Unformatted scale bill intelligent conversion method and system based on image recognition

Abstract

The invention discloses an unformatted scale bill intelligent conversion method and system based on image recognition, relates to the technical field of character intelligent recognition, and solves the technical problems that the accuracy of subsequent OCR recognition is unstable due to the fact that the inclination angle detection precision is low and noise filtering and contrast enhancement effects are limited by scenes. According to the method, table line detection and text line analysis are adopted for the structured / no-table scale bill, the inclination angle detection precision is improved, layered denoising and adaptive contrast enhancement are achieved, the text recognition accuracy in a complex scene is improved, PaddleOCR is dynamically switched based on the Chinese proportion, a low-confidence region recheck mechanism is combined, the recognition error rate is reduced, and the recognition efficiency is improved. According to the technical scheme, a scale document field dictionary and a relation graph are constructed, intelligent term matching and error correction are achieved, the recognition problem of uncommon words and industry exclusive vocabularies is solved, a hierarchical rule system of basic verification, association verification and compliance verification is adopted, format, logic and industry / enterprise compliance full dimensions are covered, and the verification coverage rate is increased.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of character intelligent recognition, in particular to a non-formatted pound single intelligent conversion method and system based on image recognition. BACKGROUND

[0002] In the fields of logistics, freight, warehousing, etc., the pound single is the core voucher of the weight of goods, transaction amount, and transportation information, and its processing efficiency and data accuracy directly affect the business flow and financial settlement.

[0003] With the development of OCR technology and automatic systems, some enterprises introduce image preprocessing and character recognition tools, but the existing solutions mostly use single engine recognition and fixed rule verification, which is difficult to meet the actual needs of various types of pound singles, mixed information, and complex business scenarios, and still requires a large amount of manual intervention, which restricts the improvement of automatic processing level. Traditional pound single processing relies on manual input and review, which has the following limitations: Firstly, manual recognition of handwritten or blurred pound singles is time-consuming, especially when dealing with a large number of batches, which is prone to backlog. Secondly, the accuracy is low: limited by human experience, data deviation is caused by character recognition errors and format verification omissions. Thirdly, there is a lack of unified verification standard, which is difficult to fully cover industry standards, enterprise systems, and policies and regulations. Fourthly, the pound single image is often difficult to recognize due to shooting conditions, and the existing preprocessing method is not suitable for complex scenes. SUMMARY

[0004] To overcome the shortcomings of the prior art, the present application provides a non-formatted pound single intelligent conversion method and system based on image recognition, which solves the problem of low detection accuracy of inclination angle, noise filtering and contrast enhancement effect limited by scene, resulting in unstable OCR recognition accuracy.

[0005] To achieve the above purpose, the present application realizes the following technical scheme: a non-formatted pound single intelligent conversion method based on image recognition, which specifically includes the following steps: Step one, through the HTTP protocol, collect non-formatted pound single pictures from the vehicle-mounted terminal of the logistics transportation vehicle, the terminal uploaded by manual, or receive the pound single picture file through the FTP server; Step two, convert the RGB three-color channel to gray value by weighted average method for the collected pound single picture, and calculate the global threshold and local threshold by Otsu algorithm to obtain the gray image; Step three, perform inclination correction on the obtained gray image by probability Hough transform, and perform denoising by median filtering, and automatically calculate the Gamma value of the whole image brightness to obtain the preprocessed pound single picture; Step four, based on the proportion of text in the pre-processed pound single picture, different recognition engines are selected, and a pound single field dictionary is constructed combined with historical data; Step five, the structured data in the pre-processed pound single picture is checked by adopting multi-level rule checking to generate error reasons, the error reasons are classified, and corresponding processing information is generated.

[0006] As a further scheme of the present application, the specific way of obtaining the gray image is: The RGB three-color channel is converted into a gray value by using a weighted average method, and the formula is gray value = 0.299 * R + 0.587 * G + 0.114 * B, wherein R, G and B represent red, green and blue channel values respectively; for the pound single containing a red or blue seal, the blue channel is extracted, and the B channel is fused with the basic gray image in a 7:3 ratio; The Otsu algorithm is used to calculate the global threshold T1, the image is segmented into 8*8 sub-blocks, and the local threshold T2 is obtained by subtracting an offset of 2-5 from the Gaussian weighted average value of the pixels in the sub-block; if T2 < 0.8 * T1, T2 is used for checking; if T2 > 1.2 * T1, T1 is used for checking; otherwise, a weighted threshold T = 0.6 * T1 + 0.4 * T2 is used.

[0007] As a further scheme of the present application, the specific way of obtaining the pre-processed pound single picture is: If it is a structured pound single, the table lines are extracted by the probability Hough transform, the line segments shorter than 50 pixels are filtered, and the angle with the highest frequency of occurrence is taken as the tilt angle; if it is a table-free pound single, the MSER algorithm is used to locate the text area, the effective area with an area of more than 200 pixels is screened, the tilt angle is determined by fitting the main direction of the text line and removing outliers, the image is rotated according to the tilt angle, the edge is cropped, and the picture quality is optimized by selecting the bilinear interpolation according to the resolution; For random noise, 3*3 median filtering combined with morphological opening operation is used for cleaning; for Gaussian noise, bilateral filtering or non-local mean filtering is used for cleaning; for characteristic noise, non-local mean filtering and iterative repair are used for cleaning, the Gamma value is calculated according to the brightness mean value of the image, and histogram equalization is performed on the image blocks.

[0008] As a further scheme of the present application, the specific way of constructing the pound single field dictionary is: Text recognition is performed on the pre-processed pound single picture, the text in the picture is converted into editable text, the Chinese proportion of the pound single text is detected, PaddleOCR is used when the proportion is > 60%, and Tesseract is used when the proportion is ≤ 60%; the image area is divided and the recognition confidence is calculated, and another engine is triggered for re-inspection for the area with a confidence of < 85%; Based on historical data, industry documents and enterprise knowledge base, after data cleaning, term extraction, classification and word vector clustering, a regularized dictionary containing entity relationship is constructed.

[0009] As a further scheme of the present application, the multi-level rule verification includes basic verification, correlation verification and compliance verification, wherein the basic verification includes format verification and value range verification, the correlation verification includes numerical value and entity relationship, and the compliance verification includes industry and enterprise rules.

[0010] As a further scheme of the present application, the specific way of classifying error causes is: The error causes are obtained and error severity classification processing is performed, and a fatal error is classified when a business logic or a regulation requirement is seriously violated, data is completely invalid, and immediate processing is required, a repairable error is classified when a business key link is affected but can be remedied by a temporary measure, a business operation is not affected, and there is a potential risk or optimization space.

[0011] As a further scheme of the present application, the specific way of generating corresponding processing information is: For a fatal error, all subsequent business processes are suspended, data modification authority is locked, error fields are highlighted in a structured data report, specific violation descriptions are attached, and data sources are automatically associated; For a warning error, data is allowed to enter a subsequent link, an orange warning mark is marked on the interface, a risk warning sheet is automatically generated, deviation contents are listed, and the sheet is pushed to a business auditor; For a repairable error, the system corrects the error according to preset rules, the system automatically corrects is noted in a data report, original values and corrected values are recorded, and after repair, basic rule verification is automatically triggered.

[0012] The non-formatted pound sheet intelligent conversion system based on image recognition comprises: An image acquisition module is configured to acquire non-formatted pound sheet pictures through HTTP or FTP mode; An image preprocessing module is configured to convert a color image into a gray image by using OpenCV combined with a self-defined algorithm, highlight a text area by using a binarization processing, identify noise points, remove small noise points by using value filtering, clean residual noise points by using morphological operation, rotate and correct an inclined pound sheet image, enhance contrast by using a gray histogram equalization method, obtain a preprocessed pound sheet picture, and perform text recognition on the preprocessed pound sheet picture by using PaddleOCR, and convert the text in the picture into editable text. A structured analysis module is configured to analyze the editable text obtained by OCR recognition by using a regular expression combined with a natural language processing technology, extract key data and understand semantics thereof. Data verification module, using Drools rule engine, carries out verification to the structured data obtained by parsing, ensures the accuracy and compliance of data.

[0013] The application provides a non-formatted pound single intelligent conversion method and system based on image recognition. The application improves the tilt angle detection precision by adopting table line detection and character line analysis for structured / tableless pound single respectively, improves the character recognition accuracy in complex scenes through layered denoising and adaptive contrast enhancement, reduces the recognition error rate based on the dynamic switching of PaddleOCR and Tesseract engines combined with the low-confidence area re-inspection mechanism, constructs a pound single field dictionary and a relationship graph, realizes intelligent matching and error correction of terms, and solves the recognition problem of rare words and industry-specific words.

[0014] The application adopts a layered rule system of basic verification, correlation verification and compliance verification, covers formats, logic and industry / enterprise compliance in all dimensions, improves the verification coverage, supports on-demand loading of the modular rule engine, avoids invalid verification, improves the processing efficiency, classifies the processing according to fatal errors, warning errors and repairable errors, realizes instant blocking of fatal errors, warning error early warning and limited period processing, automatic correction of repairable errors, reduces manual intervention, ensures dynamic compliance through the secondary review mechanism of key fields combined with real-time data, reduces business disputes caused by data errors, is compatible with structured, tableless pound single, multiple industry scenarios and enterprise customization requirements, the rules can be dynamically updated without interrupting the system, and the expansion of the scheme is improved. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 The application provides a step method diagram; Figure 2 The application provides a system block diagram. DETAILED DESCRIPTION

[0016] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the application.

[0017] Embodiment one Please refer to Figure 1 The application provides a non-formatted pound single intelligent conversion method based on image recognition, which specifically includes the following steps: Step one, collect non-formatted scale ticket pictures from the vehicle terminal of the logistics transport vehicle or the terminal uploaded by manual, and the specific representation is that the corresponding information transmission processing is performed based on the HTTP protocol when the picture is transmitted, or the scale ticket picture file can be received through the FTP server to ensure complete acquisition of the picture information.

[0018] Step two, process the collected scale ticket pictures by using the OpenCV library combined with the self-developed custom algorithm, and the processing operation includes grayscale and binarization, and the specific processing method is as follows: Convert the RGB three-color channel to a grayscale value by using the weighted average method, according to the formula Grayscale value = 0.299xR + 0.587xG + 0.114xB, wherein R represents the red channel value, G represents the green channel value, and B represents the blue channel value. For the common red / blue seal in the scale ticket, reduce the interference by using the channel separation technology, extract the blue channel (B channel), and fuse the B channel with the basic grayscale image at a ratio of 7:3 to enhance the contrast of the text area. Then, convert the grayscale image into an image containing only black and white colors, automatically calculate the best threshold value as the global threshold value T1 by using the Otsu algorithm, divide the image into 8x8 or 16x16 sub-blocks, calculate the threshold value for each sub-block separately, and the specific calculation method is to take the Gaussian weighted average value of the pixels in the sub-block as the reference, subtract the offset (usually 2-5) as the threshold value as the local threshold value T2, and compare the obtained threshold values. If the local threshold value T2 < global threshold value T1x0.8, use the local threshold value T2 as the standard (dark area enhancement), if the local threshold value T2 > global threshold value T1x1.2, use the global threshold value T1 as the standard (light area suppression), and for other cases, use the weighted average for calculation, and the specific T = T1x0.6 + T2x0.4. Step three, correct the obtained grayscale image, obtain the grayscale image and classify it, if the grayscale image corresponds to a structured scale ticket, there are obvious table lines (horizontal / vertical line proportion ≥ 30%), then use the probabilistic Hough transform to extract the table lines, filter the short line segments with a length < 50 pixels, calculate the angle distribution of all valid lines, take the angle value with the highest frequency as the inclination angle, determine the inclination angle of the scale ticket, if the grayscale image corresponds to a non-table scale ticket, the specific representation is that the text line distribution is dominant, and there are no regular lines, use the MSER algorithm to locate the text area, screen the connected domain with an area > 200 pixels, fit the minimum bounding rectangle of the text area, extract the long side direction of the rectangle as the main direction of the text line, and remove the outliers (such as inclined single characters) by using the RANSAC algorithm to fit the overall inclination angle. According to the tilt angle, the rotation transformation is performed to ensure that the text lines are horizontal or the table lines are vertical, the effective content area is identified through boundary detection, the black edges generated by automatic cropping and rotation are removed, then the bilinear interpolation or Lanczos interpolation algorithm is used for interpolation optimization, and specific conventional scenes use bilinear interpolation, and high-resolution images (≥300 dpi) use Lanczos interpolation to obtain the angle-corrected picture. Then, the angle-corrected picture is denoised to obtain all noise points corresponding to the angle-corrected picture, and the noise points are classified to obtain random noise points (represented as scattered and distributed black and white small points, usually generated by a shooting device or a transmission process), Gaussian noise (represented as a blur effect on the image, commonly found in low-light environments), and characteristic noise (such as seal interference (block color blocks), creases (linear interference), and stains (irregular dark areas)), the obtained random noise points are cleaned by 3x3 median filtering to remove small noise points, and then morphological operation (opening operation) is used to clean residual noise points, Gaussian noise is cleaned by bilateral filtering or non-local mean filtering, and characteristic noise is cleaned by non-local mean filtering and iterative repair. Then, the processed gray-scale picture is subjected to contrast enhancement processing, and the Gamma value is automatically calculated according to the overall brightness of the image. The Gamma value refers to a parameter in image processing and printing technology that describes the non-linear relationship between input signal and output brightness. For dark images (brightness mean <100), use Gamma<1 to improve brightness; for bright images (brightness mean >200), use Gamma>1 to reduce brightness. The image is subjected to block processing (such as 8x8 grid), and the histogram of each block is equalized to improve the local detail contrast.

[0019] Step four, using PaddleOCR (text recognition and document parsing) or Tesseract (open source text recognition engine) as an OCR recognition engine, the pre-processed pound sheet picture is subjected to text recognition, and the text in the picture is converted into editable text, and the specific processing method is as follows: The pre-processed pound sheet picture is obtained, and the corresponding language detection result is obtained. At the same time, according to the obtained language detection result, the engine is switched. Specifically, if the Chinese proportion in the language detection result is >60%, PaddleOCR is selected as the recognition engine, otherwise if the Chinese proportion is ≤60%, Tesseract is selected as the recognition engine. At the same time, the pound sheet picture is subjected to region division to obtain the division region, and the recognition confidence corresponding to the division region is obtained. The recognition confidence is recognized by the CRNN model of PaddleOCR and the LSTM model of Tesseract. This is prior art and will not be described in detail. Then, the obtained recognition confidence is judged. If the recognition confidence of the division region is <85%, another engine recheck is automatically triggered. Then build the pound single field dictionary, first get the historical pound single data, industry standard documents and enterprise internal knowledge base, and clean the obtained data, including removing duplicates, unifying case, standardizing units, extracting information through regular expressions, and using the BERT model to classify material names, while discovering semantically related terms through word vector clustering, then building the corresponding relationship graph according to the obtained terms to generate the pound single field dictionary.

[0020] Step five, get the structured data in the pound single picture and check its format and value range through preset rules. Specifically, for format checking, the license plate number needs to match the combination of "province abbreviation + letter + 5 digits / letters" (such as "Yue A8B7C6"), the date needs to conform to the "year-month-day" or "year / month / day" format, and the weight value needs to be positive with 1-3 decimal places. For value range checking, "net weight" needs to be greater than 0 and less than "gross weight", "unit price" needs to match the reasonable range of the cargo type (such as steel price ≥ 3000 yuan / ton, coal price ≥ 500 yuan / ton), and "transportation time" needs to be after "weighing time"; Based on the business logic relationship between multiple fields, identify data contradictions or logical vulnerabilities, value association: such as "net weight = gross weight - tare weight" (allowable error ≤ 0.1 tons), "total amount = net weight × unit price" (allowable error ≤ 0.5%), "vehicle rated load" needs to be greater than "net weight" (overload needs special marking); Entity association: such as "shipping unit" and "receiving unit" cannot be the same enterprise, "license plate number" needs to match "transportation route start / end point" (such as Jizhuan license plate corresponds to Hebei area), "cargo type" needs to match "packing unit" (such as liquid cargo unit should be "liter" "cubic meter"); Combine industry standards, enterprise systems and external policies to achieve deep compliance verification, and the specific verification methods are as follows: Industry-specific rules: such as dangerous goods pound single needs to include "dangerous goods number" "transportation permit number", and the permit validity period needs to cover the transportation time; import goods pound single needs to associate "customs declaration number" and check the format validity.

[0021] Enterprise custom rules: such as pound single with core customers (such as "XX automobile factory") needs to include "order number" "batch number", and when the daily transportation volume exceeds the threshold (such as ≥ 5000 tons), automatically trigger multi-level approval process; Meanwhile, different types of check rules are split into independent modules (such as "basic format module", "steel industry module", "dangerous goods special module"), and corresponding modules are automatically loaded according to the type of the bill (such as highway freight, railway freight, customs declaration), so as to avoid invalid rule execution, and the rules are executed in the order of basic check, correlation check and business compliance. If the previous layer of check fails, the subsequent process is directly terminated and the specific error reason (such as "tare weight format error: should be a positive number") is returned, and a "secondary check" mechanism is set for key fields (such as "net weight" and "total amount"): after the first check passes, the rule engine is called again after 5-10 seconds (simulating manual review delay) to review in combination with the latest business data (such as real-time exchange rate and price fluctuations), so as to ensure dynamic compliance. Based on the check result, the corresponding error reason is obtained, and error severity classification processing is performed to obtain fatal errors, warning errors and repairable errors. Specifically, a fatal error refers to a serious violation of business logic or regulatory requirements, resulting in completely invalid data that must be handled immediately. A warning error refers to an error that affects a key business link but can be remedied through temporary measures without affecting business operation, but may have potential risks or optimization space. A repairable error refers to an error that can be repaired, and the specific check processing methods are as follows: For fatal errors, all subsequent business processes are suspended, data modification permissions are locked, error fields are highlighted in structured data reports, specific violation descriptions are attached, and data sources are automatically associated. For warning errors, the data is allowed to enter subsequent links (such as OCR recognition and system archiving), but an orange warning mark is displayed on the interface, a risk warning sheet is automatically generated, the deviation content is listed (such as "unit price deviates from the average by 12%, suggest verifying market conditions"), and the sheet is pushed to business auditors. For repairable errors, the system corrects the error according to the preset rules, notes "system automatic correction" in the data report, records the original value and the corrected value, and automatically triggers basic rule check after repair to confirm that the correction result meets the standard.

[0022] Embodiment Two Please refer to Figure 2 The present application provides an image recognition-based non-formatted bill intelligent conversion system, which includes an image acquisition module, an image preprocessing module, a structured analysis module and a data check module, and combines Figure 2 It can be known that the information between the above functional modules is unidirectional transmission.

[0023] The image acquisition module is used to acquire non-formatted bill pictures through HTTP or FTP mode. An image preprocessing module uses OpenCV combined with a self-defined algorithm to perform grayscale, binarization, noise removal, tilt correction, and contrast enhancement on the collected pound sheet pictures to obtain preprocessed pound sheet pictures, and uses PaddleOCR or Tesseract to perform text recognition on the preprocessed pound sheet pictures to convert the text in the pictures into editable text; A structured analysis module analyzes the editable text obtained by OCR recognition through regular expressions combined with natural language processing (NLP) technology, extracts key data, and understands the semantics thereof; A data verification module uses a Drools rule engine to verify the structured data obtained by analysis to ensure the accuracy and compliance of the data.

[0024] Some data in the above formula are calculated by taking their numerical values, and are not calculated by substituting parameter units. The contents not described in detail in the specification all belong to the prior art known to those skilled in the art.

[0025] The above embodiments are only used to illustrate the technical method of the present application and not to limit it. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical method of the present application.

Claims

1. A non-formatted pound sheet intelligent conversion method based on image recognition, characterized in that, The method specifically includes the following steps: Step 1: Collect unformatted weighbridge slip images from the on-board terminal of the logistics transport vehicle or the terminal uploaded manually via HTTP protocol, or receive weighbridge slip image files via FTP server. Step 2: Convert the RGB three-color channels of the collected weighbridge image to grayscale values ​​using a weighted average method. At the same time, use the Otsu algorithm to calculate the global threshold and local threshold to obtain the grayscale image. Step 3: Apply probabilistic Hough transform to the obtained grayscale image for tilt correction, and perform noise reduction through median filtering. At the same time, automatically calculate the Gamma value of the overall image brightness to obtain the preprocessed bill image. Step 4: Select different recognition engines based on the text ratio in the preprocessed weighbridge image, and build a weighbridge domain dictionary by combining historical data; Step 5: Use multi-level rule verification to verify the structured data in the preprocessed weighbridge image, generate error reasons, classify the error reasons, and generate corresponding processing information.

2. The image recognition based unformatted pounder sheet intelligent conversion method according to claim 1, characterized in that, The specific method for obtaining the grayscale image is as follows: The RGB three-color channels are converted to grayscale values ​​using a weighted average method. The formula is grayscale value = 0.299 × R + 0.587 × G + 0.114 × B, where R, G, and B represent the red, green, and blue channel values, respectively. For the bill containing red, the blue channel is extracted, and the B channel is blended with the base grayscale image at a 7:3 ratio. The global threshold T1 is calculated using the Otsu algorithm. The image is divided into 8×8 sub-blocks. The local threshold T2 is calculated by subtracting an offset of 2-5 from the Gaussian weighted average of the pixels in the sub-block. If T2 < 0.8 × T1, T2 is used for verification. If T2 > 1.2 × T1, T1 is used for verification. In other cases, a weighted threshold of T = 0.6 × T1 + 0.4 × T2 is used.

3. The image recognition based unformatted pounder sheet intelligent conversion method according to claim 1, characterized in that, The specific method for obtaining the preprocessed weight list image is as follows: For structured order sheets, table lines are extracted using probabilistic Hough transform, line segments shorter than 50 pixels are filtered out, and the angle with the highest frequency of occurrence is taken as the tilt angle. For order sheets without tables, the MSER algorithm is used to locate the text area, effective areas with an area exceeding 200 pixels are selected, the main direction of the text lines is fitted, and outliers are removed to determine the tilt angle. The image is rotated according to the tilt angle, and after cropping the edges, bilinear interpolation is selected to optimize the image quality based on the resolution. Random noise is cleaned using 3×3 median filtering combined with morphological opening operation; Gaussian noise is cleaned using bilateral filtering or nonlocal mean filtering; distinctive noise is cleaned using nonlocal mean filtering and iterative repair. The Gamma value is calculated based on the average image brightness, and histogram equalization is performed on the image blocks.

4. The image recognition based unformatted pounder sheet intelligent conversion method according to claim 1, characterized in that, The specific method for constructing the list domain dictionary is as follows: The preprocessed weighbridge image is subjected to text recognition, and the text in the image is converted into editable text. The proportion of Chinese characters in the weighbridge text is detected. When the proportion is >60%, PaddleOCR is used, and when the proportion is ≤60%, Tesseract is used. The image region is divided and the recognition confidence is calculated. For regions with a confidence of <85%, another engine is triggered for re-examination. Based on historical data, industry documents and enterprise knowledge base, after data cleaning, term extraction, classification and word vector clustering, a regularized dictionary containing entity relationship is constructed.

5. The image recognition based unformatted pounder sheet intelligent conversion method according to claim 1, wherein, The multi-level rule verification includes basic verification, correlation verification and compliance verification, wherein the basic verification includes format verification and value range verification, the correlation verification includes numerical value and entity relationship, and the compliance verification includes industry and enterprise rules.

6. The image recognition based unformatted pounder sheet intelligent conversion method according to claim 1, wherein, The specific way of classifying error causes is: The error causes are obtained and error severity classification processing is performed, and the specific way of classifying error causes is:

7. The image recognition based unformatted pounder sheet intelligent conversion method according to claim 1, wherein, The error causes are obtained and error severity classification processing is performed, and the specific way of classifying error causes is: The specific way of generating corresponding processing information is: For fatal errors, suspend all subsequent business processes, lock data modification permissions, highlight error fields in structured data reports, and attach specific violation descriptions, while automatically associating data sources; For warning errors, allow data to enter subsequent steps, but mark orange warning signs on the interface, automatically generate risk warning sheets, list deviation content, and push to business auditors; 8. An image recognition based unformatted pound sheet intelligent conversion system for performing the image recognition based unformatted pound sheet intelligent conversion method of any one of claims 1-7, characterized in that, For repairable errors, the system corrects errors according to preset rules, notes system automatic correction in data reports, records original and corrected values, and automatically triggers basic rule verification after repair. It includes: An image acquisition module for acquiring non-formatted pound sheet pictures through HTTP or FTP; An image preprocessing module that uses OpenCV combined with custom algorithms to convert color images to grayscale images, then highlights the text area through binarization processing and identifies noise points, removes small noise points through value filtering, cleans up residual noise points through morphological operations, rotates and corrects inclined pound sheet images, enhances contrast using the gray histogram equalization method, obtains preprocessed pound sheet pictures, and uses PaddleOCR to recognize text from preprocessed pound sheet pictures, converts the text in the pictures to editable text; A structured analysis module that uses regular expressions combined with natural language processing techniques to analyze the editable text obtained by OCR recognition, extracts key data and understands its semantics; A data verification module that uses the Drools rule engine to verify the structured data obtained by analysis to ensure data accuracy and compliance.

Images