A method for automatic archiving of container image data based on computer vision recognition
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANGHAI BOX CLOUD LOGISTICS TECH CO LTD
- Filing Date
- 2023-09-26
- Publication Date
- 2026-06-30
Smart Images
Figure CN117197581B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of container logistics technology, specifically a method for automatically archiving container image data based on computer vision recognition. Background Technology
[0002] A shipping container is a large cargo container with a certain strength, rigidity and specifications, designed for repeated use. When using a shipping container to transport goods, the goods can be loaded directly at the shipper's warehouse and unloaded at the consignee's warehouse. When changing vehicles or ships en route, there is no need to remove the goods from the container for repackaging.
[0003] The standardization of container products and the resulting transportation system enables the standardization of a massive vessel weighing tens of tons. Based on this, a global logistics system integrating ships, ports, shipping routes, highways, transit stations, bridges, tunnels, and multimodal transport can be gradually realized. This is indeed one of the greatest miracles created by mankind in history.
[0004] With continuous economic development and deepening exchanges between countries, freight transportation has become particularly important in economic exchanges between nations. Especially given my country's vast territory, the distances to reach western regions via sea or ports are relatively long, making rail transport advantageous in terms of time and cost, attracting many European customers.
[0005] Since its inception in 2011, the China-Europe Railway Express has opened routes in many cities across my country, with both the number and frequency of trains steadily increasing. There is still much room for cooperation in new areas to explore, and the development prospects are very broad.
[0006] China Railway has put forward requirements for container loading and photographing of containers in the safe operation management of China-Europe freight train import and export trains (see the "Notification Letter from China Railway International Multimodal Transport Co., Ltd. on Strengthening the Safe Operation Management of Overseas Agent Transportation for China-Europe (Central Asia) Return Trains" issued by China Railway International Multimodal Transport Co., Ltd.). All train operations are required to submit photos of container loading for relevant departments of China Railway to review the safety of the goods and the loading of the goods.
[0007] However, the collection, processing, classification, archiving, and transmission of traditional container image data are no different from the processing of ordinary images, requiring multiple stages in the supply chain or logistics chain. This process involves numerous stages, many personnel, and a long timeframe. The standards for manual review and processing are varied, necessitating a significant amount of manual work and a considerable time commitment.
[0008] Based on the above reasons, this invention designs a method for automatically archiving container image data based on computer vision recognition. By training a computer vision recognition model, the method allows users to directly take photos via mobile devices, bypassing intermediate links such as freight companies, freight forwarders, and train platforms. This avoids the rejection of non-compliant loading photos and directly guides users to collect container loading photos that comply with relevant China Railway regulations, meeting the requirements of China Railway's review conditions and safety requirements. This greatly improves the efficiency of container loading review and reduces various costs such as time and manpower. Summary of the Invention
[0009] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for automatically archiving container image data based on computer vision recognition. By training a computer vision recognition model, the method allows workers / drivers to take photos directly, bypassing intermediate links such as freight and train services, and directly meeting the photo requirements of China Railway's relevant train services. This satisfies the review conditions and security requirements, greatly improves transportation efficiency, reduces time, manpower and other costs, and reduces the learning cost for non-professional users.
[0010] To achieve the above objectives, the present invention provides a method for automatically archiving container image data based on computer vision recognition, comprising the following steps:
[0011] S1 trains a computer vision-based container image recognition model, which recognizes the container body, container condition, and container character recognition (OCR) to identify container body parts, container status, container number, and container type. It also sets up a container number and type verification mechanism based on the ISO container standard to verify the accuracy of the container number and type collected by visual recognition.
[0012] S2, identify the uploaded container image data, obtain the container number, container type, container body location and status corresponding to the image, and create a data document for the container based on the identified container number, container type, current date and other reference data;
[0013] S3: Users upload container image data using their mobile devices. The container image recognition model in S1 identifies the container's position and status, and automatically archives and stores the container images according to the rules of the container's position and status.
[0014] Automatic archiving sorts containers based on container number, container type, image, location, date, and status, which is used for subsequent sorting and filtering based on image content;
[0015] S4 classifies the visual recognition values of the container, and the system will issue warnings to users by setting thresholds;
[0016] The threshold is set based on image clarity, rejection of secondary shooting, box recognition, and compliance.
[0017] The identification of reproduced images is performed using moiré patterns, perspective principles, image quality, and feature analysis, specifically:
[0018] Moiré pattern method: This method determines whether a photo has been copied by detecting the moiré pattern generated by the difference between the grid in the copied photo and the grid in the original photo.
[0019] Perspective principle method: Determine whether a photo has been copied by detecting perspective distortion in the cube shape of the container in the copied photo due to the camera not being parallel to the plane of the photo;
[0020] Image quality method: Determine whether a photo has been copied by detecting changes in image quality such as blurriness, color difference, or noise.
[0021] The specific steps of S1 are as follows:
[0022] S1-1, Data Preprocessing: Prepare and preprocess the collected container image data, converting the image data into a digital form that the model can understand, and performing standardization processing, specifically:
[0023] S1-1-1, Data Collection: Use deep learning models to collect data from local or remote data sources, and perform corresponding data loading and preprocessing;
[0024] S1-1-2, Data Cleaning: Using deep learning models to handle missing values, outliers, and duplicate values;
[0025] S1-1-3, Data Transformation: Convert container images into pixel value matrices using pre-trained models in deep learning models or custom models;
[0026] S1-1-4, Data Standardization: Use deep learning models to scale the data and normalize it to a normal distribution, etc.
[0027] S1-1-5, Data Augmentation: Use deep learning models to increase the diversity of pre-trained data to improve the model's generalization ability;
[0028] S1-1-6, Label Encoding: Use sparse coding to convert category labels into numerical form. If it is a binary classification, convert the sparse coding into a dense vector; if it is a multi-class classification, automatically perform one-hot coding through the loss function.
[0029] S1-2, Model Training: The deep learning model preprocessed in S1-1 is used in combination with the suspended chain convolutional neural network computer vision recognition model for training.
[0030] The specific steps of the suspended chain computer vision recognition model are as follows:
[0031] S1-2-1, capturing images of overhead conveyor containers using a high-definition camera;
[0032] S1-2-2, Image preprocessing: Denoise the acquired image, enhance contrast, and perform binarization to improve image quality and recognition accuracy;
[0033] S1-2-3, Feature Extraction: Extract the contour, edge, and texture features of the chain using image processing algorithms;
[0034] S1-2-4, Feature matching: Match the extracted features with the pre-trained model to determine the shape, position, and size of the chain;
[0035] S1-2-5, Result Output: Output the recognition result to the control system to achieve automated control;
[0036] S1-3, Model Evaluation: The trained model is applied to the actual test dataset, and the accuracy of the model is evaluated by accuracy, precision, and recall.
[0037] Accuracy, precision, and recall were calculated by comparing unlabeled, container-related image data with the dataset after model training.
[0038] S1-4, Model Deployment: Convert the trained model into a format suitable for cloud deployment and deploy it to the cloud. Cloud deployment facilitates the collection and aggregation of misidentified or incorrectly identified data. After collecting sufficient data, the labeling, training, and deployment processes can be repeated, allowing the intelligent recognition model to be optimized in applications.
[0039] The numerical form in S1-1 is a matrix of pixel values.
[0040] In S1-1-4, data standardization scales the data to 0 to 1, normalizes it to a mean of 0, and sets the standard deviation to 1.
[0041] The data augmentation tools in S1-1-5 are:
[0042] RandomCropAug enhancement;
[0043] Random Horizontal Flip Aug
[0044] Random Vertical Flip Aug.
[0045] After evaluating the accuracy of the model in S1-3, a second cleaning operation is performed on samples that cannot be identified or are incorrectly identified.
[0046] The storage of S2 images does not rely on preset documents, but is generated autonomously based on the container number information identified from the contents of the container tube image;
[0047] S3 identifies the container type, status, and loaded cargo, and classifies the container types according to their functions, including refrigerated containers, flat rack containers, open top containers, flatbed containers, and side-opening containers. The images show the specific state of the container, such as closed, open, loaded with a small amount of cargo, fully loaded, partially closed, and closed and sealed; as well as the status of the cargo inside the container: stable loading, off-center loading, cargo reinforced, cargo reinforced with air bags, tension straps, triangular wood or wooden strips, container mesh reinforcement, and automatic storage.
[0048] The application of location information in the user image data uploading process involved in S3: the movement of containers requires container trucks. If the location is in a restricted or prohibited area for container trucks during the image uploading process, it can be determined that the image data was not collected and uploaded in real time, and the authenticity of the image is questionable.
[0049] The container type images in S3 are dry cargo containers, refrigerated containers, flat rack containers, open top containers, flatbed containers, and side-opening containers; the status images are closed, open, empty, loaded with a small amount of goods, full of goods, half-closed, closed and sealed, as well as images of stable loading, off-center loading, goods unsecured, goods secured with air bags, secured with tension straps, secured with wire ropes, secured with triangular wood or wooden strips, and secured with container mesh.
[0050] Compared with existing technologies, this invention trains a computer vision recognition model that allows staff / drivers to directly take photos via their mobile devices, bypassing intermediate links such as freight companies, freight forwarders, and train platforms. This avoids the rejection of non-compliant loading photos and directly guides users to collect container loading photos that comply with China Railway's relevant regulations, meeting China Railway's review conditions and safety requirements. This greatly improves the efficiency of container loading review, reduces time, manpower, and other costs, and reduces the learning cost for non-professional users. Attached Figure Description
[0051] Figure 1 This is a schematic diagram of the overall process of the present invention.
[0052] Figure 2 This is a schematic diagram of the image acquisition process of the present invention.
[0053] Figure 3 This diagram illustrates the identification, classification, and prompting processes during the mobile data collection process of this invention.
[0054] Figure 4 This is a schematic diagram illustrating the actual effect of image acquisition, system backend recognition, and evaluation in this invention. Implementation
[0055] The present invention will now be further described with reference to the accompanying drawings.
[0056] See Figures 1-4 This invention provides a method for automatically archiving container image data based on computer vision recognition:
[0057] Includes the following steps:
[0058] S1: Set up and train a container image recognition model based on computer vision recognition, which includes container body, container condition, and container character recognition (OCR) to identify container body parts, container status, container number, and container type; and set up a container number and container type verification mechanism based on the ISO standard labeling of containers to verify the accuracy of the container number and container type collected by visual recognition.
[0059] S1.1 The process of training a container visual recognition model using a general large model includes:
[0060] S1.1.1 Data Preprocessing: First, prepare and preprocess the collected image data. This includes converting the images into a numerical representation that the model can understand (such as a pixel value matrix) and performing necessary standardization (such as scaling and normalization). The specific process is as follows:
[0061] S1.1.1.1 Data Collection: Use the PaddleHub or Paddle Dataset libraries to collect data from local or remote data sources. These libraries provide convenient data loading and preprocessing capabilities.
[0062] S1.1.1.2 Data Cleaning: Use the DataFrame and Series operations provided by general large models (such as PaddlePaddle) to handle missing values, outliers, and duplicate values. For example, use the dropna() function to delete rows or columns containing missing values, and use the drop_duplicates() function to remove duplicate values.
[0063] S1.1.1.3 Data Transformation: For image data, use a pre-trained model in PaddleHub or a custom model to convert the image into a matrix of pixel values.
[0064] S1.1.1.4 Data Standardization: Standardize the data using the transforms operations provided by the general large model (PaddlePaddle). For example, use the StandardScaler() function to scale the data to between 0 and 1, and use the Normalization() function to normalize the data to a normal distribution with a mean of 0 and a standard deviation of 1.
[0065] S1.1.1.5 Data Augmentation: Use data augmentation tools in PaddleHub to perform data augmentation operations: RandomCropAug, RandomHorizontalFlipAug, and RandomVerticalFlipAug. These tools can increase the diversity of training data and improve the model's generalization ability.
[0066] RandomCropAug is a random cropping augmentation method that randomly crops a sub-image of a certain size from the original image as a new input sample. By randomly cropping, the diversity of training samples can be increased, allowing the model to better adapt to inputs of various scales and shapes.
[0067] RandomHorizontalFlipAug is a random horizontal flip augmentation method. This augmentation method is suitable for some symmetrical problems, such as object classification and face recognition.
[0068] RandomVerticalFlipAug is a random vertical flip enhancement method that flips the original image along the vertical axis to obtain a new flipped image as the input sample. Similar to random horizontal flip enhancement, this enhancement method is also suitable for some symmetrical problems.
[0069] S1.1.1.6 Label Encoding: For classification problems, the sparse encoding (SparseTensor) provided by PaddlePaddle is used to convert the class labels into numerical form. For binary classification problems, the to_dense() function is used to convert the sparse encoding into a dense vector; for multi-class classification problems, the CategoricalCrossEntropy() loss function is used to automatically perform one-hot encoding.
[0070] S1.1.2 Model Training: Next, the model is trained using the general-purpose large model tool (PaddlePaddle) and the suspended chain computer vision recognition model. This involves inputting image data into the model and using labels or annotations to guide the model in learning to recognize specific objects or features in the images.
[0071] S1-2-1, capturing images of the suspension chain using a high-definition camera;
[0072] S1-2-2, Image preprocessing: Denoise the acquired image, enhance contrast, and perform binarization to improve image quality and recognition accuracy;
[0073] S1-2-3, Feature Extraction: Extract the contour, edge, and texture features of the chain using image processing algorithms;
[0074] S1-2-4, Feature matching: Match the extracted features with the pre-trained model to determine the shape, position, and size of the chain;
[0075] S1-2-5, Result Output: Output the recognition result to the control system to achieve automated control;
[0076] S1.1.3 Model Evaluation: After training, evaluate the model's performance. This is done by applying the model to the test dataset and using evaluation metrics such as accuracy, precision, and recall to measure the model's accuracy.
[0077] Accuracy, precision, and recall are calculated by comparing unlabeled encoded image data with the dataset after model training.
[0078] After evaluating the accuracy of the model, a second cleaning operation is performed on samples that cannot be identified or are incorrectly identified.
[0079] S1.1.4 Model Deployment: Deploy the trained model to the actual application. Export the model as an executable file or API for running on a server or mobile device.
[0080] S2: The system identifies the uploaded container image data, obtains the container number, container type, container body location, and container status corresponding to the image, and creates a data document for the container based on the identified container number, container type, current date, and other reference data.
[0081] S3: The container image is identified by a container visual recognition model to identify the container position and status, and the container images are automatically archived and stored according to the rules of container position and status.
[0082] Automatic archiving sorts containers based on container number, container type, image, location, date, and status, which is used for subsequent sorting and filtering based on image content;
[0083] S4: Classify the visual recognition values of containers, and issue warnings to users by setting thresholds; the thresholds are set based on image clarity, rejection of secondary shooting, container recognition accuracy, and compliance.
[0084] The threshold is set based on image clarity, identification of copied images, box recognition accuracy, and compliance.
[0085] The identification of reproduced images is performed using moiré patterns, perspective principles, image quality, and feature analysis, specifically:
[0086] Moiré pattern method: This method determines whether a photo has been copied by detecting the moiré pattern generated by the difference between the grid in the copied photo and the grid in the original photo.
[0087] Perspective principle method: Determine whether a photo has been copied by detecting perspective distortion in the cube shape of the container in the copied photo due to the camera not being parallel to the plane of the photo;
[0088] Image quality method: Determine whether a photo has been copied by detecting changes in image quality such as blurriness, color difference, or noise.
[0089] The specific steps of S1 are as follows:
[0090] S1-1, Data Preprocessing: Prepare and preprocess the collected container image data, converting the image data into a digital form that the model can understand, and performing standardization processing, specifically:
[0091] S1-1-1, Data Collection: Use deep learning models to collect data from local or remote data sources, and perform corresponding data loading and preprocessing;
[0092] S1-1-2, Data Cleaning: Use the deep learning model to handle missing values, outliers, and duplicate values;
[0093] S1-1-3, Data Transformation: Convert the container image into a pixel value matrix using a pre-trained model or a custom model in the deep learning model;
[0094] S1-1-4, Data Standardization: The data is scaled and normalized to a normal distribution using the deep learning model described above;
[0095] S1-1-5, Data Augmentation: Use the deep learning model to increase the diversity of the pre-trained data to improve the model's generalization ability;
[0096] S1-1-6, Label Encoding: Use sparse coding to convert category labels into numerical form. If it is a binary classification, convert the sparse coding into a dense vector; if it is a multi-class classification, automatically perform one-hot coding through a loss function.
[0097] S1-2, Model Training: The deep learning model preprocessed in S1-1 is used in combination with the catenary computer vision recognition model for training.
[0098] The specific steps of the suspended chain computer vision recognition model are as follows:
[0099] S1-2-1, Acquire high-resolution images of overhead conveyor containers;
[0100] S1-2-2, Image preprocessing: Denoise the acquired image, enhance contrast, and perform binarization to improve image quality and recognition accuracy;
[0101] S1-2-3, Feature Extraction: Extract the contour, edge, and texture features of the chain using image processing algorithms;
[0102] S1-2-4, Feature matching: Match the extracted features with the pre-trained model to determine the shape, position, and size of the chain;
[0103] S1-2-5, Result Output: Output the recognition result to the control system to achieve automated control;
[0104] S1-3, Model Evaluation: The trained model is applied to the actual test dataset, and the accuracy of the model is evaluated by accuracy, precision, and recall.
[0105] The accuracy, precision, and recall are calculated by comparing the unlabeled, container-related image data with the dataset after model training.
[0106] S1-4, Model Deployment: Convert the trained model into a format suitable for cloud deployment and deploy it to the cloud. Cloud deployment facilitates the collection and aggregation of misidentified or incorrectly identified data. After collecting sufficient data, repeat the labeling, training, and deployment process to optimize the intelligent recognition model in applications.
[0107] The numerical form in S1-1 is a matrix of pixel values.
[0108] In S1-1-4, data standardization scales the data to 0 to 1, normalizes it to a mean of 0, and sets the standard deviation to 1.
[0109] The data augmentation tools in S1-1-5 are RandomCropAug, RandomHorizontalFlipAug, and RandomVerticalFlipAug.
[0110] After evaluating the accuracy of the model in S1-3, a second cleaning operation is performed on samples that cannot be identified or are incorrectly identified.
[0111] The storage of S2 images does not rely on preset documents, but is generated autonomously based on the container number information identified from the contents of the container tube image;
[0112] S3 identifies the container type, status, and loaded cargo, and classifies the container types according to their functions, including refrigerated containers, flat rack containers, open top containers, flatbed containers, and side-opening containers. The images show the specific state of the container, such as closed, open, loaded with a small amount of cargo, fully loaded, partially closed, and closed and sealed; as well as the status of the cargo inside the container: stable loading, off-center loading, cargo reinforced, cargo reinforced with air bags, tension straps, triangular wood or wooden strips, container mesh reinforcement, and automatic storage.
[0113] The application of location information in the user image data uploading process involved in S3: the movement of containers requires container trucks. If the location is in a restricted or prohibited area for container trucks during the image uploading process, it can be determined that the image data was not collected and uploaded in real time, and the authenticity of the image is questionable.
[0114] The container types and images in S3 are dry cargo containers, refrigerated containers, flat rack containers, open-top containers, flatbed containers, and side-opening containers, respectively; the status images are closed, open, empty, loaded with a small amount of goods, full of goods, half-closed, closed and sealed, as well as images of stable loading, off-center loading, goods unsecured, goods secured with air bags, secured with tension straps, secured with wire ropes, secured with triangular wood or wooden strips, and secured with container mesh.
[0115] The above are merely preferred embodiments of the present invention, intended only to aid in understanding the method and core ideas of this application. The scope of protection of the present invention is not limited to the above embodiments; all technical solutions falling within the scope of the present invention's concept are within its protection. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.
[0116] This invention comprehensively solves the problems in existing technologies regarding the collection, organization, classification, archiving, and transmission of container image data, which involve numerous personnel, long processing times, complex standards for manual review and organization, and require a significant amount of manual work and time. By training a computer vision recognition model, it enables users to directly capture container loading photos that comply with relevant China Railway regulations, bypassing intermediate links such as freight companies, freight forwarders, and train platforms. This avoids the rejection of non-compliant photos, meets the review conditions and security requirements of China Railway, greatly improves the efficiency of container loading review, reduces time, manpower, and other costs, and minimizes the learning costs for non-professional users.
Claims
1. A method for automatic archiving of container image data based on computer vision recognition, characterized by, Includes the following steps: S1, train a computer vision-based container image recognition model. The recognition content includes container body, container condition, and container character recognition (OCR), used to identify container body parts, container status, container number, and container type; and set a container number and type verification mechanism based on the ISO container standard to verify the accuracy of the container number and type collected by visual recognition. S2, identify the uploaded container image data, obtain the container number, container type, container body position and status corresponding to the image, and create a data document for the container based on the identified container number, container type and current date reference data; S3, using the container image recognition model in S1, identifies the container position and status of the container image, and automatically archives and stores the container image according to the rules of container position and status. The automatic archiving sorts containers based on container number, container type, image, container location, date, and status, which is used for subsequent sorting and filtering based on image content. The identification of container status and loaded cargo specifically includes: identifying whether the cargo inside the container is stably loaded or unevenly loaded, and identifying the cargo reinforcement method as being reinforced with air bags, tension straps, triangular wood or wooden strips, or container mesh; obtaining the location information during the process of user-uploaded container image data; if the location information is in a restricted or prohibited area for container trucks, it is determined that the image data was not collected and uploaded in real time, and the authenticity of the image is questionable. S4 classifies the visual recognition values of the container, and the system will issue warnings to users by setting thresholds; The threshold is set based on image clarity, identification of re-photographing, box recognition accuracy, and compliance. The identification and reproduction are performed using moiré patterns, perspective principles, image quality, and feature analysis, specifically: Moiré pattern method: This method determines whether a photo has been copied by detecting the moiré pattern generated by the difference between the grid in the copied photo and the grid in the original photo. Perspective principle method: Determine whether a photo has been copied by detecting perspective distortion in the cube shape of the container in the copied photo due to the camera not being parallel to the plane of the photo; Image quality method: Determine whether a photo has been copied by detecting changes in image quality such as blurriness, color difference, or noise.
2. The method for automatic archiving of container image data based on computer vision recognition according to claim 1, characterized in that, The specific steps in S1 for training the computer vision-based container image recognition model include: S1-1, Data Preprocessing: Prepare and preprocess the collected container image data, converting the image data into a digital form that the model can understand, and performing standardization processing, specifically: S1-1-1, Data Collection: Use deep learning models to collect data from local or remote data sources, and perform corresponding data loading and preprocessing; S1-1-2, Data Cleaning: Use the deep learning model to handle missing values, outliers, and duplicate values; S1-1-3, Data Transformation: Convert the container image into a pixel value matrix using a pre-trained model or a custom model in the deep learning model; S1-1-4, Data Standardization: The data is scaled and normalized to a normal distribution using the deep learning model described above; S1-1-5, Data Augmentation: Use the deep learning model to increase the diversity of the pre-trained data to improve the model's generalization ability; S1-1-6, Label Encoding: Use sparse coding to convert category labels into numerical form. If it is a binary classification, convert the sparse coding into a dense vector; if it is a multi-class classification, automatically perform one-hot coding through a loss function. S1-2, Model Training: The deep learning model preprocessed in S1-1 is used in combination with the catenary computer vision recognition model for training. The specific steps of the suspended chain computer vision recognition model are as follows: S1-2-1, Acquire high-resolution images of overhead conveyor containers; S1-2-2, Image preprocessing: Denoise the acquired image, enhance contrast, and perform binarization to improve image quality and recognition accuracy; S1-2-3, Feature Extraction: Extract the contour, edge, and texture features of the chain using image processing algorithms; S1-2-4, Feature matching: Match the extracted features with the pre-trained model to determine the shape, position, and size of the chain; S1-2-5, Result Output: Output the recognition result to the control system to achieve automated control; S1-3, Model Evaluation: The trained model is applied to the actual test dataset, and the accuracy of the model is evaluated by accuracy, precision, and recall. The accuracy, precision, and recall are calculated by comparing unlabeled container-related image data with the dataset after model training. S1-4, Model Deployment: Convert the trained model into a format suitable for cloud deployment and deploy it to the cloud.
3. The method for automatic archiving of container image data based on computer vision recognition according to claim 2, characterized in that, The digital form in S1-1 is a pixel value matrix.
4. The method for automatically archiving container image data based on computer vision recognition according to claim 2, characterized in that, In S1-1-4, data standardization scales the data to 0 to 1, and normalization is performed to a mean of 0 and a standard deviation of 1.
5. The method for automatically archiving container image data based on computer vision recognition according to claim 2, characterized in that, The data augmentation tools in S1-1-5 are RandomCropAug, RandomHorizontalFlipAug, or RandomVerticalFlipAug.
6. The method for automatically archiving container image data based on computer vision recognition according to claim 1, characterized in that, After evaluating the accuracy of the model in steps S1-3, a second cleaning operation is performed on samples that cannot be identified or are incorrectly identified.
7. The method for automatically archiving container image data based on computer vision recognition according to claim 1, characterized in that, The storage of the S2 image does not depend on a preset document, but is generated autonomously based on the container number information identified from the contents of the container tube image.
8. The method for automatically archiving container image data based on computer vision recognition according to claim 1, characterized in that, In step S4, the visual recognition values of the container are classified. By setting a threshold, the user determines that the container image data with the characteristics of being copied is not collected and uploaded in real time, and the authenticity of the image is questionable.
9. The method for automatically archiving container image data based on computer vision recognition according to claim 1, characterized in that, The box types in S3 are dry goods boxes, refrigerated boxes, frame boxes, top-opening boxes, flat-panel boxes, and side-opening boxes; The status images include images of closed doors, open doors, empty containers, containers loaded with a small amount of goods, containers fully loaded, partially closed doors, closed and sealed doors, stable loading, unbalanced loading, goods secured with air bags, secured with tension straps, secured with wire ropes, secured with triangular blocks or wooden strips, and containers secured with mesh.