An information processing method for port container sorting

Abstract

The application discloses an information processing method for port container number tallying, and relates to the technical field of information processing; the method comprises the following steps: deploying a visual network at a key node under a target area, acquiring a container number related data set, using an embedded lightweight convolutional neural network model to score the quality of each original image, outputting a comprehensive quality evaluation value, comparing the comprehensive quality evaluation value with an evaluation threshold, and selecting whether to trigger a retake adjustment instruction according to the comparison result; the technical key points are that the operation of the overall scheme effectively solves the problem that the efficiency and resources cannot be balanced due to the unified processing of all low confidence results in the traditional tallying scheme, realizes layered processing, embodies an optimized scheme based on multi-source evidence fusion and setting layers, and makes the overall scheme have rigor and flexibility, thereby greatly improving the overall recognition success rate and automation level in a complex degradation scene.

Description

Technical Field

[0001] This invention relates to the field of information processing technology, specifically to an information processing method for tallying container numbers at ports. Background Technology

[0002] Information processing in port container tallying refers to the use of various technologies to distinguish and manage container numbers and related data. Current technologies typically use cameras for data collection and image processing technology to identify the numbers on the containers. At the same time, data verification and updates are performed in conjunction with the operating system and other business systems at the terminal or port to ensure the accuracy and timeliness of logistics information.

[0003] However, existing port container number recognition methods mostly rely on a single camera to capture images and directly input them into OCR for recognition, lacking pre-assessment and intervention mechanisms for image quality. When encountering rain, fog, oil stains, reflections, or obstructions from stacked containers, the captured images are often blurry, overexposed, or partially missing, yet the system still forces recognition, resulting in a high false recognition rate. For example, misidentifying AABC as AABO or recognizing 5 as S, the current recognition system cannot determine the reliability of the result and can only output it as is. More seriously, for these low-quality recognition results, regardless of the error type, they are all uniformly handed over to manual review or supporting system review, without prioritization or auxiliary judgment methods, resulting in an ineffective balance between resources and efficiency. At the same time, traditional solutions lack the ability to automatically verify container number rules, let alone use historical data and knowledge graphs for semantic reasoning. Once characters are blurry or missing, there is no way to effectively complete them. All these problems combined make the existing design system appear more rigid and inefficient, making it difficult to support the high-efficiency and fully automated operation requirements of modern ports. Summary of the Invention

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A method for information processing of port container tallying numbers, the method comprising:

[0006] A visual network is deployed at key nodes in the target area to acquire a dataset related to container numbers. An embedded lightweight convolutional neural network model is used to score the quality of each original image frame, outputting a comprehensive quality assessment value. The comprehensive quality assessment value is compared with an evaluation threshold, and a reshoot adjustment instruction is triggered based on the comparison result. Under the condition of acquiring multiple frames of images of the same target container at different times and angles, an image registration and fusion algorithm based on feature matching is used to output a reconstructed dataset related to container numbers. A two-level deep neural network is deployed to complete the container number recognition action.

[0007] Based on the completion of the box number recognition action, a pre-built heterogeneous OCR model cluster is used to recognize each standard character in parallel, obtain three sets of recognition results, and execute a revision and dynamic weighted fusion strategy based on the recognition results. The revised first confidence, second confidence and third confidence are compared with the standard threshold range.

[0008] If the corrected first confidence level exceeds the upper limit of the standard threshold range, the recognition result is adopted;

[0009] If the corrected first confidence level is within the standard threshold range, the recognition results of the heterogeneous OCR model cluster are weighted and voted out. If the corrected first confidence level is lower than the lower limit of the standard threshold range, a secondary judgment mechanism is executed, comparing the corrected second and third confidence levels with the lower limit of the standard threshold range. If both are lower than the lower limit of the standard threshold range, they are marked as abnormal characters; otherwise, they are marked as ambiguous characters.

[0010] Furthermore, the container number-related dataset includes at least: images of the target container taken from several angles; each original image frame is the corresponding target container image, and the process of quality scoring for each original image frame is as follows: weighted summation based on the scoring dimensions to obtain a comprehensive quality assessment value; wherein, the scoring dimensions include at least: illumination uniformity, motion blur, key area coverage, and character clarity.

[0011] Furthermore, the process of running the feature matching-based image registration and fusion algorithm is as follows:

[0012] SIFT is used to extract stable feature points from each image, and RANSAC algorithm is used to remove mismatches to achieve subpixel-level alignment. The ESRGAN-Port super-resolution model based on generative adversarial network is applied to fuse and reconstruct the multi-frame low-resolution images after subpixel-level alignment into a single high-resolution image HR. The ESRGAN-Port model is trained on a dedicated dataset containing several real container images. During the reconstruction process, the ESRGAN-Port model prioritizes enhancing high-frequency details in the container number region of the target container through an attention mechanism.

[0013] Furthermore, the two-level deep neural network includes: a container region detection model YOLOv7-Cargo and a character segmentation network SegNet-CR; a single high-resolution image HR is input to the container region detection model YOLOv7-Cargo, and the output is: the minimum bounding rectangle of the container number region on the target container; the character segmentation network SegNet-CR is used to segment the detected container number and obtain several standard characters.

[0014] Furthermore, the heterogeneous OCR model cluster includes at least: a primary heterogeneous model, a first sub-heterogeneous model, and a second sub-heterogeneous model; wherein, the primary heterogeneous model adopts a deep convolutional neural network CRNN-CargoNet; the first sub-heterogeneous model adopts a Transformer-based visual recognition model ViT-CR; the second sub-heterogeneous model adopts a lightweight model MobileOCR-Edge; the confidence level corresponding to the primary heterogeneous model is labeled as the first confidence level, the confidence level corresponding to the first sub-heterogeneous model is labeled as the second confidence level, and the confidence level corresponding to the second sub-heterogeneous model is labeled as the third confidence level.

[0015] Furthermore, the process of implementing the correction and dynamic weighted fusion strategy also includes: using the comprehensive quality assessment value as a basis, calculating the mean of the comprehensive quality assessment value for each frame within a preset period to obtain the comprehensive quality mean Q, constructing a linear decay model based on the Q value, inputting the original confidence level and the comprehensive quality mean Q, and outputting: the confidence level after first-level correction;

[0016] The constraints for the first-level corrected confidence level are set as follows:

[0017] Condition 1: The confidence level after the first-level correction exceeds the set threshold;

[0018] Condition 2: The difference between the confidence levels after the first-level correction does not exceed the error range; when the confidence level after the first-level correction meets the constraint condition, it means that the effect is satisfactory, and the confidence level after the first-level correction shall be used as the standard; otherwise, the second-level correction mechanism shall be triggered.

[0019] Query the historical accuracy database to obtain the corresponding historical average accuracy; construct a piecewise affine transformation calibration model, input the first-level corrected confidence level and the corresponding historical average accuracy, and output the second-level corrected confidence level.

[0020] Furthermore, the function upon which the piecewise affine transformation calibration model is based is: P2 = f(P1, A_hist); where P2 represents the confidence level after second-order correction, P1 represents the confidence level after first-order correction, and A_hist represents the corresponding historical average accuracy. The specific expansion formula of the function upon which the piecewise affine transformation calibration model is based is: when P1 is not lower than A_hist, then P2 = A_hist + u*(P1 - A_hist); where u represents the scaling factor, and u∈(0,1); when P1 is lower than A_hist, then P2 = P1.

[0021] Furthermore, the method also includes: when abnormal characters are detected, a dynamic error correction and completion mechanism based on context and business rules is triggered; when ambiguous characters are detected, a port-level container number knowledge graph and association reasoning mechanism is triggered.

[0022] Furthermore, the dynamic error correction and completion mechanism based on context and business rules proceeds as follows: The container master code database is invoked to verify the validity of the first few letter combinations; cross-validation is performed after obtaining pre-provided information; mathematical verification is performed using the container number checksum algorithm, converting the first n characters into numerical values, calculating the remainder modulo n+1 after weighted summation, and checking if it matches the (n+1)th character; if not, all character combinations are traversed to calibrate the error position until calibration passes; combining the spatiotemporal context, if the container numbers of adjacent target containers within the same target area are consecutive, and if the container numbers of the preceding and following containers have been confirmed, the intermediate container number is inferred.

[0023] Furthermore, the process of establishing a port-level container number knowledge graph and association reasoning mechanism is as follows: construct a container number knowledge graph under the target area, with the container number as the entity node, and the associated attributes include at least: the company, vessel, voyage, port of origin, port of destination, cargo type, historical port entry and exit records, and RFID tag ID; call the container number knowledge graph to perform association reasoning and output the association reasoning results.

[0024] This invention provides an information processing method for port container tallying, which has the following beneficial effects:

[0025] 1) This solution deploys a visual network, enabling the method to acquire container numbers and related information in complex environments, ensuring the quality and integrity of the dataset. At the same time, it uses an embedded lightweight convolutional neural network model to evaluate the quality of images and triggers re-shooting and adjustment instructions when necessary, further ensuring that the images entering the subsequent processing flow meet the basic requirements for recognition. This solution not only improves the accuracy and reliability of data acquisition, but also reduces the problem of data loss due to harsh environments or technical limitations.

[0026] 2) This scheme determines whether to trigger a reshoot adjustment command by comparing the comprehensive quality assessment value with the assessment threshold. Based on this, it further derives the comprehensive quality mean and constructs a linear decay model to perform a first-level correction on the original confidence score. On the one hand, this ensures that additional measures are only taken to guarantee the quality and integrity of data acquisition when the image quality is poor, avoiding unnecessary waste of resources. On the other hand, by performing a conservative calibration of the original confidence score based on image quality, it can more accurately reflect the true reliability of the recognition results, enhance the robustness of the decision-making process, and effectively solve the problem of high misjudgment rate caused by neglecting image quality in traditional schemes. This achieves the goal of maintaining high-precision recognition even under complex working conditions.

[0027] 3) This solution introduces a first-level correction based on Q-value and a second-level correction supported by historical accuracy. The combination of the two results in a revision and dynamic weighted fusion strategy, as well as error correction and information reasoning recognition methods based on decision results. This not only enables rapid response and calibration of the initial recognition results, but also allows for in-depth mining of historical data when necessary to achieve more accurate secondary correction. Through the application of associative reasoning and knowledge graphs, a leap from character recognition to semantic understanding is achieved, which greatly improves the final recognition success rate of abnormal or fuzzy characters corresponding to box numbers.

[0028] 4) The differentiation mechanism given above effectively solves the problem of unbalanced efficiency and resources caused by uniform processing of all low-confidence results in the traditional sorting scheme. It realizes the function of hierarchical processing, reflects the optimal scheme based on multi-source evidence fusion and hierarchical setting, and makes the overall scheme both rigorous and flexible, greatly improving the overall recognition success rate and automation level in complex degradation scenarios. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the overall process of an information processing method for port container number tallying in this invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0031] Please see Figure 1 This embodiment provides an information processing method for container number tallying in ports. This method processes container numbers and related data information in port areas. The specific steps of this processing method are as follows:

[0032] S1. Collection and preprocessing of container number information for multimodal sensing:

[0033] S1.1 Deployment Information Collection:

[0034] Deploy a visual network at key nodes in the target area to obtain a dataset related to container numbers;

[0035] The target area is the selected port area. Key nodes within the target area include: port gantry cranes, quay cranes, container transport channels, and container stacking areas. The deployed visual network includes: an array of high-definition cameras, LiDAR, inductive loops, and RFID, ensuring the target area is fully covered by the visual network. The purpose is to ensure the comprehensiveness and completeness of the required information or data. The container number-related dataset contains images of containers taken from several angles to ensure that the appearance of the target container is captured without blind spots. The specific content of the container images can be selected as: the front, sides, bottom, and top markings of the complete target container.

[0036] Specifically, for each high-definition camera in the visual network, each high-definition camera should be equipped with autofocus, zoom and wide-angle lenses to ensure that container images can be acquired under different lifting heights and different container stacking layers; each high-definition camera should have anti-salt spray and anti-vibration design to adapt to the harsh environment of port areas such as high humidity and strong winds, and ensure the quality and integrity of the data set related to container numbers from a physical level.

[0037] Visual networks trigger linkage mechanisms during actual operation:

[0038] When the target container enters the target area, the ground loop is activated for initial positioning, and the RFID reads the equipment number carrying the target container and connects to the network to associate it with pre-provided information, including: bill of lading number, vessel name and voyage number, loading and unloading plan, etc. At the same time, the lidar scans the outline of the target container and triggers the high-definition camera at the corresponding key node to continuously capture images of the target container. Among them, no less than 3 images from different angles are collected at a time to ensure that the container number area on the target container is presented more completely and clearly in at least 2 images. The application of this linkage mechanism solves to a certain extent the defects of traditional single-point fixed cameras that are easily obstructed and have limited viewing angles. Through the joint collaborative operation of multiple devices, the complete acquisition of container number-related datasets under complex working conditions can be significantly improved.

[0039] S1.2 Image Information Evaluation:

[0040] Based on the acquired container number-related dataset, an embedded lightweight convolutional neural network model is used to score the quality of each original image frame. The scoring dimensions for quality scoring include: illumination uniformity, motion blur, key area coverage, and character clarity. The weighted sum of each scoring dimension is calculated to output a comprehensive quality assessment value. The comprehensive quality assessment value is then compared with an assessment threshold, and a reshoot adjustment instruction is triggered based on the comparison result.

[0041] Among them, illumination uniformity is obtained using local standard deviation analysis to avoid overexposed / underexposed areas; motion blur is obtained based on gradient magnitude variance calculation; key area coverage is obtained by locating the target box number area using a pre-trained box ROI detector and calculating its pixel ratio; character sharpness is obtained by extracting edges using the Sobel operator and then calculating the sharpness index; after weighted summation of illumination uniformity, motion blur, key area coverage, and character sharpness, the required comprehensive quality assessment value can be obtained; the assessment threshold is preset in advance according to actual needs; when the limit of the comprehensive quality assessment value is within 1, the corresponding assessment threshold is 1 / 7, that is, 0.7; Taking an evaluation threshold of 0.7 as an example, if the overall quality evaluation value exceeds the evaluation threshold, no response action is taken; if the overall quality evaluation value does not exceed the evaluation threshold, a reshoot adjustment instruction is triggered: the image is automatically marked as low quality, and the visual network parameters are adjusted synchronously before reshooting. The visual network parameters correspond to the high-definition camera's increased exposure compensation and focus switching. This step effectively filters out blurred or distorted images caused by rain, fog, oil stains, reflections, or rapid movement, ensuring that all images entering the subsequent processing flow meet the minimum requirements for recognition and reducing unnecessary computational resource consumption. Compared with traditional fixed threshold filtering, this solution achieves more accurate quality judgment by adaptively learning image degradation patterns in port-specific environments.

[0042] Explanation: By deploying a visual network, this method can acquire container numbers and related information in complex environments, ensuring the quality and integrity of the dataset. At the same time, an embedded lightweight convolutional neural network model is used to evaluate the quality of images and trigger re-shooting and adjustment instructions when necessary, further ensuring that the images entering the subsequent processing flow meet the basic requirements for recognition. This approach not only improves the accuracy and reliability of data acquisition, but also reduces data loss due to harsh environments or technical limitations.

[0043] S1.3 Multi-source image fusion and reconstruction:

[0044] Under the condition of acquiring multiple frames of images of the same target container at different times and angles, an image registration and fusion algorithm based on feature matching is used to output the reconstructed container number related dataset.

[0045] The process of running the feature matching-based image registration and fusion algorithm is as follows:

[0046] S1.3.1. Stable feature points are extracted from each image using SIFT, and mismatches are eliminated using the RANSAC algorithm to achieve sub-pixel alignment. S1.3.2. The ESRGAN-Port super-resolution model based on generative adversarial networks is applied to fuse and reconstruct a single high-resolution image (HR) from multiple frames of low-resolution images after sub-pixel alignment. The high resolution mentioned here is at least 4k. S1.3.3. The ESRGAN-Port model is trained on a dedicated dataset containing several real container images. Specifically, during reconstruction, the ESRGAN-Port model prioritizes enhancing high-frequency details in the container number area of ​​the target container through an attention mechanism and suppresses background noise. The output single high-resolution image (HR) not only improves character edge clarity but also partially recovers slightly obscured stroke information to ensure the input quality of subsequent data. These steps combine multi-view redundant information with deep learning super-resolution technology to solve the problem of insufficient resolution caused by distance or occlusion in a single camera.

[0047] S1.4, Container Number Area Location and Segmentation:

[0048] On the reconstructed container number-related dataset, a two-level deep neural network is deployed to complete the container number recognition action;

[0049] The two-level deep neural network includes: a container region detection model YOLOv7-Cargo and a character segmentation network SegNet-CR. The YOLOv7-Cargo model takes a 4K image as input and outputs the minimum bounding rectangle of the container number region on the target container. The YOLOv7-Cargo model is an existing model with a localization accuracy exceeding 99%. The SegNet-CR network performs pixel-level segmentation on the detected container number, resulting in several standard characters, such as a 4-digit container owner code + a 6-digit... The code consists of a number plus a 1-bit checksum. SegNet-CR uses a U-Net architecture with a ResNet-34 encoder and a decoder that combines dilated convolutions to expand the receptive field, outputting a binary mask for each character. To address issues such as character adhesion, breakage, or oil stains, a conditional random field post-processing is introduced. Topological constraints are used to optimize the segmentation results by utilizing fixed spacing and arrangement rules between characters, thus standardizing the size of the segmented single-character image. This step, through a dedicated network trained end-to-end, achieves robust localization and accurate segmentation of the box number region in complex backgrounds. Compared to the traditional edge detection combined with projection method, it reduces the error rate by at least 60% in handling adhered characters, demonstrating significant effectiveness.

[0050] S2. Character recognition and confidence assessment and correction:

[0051] Based on the completed box number recognition action, a pre-constructed heterogeneous OCR model cluster is used to recognize each standard character in parallel. The heterogeneous OCR model cluster includes a main heterogeneous model, a first sub-heterogeneous model, and a second sub-heterogeneous model. The main heterogeneous model is a deep convolutional neural network CRNN-CargoNet; the first sub-heterogeneous model is a Transformer-based visual recognition model ViT-CR; and the second sub-heterogeneous model is a lightweight model MobileOCR-Edge. Three sets of recognition results are obtained, and a revision and dynamic weighted fusion strategy is executed based on the recognition results.

[0052] Specifically, the deep convolutional neural network CRNN-CargoNet adopts a CNN+BiLSTM+CTC architecture and is optimized for sequence recognition; the Transformer-based visual recognition model ViT-CR divides character images into 16×16 patch sequences for input, uses a self-attention mechanism to capture global context, and is more robust when dealing with slightly deformed or partially rotated characters; the partial rotation limit is within ±15°; MobileOCR-Edge is compressed from CRNN-CargoNet based on knowledge distillation technology, with 15% of the parameters of the original model, and is deployed on edge computing nodes for real-time preliminary recognition;

[0053] The three sets of identification results obtained are three sets of confidence levels;

[0054] The confidence level of the deep convolutional neural network CRNN-CargoNet is labeled as the first confidence level, the confidence level of the Transformer-based visual recognition model ViT-CR is labeled as the second confidence level, and the confidence level of the lightweight model MobileOCR-Edge is labeled as the third confidence level.

[0055] The process of implementing the correction and dynamic weighted fusion strategy is as follows:

[0056] S2.1, First-level correction: Based on the comprehensive quality assessment value obtained in S1.2, calculate the mean of the comprehensive quality assessment value for each frame within a preset period to obtain the comprehensive quality mean Q. Construct a linear decay model based on the Q value. Input the original confidence level and the comprehensive quality mean Q, and output the confidence level after first-level correction. The confidence level after first-level correction is the product of the original confidence level and the comprehensive quality mean Q. Logical explanation: The confidence level of this model is proportional to the image quality. Through conservative calibration using a product approach, it can be performed on edge nodes.

[0057] Explanation: In S1.2, the comparison between the comprehensive quality assessment value and the assessment threshold is used to determine whether to trigger a reshoot adjustment instruction. Based on this, the comprehensive quality mean is further derived, and a linear decay model is constructed to perform a first-level correction on the original confidence score. On the one hand, this ensures that additional measures are only taken when the image quality is poor to guarantee the quality and integrity of data acquisition, avoiding unnecessary waste of resources. On the other hand, by performing a conservative calibration of the original confidence score based on image quality, the true reliability of the recognition results can be more accurately reflected, enhancing the robustness of the decision-making process.

[0058] The above method effectively solves the problem of high misjudgment rate caused by neglecting image quality in traditional solutions, and achieves the goal of maintaining high-precision recognition even under complex working conditions. At the same time, the strategy adopted reflects the design concept of combining intelligent perception and adaptive processing, which not only improves the success rate of a single recognition task, but also improves the efficiency and response speed of the overall solution by optimizing resource allocation. This ensures that reliable cargo handling decisions can be made to a certain extent even under low-quality image conditions, reflecting the automation level and accuracy of port container number cargo handling.

[0059] S2.2 Effect Evaluation and Compliance Judgment: The constraints for the first-level corrected confidence level are set as follows: Condition 1, the first-level corrected confidence level exceeds the set threshold; Condition 2, the difference between the first-level corrected confidence levels does not exceed the error range; wherein, the first-level corrected confidence level includes three confidence levels; when the first-level corrected confidence level meets the constraints, it indicates that the effect has met the standard, and the first-level corrected confidence level shall be used as the standard; when the first-level corrected confidence level does not meet the constraints, the second-level correction mechanism is triggered; wherein, the first-level corrected confidence level is denoted as P1;

[0060] S2.3, Secondary Correction Mechanism: Query the historical accuracy database, which is used to build a multi-dimensional lookup table based on the comprehensive quality assessment value bin, model type, and character position (first / middle / last character);

[0061] For example, the historical average accuracy of the query entry (Model=CRNN-CargoNet, Q_bin=[0.80, 0.85), Position=Middle) is A_hist=89.5%, which is only an example for reference.

[0062] A piecewise affine transformation calibration model is constructed, taking the first-level corrected confidence level and the corresponding historical average accuracy as input, and outputting the second-level corrected confidence level. The function upon which the piecewise affine transformation calibration model is based is: P2 = f(P1, A_hist); where P2 represents the second-level corrected confidence level, and A_hist represents the corresponding historical average accuracy. The specific expansion formula of the function is: when P1 is not lower than A_hist, then P2 = A_hist + u*(P1 - A_hist); where u represents the scaling factor, and u∈(0,1). In this embodiment, the value of u is usually 0.5 to prevent overestimation; when P1 is lower than A_hist, then P2 = P1, conservatively maintaining the accuracy; the second-level corrected confidence level is used as the standard.

[0063] Whether the confidence level is adjusted to the first level or the confidence level is adjusted to the second level, the final confidence levels are the required adjusted first confidence level, adjusted second confidence level, and adjusted third confidence level.

[0064] This scheme introduces a first-level correction based on Q-values ​​and a second-level correction supported by historical accuracy. The combination of these two approaches yields a revision and dynamic weighted fusion strategy, along with error correction and information reasoning recognition methods based on decision results. This not only enables rapid response and calibration of initial recognition results but also allows for in-depth mining of historical data when necessary to achieve more accurate secondary corrections. Through the application of associative reasoning and knowledge graphs, it achieves a leap from character recognition to semantic understanding, significantly improving the final recognition success rate of abnormal or ambiguous characters corresponding to box numbers. From the overall structure of the scheme, it constitutes an efficient and intelligent information processing closed loop, effectively solving the problems of misjudgment and omission that may occur in traditional information recognition schemes.

[0065] S2.4, Implement the weighted fusion strategy:

[0066] The corrected first confidence level is compared with the standard threshold range. If the corrected first confidence level exceeds the upper limit of the standard threshold range, the recognition result is adopted. If the corrected first confidence level is within the standard threshold range, the recognition results of the main heterogeneous model, the first sub-heterogeneous model, and the second sub-heterogeneous model are weighted and voted on for output. If the corrected first confidence level is lower than the lower limit of the standard threshold range, a secondary judgment mechanism is executed. The corrected second confidence level and the corrected third confidence level are both compared with the lower limit of the standard threshold range. If both are lower than the lower limit of the standard threshold range, they are marked as abnormal characters, and step S3.1 is executed. Otherwise, they are marked as ambiguous characters, and step S3.2 is executed.

[0067] Note: The above-described correction and dynamic weighted fusion strategy prioritizes Q-value for rapid calibration and only calls historical data when necessary, enabling fast processing. In some complex scenarios, compared to a single correction method, it avoids overconfidence in low-quality images, thus improving the accuracy of the final decision. Most recognition tasks meet the standard after the first-level correction, eliminating the need to query the historical database, reducing system latency and optimizing resources. The constructed two-layer confidence calibration system enables the overall information processing solution to possess both real-time response capabilities and deep learning capabilities, making it a key innovation in the field of intelligent cargo handling in ports.

[0068] S3, Error Correction and Completion with Information Reasoning and Recognition:

[0069] S3.1 Dynamic error correction and completion mechanism based on context and business rules:

[0070] S3.1.1 Call the container master code database to verify the validity of the first few letter combinations; S3.1.2 Obtain the pre-provided information from S1.1 and perform cross-validation accordingly; S3.1.3 Apply the container number check code algorithm for mathematical verification, convert the first n characters into numerical values ​​according to the rules, calculate the remainder modulo n+1 after weighted summation, and check whether it matches the (n+1)th digit; if it does not match, iterate through all possible character combinations to calibrate the error digits until calibration passes; in this embodiment, the value of n is 10; S3.1.4 Combine the spatiotemporal context, if the container numbers of adjacent target containers in the same area are consecutive, and the container numbers of the preceding and following containers have been confirmed, then infer the intermediate container number.

[0071] The container master code database is a known database, for example, the mapping between the first 4 letters of a company's code and its corresponding company name, used to verify the validity of the first 4 letter combination. If the first 3 letters are determined and the 4th letter is ambiguous, the 4th letter is tried first, rather than other letters. In S3.1.2, if a container number is known in advance and the recognition result is the same as the known result, only the 6th letter is ambiguous, the 6th letter can be directly completed according to the known result. In S3.1.3, for example, if the recognition result is XXXX123456 and the verification fails, the 6 in the 10th letter is replaced with any number until the verification passes. The above-mentioned dynamic error correction and completion method based on context and business rules integrates industry knowledge base, business data flow and mathematical rules, which improves the final recognition success rate of abnormal characters corresponding to container numbers from the original, for example, 70% by at least 20%.

[0072] S3.2 Establish a port-level container number knowledge graph and association reasoning mechanism:

[0073] S3.2.1 Construct a container number knowledge graph for the target region. This knowledge graph uses container numbers as entity nodes and associates attributes including: company, vessel, voyage number, port of origin, port of destination, cargo type, historical port entry and exit records, maintenance records, RFID tag ID, etc.; S3.2.2 Call the knowledge graph to perform association reasoning and output the association reasoning results.

[0074] For example, if the first six digits of a container number are identified as XXXX12, a query of the knowledge graph reveals that containers starting with XXXX12 on recently docked shipping vessels all originated from Port A and mostly carried electronic products. If the current container's appearance is similar to such cargo containers, then XXXX12 will be the final result. Conversely, if the container appears next to a bulk carrier carrying ore, an anomaly warning will be triggered, and S3.1 will continue to be executed. The knowledge graph also records historical recognition error patterns, enabling continuous learning and updates, achieving a leap from character recognition to semantic understanding, thereby enabling more complete and comprehensive information processing.

[0075] In this scheme, both the corrected second and third confidence scores are compared with the lower limit of the standard threshold range. If both are lower than the lower limit, the character is marked as an abnormal character, and step S3.1 is executed; otherwise, it is marked as an ambiguous character, and step S3.2 is executed. This design reflects a refined classification and differentiated response strategy for the reasons for recognition failure. On the one hand, when the confidence scores of the heterogeneous OCR model cluster are all low, it indicates that the current image has serious degradation and can be attributed to systematic recognition failure. Therefore, it prioritizes relying on external strong constraint information for deterministic repair. The industry rules and mathematical verification provided by S3.1 can directly complete or correct the box number with a high probability, achieving the effect of quickly restoring key information. On the other hand, When only the primary heterogeneous model has low confidence due to its sensitivity to quality, but other models still have a certain level of confidence, it indicates that although the characters are blurry, they still have some discernible features, which is a local uncertainty problem. In this case, the knowledge graph reasoning of S3.2 can combine semantic context to make intelligent inferences, thereby achieving flexible decision-making in the absence of clear rules. The differentiation mechanism given above effectively solves the problem of unbalanced efficiency and resources caused by uniformly processing all low-confidence results in the traditional inventory sorting scheme. It realizes the role of hierarchical processing, reflects the optimal solution based on multi-source evidence fusion and hierarchical setting, and makes the overall scheme both rigorous and flexible, greatly improving the overall recognition success rate and automation level in complex degradation scenarios.

[0076] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0077] 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; 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, depending on actual needs.

[0078] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for information processing of container number tallying in ports, characterized in that: The method includes: A visual network is deployed at key nodes in the target region to acquire a dataset related to container numbers. An embedded lightweight convolutional neural network model is used to score the quality of each original image frame, outputting a comprehensive quality assessment value. This comprehensive quality assessment value is compared with an evaluation threshold, and a reshoot adjustment command is triggered based on the comparison result. Under the condition of acquiring multiple frames of images of the same target container at different times and angles, an image registration and fusion algorithm based on feature matching is used to output a reconstructed dataset related to container numbers. A two-level deep neural network is deployed to complete the container number recognition action. The dataset related to container numbers includes at least: images of the target container taken from several angles. Each original image frame is the corresponding target container image. The process of scoring the quality of each original image frame is as follows: a weighted sum is performed based on the scoring dimensions to obtain a comprehensive quality assessment value. The scoring dimensions include at least: illumination uniformity, motion blur, key area coverage, and character clarity. The process of running the image registration and fusion algorithm based on feature matching is as follows: SIFT is used to extract stable feature points from each image, and RANSAC algorithm is used to remove mismatches to achieve subpixel-level alignment. The ESRGAN-Port super-resolution model based on generative adversarial networks is applied to fuse the subpixel-aligned multi-frame low-resolution images (LR) into a single high-resolution image (HR). The ESRGAN-Port model is trained on a dedicated dataset containing several real container images. During reconstruction, the ESRGAN-Port model prioritizes enhancing high-frequency details in the container number region of the target container through an attention mechanism. A two-level deep neural network is used: the YOLOv7-Cargo container region detection model and the SegNet-CR character segmentation network. The single high-resolution image (HR) is input to the YOLOv7-Cargo model, and the output is the minimum bounding rectangle of the container number region on the target container. The SegNet-CR character segmentation network is used to segment the detected container numbers, deriving several standard characters. Based on the standard characters obtained from the box number recognition process, a pre-constructed heterogeneous OCR model cluster performs parallel recognition on each standard character, obtaining three sets of recognition results. A revision and dynamic weighted fusion strategy is then executed based on these results, comparing the revised first, second, and third confidence levels with the standard threshold range. The heterogeneous OCR model cluster includes at least: a main heterogeneous model, a first sub-heterogeneous model, and a second sub-heterogeneous model. The main heterogeneous model uses a deep convolutional neural network CRNN-CargoNet; the first sub-heterogeneous model uses a Transformer-based visual recognition model ViT-CR; and the second sub-heterogeneous model uses a lightweight model MobileOCR-Edge. The confidence level corresponding to the main heterogeneous model is labeled as the first confidence level, the confidence level corresponding to the first sub-heterogeneous model is labeled as the second confidence level, and the confidence level corresponding to the second sub-heterogeneous model is labeled as the third confidence level. If the corrected first confidence level exceeds the upper limit of the standard threshold range, the recognition result is adopted; If the corrected first confidence level is within the standard threshold range, the recognition results of the heterogeneous OCR model cluster are weighted and voted out. If the corrected first confidence level is lower than the lower limit of the standard threshold range, a secondary judgment mechanism is executed, comparing the corrected second and third confidence levels with the lower limit of the standard threshold range. If both are lower than the lower limit of the standard threshold range, they are marked as abnormal characters; otherwise, they are marked as ambiguous characters.

2. The information processing method for port container number tallying according to claim 1, characterized in that: The process of implementing the correction and dynamic weighted fusion strategy also includes: using the comprehensive quality assessment value as a basis, calculating the mean of the comprehensive quality assessment value for each frame within a preset period to obtain the comprehensive quality mean Q, constructing a linear attenuation model based on the Q value, inputting the original confidence level and the comprehensive quality mean Q, and outputting: the confidence level after first-level correction; The constraints for the first-level corrected confidence level are set as follows: Condition 1: The confidence level after first-level correction exceeds the set threshold; Condition 2: The difference between the confidence levels after the first-level correction does not exceed the error range; when the confidence level after the first-level correction meets the constraint condition, it means that the effect is satisfactory, and the confidence level after the first-level correction shall be used; otherwise, the second-level correction mechanism shall be triggered. Query the historical accuracy database to obtain the corresponding historical average accuracy; construct a piecewise affine transformation calibration model, input the first-level corrected confidence level and the corresponding historical average accuracy, and output the second-level corrected confidence level.

3. The information processing method for port container number tallying according to claim 2, characterized in that: The function upon which the piecewise affine transformation calibration model is based is: P2 = f(P1, A_hist); where P2 represents the confidence level after second-order correction, P1 represents the confidence level after first-order correction, and A_hist represents the corresponding historical average accuracy. The specific expansion formula of the function upon which the piecewise affine transformation calibration model is based is: when P1 is not lower than A_hist, then P2 = A_hist + u*(P1 - A_hist); where u represents the scaling factor, and u∈(0,1); when P1 is lower than A_hist, then P2 = P1.

4. The information processing method for port container number tallying according to claim 1, characterized in that: The method also includes: when abnormal characters are detected, a dynamic error correction and completion mechanism based on context and business rules is triggered; when ambiguous characters are detected, a port-level container number knowledge graph and association reasoning mechanism is triggered.

5. The information processing method for port container number tallying according to claim 4, characterized in that: The dynamic error correction and completion mechanism based on context and business rules is as follows: The container master code database is called to verify the validity of the first few letter combinations; cross-validation is performed after obtaining the pre-provided information; mathematical verification is performed using the container number check code algorithm, converting the first n characters into numerical values, calculating the remainder modulo n+1 after weighted summation, and checking if it matches the (n+1)th character; if it does not match, all character combinations are traversed to calibrate the error position until calibration passes; combining the spatiotemporal context, if the container numbers of adjacent target containers within the same target area are consecutive, and if the container numbers of the preceding and following containers have been confirmed, the intermediate container number is inferred.

6. The information processing method for port container number tallying according to claim 4, characterized in that: The process of establishing a port-level container number knowledge graph and association reasoning mechanism is as follows: construct a container number knowledge graph under the target area, with the container number as the entity node, and the associated attributes include at least: the company, vessel, voyage, port of origin, port of destination, cargo type, historical port entry and exit records, and RFID tag ID; call the container number knowledge graph to perform association reasoning and output the association reasoning results.

Images