Scoring method of bank receipt recognition system

Abstract

The invention discloses a scoring method of a bank receipt recognition system. The scoring method comprises the following steps: S1, collecting and preprocessing receipt data; s2, based on the receipt template library, automatically positioning a key information area on the receipt, determining core fields needing to be evaluated, and for each core field, constructing the following evaluation features: an accuracy feature, an integrity feature, a definition correlation feature and a format normalization feature; s3, historical receipt recognition cases are collected, an artificial verification result and corresponding feature data of each case are marked, a training sample set is formed, and a machine learning algorithm is combined with a natural language processing model to train a scoring model; s4, for a field set which should be extracted by the receipt, calculating a score according to a ratio of the number of fields which should be extracted to the number of actually extracted fields; and S5, generating an evaluation report containing the single-field score, the comprehensive score and the score deduction reason, wherein the evaluation report supports charting display.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of finance, in particular to a scoring method of a bank receipt identification system. BACKGROUND

[0002] With the rapid development of financial technology, bank receipts, as important evidence of fund transactions, play a key role in enterprise financial reconciliation, audit verification, credit approval, and other scenarios. To improve the efficiency of processing receipts, various bank receipt identification systems have emerged. These systems use optical character recognition (OCR), deep learning, and other technologies to automatically extract key information from receipts, such as transaction amounts, transaction times, transaction account numbers, and summary descriptions, supporting the digital transformation of financial automation, supply chain financial risk control, and other businesses.

[0003] However, the performance evaluation of existing bank receipt identification systems still has significant pain points:

[0004] On the one hand, receipt styles are complex and diverse, with differences in format, font, and seal position between different banks. The same bank's receipts may also have varying image quality due to differences in printing equipment and paper quality (such as blurring, wrinkling, and stains), which can cause errors in information extraction (such as amount recognition bias, account number omission, and summary semantic misunderstanding) in actual application. However, there is currently a lack of quantitative evaluation mechanism for recognition results, making it difficult to intuitively determine the reliability of individual recognition results, resulting in inefficient manual review of all recognition content.

[0005] On the other hand, different recognition systems have different technical routes (such as rule matching based on traditional OCR and feature learning based on small sample models), and their recognition capabilities in specific scenarios (such as fault tolerance for blurred receipts and semantic understanding accuracy for complex summaries) vary. However, there is no uniform scoring standard in the industry, making it impossible to objectively compare the comprehensive performance of different systems, leading to a lack of data support for enterprises when selecting recognition systems and difficulty in matching their business needs for recognition accuracy.

[0006] In addition, in scenarios with high accuracy requirements such as financial risk control and auditing, errors in recognition results can cause financial risks (such as financial losses due to reconciliation errors) or compliance risks (such as misjudgment of audit evidence). In existing technologies, a binary "correct / incorrect" judgment cannot reflect the risk level of recognition results, making it difficult to meet the differentiated needs of different business scenarios for recognition reliability.

[0007] Therefore, constructing a mechanism capable of quantifying the recognition result of the bank return slip recognition system, realizing reliable evaluation of a single recognition result, performance comparison of different systems, and risk level division, becomes a key requirement for improving return slip processing efficiency and reducing business risk. SUMMARY

[0008] The present application provides a bank return slip recognition system scoring method to solve the above technical problems existing in the prior art.

[0009] To achieve the above purpose, the present application provides a bank return slip recognition system scoring method, which comprises:

[0010] S1: return slip data acquisition and preprocessing: collecting bank return slip image data to be evaluated, and synchronously acquiring structured recognition results output by the return slip recognition system;

[0011] S2: based on the return slip template library, automatically positioning the key information area on the return slip, the key information area including the transaction amount column, the payer account column and the seal area, determining the core fields to be evaluated, and for each core field, the following evaluation features are constructed respectively: accuracy feature, integrity feature, clarity correlation feature and format specification feature;

[0012] S3: collect historical return slip recognition cases, label the artificial verification results and corresponding feature data of each case to form a training sample set,

[0013] Using machine learning algorithm combined with natural language processing model, training scoring model:

[0014] For numerical fields, the score is quantified by calculating the deviation rate of the recognition value and the standard value: |recognition value-standard value| / standard value,

[0015] For character fields, the matching score is calculated by edit distance,

[0016] For non-text information, the seal integrity and clarity are evaluated by the image recognition model to convert into corresponding scores,

[0017] Based on the importance of different fields in the business scenario, the weight coefficient is configured by using the analytic hierarchy process or business rules to ensure that the scoring result meets the actual demand;

[0018] S4: for the field set to be extracted from the return slip, the score is calculated according to the ratio of the number of fields to be extracted to the number of actually extracted fields,

[0019] If all the fields to be extracted are successfully extracted and the corresponding values are accurate, the return slip score is 1; if any of the fields to be extracted is not extracted or the content of the extracted field is incorrect, the return slip score is 0;

[0020] S5: generating an evaluation report containing the single field score, the comprehensive score and the deduction reason, and the evaluation report supports graphical display.

[0021] In an optional embodiment of the present application, in step S1, the bank return slip image data includes a paper return slip scan and an electronic return slip screenshot, and the structured recognition result includes the transaction amount, the account number and the abstract.

[0022] In an optional embodiment of the present application, in step S1, further comprising data cleaning and standardization:

[0023] The collected return slip image is subjected to noise reduction, tilt correction and edge cropping to remove image interference information.

[0024] The recognition result is subjected to format unification.

[0025] In an optional embodiment of the present application, in step S2, the return slip template library covers the mainstream bank return slip format.

[0026] In an optional embodiment of the present application, in step S2, each evaluation feature is evaluated according to the following requirements:

[0027] Accuracy feature: evaluating the matching degree of the recognition result and the standard result manually labeled, including character matching rate and semantic consistency;

[0028] Completeness feature: evaluating whether the key field is missing, including whether the last 4 digits of the recognized account number are complete;

[0029] Clarity correlation feature: evaluating the correlation coefficient of the clarity of the return slip image and the stability of the recognition result, the clarity of the return slip image including resolution and blurriness;

[0030] Format specification feature: evaluating whether the recognition result conforms to the field format rule, including whether the amount contains the currency symbol and whether the account number is compliant.

[0031] In an optional embodiment of the present application, in step S3, the manual verification result includes correct, partially incorrect and completely incorrect.

[0032] In an optional embodiment of the present application, in step S3, the machine learning algorithm includes random forest and / or gradient boosting tree.

[0033] In an optional embodiment of the present application, in step S4, the total number of fields that must be extracted from the bank receipt is determined according to the type of the bank receipt and the business rules, for the number of fields to be extracted,

[0034] The actual number of extracted fields is the number of valid fields that are successfully extracted, that is, the fields are not only identified by name, but also the corresponding values are verified as valid information.

[0035] In an optional embodiment of the present application, the scoring method of the bank receipt recognition system further comprises step S6:

[0036] Feedback data collection: record the correction results after manual review, establish a closed-loop data chain of "recognition result - score - correction result",

[0037] Model dynamic updating: periodically retrain the scoring model with new data, adjust the field weight and scoring threshold, and ensure that the scoring system is suitable for changes in bank receipt styles and upgrades in recognition technology.

[0038] The scoring method of the bank receipt recognition system provided by the present application has the following beneficial technical effects:

[0039] 1. Improve the efficiency of bank receipt processing and reduce labor costs

[0040] Through the extraction completeness of key information by the field scoring quantification recognition system, high completeness recognition results (such as field scores close to 1.00) can be quickly screened out. Such results do not need to be manually reviewed field by field, and directly enter the subsequent business process (such as financial accounting and reconciliation). Low completeness results only need to be checked for missing or incorrect fields, greatly reducing invalid manual operations, significantly improving the efficiency of bank receipt processing, and reducing labor costs.

[0041] 2. Clear recognition quality standards, optimize system selection and iteration

[0042] The bank receipt score (1 point or 0 point) provides a rigid judgment standard for the overall validity of the recognition result, which facilitates enterprises or financial institutions to objectively compare the performance of different bank receipt recognition systems — by statistics of the proportion of "1 point bank receipt" output by different systems, the system adaptability can be accurately evaluated, providing data support for selection; at the same time, the field score can locate the weak links of the recognition system (such as low extraction rate of certain fields), indicating the direction for system iteration and optimization, and promoting continuous upgrading of recognition technology.

[0043] 3. Strengthen business risk control and ensure data reliability

[0044] The full-field matching return slip scoring mechanism ensures that the return slip information entering the core business process (such as audit, credit approval) is complete and accurate, avoiding financial risks (such as account reconciliation deviation) and compliance risks (such as invalid audit evidence) caused by missing or incorrect key fields; and the field scoring can early warning identify vulnerabilities, help business personnel prevent operational risks caused by incomplete information, and improve data reliability.

[0045] 4. Adapt to various business scenarios and enhance system flexibility

[0046] Based on the dynamic definition of "fields to be extracted" in the return slip template library, the scoring system can adapt to the format differences of different banks and different types of return slips (such as transfer return slips and collection return slips), and meet the personalized needs of enterprise finance, bank public business, and supply chain finance in multiple scenarios; at the same time, the linkage design of scoring results and business processes (such as automatic triggering of review or entry) further enhances the landing and flexibility of the system in actual business.

[0047] 5. Promote digital closed loop of return slip processing and help intelligent transformation

[0048] Through the closed loop feedback of scoring results and manual review data, the system continuously optimizes the scoring model, forming a digital link of "identification - scoring - correction - iteration", not only improving the processing capacity of single link, but also promoting the intelligentization of the whole process of return slip from collection, identification to application, providing key technical support for the automation and intelligent transformation in the field of financial technology. BRIEF DESCRIPTION OF DRAWINGS

[0049] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of these drawings.

[0050] Figure 1 The flowchart of the scoring method of the bank return slip identification system of an embodiment of the present application;

[0051] Figure 2 A bank return slip. DETAILED DESCRIPTION

[0052] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0053] The key point of the present application is to construct a set of standardized and quantitative scoring mechanism for the bank return single recognition system. The core innovation and key technical points are as follows:

[0054] 1. Double-layer scoring system design

[0055] The score is divided into two levels of "field score" and "return single score":

[0056] The field score takes the ratio of "number of fields that should be extracted" to "number of fields actually extracted" as the core, and accurately quantifies the extraction integrity of key information (such as bank name, payee information, etc.) by the recognition system;

[0057] The return single score takes "full field matching" as the judgment standard (1 point or 0 point), directly reflects the overall effectiveness of the return single recognition result, and meets the rigid demand of "full match is effective" in financial, auditing and other scenarios.

[0058] 2. Dynamic adaptive field definition mechanism

[0059] Based on the return single template library (covering different banks and different types of return single formats), the "set of fields that should be extracted" can be dynamically determined according to the type of return single and business rules, solving the problem of difficult scoring standard solidification caused by the variety of bank return single styles and the non-uniformity of key information, and ensuring the adaptability of the score.

[0060] 3. Scoring and business scenario linkage capability

[0061] The scoring result is not only a simple numerical output, but also can directly drive the business process (such as "high integrity field result automatically skipping review" and "full field matching result directly entering account"), realizing the closed loop from "scoring evaluation" to "efficiency improvement", which is different from the traditional scoring tool that only judges the result.

[0062] 4. Iterative scoring optimization mechanism

[0063] By collecting the correction data of manual review, continuously updating the sample set and optimizing the scoring model, the system can adapt to dynamic scenarios such as changes in return single format and recognition technology upgrade, ensuring the long-term accuracy and applicability of the score.

[0064] These key points collectively solve the core pain points in the existing bank return single recognition field, such as "non-uniform evaluation standard, disconnection between result and business, and difficulty in adapting to various scenarios", providing a practical technical solution for performance evaluation and business process optimization of the recognition system.

[0065] Figure 1 The flowchart of the scoring method of the bank return single recognition system of an embodiment of the present application is as follows:Figure 1 As shown, the scoring method of the bank statement recognition system provided by the present invention includes:

[0066] S1: Bank receipt data acquisition and preprocessing: Acquire image data of bank receipts to be evaluated, and simultaneously obtain the structured recognition results output by the receipt recognition system;

[0067] In this embodiment, the bank receipt image data includes, for example, scanned copies of paper receipts and screenshots of electronic receipts. The structured recognition results include the transaction amount, account number, and summary. Scanned copies of paper receipts can be collected using a scanning device, while screenshots of electronic receipts can be transmitted via an API interface.

[0068] This embodiment further includes data cleaning and standardization:

[0069] The collected return receipt images are subjected to noise reduction, tilt correction, and edge cropping to remove image interference information;

[0070] Standardize the format of the identification results (e.g., use "YYYY-MM-DD" format for dates and retain two decimal places for amounts) to ensure consistency in subsequent evaluation data.

[0071] S2: Based on the receipt template library, automatically locate the key information areas on the receipt. The key information areas include the transaction amount field, the payer account field, and the seal area. Determine the core fields to be evaluated, and construct the following evaluation features for each core field: accuracy feature, completeness feature, clarity association feature, and format standardization feature.

[0072] Step S2 is used to extract key information and construct features. The receipt template library covers the mainstream bank receipt formats.

[0073] In this embodiment, each evaluation feature is evaluated according to the following requirements:

[0074] Accuracy features: Evaluate the degree of matching between the recognition results and the standard results of manual annotation, including character matching rate and semantic consistency;

[0075] Completeness features: Assess whether key fields are missing, including whether the last four digits of the account number are complete;

[0076] Clarity-related features: The correlation coefficient between the clarity of the receipt image and the stability of the recognition result is evaluated. The clarity of the receipt image includes resolution and blur.

[0077] Format compliance features: Evaluate whether the recognition results conform to the field format rules, including whether the amount contains currency symbols and whether the number of digits in the account is compliant.

[0078] S3: Collect historical receipt recognition cases, label the manual verification results of each case (manual verification results include correct, partially incorrect, and completely incorrect) and corresponding feature data, forming a training sample set.

[0079] The scoring model is trained using machine learning algorithms (random forest and / or gradient boosting tree) combined with a natural language processing model.

[0080] For numerical fields, the score is quantified by calculating the deviation rate between the identified value and the standard value: |identified value - standard value| / standard value.

[0081] For character fields, a matching score is calculated using edit distance (a metric that measures string differences).

[0082] For non-textual information, an image recognition model is used to evaluate the integrity and clarity of the seal, converting it into a corresponding score.

[0083] Based on the importance of different fields in business scenarios (such as the transaction amount having a higher weight than the summary in financial reconciliation), the weight coefficients are configured using the Analytic Hierarchy Process (AHP) or business rules to ensure that the scoring results match actual needs.

[0084] S4: For the set of fields to be extracted from the receipt, calculate the score based on the ratio of the number of fields to be extracted to the number of fields actually extracted.

[0085] First, clarify the "number of fields to be extracted". Based on the receipt type (such as transfer receipt or payment receipt) and business rules, determine the total number of fields (key values) that must be extracted from the receipt. For example, the fields to be extracted from a bank transfer receipt include "Bank Name", "Payee Name", "Payee Account", "Payer Name", "Payer Account", "Transaction Amount", and "Transaction Time", totaling 7 fields. Therefore, the number of fields to be extracted is 7.

[0086] The actual number of fields extracted is the number of valid fields (key-value pairs) successfully extracted. This means that not only are the field names identified, but their corresponding values ​​are also verified as valid information (e.g., "Payee Name" is identified and the content is accurate). For example, if 6 out of the 7 fields to be extracted are successfully extracted, then the actual number of fields extracted is 6. Field score = Actual number of fields extracted / Number of fields to be extracted (the result is rounded to two decimal places, with a maximum of 1.00 and a minimum of 0).

[0087] The judgment criterion is to match all fields. If all fields that should be extracted are successfully extracted and the corresponding values ​​are accurate (i.e., the name and content of each field are consistent with the standard result), the order score is 1 point. If any field that should be extracted is not extracted, or the content of the extracted field is incorrect, the order score is 0 points.

[0088] S5: Generates an evaluation report that includes individual field scores, overall scores, and reasons for deductions (e.g., "10 points deducted for semantic deviation in the abstract"). The evaluation report supports graphical display (e.g., radar charts of scores for each field), making it easier for users to intuitively understand the reliability of the recognition results.

[0089] Business scenario linkage:

[0090] For results rated "Excellent", manual review is automatically skipped, and the data is directly entered into the financial system to complete the accounting.

[0091] For results rated "Medium" or "Poor", mark high-risk fields and push them to the manual review interface, highlighting key areas for verification;

[0092] By accumulating scoring data from different recognition systems, a performance comparison report is generated to provide a basis for enterprises to select recognition tools.

[0093] In this embodiment, the scoring method of the bank receipt recognition system further includes step S6:

[0094] Feedback data collection: Record the correction results after manual review, and establish a closed-loop data chain of "identification result - scoring - correction result".

[0095] Dynamic model updates: The scoring model is retrained regularly with new data, and field weights and scoring thresholds are adjusted (e.g., for a bank's new receipt format, the weights of corresponding fields are increased) to ensure that the scoring system adapts to changes in receipt styles and upgrades in recognition technology.

[0096] Case 1: Extracting specified information from a receipt

[0097] like Figure 2 The image shown is a bank statement. Suggested answer:

[0098] {

[0099] Bank Name: Industrial and Commercial Bank of China

[0100] "Transaction Date": "2025-01-07",

[0101] "Serial Number": "3396****"

[0102] Payee Name: Anhui west** Watch Co., Ltd.

[0103] "Payer's Account": "1313202009200******",

[0104] Payee Name: Suzhou **** Technology Co., Ltd.

[0105] "Payer's Account": "512915820******",

[0106] Amount: 2679.00

[0107] Purpose: "",

[0108] Abstract: "Cross-line",

[0109] Postscript:"",

[0110] "Remark":""

[0111] }

[0112] The model answers as follows:

[0113] {

[0114] Bank Name: Industrial and Commercial Bank of China

[0115] "Receipt Type": "Payment"

[0116] "Transaction Date": "2025-01-07",

[0117] "Serial Number": "3396****"

[0118] Payee Name: Suzhou **** Technology Co., Ltd.

[0119] "Payee Account": "512915820******",

[0120] Payer's Name: Anhui west** Watch Co., Ltd.

[0121] "Payer's Account": "1313202009200******",

[0122] Amount: 2679.00

[0123] Purpose: "Cross-line"

[0124] Postscript: Payment Transaction Number: 72111845; Message Type: IBP101 Online Banking Credit Transaction Message; Entrustment Date: 2025-01-07; Transaction Type: Remittance.

[0125] }

[0126] Evaluation process:

[0127] I. Data Collection and Preprocessing of Return Receipts (Step 1)

[0128] Data collection: Obtain the receipt image (electronic receipt screenshot with desensitized characters) through a scanning device, and synchronously access the structured results output by the recognition system (such as "Bank name: Industrial and Commercial Bank of China", "Amount: 2679.00");

[0129] Data cleaning: Denoise the image (remove edge noise), correct the tilt (ensure the text is horizontal), and crop redundant areas (exclude content unrelated to "Important Notice");

[0130] Data standardization: Unify "Timestamp 2025-01-07-15.06.07.733773" and "Accounting date January 7, 2025" into "2025-01-07", convert "¥2,679.00 yuan" to "2679.00", and retain the format of desensitized fields (such as "3396****").

[0131] II. Key Information Extraction and Feature Construction (Step 2)

[0132] Field positioning: Rely on the receipt template library (adapted to ICBC transfer receipts) to locate the positions of 6 fields - the bank name is centered at the top of the receipt, the receipt type is judged through the "Payer / Payee" column, the transaction date is taken from the "Entrustment date" in the postscript, the serial number is marked at the "Transaction serial number" in the middle, the amount is in a prominent position in the middle, and the summary is below the amount;

[0133] Feature verification: Verify accuracy (such as the bank name is exactly the same as the manually marked "Industrial and Commercial Bank of China"), integrity (no field is missing), and format compliance (the date conforms to "YYYY-MM-DD", and the desensitized format of the serial number is compliant).

[0134] III. Scoring Model Training and Weight Allocation (Step 3)

[0135] Model quantization: For numerical fields (amount), calculate the deviation rate =|2679.00 - 2679.00| / 2679.00 = 0, getting 100% score; for character fields (bank name, summary, etc.), calculate the edit distance (such as the edit distance between the summary "Inter-bank transfer" and the standard value = 0), all getting 100% score;

[0136] IV. Calculation of Recognition Result Score (Step 4)

[0137] Confirm that all 6 fields meet the requirements of "field name is recognized + corresponding value is verified as valid", and count them into the "actual number of extracted fields"; on the contrary, incorrect fields (such as the payee's account name confusing the payer / payee) are not counted because the feature verification in Step 2 fails and the quantization score in Step 3 does not meet the standard.

[0138] Therefore, the final field accuracy is 6 / 12.

[0139] The scoring method of the bank receipt recognition system provided by this invention has the following beneficial technical effects:

[0140] 1. Improve the efficiency of receipt processing and reduce labor costs.

[0141] By using a field scoring quantification system to identify the completeness of key information, highly complete identification results (such as field scores close to 1.00) can be quickly filtered out. These results do not require manual field-by-field verification and can be directly entered into subsequent business processes (such as financial accounting and reconciliation). Meanwhile, results with low completeness only require key verification of missing or incorrect fields, which greatly reduces unnecessary manual operations, significantly improves the efficiency of return processing, and reduces labor costs.

[0142] 2. Clearly identify quality standards and optimize system selection and iteration.

[0143] The receipt score (1 or 0 points) provides a rigid standard for judging the overall validity of the recognition results, making it easy for enterprises or financial institutions to objectively compare the performance of different receipt recognition systems. By statistically analyzing the proportion of "1-point receipts" output by different systems, the system's adaptability can be accurately assessed, providing data support for selection. At the same time, field scores can pinpoint the weak links of the recognition system (such as low extraction rates of certain fields), pointing the way for system iteration and optimization, and promoting continuous upgrades in recognition technology.

[0144] 3. Strengthen business risk control and ensure data reliability

[0145] The full-field matching receipt scoring mechanism ensures that receipt information entering core business processes (such as auditing and credit approval) is complete and accurate, avoiding financial risks (such as reconciliation discrepancies) and compliance risks (such as invalid audit evidence) caused by missing or incorrect key fields; while field scoring can provide early warnings to identify vulnerabilities, helping business personnel to take targeted precautions against operational risks caused by incomplete information and improve data reliability.

[0146] 4. Adapt to diverse business scenarios and enhance system flexibility

[0147] Based on the dynamic definition of "fields to be extracted" from the receipt template library, the scoring system can adapt to the format differences of different banks and different types of receipts (such as transfer receipts and payment receipts), meeting the personalized needs of multiple scenarios such as corporate finance, corporate banking, and supply chain finance. At the same time, the linkage design between the scoring results and business processes (such as automatic triggering of review or accounting) further enhances the system's practicality and flexibility in actual business.

[0148] 5. Promote the digital closed loop of receipt processing to facilitate intelligent transformation.

[0149] This system continuously optimizes the scoring model through closed-loop feedback between scoring results and manually reviewed data, forming a digital link of "identification-scoring-correction-iteration". This not only improves the processing capacity of each link, but also promotes the intelligentization of the entire process of receipt collection, identification and application, providing key technical support for the automation and intelligent transformation of the financial technology field.

[0150] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of one embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing the present invention.

[0151] Those skilled in the art will understand that the modules in the apparatus of the embodiments can be distributed in the apparatus of the embodiments as described in the embodiments, or they can be located in one or more devices different from this embodiment with corresponding changes. The modules of the above embodiments can be combined into one module, or they can be further divided into multiple sub-modules.

[0152] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A scoring method for a bank receipt recognition system, characterized in that, include: S1: Bank receipt data acquisition and preprocessing: Acquire image data of bank receipts to be evaluated, and simultaneously obtain the structured recognition results output by the receipt recognition system; S2: Based on the receipt template library, automatically locate the key information areas on the receipt. The key information areas include the transaction amount field, the payer account field, and the seal area. Determine the core fields to be evaluated, and construct the following evaluation features for each core field: accuracy feature, completeness feature, clarity association feature, and format standardization feature. S3: Collect historical receipt recognition cases, label the manual verification results and corresponding feature data for each case, and form a training sample set. The scoring model is trained using machine learning algorithms combined with natural language processing models. For numerical fields, the score is quantified by calculating the deviation rate between the identified value and the standard value: |identified value - standard value| / standard value. For character fields, the matching score is calculated using edit distance. For non-textual information, an image recognition model is used to evaluate the integrity and clarity of the seal, converting it into a corresponding score. Based on the importance of different fields in the business scenario, the weight coefficients are configured using the analytic hierarchy process or business rules to ensure that the scoring results match the actual needs. S4: For the set of fields to be extracted from the receipt, calculate the score based on the ratio of the number of fields to be extracted to the number of fields actually extracted. The judgment criterion is to match all fields. If all fields that should be extracted are successfully extracted and the corresponding values ​​are accurate, the order score is 1 point. If any field that should be extracted is not extracted, or if the content of the extracted field is incorrect, the order score is 0 points. S5: Generates an assessment report that includes single-field scores, overall scores, and reasons for deductions. The assessment report supports graphical display.

2. The scoring method of the bank receipt identification system according to claim 1, characterized in that, In step S1, the bank receipt image data includes scanned copies of paper receipts and screenshots of electronic receipts, and the structured recognition results include transaction amount, account number, and summary.

3. The scoring method of the bank receipt identification system according to claim 1, characterized in that, Step S1 further includes data cleaning and standardization: The collected return receipt images are subjected to noise reduction, tilt correction, and edge cropping to remove image interference information; Standardize the format of the recognition results.

4. The scoring method of the bank receipt identification system according to claim 1, characterized in that, In step S2, the receipt template library covers the receipt formats of mainstream banks.

5. The scoring method of the bank receipt identification system according to claim 1, characterized in that, In step S2, each evaluation feature is evaluated according to the following requirements: Accuracy features: Evaluate the degree of matching between the recognition results and the standard results of manual annotation, including character matching rate and semantic consistency; Completeness features: Assess whether key fields are missing, including whether the last four digits of the account number are complete; Clarity-related features: The correlation coefficient between the clarity of the receipt image and the stability of the recognition result is evaluated. The clarity of the receipt image includes resolution and blur. Format compliance features: Evaluate whether the recognition results conform to the field format rules, including whether the amount contains currency symbols and whether the number of digits in the account is compliant.

6. The scoring method of the bank receipt identification system according to claim 1, characterized in that, In step S3, the manual verification results include correct, partially incorrect, and completely incorrect.

7. The scoring method of the bank receipt identification system according to claim 1, characterized in that, In step S3, the machine learning algorithm includes random forest and / or gradient boosting tree.

8. The scoring method of the bank receipt identification system according to claim 1, characterized in that, In step S4, regarding the number of fields to be extracted, the total number of fields that must be extracted from the receipt is determined based on the receipt type and business rules. The actual number of fields extracted is the number of valid fields successfully extracted, meaning that not only are the names of the fields identified, but their corresponding values ​​are also verified as valid information.

9. The scoring method of the bank receipt identification system according to claim 1, characterized in that, Further includes step S6: Feedback data collection: Record the correction results after manual review, and establish a closed-loop data chain of "identification result - scoring - correction result". Dynamic model updates: The scoring model is retrained regularly with new data, and field weights and scoring thresholds are adjusted to ensure that the scoring system adapts to changes in receipt styles and upgrades in recognition technology.

Images