Intelligent logistics OCR intelligent receipt reading system

Abstract

The invention belongs to the technical field of crossing of intelligent logistics and optical character recognition, and particularly relates to an intelligent logistics OCR intelligent receipt reading system, which comprises an image acquisition module used for acquiring a paper scale receipt image by adopting a high-definition industrial camera matched with a light supplementing device, supporting automatic focusing and tilt correction, recording acquisition time and an equipment number at the same time, and sending the acquired image to a server; basic data is provided for subsequent traceability; the image preprocessing module is used for converting an original image into a high-contrast and interference-free standardized image through the technical means of gray processing, texture repairing and illumination normalization. According to the invention, through double accurate guarantee of context-associated line positioning and multi-modal feature fusion recognition, low error of a recognition result is realized.

Description

Technical Field

[0001] This invention relates to the field of intersection technology of smart logistics and optical character recognition, specifically a smart logistics OCR smart document reading system. Background Technology

[0002] In a smart logistics system, paper weighbridge slips are core documents recording cargo weight, transportation mileage, consignor and consignee information, and the basis for cost calculation. The accuracy and efficiency of their data entry directly impact logistics scheduling, financial settlement, and traceability management. Currently, the logistics industry primarily relies on two methods for processing paper weighbridge slips:

[0003] Manual data entry: Staff manually read the contents of the weighbridge slips and enter them into the system. This method has problems such as low efficiency (the average processing time for a single weighbridge slip is about 3-5 minutes), high human error rate (the error rate exceeds 8% due to fatigue and misreading of handwriting), and high labor costs (requiring dedicated personnel to be on duty at the weighbridge station). It is difficult to meet the document processing needs during peak logistics periods.

[0004] Traditional OCR recognition methods: Some companies have introduced general OCR systems to process weighbridge slips, but existing systems have significant shortcomings in real-world logistics scenarios. Specific problems and examples are as follows:

[0005] The problem of misaligned line recognition: In logistics transportation, weighbridge slips are prone to deformation due to folding, squeezing, or tilting. Traditional OCR only divides characters based on a single line boundary, failing to associate contextual logical relationships, which easily leads to field misalignment. For example, on a freight station's weighbridge slip, "Gross Weight: 3500kg" and "Net Weight: 3200kg" are adjacent. Due to document folds, the two lines partially overlap. Traditional OCR recognizes this as "Gross Weight: 350kg Net Weight: 00kg," incorrectly dividing the numerical values ​​and field correspondence, leading to subsequent weight calculation errors.

[0006] The problem of blurry handwriting recognition: Weighbridge slips often have blurred characters due to ink smudges, paper wear, and illegible handwriting. Traditional OCR relies on only a single image feature for recognition and cannot combine semantics and historical data for correction. For example, a logistics company's handwritten weighbridge slip number "NO.20240618-008" has blurred outlines of "6" and "8" due to ink smudges. Traditional OCR recognizes it as "NO.20240313-003", resulting in a mismatch between the document and the shipping order, affecting cargo traceability.

[0007] Format adaptation issues: Weighbridge slips generated by different freight stations and weighing equipment have significantly different formats (e.g., the "Shipper" field position: freight station A uses the upper left corner, while freight station B uses the upper right corner; field separation method: freight station C uses horizontal lines, while freight station D uses boxes). Traditional OCR requires training a separate template for each format. When a new format is introduced, 500+ samples need to be re-annotated and trained for 1-2 weeks, resulting in low adaptation efficiency. For example, when a logistics company adds a weighbridge slip from freight station E, because the "Unit of Weight" field has been moved from after "kg" to before the value (e.g., "kg2800"), traditional OCR, due to not matching the template, identifies "kg" as "abbreviated name of the consignee," leading to the loss of crucial weight information.

[0008] Based on the above, a smart logistics OCR smart document recognition system is invented. Summary of the Invention

[0009] This invention provides the following technical solution:

[0010] A smart logistics OCR smart document reading system, comprising:

[0011] The image acquisition module is used to acquire images of paper weighing slips using a high-definition industrial camera and a supplementary lighting device. It supports automatic focusing and tilt correction, and records the acquisition time and device number to provide basic data for subsequent traceability.

[0012] The image preprocessing module is used to transform the original image into a high-contrast, interference-free standardized image through grayscale processing, texture restoration, and illumination normalization techniques.

[0013] The context-related row positioning module is driven by both the field semantic logic library and historical position data. It first initially segments the row area, and then corrects misaligned rows through field association rules and position constraints to ensure that each field and its corresponding value belongs to the correct row area, providing accurate row-level data range for character recognition.

[0014] The multimodal feature fusion character recognition module is used to fuse the image features, semantic features and historical data features of characters, and through multi-dimensional feature complementarity verification, it can achieve accurate recognition of blurry and illegible characters and output correct recognition results at the single character level.

[0015] The identification result verification and anomaly handling module is used to automatically detect business logic anomalies in the identification results based on the logistics business rule base, and to perform differentiated processing through minor, moderate and severe classification strategies to intercept non-compliant data.

[0016] The adaptive layout learning module is used to automatically match fields of new layout weight slips by building a general field feature library;

[0017] The model iteration and optimization module is used to automatically collect error samples from each module, combine them with the results of manual review to label them with correct labels, and regularly conduct incremental training to update model parameters. At the same time, it monitors the recognition accuracy of each scene in real time and triggers abnormal iteration processes.

[0018] As a preferred embodiment of the intelligent logistics OCR intelligent document recognition system described in this invention, the image preprocessing module includes:

[0019] The stain removal module is used to eliminate paper stains and ink spots through grayscale thresholding and morphological operations.

[0020] The wrinkle repair module is used to identify wrinkled areas based on image texture features and restore a smooth image using an interpolation algorithm;

[0021] The illumination normalization module is used to balance the differences in brightness and darkness of the image using the Retinex algorithm, thereby improving the contrast between characters and the background.

[0022] As a preferred embodiment of the intelligent logistics OCR intelligent document recognition system described in this invention, the context-related line positioning module includes:

[0023] The weighbridge field logic library construction module is used to build the weighbridge field logic library in order to sort out the semantic association rules of the core fields of logistics weighbridges;

[0024] The row segmentation and correction module is used to first segment row regions using the projection method, and then correct misaligned rows based on the field logic library and historical row position data.

[0025] As a preferred embodiment of the intelligent logistics OCR intelligent document recognition system described in this invention, the multimodal feature fusion character recognition module includes:

[0026] The image feature module is used to extract the contour, edge, and grayscale distribution features of blurred characters, and generate feature vectors through a convolutional neural network.

[0027] The semantic feature module is used to constrain semantics based on field type and filter out unreasonable recognition results.

[0028] The historical data feature module is used to build a historical identification database for logistics weighbridge slips and corrects fuzzy characters through cosine similarity matching.

[0029] As a preferred embodiment of the intelligent logistics OCR intelligent document recognition system described in this invention, the recognition result verification and anomaly handling module includes:

[0030] The business rule base construction module is used to organize the business constraint logic of the core fields of logistics weighbridge slips and form a configurable rule base.

[0031] The anomaly detection and classification module first compares the structured data output by multimodal recognition with the business rule base and automatically marks the abnormal data; then, it classifies the anomalies into three levels—minor, moderate, and severe—based on the degree of impact, and corresponding to different processing strategies.

[0032] The exception log and feedback module is used to automatically record all exception data, form a structured exception log library, and periodically synchronize it to the model iteration and optimization module to provide direction for model optimization.

[0033] As a preferred embodiment of the intelligent logistics OCR intelligent document recognition system described in this invention, the rule base includes numerical rationality rules, field integrity rules, and cross-document association rules.

[0034] The numerical rationality rules are based on setting reasonable ranges and numerical logical relationships for weight, volume, and unit price for different types of goods.

[0035] The field integrity rules define a list of required fields and field format rules.

[0036] The cross-document association rule is to connect to the logistics order system, obtain the order data to be matched, and verify the consistency between the weighbridge identification result and the order data.

[0037] As a preferred embodiment of the intelligent logistics OCR intelligent document recognition system described in this invention, the adaptive layout learning module includes:

[0038] The general field feature library construction module is used to extract the text features and semantic features of the core fields common to all logistics weighbridges and establish a general feature library.

[0039] The new format matching module is used to automatically extract field features when entering new format weight slips, match them with a general field feature library, and determine the field type and corresponding position.

[0040] The incremental fine-tuning module is used to fine-tune local parameters of the model when only 10-20 new format samples need to be labeled.

[0041] As a preferred embodiment of the intelligent logistics OCR intelligent document recognition system described in this invention, the model iteration optimization module includes:

[0042] The automatic error sample labeling module is used to compare the recognition results with the results of manual review, automatically mark the error samples, and label them with the correct labels;

[0043] The incremental training module is used to incrementally train the context-related line localization module and the multimodal feature fusion character recognition module every week based on newly added error samples and new scene data, and update the model parameters.

[0044] The accuracy monitoring module is used to statistically analyze the recognition accuracy in various scenarios in real time.

[0045] Compared with existing technologies:

[0046] 1. Through fully automated processing (from image acquisition and preprocessing to structured output without manual intervention), it can achieve efficient processing of weighbridge slips.

[0047] 2. With the dual guarantee of context-related line localization and multimodal feature fusion recognition, it can achieve low error in recognition results.

[0048] 3. By replacing the manual reading and data entry process, it can reduce the labor costs of having dedicated personnel on duty at the weigh station.

[0049] 4. Through the context-related row positioning mechanism of logical association and position constraints, it can automatically correct misaligned rows and ensure that fields and values ​​are correctly matched.

[0050] 5. By employing a multimodal recognition strategy that integrates image features, semantic features, and historical data features, it can achieve accurate recognition of fuzzy characters.

[0051] 6. The design, which utilizes a general field feature library and fine-tunes with a small number of incremental samples, enables rapid adaptation to new versions of weighbridge slips.

[0052] 7. Through a closed-loop mechanism of automatic labeling of erroneous samples, incremental training, and accuracy monitoring, the model can continuously improve its recognition accuracy as the real scene changes. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of the overall framework of the present invention;

[0054] Figure 2 This is a schematic diagram of the image preprocessing module framework of the present invention;

[0055] Figure 3 This is a schematic diagram of the context-related row positioning module framework of the present invention;

[0056] Figure 4 This is a schematic diagram of the multimodal feature fusion character recognition module framework of the present invention;

[0057] Figure 5 This is a schematic diagram of the identification result verification and anomaly handling module framework of the present invention;

[0058] Figure 6 This is a schematic diagram of the adaptive layout learning module framework of the present invention;

[0059] Figure 7 This is a schematic diagram of the model iterative optimization module framework of the present invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0061] This invention provides a smart logistics OCR smart document recognition system. Please refer to [link / reference]. Figures 1-7 ,include:

[0062] The image acquisition module is used to acquire images of paper weighing slips using a high-definition industrial camera (resolution ≥ 12 million pixels) and a supplementary lighting device. It supports automatic focusing and tilt correction (correction accuracy ±0.5°) and records the acquisition time and device number to provide basic data for subsequent traceability.

[0063] The image preprocessing module is used to transform the original image into a high-contrast, interference-free standardized image through grayscale processing, texture restoration, and illumination normalization techniques.

[0064] The context-related row positioning module is driven by both the field semantic logic library and historical position data. It first initially segments the row area, and then corrects misaligned rows through field association rules and position constraints to ensure that each field and its corresponding value belongs to the correct row area, providing accurate row-level data range for character recognition.

[0065] The multimodal feature fusion character recognition module is used to fuse the image features, semantic features and historical data features of characters, and through multi-dimensional feature complementarity verification, it can achieve accurate recognition of blurry and illegible characters and output correct recognition results at the single character level.

[0066] The identification result verification and anomaly handling module is used to automatically detect business logic anomalies in the identification results based on the logistics business rule base, and to perform differentiated processing through minor, moderate and severe classification strategies to intercept non-compliant data.

[0067] The adaptive layout learning module is used to automatically match fields of new layout weight slips by building a general field feature library;

[0068] The model iteration and optimization module is used to automatically collect error samples from each module, combine them with the results of manual review to label them with correct labels, and regularly conduct incremental training to update model parameters. At the same time, it monitors the recognition accuracy of each scene in real time and triggers abnormal iteration processes.

[0069] The image preprocessing module includes:

[0070] The stain removal module is used to eliminate paper stains and ink spots through grayscale thresholding and morphological operations.

[0071] The wrinkle repair module is used to identify wrinkled areas based on image texture features and restore a smooth image using an interpolation algorithm;

[0072] The illumination normalization module is used to balance the differences in brightness and darkness of the image using the Retinex algorithm, thereby improving the contrast between characters and the background.

[0073] The context-related row location module includes:

[0074] The weighbridge field logic library construction module is used to build the weighbridge field logic library to sort out the semantic association rules of the core fields of logistics weighbridge (such as "shipper", "receiver", "gross weight", "net weight" and "weighing time"). For example, the value of "gross weight" must be greater than the value of "net weight" and the format of "weighing time" is "YYYY-MM-DDHH:MM").

[0075] The row segmentation and correction module is used to first segment the row area using the projection method, and then correct misaligned rows based on the field logic library and historical row position data (such as the y-coordinate of the "gross weight" row is usually 5-15mm smaller than the y-coordinate of the "net weight" row). For example, if the initial segmentation misclassifies "3200kg" of "net weight: 3200kg" to the "gross weight" row, the system will automatically return "3200kg" to the "net weight" row using the logical rule of "gross weight > net weight".

[0076] The multimodal feature fusion character recognition module includes:

[0077] The image feature module is used to extract the contour, edge, and grayscale distribution features of blurred characters, and generate feature vectors through a convolutional neural network (CNN).

[0078] The semantic feature module is used to constrain semantics based on field type (e.g., the "weight" field must be followed by a number + "kg" / "t", and the "number" field must be "NO. + number / letter combination"), and filter out unreasonable recognition results;

[0079] The historical data feature module is used to build a historical identification database for logistics weighbridge slips (storing 100,000+ valid weighbridge slip data). It corrects fuzzy characters through cosine similarity matching (e.g., comparing the fuzzy "6" with the "6" and "8" features in the "weighing time" field of historical data, and combining the constraint of "month ≤ 12" to determine the character as "6").

[0080] The recognition result verification and anomaly handling module includes:

[0081] The business rule base construction module is used to organize the business constraint logic of the core fields of logistics weighbridge slips and form a configurable rule base.

[0082] The rule base includes rules for numerical rationality, rules for field integrity, and rules for cross-document association.

[0083] The numerical reasonableness rules are based on setting reasonable ranges for weight, volume, and unit price for different types of goods (such as general goods, dangerous goods, and fresh produce) (e.g., gross weight of general goods ≤ 50,000 kg, transportation distance of fresh produce ≤ 1,000 km) and numerical logical relationships (e.g., "freight cost = weight × unit price" error ≤ 5%).

[0084] The field integrity rules define a list of required fields (e.g., "shipper name + recipient name + gross weight + weighing time" are required, and missing any one of them will be considered an anomaly) and field format rules (e.g., "contact number" must conform to the 11-digit mobile phone number format, and "weighing time" must be later than "order creation time").

[0085] The cross-document association rule is to connect to the logistics order system, obtain the order data to be matched (such as order number and estimated weight), and verify the consistency between the weighbridge identification result and the order data (such as the weighbridge "order number" not existing in the order system, or the weighbridge "gross weight" deviating from the order "estimated weight" by more than 20%).

[0086] The anomaly detection and classification module first compares the structured data output by multimodal recognition (such as "gross weight: 60000kg, cargo type: general cargo" and "contact number: 12345678") with the business rule base and automatically marks the abnormal data; then, according to the degree of impact, the anomalies are divided into three levels: minor, moderate and severe, corresponding to different processing strategies.

[0087] The minor deviations are non-core field format deviations (such as missing one street address in "recipient address"). The error log will be automatically recorded, without interrupting the subsequent process, and a reminder will be pushed to the staff for review.

[0088] The term "moderate" refers to a deviation in the core field values ​​that can be corrected (e.g., "gross weight: 55000kg" exceeds the 50000kg limit for ordinary goods, but the goods in the order system are actually "heavy equipment"). The weight range is automatically adjusted by matching the goods type, and the goods are re-evaluated as normal.

[0089] The serious issue is the absence of core fields or key logical errors (such as the absence of "shipper name" or "gross weight = 2000kg, net weight = 3000kg", which violates the "gross weight > net weight" logic). The system will immediately block the process, generate an abnormal work order, and attach the reason for the abnormality (such as "net weight > gross weight, which violates the weight logic") and correction suggestions (such as "re-collect the weighbridge image or manually enter the net weight"), and push it to the staff for processing.

[0090] The exception log and feedback module is used to automatically record all exception data (including exception type, triggering rule, processing result, processing personnel, and processing time) to form a structured exception log library, and periodically synchronize it to the model iteration and optimization module to provide direction for model optimization (e.g., if the weight range rule of a certain type of goods frequently triggers exceptions, it prompts that the rule library needs to be updated or the character recognition model needs to be optimized).

[0091] The adaptive layout learning module includes:

[0092] The general field feature library construction module is used to extract the text features (TF-IDF vector) and semantic features (Word2Vec vector) of the core fields (such as "shipper" and "gross weight") common to all logistics weighbridges, and to build a general feature library;

[0093] The new format matching module is used to automatically extract field features when entering new format weight slips, match them with a general field feature library, and determine the field type and corresponding position (e.g., matching the "shipping unit" field in the new format with the "shipping party" field in the general library).

[0094] The incremental fine-tuning module is used to fine-tune local parameters of the model (such as row segmentation threshold and field positioning coordinates) when only 10-20 new format samples need to be labeled.

[0095] The model iterative optimization module includes:

[0096] The automatic error sample labeling module is used to compare the recognition results with the manual review results, automatically mark error samples (such as misaligned lines or fuzzy recognition errors), and label them with the correct labels;

[0097] The incremental training module is used to incrementally train the context-related line localization module and the multimodal feature fusion character recognition module weekly based on newly added error samples and new scenario data (such as weighbridge slips from newly added freight stations), and update the model parameters.

[0098] The accuracy monitoring module is used to statistically analyze the recognition accuracy in various scenarios in real time (such as misaligned scenarios, blurred scenarios, and new layout scenarios). When the accuracy of a certain scenario is lower than 95%, an emergency iteration process is triggered.

[0099] In practical use, the specific operating steps for those skilled in the art are as follows:

[0100] S1: The image acquisition module uses a high-definition industrial camera (resolution ≥ 12 million pixels) and a supplementary lighting device to acquire images of paper weighing slips, and supports automatic focus and tilt correction (correction accuracy ±0.5°). At the same time, it records the acquisition time and equipment number to provide basic data for subsequent traceability.

[0101] S2: The image preprocessing module uses grayscale processing, texture restoration, and illumination normalization techniques to transform the original image into a high-contrast, interference-free standardized image.

[0102] S3: The context-related row positioning module is driven by both the field semantic logic library and historical position data. It first initially segments the row area, and then corrects misaligned rows through field association rules and position constraints to ensure that each field and its corresponding value belong to the correct row area, providing accurate row-level data range for character recognition.

[0103] S4: By fusing the image features, semantic features and historical data features of characters through the multimodal feature fusion character recognition module, and through multi-dimensional feature complementarity verification, it achieves accurate recognition of fuzzy and illegible characters and outputs correct recognition results at the single character level;

[0104] S5: Based on the logistics business rule base, the identification result verification and anomaly handling module automatically detects business logic anomalies in the identification results and performs differentiated processing through minor, moderate and severe classification strategies to intercept non-compliant data.

[0105] S6: By utilizing the adaptive layout learning module to build a general field feature library, the fields of the new layout of the bill of lading can be automatically matched.

[0106] S7: The model iteration optimization module automatically collects error samples from each module and labels them correctly based on the results of manual review. It also regularly performs incremental training to update model parameters and monitors the recognition accuracy of each scene in real time, triggering abnormal iteration processes.

[0107] Although the present invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the disclosed embodiments can be combined with each other in any manner. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A smart logistics OCR smart document reading system, characterized in that, Comprise: An image acquisition module for acquiring paper pound sheet images using a high-definition industrial camera in combination with a light supplementing device, supporting automatic focusing and tilt correction, and recording acquisition time and equipment number to provide basic data for subsequent tracing; An image preprocessing module for converting the original image into a high-contrast, interference-free standardized image through gray scale processing, texture repair, and illumination normalization techniques; A context-associated line positioning module for driving with a field semantic logic library and historical position data, preliminarily segmenting line areas, and correcting misaligned lines through field association rules and position constraints to ensure that each field and corresponding value belong to the correct line area, providing accurate line-level data range for character recognition; A multi-modal feature fusion character recognition module for fusing image features, semantic features, and historical data features of characters, and achieving accurate recognition of fuzzy and sloppy characters through multi-dimensional feature complementary verification, and outputting single character-level correct recognition results; An identification result verification and exception handling module for automatically detecting business logic exceptions in the identification results based on a logistics business rule library, and performing differential processing through slight, moderate, and severe classification strategies to intercept non-compliant data; An adaptive format learning module for realizing automatic matching of fields of new format pound sheets by constructing a general field feature library; A model iteration optimization module for automatically collecting error samples of each module, combining manual review results to label correct tags, periodically carrying out incremental training to update model parameters, and simultaneously monitoring the recognition accuracy of each scene to trigger the abnormal iteration process.

2. The intelligent logistics OCR intelligent document reading system according to claim 1, characterized in that, The image preprocessing module comprises: A stain removal module for eliminating paper stains and ink spots through gray scale threshold segmentation and morphological operation; A wrinkle repair module for identifying wrinkle areas according to image texture features and restoring flat images through interpolation algorithm; An illumination normalization module for balancing image light and dark differences using Retinex algorithm to improve character and background contrast. 3.The smart logistics OCR smart document reading system of claim 1, wherein, The context-associated line positioning module comprises: A pound sheet field logic library construction module for constructing a pound sheet field logic library to sort out the semantic association rules of core fields of logistics pound sheets; A line segmentation and correction module for preliminarily segmenting line areas through projection method, and correcting misaligned lines based on field logic library and historical line position data.

4. The intelligent logistics OCR intelligent document reading system according to claim 1, characterized in that, The multi-modal feature fusion character recognition module comprises: An image feature module for extracting contour, edge, and gray scale distribution features of fuzzy characters, and generating feature vectors through convolutional neural network; A semantic feature module for filtering unreasonable recognition results according to field type constraint semantics; A historical data feature module for constructing a logistics pound sheet historical recognition database, and correcting fuzzy characters through cosine similarity matching. 5.The smart logistics OCR smart document reading system of claim 1, wherein, The identification result verification and exception handling module comprises: A business rule library construction module for sorting out business constraint logic of core fields of logistics pound sheets to form a configurable rule library; An abnormality detection and grading processing module is configured to compare the structured data of the multi-modal recognition output with a business rule library to automatically mark abnormal data, and then divide the abnormality into three levels of slight, moderate and severe according to the influence degree, corresponding to different processing strategies; An abnormality log and feedback module is configured to automatically record all abnormal data to form a structured abnormality log library, and periodically synchronize to the model iteration optimization module to provide a direction for model optimization.

6. The intelligent logistics OCR intelligent document reading system according to claim 5, characterized in that, The rule library includes numerical reasonableness rules, field integrity rules and cross-invoice association rules; The numerical reasonableness rules are based on different types of goods to set reasonable intervals of weight, volume and unit price, and numerical logical relationships; The field integrity rules define a list of mandatory fields and field format rules; The cross-invoice association rules are connected to a logistics order system to obtain order data to be matched, and verify the consistency of the scale invoice recognition result and the order data. 7.The smart logistics OCR smart document reading system of claim 1, wherein, The adaptive layout learning module includes: A general field feature library construction module is configured to extract text features and semantic features of core fields common to all logistics scale invoices, and establish a general feature library; A new format matching module is configured to automatically extract field features of a new format scale invoice when the new format scale invoice is input, match the field features with the general field feature library, and determine the field type and corresponding position; An incremental fine-tuning module is configured to only need to label 10-20 new format samples to fine-tune local parameters of the model. 8.The smart logistics OCR smart document reading system of claim 1, wherein, The model iteration optimization module includes: An error sample automatic labeling module is configured to compare the recognition result with the manual review result, automatically mark error samples, and label correct labels; An incremental training module is configured to perform incremental training on the context association line positioning module and the multi-modal feature fusion character recognition module based on the newly added error samples and new scene data every week, and update the model parameters; An accuracy monitoring module is configured to real-time statistics of the recognition accuracy in each scene.

Images