-language model to constrain the accuracy of field extraction and verification judgment; L_align(θ) is the alignment loss, used to constrain the difference between model prediction and user correction, and can use the divergence loss between the predicted distribution and the user-corrected label; λ is the balancing hyperparameter, initially set to 0.1, and can be adjusted between 0.01 and 1.0 according to the actual effect. After fine-tuning, the results are evaluated on the validation set. If the metrics show no degradation, the system is hot-updated and deployed to the multimodal parsing and verification module.

[0038] The aforementioned system corresponds to an intelligent customs declaration method for international trade, which includes the following steps: Step 1: Obtain the document image data and associated form data for the batch to be declared; Step 2: Use the visual-language big data model to perform global semantic parsing on the document image data, extract business field key-value pairs and determine the recognition confidence. Step 3: By enhancing the generation of knowledge in the field of trade compliance through retrieval, cross-field logical consistency verification is performed on the values ​​of multiple fields. The confidence score formula mentioned above is used to calculate the score of the verification item, and confidence labels are generated based on the comparison of the score with the threshold. Step 4: For low-confidence fields, perform secondary recognition on the local area of ​​the document based on the attention weights or location coordinates output by the model, and update the verification results. Step 5: Based on the verification results, generate structured declaration data that conforms to the "single window" standard; Step 6: Display the differences in the visualization interface, receive user corrections for low-confidence fields, and update the declaration data; Step 7: Collect user feedback and fine-tune the multimodal perception model using the loss function described above with small sample iterations to continuously optimize model performance.

[0039] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. An intelligent customs declaration system for international trade, characterized in that, include: The data acquisition module is used to acquire the document image data and associated form data of the batch to be declared; The multimodal parsing and verification module is used to load a trained multimodal perception model, which is configured to: perform global semantic parsing on the document image data, identify and extract key-value pair information of each business field; In addition, based on pre-defined knowledge of trade compliance, the values ​​of multiple extracted business fields are subjected to cross-field logical consistency verification, generating verification results with confidence labels; The declaration data generation module is used to merge the key-value pair information with the form data based on the verification result to generate structured declaration data.

2. The intelligent customs declaration system for international trade according to claim 1, characterized in that, The multimodal perception model is a large vision-language model. The multimodal parsing and verification module is also used to inject the preset trade compliance domain knowledge into the inference context of the multimodal perception model through a retrieval-enhanced generation architecture.

3. The intelligent customs declaration system for international trade according to claim 1, characterized in that, The cross-field logical consistency verification performed by the multimodal perception model includes: identifying the business logic relationship between two or more field values ​​on the same document or related documents, and determining whether the business logic relationship conforms to the association rules stored in the trade compliance domain knowledge; the association rules include at least one of the following: the customary combination rules of trade country and currency, the mapping rules of commodity code and regulatory conditions, and the matching rules of transportation mode and port qualification.

4. The intelligent customs declaration system for international trade according to claim 1, characterized in that, The multimodal parsing and verification module is further configured to: when the confidence level of a specific field in the verification result is lower than a preset threshold, based on the attention weight or position coordinates output by the multimodal perception model, backtrack the local area in the document image data corresponding to the specific field for secondary recognition, and update the verification result.

5. An intelligent customs declaration system for international trade according to any one of claims 1-4, characterized in that, The multimodal perception model generates verification results carrying confidence labels, including: For any validation item, let the combination of field values ​​involved be {v1,v2,…,vn}, the recognition confidence of each field be {c1,c2,…,cn}, and the expected logical relationship be Rk. Then the confidence score of the validation item is calculated according to the following formula: Conf(k)=f(c1,c2,…,cn;Rk)·g(v1,v2,…,vn;Rk); In the formula, f(·) is the identification confidence factor, which represents the joint reliability of the identification confidence of each field under the expected logical relationship Rk. Its calculation method can assign different weights to the confidence of each field according to different logical relationships; g(·) is the logical conformity factor, which represents the degree of matching between the actual value combination of the field and the expected logical relationship Rk. The confidence level label is determined based on the comparison between the confidence level score and a preset threshold.

6. The intelligent customs declaration system for international trade according to claim 5, characterized in that, Also includes: The result review interaction module is used to perform differentiated rendering of the verification items on the visualization interface based on the comparison results of the confidence score and the preset threshold, and to receive user confirmation or correction instructions for at least the low confidence field. The model evolution module is used to collect user feedback for correction after processing by the result review and interaction module, forming online training samples, and to perform small-sample iterative fine-tuning of the multimodal perception model using these online training samples; wherein the loss function used for fine-tuning is: L(θ)=L_task(θ)+λ·L_align(θ); In the formula, θ is the parameter of the multimodal perception model, L_task(θ) is the original task loss, L_align(θ) is the alignment loss, which is used to constrain the difference between the model prediction result and the user correction result, and λ is the balancing hyperparameter.

7. An intelligent customs declaration method for international trade, using the intelligent customs declaration system for international trade as described in any one of claims 1-6, characterized in that, Includes the following steps: S1: Obtain the document image data and associated form data for the batch to be declared; S2: Use a trained multimodal perception model to perform global semantic parsing on the document image data, identify and extract key-value pair information of each business field; S3: Using the multimodal perception model based on preset trade compliance domain knowledge, perform cross-field logical consistency verification on the values ​​of multiple extracted business fields, and generate verification results carrying confidence labels; S4: Based on the verification result, the key-value pair information is merged with the form data to generate structured declaration data.

8. The intelligent customs declaration method for international trade according to claim 7, characterized in that, The multimodal perception model is a large vision-language model, and the preset trade compliance domain knowledge is injected into the reasoning context of the multimodal perception model in a retrieval-enhanced generation manner.

9. The intelligent customs declaration method for international trade according to claim 7, characterized in that, The generation of verification results carrying confidence labels includes: For any validation item, let the combination of field values ​​involved be {v1,v2,…,vn}, the recognition confidence of each field be {c1,c2,…,cn}, and the expected logical relationship be Rk. Then, the confidence score of the validation item is calculated according to the formula Conf(k)=f(c1,c2,…,cn;Rk)·g(v1,v2,…,vn;Rk), where f(·) is the recognition confidence factor, which can assign different weights to the confidence of each field according to different logical relationships; g(·) is the logical compliance factor. Based on the comparison between the confidence score and the preset threshold, a corresponding confidence label is assigned to the verification item.

10. The intelligent customs declaration method for international trade according to claim 9, characterized in that, Also includes: On the visual interface, differentiated rendering is performed based on the confidence score of the verification item; Receive confirmation or correction instructions from users for at least low-confidence fields, and update the structured declaration data based on the corrected data; The corrected data of the user is used as training samples, and the multimodal perception model is fine-tuned by small sample iteration using the loss function L(θ)=L_task(θ)+λ·L_align(θ), where L_task(θ) is the original task loss, L_align(θ) is the alignment loss, and λ is the balancing hyperparameter.

Images