Single-intelligent auditing method, device and equipment

Abstract

This application provides a method, apparatus, and device for intelligent verification of weighbridge slips. The method includes: acquiring a weighbridge slip image to be processed; performing data preprocessing on the weighbridge slip image to obtain a standardized image; inputting the standardized image into an intelligent verification model; performing character recognition processing on the standardized image using a recognition module based on the intelligent verification model to obtain a weighbridge slip text sequence; and processing the weighbridge slip text sequence using an verification module based on the intelligent verification model to obtain a verification result. This method aims to achieve efficient and accurate verification of weighbridge slips.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a method, apparatus, and device for intelligent verification of weighbridge slips. Background Technology

[0002] With the widespread application of new energy heavy trucks in resource-based logistics fields such as coal and sand, weighbridge slips serve as core vouchers for capacity settlement, cost accounting, and process control.

[0003] In related technologies, the weighbridge slip information is manually reviewed and processed; however, this method is inefficient and prone to errors due to human negligence.

[0004] Therefore, there is an urgent need for an efficient and accurate solution for verifying weighbridge slips. Summary of the Invention

[0005] This application provides a method, apparatus, and device for intelligent verification of weighbridge slips, which can achieve efficient and accurate verification of weighbridge slips.

[0006] In a first aspect, embodiments of this application provide a smart verification method for weighbridge slips, including:

[0007] Obtain the weighbridge image to be processed;

[0008] The weighbridge image to be processed is preprocessed to obtain a standardized image;

[0009] The standardized image is input into the intelligent auditing model. The recognition module of the intelligent auditing model performs character recognition processing on the standardized image to obtain the weight slip text sequence. The auditing module of the intelligent auditing model processes the weight slip text sequence to obtain the audit result.

[0010] In one possible implementation, the standardized image is input into an intelligent auditing model, and the recognition module of the intelligent auditing model performs character recognition processing on the standardized image to obtain a weighted list text sequence, including:

[0011] Based on the feature extraction module of the recognition module, the standardized image is processed to obtain the weight slip feature image;

[0012] Based on the attention algorithm, feature enhancement processing is performed on the text region of the bill feature image to obtain an attention feature image;

[0013] The character recognition module based on the recognition module performs character recognition processing on the attention feature image to obtain the bill of lading text sequence.

[0014] In one possible implementation, the character recognition module of the recognition module performs character recognition processing on the attention feature image to obtain a billboard text sequence, including:

[0015] The character recognition module based on the recognition module performs character recognition processing on the attention feature image to obtain an initial text sequence;

[0016] Based on the large language model, the initial text sequence is modified to obtain the bill of lading text sequence.

[0017] In one possible implementation, the audit module based on the intelligent audit model processes the weighbridge text sequence to obtain the audit result, including:

[0018] The audit module based on the intelligent audit model performs feature extraction processing on the weighbridge slip text sequence to obtain a weighbridge slip information set;

[0019] Based on the preset regular expressions of the audit module, the information in the weighbridge information set is modified to obtain a standardized weighbridge information set.

[0020] Based on the multi-level review rules of the review module, the standardized set of weighbridge information is subjected to multi-level matching processing to obtain the review result.

[0021] In one possible implementation, the weighbridge information set includes basic information, weight information, and related information; based on the preset regular expression of the review module, the information in the weighbridge information set is modified to obtain a standardized weighbridge information set, including:

[0022] If it is determined that the basic information in the weighbridge information set does not conform to the preset regular expression, then the basic information in the weighbridge information set is corrected to obtain the corrected basic information;

[0023] If it is determined that the difference between the weight information in the weighing slip information set and the preset weight information is greater than or equal to a preset threshold, then the weight information is corrected according to the preset weight information to obtain the corrected weight information;

[0024] Based on the corrected basic information, the corrected weight information, and the associated information, the standardized set of weighbridge information is determined.

[0025] In one possible implementation, the standardized weighbridge information set includes corrected basic information, corrected weight information, and associated information;

[0026] Based on the multi-level review rules of the review module, the standardized weighbridge information set is subjected to multi-level matching processing to obtain the review result, including:

[0027] Based on the multi-level review rules of the review module, the corrected basic information in the standardized weighbridge information set is matched with the basic information in the multi-level review rules to obtain the information matching degree.

[0028] Based on the multi-level review rules of the review module, the corrected weight information in the standardized weighbridge information set is matched with the logical information in the multi-level review rules to obtain the logical accuracy rate.

[0029] Based on the multi-level review rules of the review module, the associated information in the standardized weighbridge information set is matched with the associated information in the multi-level review rules to obtain the data correlation degree.

[0030] The audit result is determined based on the information matching degree, the logical correctness rate, and the data correlation degree.

[0031] In one possible implementation, determining the audit result based on the information matching degree, the logical correctness rate, and the data correlation degree includes:

[0032] If the information matching degree is determined to be greater than or equal to the first preset threshold, and the logical correctness is greater than or equal to the second preset threshold, and the data correlation degree is greater than or equal to the third preset threshold, then the review result indicates that the review has passed; otherwise, the review result indicates that the review has failed and requires manual review, and the review result is sent to the user.

[0033] In one possible implementation, the weighbridge image to be processed is preprocessed to obtain a standardized image, including:

[0034] The weighbridge image to be processed is subjected to tilt correction processing to obtain the processed weighbridge image;

[0035] The processed weighbridge image is subjected to data processing to obtain a standardized image; wherein the data processing includes one or more of the following: noise reduction processing, blur enhancement processing, occlusion processing, and size normalization processing.

[0036] Secondly, embodiments of this application provide a smart verification device for weighbridge slips, comprising:

[0037] The acquisition unit is used to acquire the weighbridge image to be processed.

[0038] The processing unit is used to preprocess the weighbridge image to be processed to obtain a standardized image;

[0039] The review unit is used to input the standardized image into the intelligent review model, and the recognition module of the intelligent review model performs character recognition processing on the standardized image to obtain the weight slip text sequence; the review module of the intelligent review model processes the weight slip text sequence to obtain the review result.

[0040] In one possible implementation, the review unit is specifically used for:

[0041] Based on the feature extraction module of the recognition module, the standardized image is processed to obtain the weight slip feature image;

[0042] Based on the attention algorithm, feature enhancement processing is performed on the text region of the bill feature image to obtain an attention feature image;

[0043] The character recognition module based on the recognition module performs character recognition processing on the attention feature image to obtain the bill of lading text sequence.

[0044] In one possible implementation, the review unit is further configured to:

[0045] The character recognition module based on the recognition module performs character recognition processing on the attention feature image to obtain an initial text sequence;

[0046] Based on the large language model, the initial text sequence is modified to obtain the bill of lading text sequence.

[0047] In one possible implementation, the review unit is further configured to:

[0048] The audit module based on the intelligent audit model performs feature extraction processing on the weighbridge slip text sequence to obtain a weighbridge slip information set;

[0049] Based on the preset regular expressions of the audit module, the information in the weighbridge information set is modified to obtain a standardized weighbridge information set.

[0050] Based on the multi-level review rules of the review module, the standardized set of weighbridge information is subjected to multi-level matching processing to obtain the review result.

[0051] In one possible implementation, the weighbridge information set includes basic information, weight information, and related information; the review unit is further configured to:

[0052] If it is determined that the basic information in the weighbridge information set does not conform to the preset regular expression, then the basic information in the weighbridge information set is corrected to obtain the corrected basic information;

[0053] If it is determined that the difference between the weight information in the weighing slip information set and the preset weight information is greater than or equal to a preset threshold, then the weight information is corrected according to the preset weight information to obtain the corrected weight information;

[0054] Based on the corrected basic information, the corrected weight information, and the associated information, the standardized set of weighbridge information is determined.

[0055] In one possible implementation, the standardized weighbridge information set includes corrected basic information, corrected weight information, and associated information; the review unit is further configured to:

[0056] Based on the multi-level review rules of the review module, the corrected basic information in the standardized weighbridge information set is matched with the basic information in the multi-level review rules to obtain the information matching degree.

[0057] Based on the multi-level review rules of the review module, the corrected weight information in the standardized weighbridge information set is matched with the logical information in the multi-level review rules to obtain the logical accuracy rate.

[0058] Based on the multi-level review rules of the review module, the associated information in the standardized weighbridge information set is matched with the associated information in the multi-level review rules to obtain the data correlation degree.

[0059] The audit result is determined based on the information matching degree, the logical correctness rate, and the data correlation degree.

[0060] In one possible implementation, the review unit is further configured to:

[0061] If the information matching degree is determined to be greater than or equal to the first preset threshold, and the logical correctness is greater than or equal to the second preset threshold, and the data correlation degree is greater than or equal to the third preset threshold, then the review result indicates that the review has passed; otherwise, the review result indicates that the review has failed and requires manual review, and the review result is sent to the user.

[0062] In one possible implementation, the processing unit is specifically used for:

[0063] The weighbridge image to be processed is subjected to tilt correction processing to obtain the processed weighbridge image;

[0064] The processed weighbridge image is subjected to data processing to obtain a standardized image; wherein the data processing includes one or more of the following: noise reduction processing, blur enhancement processing, occlusion processing, and size normalization processing.

[0065] Thirdly, embodiments of this application provide an electronic device, including: a memory and a processor;

[0066] The memory stores computer-executed instructions;

[0067] The processor executes computer execution instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0068] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0069] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0070] This application provides a method, apparatus, and device for intelligent verification of weighbridge slips. It acquires a weighbridge slip image, preprocesses it to obtain a standardized image with upright orientation, clear text, and uniform size, and then inputs the standardized image into an intelligent verification model trained on a neural network from historical weighbridge slip image data. The model's recognition module performs character recognition to obtain the weighbridge slip text sequence, and the verification module processes the text sequence to output the verification result. This method utilizes standardized preprocessing to eliminate the negative impact of image tilt, blurriness, and inconsistent sizes on recognition accuracy. Based on the trained neural network model, it automates the process of character extraction and compliance verification, achieving efficient, accurate, and intelligent weighbridge slip verification. This effectively reduces the human resource input and subjective errors of manual verification, improves the efficiency and accuracy of weighbridge slip verification, and ensures the compliance and consistency of transportation measurement information. Attached Figure Description

[0071] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0072] Figure 1 A flowchart illustrating a smart verification method for weighbridge slips provided in this application embodiment. Figure 1 ;

[0073] Figure 2 A flowchart illustrating a smart verification method for weighbridge slips provided in this application embodiment. Figure 2 ;

[0074] Figure 3A flowchart illustrating step S203 in a smart verification method for weighbridge slips provided in this application embodiment;

[0075] Figure 4 A flowchart illustrating step S204 in a smart verification method for weighbridge slips provided in this application embodiment;

[0076] Figure 5 A schematic diagram of the structure of a smart weighbridge verification device provided in this application embodiment. Figure 1 ;

[0077] Figure 6 A schematic diagram of the structure of a smart weighbridge verification device provided in this application embodiment. Figure 2 ;

[0078] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0079] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0080] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application.

[0081] In the field of new energy heavy-duty truck material transportation management, weighbridge slip verification is a crucial step in weighing settlement, transportation accounting, and process management. This scenario typically occurs at transportation nodes such as mining areas, factory areas, and logistics parks. Weighbridge slip images can be obtained from photos, scans, or electronic screenshots of paper documents, and information extraction and verification processing must be completed in the transportation management platform.

[0082] In existing technologies, the process typically involves first acquiring an image of the weighbridge slip, then using Optical Character Recognition (OCR) technology to extract textual information such as vehicle number and weight. Afterward, manual or rule-based procedures are used to verify the completeness of the fields to assist in transportation settlement and business record-keeping. The basic principle of this type of solution is to convert the document content into text through image recognition, and then determine whether there are missing items, errors, or obvious anomalies based on preset rules.

[0083] However, weighbridge slips come from diverse sources, have significantly different formats, and often suffer from issues such as tilting, blurring, uneven lighting, and occlusion. Fixed templates or OCR solutions with weak adaptability struggle to provide stable recognition, easily leading to text omissions, misidentifications, and field misalignments. Meanwhile, auditing methods relying solely on simple integrity checks are insufficient for thorough review and assessment of the recognized text, resulting in low accuracy in complex transportation scenarios and necessitating significant manual intervention. This, in turn, impacts audit efficiency, settlement timeliness, and consistency in business processing.

[0084] Therefore, this application provides an intelligent verification method for weighbridge slips. By acquiring the weighbridge slip image to be processed and performing data preprocessing on the weighbridge slip image to obtain a standardized image with upright orientation, clear text, and uniform size, the standardized image is then input into an intelligent verification model trained based on historical weighbridge slip image data. The recognition module performs character recognition processing to obtain a weighbridge slip text sequence, and the verification module processes the weighbridge slip text sequence to obtain the verification result, thereby improving the automation level and processing effect of weighbridge slip verification.

[0085] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0086] Figure 1 A flowchart illustrating a smart verification method for weighbridge slips provided in this application embodiment. Figure 1 ,like Figure 1 As shown, the method includes:

[0087] S101. Obtain the image of the weighbridge slip to be processed.

[0088] For example, the weighbridge image to be processed is the raw input data processed by the method. It is used to carry textual content such as vehicle number, gross weight, tare weight, net weight, weighing time, cargo name, transportation unit, and consignee and consignee units during the transportation of goods. Subsequent data preprocessing, character recognition processing and review processing are all based on this image.

[0089] The weighbridge slip image to be processed can be a bitmap file generated by scanning a paper weighbridge slip, a photo file generated by taking a picture of a paper weighbridge slip with a mobile terminal, or a screenshot file of the electronic weighbridge slip page in the business system.

[0090] The implementing entity can be an application server deployed in a transportation management platform, or an edge computing device, a fleet management terminal, or an auditing program installed on a mobile device.

[0091] In one possible implementation, the acquisition process can be completed through three channels: First, drivers, weighbridge operators, or on-site managers can upload image files via mobile applications. Upon receiving the upload request, the system records the file identifier, upload time, user identifier, and associated transit order number, and stores the image in object storage or a file server. Second, the system can directly read the scanned output file through the acquisition program of a scanner or all-in-one machine and convert it into a unified image format, such as JPEG (Joint Photographic Experts Group), PNG (Portable Network Graphics), or TIFF (Tagged Image File Format). Third, the system can receive electronic screenshots or exported image files through data interfaces with transportation business platforms, metering systems, or office systems.

[0092] To ensure the execution of subsequent processing steps, after acquiring the weighbridge image, basic checks can be performed on the file integrity, including determining whether the file is decodable, whether the image resolution meets the minimum recognition requirements, whether the color channels are normal, whether the file size exceeds the processing threshold, and whether there are blank or pure black pages. For photographed images, shooting direction information, device model information, and geolocation markers from the time of upload can also be received simultaneously for subsequent anomaly tracing, but this part does not change the core processing logic of this method.

[0093] By establishing a unified image access method for paper scans, mobile phone photos, and electronic screenshots, the content of weighbridge slips from different sources is transformed into computable image data, enabling subsequent steps to be processed based on a unified input, thus addressing the diverse sources of weighbridge slips in complex transportation scenarios.

[0094] In one possible embodiment, after receiving the weighbridge image, the system can also generate a unique business serial number, and associate the serial number with the image storage path, transportation task number, vehicle number and uploading entity, so that the subsequent output review results can be written back to the transportation management platform and complete the process loop.

[0095] S102. Perform data preprocessing on the weighbridge image to be processed to obtain a standardized image.

[0096] For example, data preprocessing is used to convert the weighbridge image to be processed into a standard input suitable for processing by a neural network model. The standardized image is the result image after image correction, sharpening and size unification.

[0097] Since the weighbridge images to be processed may have issues such as tilted shooting, uneven edge cropping, background interference, uneven lighting, local blurring, stains, and inconsistent resolution, directly inputting them into the model can easily cause character region misalignment, blurred text boundaries, and unstable field recognition. Therefore, it is necessary to complete the unified correction of image quality in this step.

[0098] In practice, data preprocessing can involve adjusting the orientation of the weighbridge image, improving its clarity, suppressing noise or background interference, and standardizing its size to output a standardized image with correct orientation, clear text, and uniform size. The function of standardized images is to reduce the interference of input differences on the recognition and verification modules.

[0099] By converting bill images with different shooting conditions and different layout quality into image data of uniform scale and orientation, the character shape and spatial layout are made comparable at the model input, thereby reducing the probability of omission, misidentification and field misalignment.

[0100] In one possible embodiment, data preprocessing can be performed according to a preset processing flow, and a processed image copy or quality index can be recorded at the processing node for anomaly backtracking and model tuning. It should be understood that the above image processing method can adopt equivalent implementations, as long as it can output standardized images with upright orientation, clear text, and uniform size.

[0101] S103. The standardized image is input into the intelligent auditing model. The recognition module of the intelligent auditing model performs character recognition processing on the standardized image to obtain the weight slip text sequence. The auditing module of the intelligent auditing model processes the weight slip text sequence to obtain the audit result. For example, the intelligent auditing model is the core processing unit of this method, consisting of a recognition module and an auditing module. The intelligent auditing model is trained on a neural network model using historical weight slip image data.

[0102] Historical weighbridge image data includes historical weighbridge image samples from different mining areas, factory areas, logistics parks or carriers and their corresponding annotation information. The annotation information includes at least the text content actually recorded in the image, and may also include, for example, audit status labels, field integrity labels or business exception labels.

[0103] Neural network models can adopt network structures suitable for baggage label image processing and text review. Through the training process, parameters are continuously adjusted so that the model learns the character distribution features, layout features, and text review rules in baggage label images.

[0104] For example, the standardized image is first loaded into memory according to the model input specifications, and after normalized tensor transformation, it is sent to the recognition module. The recognition module is used to perform character recognition processing, that is, to convert the visible text content in the image into text form, thereby forming the weighbridge slip text sequence. The weighbridge slip text sequence is used to represent the sequence of text content recorded in the weighbridge slip image to be processed. Its content can include continuous text from the entire weighbridge slip, or it can be expressed in the form of field fragments concatenated within the model, as long as it can reflect the text recorded in the original image. For positions with low confidence in the recognition result, sequence correction can be performed by combining contextual relationships, but the basis for correction still comes from the output of image features by the recognition module. After obtaining the weighbridge slip text sequence, the review module processes the text sequence and outputs the review result. The review result can be represented using structured status data, such as review passed, manual review required, or review failed, and can be accompanied by an exception reason code, exception field name, and model confidence level, so as to write back to the transportation management platform.

[0105] Based on the above analysis, this step integrates recognition and review within the same intelligent review model framework. The front end trains on historical weighbridge image data to develop character recognition capabilities for images of various formats and resolutions. The back end further converts the recognized text into business review results, thus handling not only image-to-text conversion but also text-to-review conclusion generation. Compared to simply performing OCR output followed by basic manual verification, this step completes character extraction and review judgment within a unified model chain, providing more stable automatic review results for weighbridge text content in complex scenarios.

[0106] During the training phase, the historical bill of lading image data covers various layout, lighting, and shooting conditions. The neural network model iteratively optimizes the parameters through the loss function, so that the model gradually converges the mapping relationship between the bill of lading images and the review results. Therefore, it can maintain good adaptability to bill of lading images from different sources during the actual inference phase.

[0107] In one possible embodiment, in addition to the status value, the audit result can also generate an audit record object, recording the original image identifier, standardized image identifier, recognized text, audit conclusion, anomalies, and processing time, for use in settlement process documentation and subsequent review. It should be understood that the above example is merely illustrative and not limiting; any implementation method that can perform character recognition on standardized images and output audit results for the recognized text based on a trained neural network model can be used to implement this step.

[0108] In one possible embodiment, the training of the intelligent audit model is a data-driven supervised learning process. Its core lies in using a large number of historical weighbridge images and their corresponding accurate text annotations to enable the neural network to learn to extract text information from the images and output the correct weighbridge text sequence.

[0109] At the beginning of training, tens of thousands of historical weighbridge images taken or scanned in real business scenarios were collected. These images should cover complex situations such as different angles, lighting conditions, degree of wrinkles and stamp stains. The complete text content in each image was labeled word by word by manual or automated tools to generate one-to-one corresponding text sequence labels, which constituted the training dataset.

[0110] The standardized images, after grayscale conversion, tilt correction, and size normalization preprocessing, are input into the neural network model in batches. The model's feature extraction module automatically captures low-level visual features such as text strokes and table lines through multi-layer convolution operations, and abstracts them into high-level semantic features layer by layer. At the same time, the attention mechanism guides the network to focus on the text region while suppressing irrelevant information such as seals and backgrounds, and finally generates attention feature images.

[0111] The character recognition module encodes the attention feature images into feature sequences and uses a sequence modeling network and decoder to predict character categories step-by-step, outputting the predicted order text sequence. During training, the system compares the model-predicted text sequence with manually labeled real text sequences, calculates the difference between the two by combining a temporal classification loss function or a cross-entropy loss function, and uses a backpropagation algorithm to propagate the error gradient layer by layer, updating all network weight parameters of the feature extraction module and the character recognition module.

[0112] This process iterates thousands of times over a vast amount of historical data, continuously optimizing and adjusting the convolutional kernel parameters, attention weight matrix, and gating parameters of the sequence modeling units until the model's text recognition accuracy on the independent validation set converges to a preset standard, such as a character accuracy stabilizing above 99%. Finally, the trained neural network model is solidified into an intelligent verification model. At this point, the model has internalized visual knowledge implicit in the historical data, such as font style, table structure, character layout patterns, and seal occlusion patterns, into its parameters, enabling it to perform end-to-end, high-precision text extraction and verification of newly encountered weight slip images.

[0113] This application provides an intelligent verification method for weighbridge slips. It acquires weighbridge slip images, preprocesses them to obtain standardized images with upright orientation, clear text, and uniform size, and then inputs these standardized images into an intelligent verification model trained on a neural network using historical weighbridge slip image data. The model's recognition module performs character recognition to obtain the weighbridge slip text sequence, and the verification module processes the text sequence to output the verification result. This method utilizes standardized preprocessing to eliminate the negative impact of image tilt, blurriness, and inconsistent sizes on recognition accuracy. Based on the trained neural network model, it automates the process of character extraction and compliance verification, achieving efficient, accurate, and intelligent weighbridge slip verification. This effectively reduces the human resource input and subjective errors of manual verification, improves the efficiency and accuracy of weighbridge slip verification, and ensures the compliance and consistency of transportation measurement information.

[0114] Figure 2 A flowchart illustrating a smart verification method for weighbridge slips provided in this application embodiment. Figure 2 ,like Figure 2 As shown, in this embodiment... Figure 1 Based on the examples, a method for intelligent verification of weighbridge slips is described in detail. The method includes:

[0115] S201. Obtain the weighbridge image to be processed.

[0116] For example, this step can be referred to step S101, and will not be described again.

[0117] S202. Perform tilt correction processing on the weighbridge image to be processed to obtain the processed weighbridge image; perform data processing on the processed weighbridge image to obtain a standardized image.

[0118] The data processing includes one or more of the following: noise reduction, blur enhancement, occlusion processing, and size normalization.

[0119] For example, the weighbridge image to be processed refers to the digital image of the weighing and measurement document obtained by means of photography, scanning, etc. Its content usually includes tabular data such as the name of goods, gross weight, tare weight, and net weight. The image is often tilted, blurry, or dirty during shooting.

[0120] Tilt correction processing refers to using image analysis to detect the tilt angle of the border or text lines on a weighbridge, and then correcting the document content to a horizontal or vertical state through rotation transformation.

[0121] Data preprocessing is the quality optimization step for the corrected image. Noise reduction suppresses interference such as photosensitive noise and compression artifacts. Blur enhancement addresses blur caused by camera shake or defocusing by sharpening or deconvolution to enhance text edges and improve legibility. Occlusion removal restores obscured key information areas covered by stamps, stains, or folds using image restoration techniques. Size normalization scales the image to a uniform preset resolution or dimensions to eliminate scale differences. The resulting standardized image is an ideal input with regular geometry, clear content, and uniform scale, suitable for direct use in text recognition or archiving.

[0122] For example, the bill image is converted to grayscale, and then edge detection is performed on the grayscale image of the bill to extract the edges of the bill and the straight lines of the table; then, the main direction straight lines in the image are detected based on the Hough transform, and the tilt angle is calculated based on the straight line fitting result. Then, an affine transformation matrix is ​​constructed, and inverse mapping and bilinear interpolation are performed on the image to rotate the bill image to a horizontal state.

[0123] After tilt correction, when processing the processed weighbridge image, Gaussian filtering can be used to smooth and suppress salt-and-pepper noise and Gaussian noise. Histogram equalization can be used to enhance text boundaries in low-contrast areas. Image inpainting algorithms can also be used to complete areas covered by stamps, stains, or obstructions. Finally, the image is uniformly scaled to a preset resolution to eliminate scale differences between images from different sources. The preset resolution can be set to a uniform width and height to facilitate the subsequent model receiving input of fixed specifications. In practical applications, other specifications can also be selected for this resolution, which is not limited in this application.

[0124] For example, the system receives a color or grayscale image containing a bill, uses an edge detection operator (such as Canny) to extract the outline segments of the bill in the image, then uses Hough transform to detect the most prominent cluster of lines, calculates the angle between the line and the horizontal or vertical direction as the tilt angle, and then performs a reverse rotation transformation based on the image center to complete the tilt correction, outputting an upright bill image. Based on this, data processing is performed sequentially as needed: first, noise reduction is performed using Gaussian filtering or bilateral filtering to smooth flat areas while preserving text edges; if the image has overall or local blur, blur enhancement is performed using an anti-sharpening mask or Wiener filtering to sharpen numbers and table lines; for areas obscured by stamps or stains, the occlusion mask is located by detecting features such as hue and saturation, and then the obscured text strokes are filled using a fast traversal method or a deep learning repair model; finally, bilinear or bicubic interpolation is used to scale the entire image to a specified size, for example, fixing the width to 800 pixels and adjusting the height proportionally, or uniformly filling it to 800×600 pixels to obtain a standardized image. The entire process is a pipeline-style process, where the output of the previous stage directly serves as the input for the next stage.

[0125] Through the above processing, tilt distortion, noise interference, blurred areas, occluded areas, and size differences in the weighbridge slip image are uniformly corrected. The output standardized image maintains consistency in orientation, clarity, and scale, and can be directly used as input for subsequent recognition and verification models, thereby improving the stability of weighbridge slip text extraction and the accuracy of verification results.

[0126] The process of performing tilt correction on the weighbridge image to obtain the processed weighbridge image includes:

[0127] The bill image to be processed is converted to grayscale to obtain a grayscale image of the bill. Based on the Canny edge detection algorithm, the text lines in the grayscale image of the bill are extracted to obtain the outline of the text lines. Based on the Hough transform detection algorithm, the outline of the text lines is fitted with a straight line to obtain the tilt angle of the text lines. With the center of the grayscale image of the bill as the rotation center, an affine transformation matrix is ​​constructed according to the tilt angle of the text lines. Based on the affine transformation matrix, the grayscale image of the bill is reverse-mapped, and bilinear interpolation is used to fill the sub-pixel positions, rotating the image to a horizontal state to obtain the processed bill image.

[0128] For example, grayscale conversion processing refers to converting a color bill image into a single-channel image containing only brightness information using a brightness weighting formula (such as using coefficients 0.299R+0.587G+0.114B) to eliminate interference from stamp color, background texture, etc. on the text structure. The generated grayscale image of the bill is the unified basis for subsequent processing.

[0129] The Canny edge detection algorithm is a multi-level edge detection operator that accurately extracts the edges of text strokes and table lines in images through Gaussian smoothing, gradient magnitude and direction calculation, non-maximum suppression, and double threshold hysteresis connection.

[0130] The outline of a line of text is a collection of connected edge segments that are generally horizontal and consist of these edge points.

[0131] The Hough transform detection algorithm maps edge points in the image space to parameter space, and finds collinear peaks through voting accumulation, thereby achieving straight line fitting of the text line contour, that is, using straight line segments to approximate the mainstream direction of the text line.

[0132] The tilt angle of the text line refers to the angle between the fitted straight line and the horizontal direction, reflecting the degree of skewness of the entire bill. The affine transformation matrix is ​​used to describe linear mappings such as rotation and translation; in this step, this matrix is ​​constructed with the center of the bill's grayscale image as the rotation center to ensure that the image rotates around its own center without shifting.

[0133] Inverse mapping is used to avoid image holes caused by forward rotation. It reverses the coordinates of each pixel in the target image to the corresponding coordinates in the source image. Bilinear interpolation uses the gray values ​​of the four integer pixels around the sub-pixel coordinates to linearly interpolate in two directions to fill in the gray values ​​at non-integer positions. Finally, it produces a package image with horizontal text lines and smooth edges.

[0134] For example, the bill image can come from photographed, scanned, or screenshotted data. It is first converted to a grayscale image and then input into an edge detection module. The edge detection module performs smoothing, gradient calculation, and non-maximum suppression on the grayscale image to extract the boundary contours of the text lines. The contour information is then input into a Hough transform module, which performs line accumulation voting according to a preset angle resolution to obtain the tilt angle corresponding to the main direction of the text lines. Using the image center as the rotation center, a two-dimensional affine transformation matrix is ​​constructed based on this tilt angle. The source coordinates of each target pixel are then calculated using a reverse mapping method. For positions where the source coordinates fall between integer pixels, bilinear interpolation is used to calculate the pixel value, thereby generating a horizontally processed bill image.

[0135] This processing method uses the overall direction of the text lines as a correction benchmark, which can rotate the tilted bill image to a horizontal state, making the interline structure more stable during subsequent text recognition and reducing character misalignment and line positioning deviation caused by tilt. After grayscale conversion, edge extraction, Hough fitting, and interpolation reconstruction, the text area of ​​the output image remains continuous, which facilitates text extraction and field matching in the subsequent review module.

[0136] For example, the original color image of the bill is converted pixel by pixel using a weighted grayscale conversion formula to output a grayscale image of the bill, thus removing color interference. Next, Canny edge detection is performed on the grayscale image: first, Gaussian filtering is used for noise reduction; then, the magnitude and direction of the gradient are calculated using a first-order partial derivative operator; subsequently, non-maximum suppression is applied to the gradient magnitude to refine the edges; finally, a dual threshold (e.g., a low threshold of 50 and a high threshold of 150) is used to filter strong edges and connect weak edges, resulting in a binary edge image that highlights the strokes of the text lines and the outlines of the table lines. Then, all foreground pixels in this edge image are used as input to the Hough transform. An angular resolution (e.g., 1 degree) and an accumulator voting threshold are set to detect several significant line segments in the image. The angles between these line segments and the horizontal direction are calculated, and after removing outliers, the mean or mode is taken to determine the tilt angle θ of the text lines. Then, an affine transformation matrix is ​​constructed based on the center coordinates (cx, cy) and angle θ of the original grayscale image. This matrix includes rotation factors cosθ, -sinθ, sinθ, cosθ, and a translation amount calculated to compensate for image position, ensuring the image center is aligned after rotation. Finally, a blank target image of the same size as the source image is created. Each pixel (x1, y1) in the target image is traversed, and the corresponding sub-pixel coordinates (x, y) in the source grayscale image are obtained through the inverse affine transformation. Bilinear interpolation is then performed using the grayscale values ​​of the four adjacent integer pixels around these coordinates, and the result is assigned to the target pixel. This point-by-point filling process generates a strictly horizontal, jagged, hole-free billboard image.

[0137] After adopting this implementation method, the directional consistency of the weighbridge image is improved, the text outline information is more suitable for subsequent recognition and processing, and the recognition error and field offset caused by tilt are reduced, thereby improving the accuracy and stability of weighbridge audit and processing.

[0138] S203. Input the standardized image into the intelligent audit model. The recognition module based on the intelligent audit model performs character recognition processing on the standardized image to obtain the weighing slip text sequence.

[0139] For example, the standardized image can be a weighbridge image that has undergone orientation correction, size normalization, and sharpening. The recognition module can consist of a convolutional feature encoding network and a sequence decoding network. The convolutional feature encoding network is used to extract global layout features and local character features of the weighbridge image, and an attention mechanism is introduced into the network to assign weights to text regions, generating an attention feature image. This attention feature image can retain the document layout information while highlighting the feature responses corresponding to text regions such as vehicle number, weight, and time. The character recognition module then performs character sequence recognition based on this feature response, outputting a weighbridge text sequence corresponding to the content of the weighbridge image. The network structure used can be a lightweight convolutional network adapted for long text recognition or a recognition network with an attention decoder. In practical applications, other models can also be selected for this component, and this application does not limit this.

[0140] Based on the above structure, the recognition module first enhances the features of the standardized image, and then performs character recognition on the enhanced text region features, so that the key information regions in the bill image are represented in a concentrated manner, reducing the interference of background, layout borders and non-text regions on the recognition results, and maintaining good integrity and continuity of the output bill text sequence.

[0141] By adopting this implementation method, the recognition module can form a clearer feature mapping for the text regions in the standardized image, improve the stability of text extraction in complex weighbridge scenarios, and provide a more accurate input weighbridge text sequence for the subsequent review module, thereby improving the automated processing effect and recognition consistency of weighbridge review.

[0142] Figure 3 This is a flowchart illustrating step S203 of a smart verification method for weighbridge slips provided in an embodiment of this application. Figure 3 As shown, step S203 includes:

[0143] S2031. The feature extraction module based on the recognition module performs feature extraction processing on the standardized image to obtain the weighbridge feature image.

[0144] Among them, the feature image representation of the weighing list is the feature map after the common visual features of each region in the standardized image are extracted equally.

[0145] For example, the feature extraction module is the front-end backbone network of the recognition module, which is usually composed of multiple layers of convolutional neural networks. Its function is to perform layer-by-layer scanning and convolution operations on the input standardized image to extract various low-level visual elements (such as edges, corners, and textures) and high-level semantic structures (such as table lines and character strokes) in the image.

[0146] After standardizing the image input recognition module, the feature extraction module can use a convolutional neural network to encode the image in multiple layers. The network can include continuous convolutional layers, normalization layers, and nonlinear activation layers to extract visual features at different scales.

[0147] The feature image of the weighing list can be a single-channel feature map or a multi-channel feature map. The spatial resolution of the feature map can be set to one-quarter, one-eighth or lower of the original image according to the network downsampling rate, so as to balance text region localization and computational efficiency.

[0148] For example, the preprocessed standardized image is fed into the feature extraction module within the recognition module. This module typically has a convolutional neural network backbone with the fully connected layers removed, such as a stack of multiple convolutional and pooling layers. When the input image passes through the first convolutional layer, a set of filters slides across the entire image with a fixed stride. Each filter performs an inner product operation on all local receptive fields of the image equally, generating a corresponding feature response map. Since the parameters of the convolutional kernels are spatially shared, the calculation rules are completely consistent regardless of whether the kernel slides to a form header, a densely populated area of ​​numbers, or a blank corner, thus achieving equal extraction of general visual features. Subsequently, through activation functions and nonlinear mappings, and after dimensionality reduction via pooling layers, the shallow feature maps are gradually combined into high-level feature maps that express more complex shapes or components. This process progresses layer by layer, but always maintains spatial equality. Finally, the output of the feature extraction module is one or more feature images with the same spatial correspondence, where each pixel position encodes the general visual features of the corresponding region of the original standardized image, without pre-cropping or weighting any region.

[0149] By extracting universal visual features equally across all regions of a standardized image through a feature extraction module, the risk of information loss due to premature spatial bias or attention clipping is completely avoided. For billboards with ever-changing layouts, varying numbers of table lines, and arbitrarily shifting field positions, this undifferentiated global feature preservation is crucial. It enables subsequent recognition modules to flexibly utilize region proposal or segmentation algorithms on a complete, intact feature base, autonomously discovering all potential text lines and key data areas without being constrained by rigid prior layout rules.

[0150] S2032. Based on the attention algorithm, feature enhancement processing is performed on the text region of the bill feature image to obtain the attention feature image.

[0151] Among them, the attention feature image representation is a differential feature map after weighted enhancement of text region features and suppression of non-text background region features in the bill feature image.

[0152] For example, the attention algorithm is a deep learning computational method that simulates the human visual attention mechanism. Its core idea is to enable the model to automatically allocate limited computational resources and assign different importance weights to different parts of the input information. In this step, it is applied to the previously generated bill feature image. Although this feature image uniformly carries the general visual features of the entire image, it has not yet distinguished which areas contain key text information and which areas are irrelevant table shading or blank backgrounds.

[0153] Text regions refer to local spatial locations in a weighbridge image where characters are printed or written, such as license plate numbers, net weight values, and date fields. Feature enhancement processing involves generating a set of spatial dimension weight coefficients using an attention algorithm and multiplying them element-wise with the original feature map. Specifically, weighted enhancement multiplies the feature values ​​corresponding to locations identified as text regions by a weight greater than 1 or a relatively amplified weight, increasing their response intensity; while suppression multiplies the feature values ​​identified as non-text background regions by a weight less than 1 or close to zero, weakening their expression.

[0154] An attention feature image is a spatially differentiated feature map—features in text regions are highlighted and sharpened, while features in background regions are suppressed and blurred, thus providing a clear focus for subsequent character recognition.

[0155] For example, the feature image of the bill is input into a lightweight attention branch network. This branch network typically consists of several convolutional or fully connected layers, and its task is to learn to predict the probability or importance score of each spatial location belonging to a text region. The input feature map is transformed layer by layer on it, and finally outputs a single-channel image whose spatial size is consistent with the input feature map. Figure 1 The attention weight map is generated, where each pixel value, after being normalized by the Sigmoid function, ranges between 0 and 1. Next, a feature recalibration operation is performed: this attention weight map is multiplied element-wise by channel in a broadcast manner with the original bill feature map. For locations where the weight value approaches 1, i.e., areas identified by the model as text regions, the original features are almost completely preserved or even doubled, thus achieving weighted enhancement; for locations where the weight value approaches 0, i.e., non-text regions identified as background, texture, or stains, the original feature values ​​are significantly attenuated to near zero, thus achieving feature suppression. The result of the multiplication is the spatially differentiated attention feature image. In this process, the attention branch network is not pre-programmed but trained end-to-end using a large amount of labeled bill data, enabling it to accurately distinguish the pattern differences between text stroke details and interfering background in the feature space.

[0156] Through the above processing, the standardized image first forms a bill of lading feature image with global layout information, and then an attention mechanism focuses on the text region and outputs an attention feature image, making the input features of the recognition module more focused on effective character information. This structure can improve the character representation ability under complex layouts, blurred backgrounds, or partial occlusion conditions, thereby improving the accuracy and consistency of bill of lading text recognition.

[0157] S2033. The character recognition module based on the recognition module performs character recognition processing on the attention feature image to obtain the bill text sequence.

[0158] For example, the character recognition module receives an attention feature image of size H×W×C. Since the subsequent sequence recognition model requires a sequence of feature vectors as input, the module first converts the two-dimensional feature map into a series of one-dimensional feature vectors through a feature serialization operation. Optionally, the conversion method involves extracting feature slices column by column along the image width with a fixed fine-grained width. The C-dimensional feature vector corresponding to each column describes the visual information of a vertical strip region in the original weight sheet image.

[0159] These sequentially arranged feature vector sequences are fed into a deep bidirectional long short-term memory network, or Transformer encoder. The core function of this network is to model long-range contextual dependencies between characters; for example, it strengthens the expectation of subsequent numbers based on the preceding term "gross weight," and corrects the recognition bias of intermediate numbers using the following "kg" unit. The bidirectional structure ensures that the prediction at each time step takes into account the contextual information of the entire sequence.

[0160] Finally, a transcription layer consisting of a fully connected layer and a softmax activation function maps the context feature vector to the probability distribution on the character set step by step. By connecting the temporal classification loss function or autoregressive decoding, duplicate predicted labels and blank placeholders are removed, and finally the probability sequence is folded and spliced ​​into a complete list text sequence.

[0161] By decoding the attention feature image through the character recognition module, an end-to-end accurate transition from visual features to semantic information is achieved, directly converting the weighbridge image into structured text data that can be directly processed by a computer, completely eliminating the inefficiency of traditional manual data entry.

[0162] In one example, the character recognition module of the recognition module performs character recognition processing on the attention feature image to obtain the initial text sequence; based on the large language model, the initial text sequence is corrected to obtain the bill of lading text sequence.

[0163] For example, the character recognition module is the back-end sequence decoder of the recognition module. It usually adopts a recurrent neural network or Transformer decoder based on the attention mechanism. Its function is to read the feature vectors of each spatial location in the attention feature image, predict the character category one by one according to the time step, and finally concatenate them into a continuous text string.

[0164] The initial text sequence is the original text content that the model decodes directly from visual features. It already has basic character shape recognition capabilities, but it may still contain partial misspellings, missing characters, or garbled characters when faced with complex situations such as stroke occlusion, blurring, and similar-looking characters.

[0165] Large language models refer to generative language models that have been pre-trained on massive amounts of text corpora. They store rich language knowledge, including word collocation habits, industry terminology, number format specifications, and contextual semantic relationships.

[0166] For example, after receiving the attention feature image, the character recognition module can perform temporal encoding on the character regions in the image based on a sequence modeling network and output the corresponding character probability distribution. Then, it generates an initial text sequence according to the maximum a posteriori rule or decoding rule. This character recognition module can adopt a structure combining convolutional feature extraction and recurrent decoding, or it can adopt a recognition structure based on attention decoding to adapt to the recognition needs of mixed text such as vehicle number, weight, date, and serial number in the weighbridge slip. In practical applications, other models of this component can also be selected, and this application does not limit this. After receiving the initial text sequence, the large language model uses it as the input text to be corrected. Combining the business context of the weighbridge slip, field expression habits, and contextual dependencies, it reconstructs the text with character misalignment, homonym misrecognition, or local missing characters, and outputs the corrected weighbridge slip text sequence. The correction process can be achieved through prompt word constraints, context rearrangement, or candidate text re-scoring to make the text result more consistent with the standardized expression of the weighbridge slip fields.

[0167] This scheme first generates an initial text sequence using a character recognition module, and then performs semantic correction using a large language model. This allows the recognition results to no longer rely solely on single character matching, but further utilize contextual relationships to complete error correction and completion, thereby improving the accuracy and consistency of the bill of lading text sequence and providing a more reliable text foundation for the subsequent review module.

[0168] S204. The audit module based on the intelligent audit model processes the weighbridge text sequence to obtain the audit result.

[0169] For example, this step can be referred to step S103, and will not be described again.

[0170] Figure 4 This is a flowchart illustrating step S204 of a smart verification method for weighbridge slips provided in an embodiment of this application. Figure 4 As shown, step S204 includes:

[0171] S2041. The audit module based on the intelligent audit model performs feature extraction processing on the weighbridge slip text sequence to obtain a weighbridge slip information set.

[0172] S2042. Based on the preset regular expression of the audit module, the information in the weighbridge information set is modified to obtain a standardized weighbridge information set.

[0173] S2043. Based on the multi-level audit rules of the audit module, the standardized set of weighbridge information is matched at multiple levels to obtain the audit results.

[0174] For example, the intelligent review model is an independent post-processing logic unit that is distinct from the pre-recognition module. It is embedded in the review module and its core capability is to perform structured understanding and logical verification of unstructured bill of lading text sequences. In essence, it can be a hybrid model that combines rules and lightweight machine learning.

[0175] The weighbridge information set refers to the structured field set extracted by the intelligent auditing model from a continuous text sequence and organized in the form of key-value pairs, such as {license plate number: Shaanxi X12345, gross weight: 40000kg}.

[0176] Preset regular expressions are a series of pre-defined string matching rule templates within the audit module, specifically used for format cleaning and standardization correction of field values. Multi-level audit rules refer to a progressive set of validation rules set according to business logic, typically including multiple levels such as format compliance, numerical logic rationality, data consistency, and historical comparison.

[0177] The audit result is a structured judgment output, which not only includes the corrected complete fields, but also the validation status, anomaly flags, and alarm information for each field.

[0178] For example, the audit module receives the weighing slip text sequence output by the preceding character recognition module. The first step of the intelligent audit model is feature extraction, where the features are no longer visual but semantic features. The model uses named entity recognition technology or a dictionary-based maximum matching algorithm, combined with field keyword anchors, to locate and extract values ​​from the text sequence. For example, it first finds the term "gross weight," then extracts the subsequent consecutive numbers and the unit 40000kg as a complete value to form the original weighing slip information set.

[0179] The review module iterates through each item in the collection, calling the corresponding preset regular expression for matching and cleaning. If the date field value is identified as May 1B, 2024, the regular expression (\d{4})year(\d{1,2})month(\d{1,2})day will fail to match or encounter an anomaly. The correction rule will attempt to change B to 1 and reformat the data. After all fields have been corrected by regularization, a standardized set of weighbridge information is obtained.

[0180] Finally, multi-level matching processing is performed: Level 1 rules verify whether the format of all fields conforms to the regular expression definition; Level 2 rules perform numerical logic verification, such as determining whether the net weight equals the gross weight minus the tare weight; Level 3 rules compare the extracted values ​​with historical weighbridge data or external system whitelists, and the final audit result is generated by combining the verification results of all levels.

[0181] In one example, the weighbridge information set includes basic information, weight information, and related information. If it is determined that the basic information in the weighbridge information set does not conform to a preset regular expression, the basic information in the weighbridge information set is corrected to obtain corrected basic information. If it is determined that the difference between the weight information in the weighbridge information set and the preset weight information is greater than or equal to a preset threshold, the weight information is corrected according to the preset weight information to obtain corrected weight information. Based on the corrected basic information, the corrected weight information, and the related information, a standardized weighbridge information set is determined.

[0182] For example, the intelligent auditing model categorizes the extracted fields into three types: basic information, weight information, and related information according to preset classification rules.

[0183] Understandably, the weighbridge information set is a structured data body obtained from the unstructured weighbridge text sequence after feature extraction. It parses the originally coherent text content into a set of key-value pairs with clear business meanings and divides them into three logical categories according to business attributes: basic information, weight information, and related information.

[0184] Basic information constitutes the identity and management dimension of the weighbridge slip, including license plate number, weighbridge slip number, carrier, driver, operator, and remarks. These fields answer the questions of whose weighbridge slip it is, who transports it, and who operates it, and are the core basis for the traceability and statistical classification of weighbridge slip circulation responsibility.

[0185] Weight information forms the core of the weighbridge, including three basic weighing values: gross weight, tare weight, and net weight, as well as the original net weight, which is used as a reference value for verifying the shipment volume. It is also equipped with time and unit fields such as weight unit, gross weight time, tare weight time, and weighing time. These data together record the complete weighing process trajectory, and there is an internal logical relationship that gross weight minus tare weight equals net weight, which is the key to subsequent business review calculations.

[0186] The associated information expands the business scenario dimensions of the weight note, including material type, transportation route and vehicle model. The material type indicates the variety of goods being transported, the transportation route records the origin and destination in some weight notes, and the vehicle model marks the specification of the transport vehicle. These fields provide support for analyzing the flow of goods, checking the matching degree of transport capacity, and counting the transportation status of different materials. Overall, the weight note information set converts the text in the image into structured data that can be directly processed by a computer, and establishes a clear business logic hierarchy through information classification, so that subsequent format verification, logic verification and business review can implement differentiated processing strategies for different types of fields, laying a structured foundation for the automatic review of weight notes and data warehousing.

[0187] For the basic information, the review module sequentially calls the preset regular expressions corresponding to each field for format matching. If the format of a certain field does not match, a correction procedure is started: based on the predefined character confusion mapping table or semantic inference rules, an attempt is made to correct the recognition error to a format accepted by the regular expression.

[0188] For the weight information, the review module first extracts it as a pure numerical value and compares it with the preset weight information. If the absolute difference between the two is less than the preset threshold, it is regarded as an acceptable low-risk error, and the system automatically overwrites the recognized value with the accurate preset weight information to complete silent correction; if the difference is greater than or equal to the preset threshold, it is determined as a serious recognition error or data abnormality, and the system will record the abnormality flag, but still perform forced correction according to the preset weight information to ensure that the subsequent business process is uninterrupted. The associated information is usually retained directly and does not participate in logical correction.

[0189] Finally, the corrected basic information, corrected weight information and unchanged associated information are recombined and packaged into a standardized weight note information set.

[0190] In a possible implementation, for key fields such as license plate number and weight value, a preset regular expression is called for pattern matching, for example, "Yu A\d{6}" or "Ji B\d{6}" is used to match the license plate format of new energy heavy trucks, and "\d+\.\d+ ton" or "\d+ ton" is used to match the weight field to verify whether the data conforms to the established writing specifications. At the same time, the system embeds context logic correction rules, the most typical of which is the automatic verification of the net weight field: calculate the difference between the gross weight and the tare weight. If the deviation between the extracted net weight and the calculation result exceeds 0.1 ton, it is determined that the extracted value may be incorrect, and the net weight value is automatically corrected to the calculated value to ensure the arithmetic self-consistency of the basic data.

[0191] In one example, the standardized weighbridge slip information set includes revised basic information, revised weight information, and related information. Based on the multi-level review rules of the review module, the revised basic information in the standardized weighbridge slip information set is matched with the basic information in the multi-level review rules to obtain the information matching degree. Based on the multi-level review rules of the review module, the revised weight information in the standardized weighbridge slip information set is matched with the logical information in the multi-level review rules to obtain the logical correctness rate. Based on the multi-level review rules of the review module, the related information in the standardized weighbridge slip information set is matched with the related information in the multi-level review rules to obtain the data correlation degree. The review result is determined based on the information matching degree, logical correctness rate, and data correlation degree.

[0192] For example, the multi-level review rules are a pre-set, hierarchical, and multi-dimensional verification rule library within the review module. It is not a single condition, but an organic system composed of three types of rules: basic information rules, logical information rules, and related information rules.

[0193] Specifically, the basic information rules include standard templates that can be matched, such as a whitelist of license plate numbers, a date format library, and a dictionary of unit names; the logical information rules specifically refer to the mathematical identities that must be satisfied between gross weight, tare weight, and net weight; and the association information rules define the cross-reference relationships between fields such as remarks, operators, and warehouses and the master data of the business system.

[0194] Information matching degree, logical correctness rate, and data relevance are the three quantitative audit indicators in this step, which respectively measure the consistency between basic information and external standards, the rigor of the inherent mathematical logic of weight data, and the cross-validation of related information and business system data.

[0195] For example, the audit module initiates the first level of matching in a multi-level audit. It retrieves the corrected basic information from the standardized set of weighbridge information and performs matching operations by calling the basic information rules in the multi-level audit rule base field by field. For example, it compares the license plate number with the company's whitelist of registered vehicles and verifies the date field against the legal date format library. For each successfully matched field, a score is accumulated, and finally, the percentage of successfully matched fields out of the total number of basic information fields is calculated to obtain the information matching degree.

[0196] The audit module initiates the second level of logical verification in a multi-level audit process. The audit module extracts the corrected weight information and substitutes the gross weight, tare weight, and net weight values ​​into the logical information rules within the multi-level audit rules for verification. The core rule checks whether |gross weight - tare weight - net weight| is less than the allowable error. If they are completely equal or the error is negligible, the logical verification passes. If the weighbridge slip contains multiple weighing records, each record is verified individually. The percentage of successful logical verifications out of the total number of logical verifications is the logical accuracy rate.

[0197] The audit module initiates the third level of cross-validation in the multi-level audit process. The audit module matches fields such as notes and warehouses in the associated information with preset external data sources in the multi-level audit rules. For example, it checks whether the warehouse actually exists or whether the shipping unit is in the list of qualified suppliers. The ratio of successfully matched associated items to the total number of associated items is the data correlation degree.

[0198] Finally, the review module performs a weighted summation or conditional judgment on the information matching degree, logical correctness rate and data relevance based on the pre-set weights and thresholds to obtain the final review result.

[0199] In one example, if the information matching degree is determined to be greater than or equal to the first preset threshold, and the logical correctness is greater than or equal to the second preset threshold, and the data correlation degree is greater than or equal to the third preset threshold, then the review result indicates that the review has passed; otherwise, the review result indicates that the review has failed and requires manual review, and the review result is sent to the user.

[0200] For example, the audit results include two distinct states: one is that the audit is passed, which means that the weighbridge data is completely reliable and can be automatically entered into the settlement and archiving process; the other is that the audit is failed and requires manual review, which means that there are one or more doubts about the weighbridge, and the system automatically intercepts and generates a pending task to be pushed to the designated auditor.

[0201] For example, after the audit module completes the calculation of quantitative indicators across three dimensions, it enters the final decision-making stage. This stage is executed by a pre-defined conditional judgment logic unit, which internally sets up three independent comparators. The first comparator receives the calculated information matching degree value and compares it with a first preset threshold. If the condition is met, it outputs a binary status bit as true; otherwise, it outputs false. The second comparator synchronously receives the logic accuracy rate value and compares it with a second preset threshold. If the condition is met, it outputs true. The third comparator receives the data correlation degree value and compares it with a third preset threshold. If the condition is met, it outputs true.

[0202] A three-input AND gate comprehensively determines the three status bits. The logic unit outputs a final approval signal only when all three inputs are true, automatically marking the weighbridge data as approved and triggering subsequent accounting or report update processes. If any one or more of the three status bits are false, the logic unit outputs an approval failure signal. In this case, the approval module immediately generates a manual review work order, packaging the standardized weighbridge information set, the specific scores for the three dimensions, details of any unmet indicators, and expected standards. This order is then sent to the preset user terminal via push notification or email, awaiting intervention from the reviewer.

[0203] In one possible implementation, if the information matching degree is determined to be greater than or equal to 95%, a binary status bit "true" is output; if the information matching degree is determined to be less than 95%, a binary status bit "false" is output; alternatively, if the information matching degree is determined to be less than 95% but greater than or equal to 85%, a binary status bit "false 1" is output; if the information matching degree is determined to be less than 85%, a binary status bit "false 2" is output; understandably, "false 1" indicates that the current weighbridge image may have some fields entered incorrectly or have recognition deviations, requiring manual review; "false 2" indicates that the current weighbridge image may be forged or completely unrelated to the current business, requiring manual review.

[0204] In one possible implementation, if the logical correctness is determined to be greater than or equal to 100%, a binary status bit "true" is output; if the information matching degree is determined to be less than 100%, a binary status bit "false" is output; alternatively, if the information matching degree is determined to be less than 100% but greater than or equal to 80%, a binary status bit "false 1" is output; if the information matching degree is determined to be less than 80%, a binary status bit "false 2" is output; understandably, "false 1" indicates that some data points in the current weighbridge image are contradictory and require manual review; "false 2" indicates that the internal data of the current weighbridge image is seriously inaccurate and requires manual review.

[0205] In one possible implementation, if the data correlation is determined to be greater than or equal to 90%, a binary status bit "true" is output; if the information matching degree is determined to be less than 90%, a binary status bit "false" is output; alternatively, if the information matching degree is determined to be less than 90% but greater than or equal to 75%, a binary status bit "false 1" is output; if the information matching degree is determined to be less than 75%, a binary status bit "false 2" is output; understandably, "false 1" indicates that the correlation of the current weighbridge image is questionable and requires manual verification; "false 2" indicates that the current weighbridge image may have been tampered with and requires manual verification.

[0206] Understandably, the system performs a comprehensive judgment based on the results of the three-level review: if all three levels of review are qualified, the entire weighbridge slip passes the review, and the data can be automatically transferred to the settlement process; if any level is unqualified, the entire weighbridge slip fails the review and is rejected from being entered into the warehouse; if there are no unqualified items but any suspicious items exist, the weighbridge slip is marked as requiring manual review, and all anomaly details are output to guide the reviewers to quickly locate the problematic fields. This hierarchical, progressive, and quantitatively evaluated automated review mechanism transforms the complex verification process that originally relied on human experience into a configurable, interpretable, and traceable machine judgment process, significantly improving the accuracy, efficiency, and business risk control capabilities of weighbridge slip review. This review module transforms the weighbridge slip text sequence into a structured set of weighbridge slip information and uses regular expressions for item-by-item matching, enabling text review to shift from overall judgment to field-level judgment, thereby improving the consistency of identification of key content such as vehicle number and weight. Since the matching rules directly affect the extracted information units, the review results can more accurately reflect the format compliance and content completeness of the weighbridge slip text sequence, thereby improving the automation processing capability and accuracy of weighbridge slip review.

[0207] This application provides an intelligent verification method for weighbridge slips. First, it acquires a weighbridge slip image and performs tilt correction. Then, it completes data processing through one or more of the following: noise reduction, blur enhancement, occlusion removal, and size normalization, resulting in a standardized image. This standardized image is then input into an intelligent verification model. The feature extraction module of the recognition module extracts attention feature images representing the feature mapping of text regions. These are then processed by the character recognition module to obtain the weighbridge slip text sequence. Finally, the verification module extracts a set of weighbridge slip information and uses preset regular expressions to match each piece of information within the set. This image preprocessing eliminates recognition interference caused by external factors such as shooting tilt, image quality loss, and partial occlusion, unifying the image input standard. Attention feature extraction then focuses on the text region, weakening the influence of irrelevant backgrounds and improving the accuracy of character recognition. Finally, regular expression matching completes automated rule verification of structured information, achieving fully automated and high-precision intelligent verification of weighbridge slips. This effectively reduces the workload and error risk of manual verification, improving the efficiency and data reliability of weighbridge slip verification.

[0208] Figure 5 A schematic diagram of the structure of a smart weighbridge verification device provided in this application embodiment. Figure 1 ,like Figure 5 As shown, the intelligent verification device 50 for weighbridge slips provided in this embodiment includes:

[0209] Acquisition unit 501 is used to acquire the weighbridge image to be processed;

[0210] Processing unit 502 is used to preprocess the weighbridge image to be processed to obtain a standardized image;

[0211] The review unit 503 is used to input the standardized image into the intelligent review model. The recognition module based on the intelligent review model performs character recognition processing on the standardized image to obtain the weighing slip text sequence. The review module based on the intelligent review model processes the weighing slip text sequence to obtain the review result.

[0212] This embodiment provides a smart verification device for weighbridge slips, which can execute the methods provided in the above-described method embodiments. Its implementation principle and technical effects are similar, and will not be described in detail here.

[0213] Figure 6 A schematic diagram of the structure of a smart weighbridge verification device provided in this application embodiment. Figure 2 ,like Figure 6 As shown, the intelligent verification device 60 for weighbridge slips provided in this embodiment includes:

[0214] Acquisition unit 601 is used to acquire the weighbridge image to be processed;

[0215] Processing unit 602 is used to preprocess the weighbridge image to be processed to obtain a standardized image;

[0216] The review unit 603 is used to input standardized images into the intelligent review model. The recognition module based on the intelligent review model performs character recognition processing on the standardized images to obtain the weighbridge slip text sequence. The review module based on the intelligent review model processes the weighbridge slip text sequence to obtain the review result.

[0217] In one possible implementation, the review unit 603 is specifically used for:

[0218] The feature extraction module based on the recognition module performs feature extraction processing on the standardized image to obtain the weighbridge feature image;

[0219] Based on the attention algorithm, feature enhancement processing is performed on the text region of the bill feature image to obtain the attention feature image;

[0220] The character recognition module based on the recognition module performs character recognition processing on the attention feature image to obtain the bill of lading text sequence.

[0221] In one possible implementation, the review unit 603 is further configured to:

[0222] The character recognition module based on the recognition module performs character recognition processing on the attention feature image to obtain the initial text sequence;

[0223] Based on the large language model, the initial text sequence is corrected to obtain the list text sequence.

[0224] In one possible implementation, the review unit 603 is further configured to:

[0225] The review module based on the intelligent review model performs feature extraction processing on the weighbridge slip text sequence to obtain a set of weighbridge slip information;

[0226] Based on the preset regular expressions of the audit module, the information in the weighbridge information set is modified to obtain a standardized weighbridge information set.

[0227] Based on the multi-level review rules of the review module, the standardized set of weighbridge information is matched at multiple levels to obtain the review results.

[0228] In one possible implementation, the weighbridge information set includes basic information, weight information, and related information; the review unit 603 is further configured to:

[0229] If it is determined that the basic information in the weighbridge information set does not conform to the preset regular expression, the basic information in the weighbridge information set is corrected to obtain the corrected basic information.

[0230] If the difference between the weight information in the weighbridge information set and the preset weight information is greater than or equal to the preset threshold, then the weight information is corrected according to the preset weight information to obtain the corrected weight information.

[0231] Based on the revised basic information, revised weight information, and related information, a standardized set of weighbridge information is determined.

[0232] In one possible implementation, the standardized weighbridge information set includes revised basic information, revised weight information, and associated information; the review unit 603 is further configured to:

[0233] Based on the multi-level review rules of the review module, the corrected basic information in the standardized weighbridge information set is matched with the basic information in the multi-level review rules to obtain the information matching degree.

[0234] Based on the multi-level review rules of the review module, the corrected weight information in the standardized weighbridge information set is matched with the logical information in the multi-level review rules to obtain the logical accuracy rate.

[0235] Based on the multi-level review rules of the review module, the associated information in the standardized weighbridge information set is matched with the associated information in the multi-level review rules to obtain the data correlation degree.

[0236] The review results are determined based on information matching degree, logical correctness rate, and data relevance.

[0237] In one possible implementation, the review unit 603 is further configured to:

[0238] If the information matching degree is greater than or equal to the first preset threshold, the logical correctness is greater than or equal to the second preset threshold, and the data correlation is greater than or equal to the third preset threshold, then the review result indicates that the review has passed; otherwise, the review result indicates that the review has failed and requires manual review, and the review result is sent to the user.

[0239] In one possible implementation, the processing unit 602 is specifically used for:

[0240] The weighbridge image to be processed is subjected to tilt correction processing to obtain the processed weighbridge image;

[0241] The processed weighbridge image is subjected to data processing to obtain a standardized image; the data processing includes one or more of the following: noise reduction processing, blur enhancement processing, occlusion processing, and size normalization processing.

[0242] This embodiment provides a smart verification device for weighbridge slips, which can execute the methods provided in the above-described method embodiments. Its implementation principle and technical effects are similar, and will not be described in detail here.

[0243] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 7 As shown, the electronic device 70 provided in this embodiment includes at least one processor 701 and a memory 702. Optionally, the electronic device 70 further includes a communication component 703. The processor 701, memory 702, and communication component 703 are connected via a bus 704.

[0244] In a specific implementation, at least one processor 701 executes computer execution instructions stored in memory 702, causing at least one processor 701 to perform the above-described method.

[0245] The specific implementation process of processor 701 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0246] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0247] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0248] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0249] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0250] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0251] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0252] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0253] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0254] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0255] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0256] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0257] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0258] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A method for intelligent verification of weighbridge slips, characterized in that, include: Obtain the weighbridge image to be processed; The weighbridge image to be processed is preprocessed to obtain a standardized image; The standardized image is input into the intelligent auditing model. The recognition module of the intelligent auditing model performs character recognition processing on the standardized image to obtain the weight slip text sequence. The auditing module of the intelligent auditing model processes the weight slip text sequence to obtain the audit result.

2. The method according to claim 1, characterized in that, The standardized image is input into the intelligent auditing model. Based on the recognition module of the intelligent auditing model, the standardized image undergoes character recognition processing to obtain the bill text sequence, including: Based on the feature extraction module of the recognition module, the standardized image is processed to obtain the weight slip feature image; Based on the attention algorithm, feature enhancement processing is performed on the text region of the bill feature image to obtain an attention feature image; The character recognition module based on the recognition module performs character recognition processing on the attention feature image to obtain the bill of lading text sequence.

3. The method according to claim 2, characterized in that, The character recognition module based on the recognition module performs character recognition processing on the attention feature image to obtain the billboard text sequence, including: The character recognition module based on the recognition module performs character recognition processing on the attention feature image to obtain an initial text sequence; Based on the large language model, the initial text sequence is modified to obtain the bill of lading text sequence.

4. The method according to claim 1, characterized in that, The audit module based on the intelligent audit model processes the weighbridge text sequence to obtain the audit results, including: The audit module based on the intelligent audit model performs feature extraction processing on the weighbridge slip text sequence to obtain a weighbridge slip information set; Based on the preset regular expressions of the audit module, the information in the weighbridge information set is modified to obtain a standardized weighbridge information set. Based on the multi-level review rules of the review module, the standardized set of weighbridge information is subjected to multi-level matching processing to obtain the review result.

5. The method according to claim 4, characterized in that, The weighbridge slip information set includes basic information, weight information, and related information; based on the preset regular expression of the review module, the information in the weighbridge slip information set is modified to obtain a standardized weighbridge slip information set, including: If it is determined that the basic information in the weighbridge information set does not conform to the preset regular expression, then the basic information in the weighbridge information set is corrected to obtain the corrected basic information; If it is determined that the difference between the weight information in the weighing slip information set and the preset weight information is greater than or equal to a preset threshold, then the weight information is corrected according to the preset weight information to obtain the corrected weight information; Based on the corrected basic information, the corrected weight information, and the associated information, the standardized set of weighbridge information is determined.

6. The method according to claim 4, characterized in that, The standardized weighbridge information set includes corrected basic information, corrected weight information, and associated information; Based on the multi-level review rules of the review module, the standardized weighbridge information set is subjected to multi-level matching processing to obtain the review result, including: Based on the multi-level review rules of the review module, the corrected basic information in the standardized weighbridge information set is matched with the basic information in the multi-level review rules to obtain the information matching degree. Based on the multi-level review rules of the review module, the corrected weight information in the standardized weighbridge information set is matched with the logical information in the multi-level review rules to obtain the logical accuracy rate. Based on the multi-level review rules of the review module, the associated information in the standardized weighbridge information set is matched with the associated information in the multi-level review rules to obtain the data correlation degree. The audit result is determined based on the information matching degree, the logical correctness rate, and the data correlation degree.

7. The method according to claim 6, characterized in that, The review result is determined based on the information matching degree, the logical correctness rate, and the data correlation degree, including: If the information matching degree is determined to be greater than or equal to the first preset threshold, and the logical correctness is greater than or equal to the second preset threshold, and the data correlation degree is greater than or equal to the third preset threshold, then the review result indicates that the review has passed; otherwise, the review result indicates that the review has failed and requires manual review, and the review result is sent to the user.

8. The method according to any one of claims 1-7, characterized in that, The weighbridge image to be processed is preprocessed to obtain a standardized image, including: The weighbridge image to be processed is subjected to tilt correction processing to obtain the processed weighbridge image; The processed weighbridge image is subjected to data processing to obtain a standardized image; wherein the data processing includes one or more of the following: noise reduction processing, blur enhancement processing, occlusion processing, and size normalization processing.

9. A smart verification device for weighbridge slips, characterized in that, include: The acquisition unit is used to acquire the weighbridge image to be processed. The processing unit is used to preprocess the weighbridge image to be processed to obtain a standardized image; The review unit is used to input the standardized image into the intelligent review model, and the recognition module of the intelligent review model performs character recognition processing on the standardized image to obtain the weight slip text sequence; the review module of the intelligent review model processes the weight slip text sequence to obtain the review result.

10. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-8.

Images