Container number and container threshold weight identification method and device based on deep learning

Abstract

The invention relates to a container number and container threshold weight identification method and device based on deep learning. The method comprises the following steps: acquiring a side image of a container body; inputting the side image of the box body into a target detection model, and detecting the coordinates of the area where each target is located and the corresponding target type; cutting out a target image from the box side image, inputting transverse box number images into a box number character block detection model, correspondingly outputting a prediction set by each transverse box number image, detecting character blocks in the prediction sets by adopting non-maximum suppression to obtain a detection set of the character blocks, cutting out character block images, and splicing the character block images to obtain a detection set of the character blocks; obtaining a transverse box number spliced image; and inputting the transverse box number spliced image and other target images into a trained character recognition model, and recognizing the box number, the ISO number and the box threshold weight limiting character. The method has the advantages that the container number and the container threshold weight can be automatically and accurately recognized, the recognition accuracy is high, the method is suitable for various complex container images, and the good recognition effect can be achieved.

Description

Technical Field

[0001] The present invention relates to the field of computer character recognition models, and in particular to a method and device for recognizing container numbers and weight limits based on deep learning. Background Art

[0002] In order to effectively eliminate the potential safety hazards brought by overweight containers to port loading and unloading operations, ship transportation, etc., it is necessary not only to solve the problem of container number recognition, but also to provide container door weight limits. The container number will be printed on all four sides of the container body. The container number is composed of a company code and a container number. The company code is composed of letters, and the container number is composed of numbers, but the entire container number is different and irregular. The format, font, and layout of the printing of different countries and container companies are all different. The weight limit information of the container generally only appears on the image of the back side of the box, that is, on the container door picture. In the current container transshipment process, the acquisition of the container number and the door weight limit mainly relies on manual recording after manual visual reading. This method not only greatly increases the labor intensity of the operator, but also inevitably increases the error rate in the data recording process, which in turn has an adverse effect on the overall efficiency of the container transshipment operation. In the prior art, the solution of recognizing characters in the image after deep learning through a multi-layer convolutional neural network has been widely used. However, the container number image on the side of the container body is different from the existing text image, please refer to Figure 1Its particularity lies in: 1. When the container number is printed on the container, there are multiple layout methods, horizontal layout, vertical layout, and there are multiple formats for the horizontal layout method, such as single-line layout, double-line layout, two adjacent characters staggered up and down, and an uncertain number of spaces between the company code and the container number. There are also gaps of different spacing between the container numbers, which undoubtedly increases the learning difficulty of the neural network used for character recognition; 2. Due to the continuous undulating steel plate structure characteristics of the container side panel, when printing the container number, each character may not be completely printed on the convex or concave surface. Taking the printing of italic characters as an example, when some characters are printed on the connecting surface of the concave and convex surfaces, the deformation of the characters in the collected image is aggravated and the spacing between the two adjacent characters changes. The adjacent characters of the overall container number are The spacing is also inconsistent, which increases the difficulty of neural network recognition and easily misidentifies the interval as some characters with similar shapes; 3. The information on the rear side door of the container is the most complete, including the container number, ISO and the weight limit information of the container door, but the frame of the door and the middle seam formed after the two door panels are closed, as well as the longitudinal reinforcement rods on some doors, all interfere with character recognition; 4. The image collectors used to collect the side panels of the container body mostly work in an open-air environment. When encountering haze and heavy rain, the visibility is low, the image clarity is poor, and rainwater is easy to stick to the lens of the image collector, so that the container number characters on the collected image are deformed and distorted. At the same time, the interval between two adjacent characters is also deformed, which is not a regular square space, and the neural network will misidentify the interval as some characters with similar shapes. The particularity of the characters in the above images will result in a very low recognition accuracy when the collected images are directly used in the character recognition model, and cannot be applied to on-site recognition. Therefore, there is an urgent need for a technical solution that can deeply extract image character features and recognize the container number characters in the image with high precision. Summary of the invention

[0003] In order to solve the above problems, the purpose of the present invention is to provide a container number and weight limit recognition method based on deep learning, which can automatically and accurately identify the container number and door weight limit, thereby significantly reducing human errors, alleviating the burden on operators, and significantly improving the efficiency and accuracy of container transshipment operations.

[0004] To achieve the above object, the present invention adopts the following technical solutions: Technical Solution 1

[0005] The container number and door weight limit recognition method based on deep learning includes the following steps: The side image of the container is collected and preprocessed, and the side image of the container includes the container number, ISO number and weight limit information of the container; the side image of the container is input into a trained target detection model to detect the coordinates of the area where each target is located and the corresponding target type; the target includes the container number, ISO number and weight limit information, and the type of the target includes horizontal container number, horizontal ISO number, vertical column container number, vertical column ISO number and container door weight limit; the target image is cropped from the side image of the container, and then the following processing is performed: if there is an image with a target type of horizontal container number in the target image, the target image of the horizontal container number is input into a trained container number character block detection model, and each horizontal container number image corresponds to a prediction set containing multiple character block information, and non-maximum suppression is used to test the character blocks in the prediction set to obtain the detection set of the character blocks that are finally retained, and the character block images are cropped from the horizontal container number image according to the detection set and spliced ​​to obtain a horizontal container number spliced ​​image; the horizontal container number spliced ​​image and other remaining target images are input into a trained character recognition model to identify the container number, ISO number and container door weight limit characters.

[0006] Based on the same inventive concept, the present invention also provides a container number and door weight limit recognition device based on deep learning.

[0007] Technical Solution 2

[0008] A container number and door weight limit identification device based on deep learning includes a memory storing an executable program and a processor, wherein the processor runs the program to execute the method steps in technical solution one.

[0009] The present invention has the following beneficial effects: 1. The present invention improves the recognition scheme of difficult-to-recognize horizontal container numbers. First, the cropped horizontal container number image is subjected to character block recognition through the improved EAST model, and the character blocks are accurately extracted. Then, the redundant blocks are removed and spliced, thereby eliminating the interference caused by unequal spacing, distorted shape spacing, multi-line format, and character up and down misalignment format to character recognition, greatly improving the character recognition accuracy, and being applicable to various complex container image scenes. Whether it is different lighting conditions or container numbers and weight limit marks in multiple font formats, a high recognition effect can be achieved.

[0010] 2. The present invention adopts a CRNN-based structure for the character recognition model, and the backbone network uses MobileNetV3, which reduces the computational complexity and parameter scale of the model and improves lightweight performance. During the training process, the optimizer uses Adam, and combines the Cosine attenuation strategy to adjust the learning rate, so that the model can continue to optimize in the later stages of training. Furthermore, L2 regularization is used to prevent overfitting, and finally the CTC loss function is used to effectively process variable-length sequences. By improving the model structure and algorithm, the accuracy and computational efficiency of character recognition are improved.

[0011] 3. The present invention uses deep learning technology to automatically identify container numbers and door weight limits through multi-layer convolutional neural networks and progressive learning mechanisms, which greatly reduces the labor burden of operators, avoids the tedious process of manual visual reading and manual recording, and improves work efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is a schematic diagram of a horizontal box number image and a side image of a part of a box door in the prior art; Figure 2 It is a schematic diagram of the process of the present invention (taking the image of the rear side of the box as an example); Figure 3 is a schematic diagram of a character block detection model of the present invention; Figure 4 Schematic diagram of the residual block structure of the present invention; Figure 5 It is a schematic diagram of the SE module structure of the present invention; Figure 6 It is the horizontal box number image spliced ​​before and after the horizontal box number image detection of the present invention. DETAILED DESCRIPTION

[0013] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments: Embodiment 1

[0014] See also Figure 2 , a container number and door weight limit recognition method based on deep learning includes the following steps: Step 10: Obtain a side image of the container through an image collector such as a camera, and then pre-process the side image of the container. The side image of the container includes the container number, ISO number, and weight limit information of the container.

[0015] Preferably, images of the four sides of the container are collected, including the left side image, the right side image, the front side image and the back side image of the container. The information and printing format contained in each side image of the container are inconsistent. By collecting images from four sides and performing character recognition on each side, and then comparing the recognition results of each side image to check whether the recognition results are consistent, the accuracy of the recognition results can be further improved.

[0016] The preprocessing process is: converting the box side image into a grayscale image, and then performing Gaussian filtering to remove image noise.

[0017] Step 20: Input the box side image into the trained target detection model to detect the coordinates of the area where each target is located and the corresponding target type. The target includes the box number, ISO number and weight limit information, and the target type includes horizontal box number, horizontal ISO number, vertical box number, vertical ISO number and box door weight limit.

[0018] The training process of the target detection model is: Data collection: The collected preprocessed box side images in four directions are called container detection datasets, and are divided into training set, test set and validation set. The image samples of the training set account for 80% of the total images, and the validation set and test set each account for 10%.

[0019] Data annotation: Use the labelImg tool to label the container image, mark the target area and target type, and generate the corresponding label file (such as VOC or COCO format).

[0020] Data Augmentation: Random data augmentation techniques (including random image translation, random image rotation, and random image tilt) are used to improve the generalization ability and accuracy of the model.

[0021] Training configuration: Through multiple rounds of training, the model hyperparameters are adjusted, including learning rate: 0.0001, batch size (BatchSize): 16, and input image size: (640, 640) to improve the detection accuracy and robustness of the model, and a trained YOLOv5 container number detection model is obtained.

[0022] S23. Input the validation set images into the trained YOLOv5 container number detection model for testing. The test output results include the rectangular box position (left, top, right, bottom) of each detected target area and the preliminary target category (horizontal box number, horizontal ISO number, vertical box number, vertical ISO number, box door weight limit).

[0023] Step 30: crop a target image from the box side image according to the target area coordinates, including a horizontal box number image, a horizontal ISO number image, a vertical box number image, a vertical ISO number image and a box door weight limit image.

[0024] Step 40: Process the cropped images according to the target type: The horizontal ISO number image and the box gate weight limit image can be used as the input of the character recognition model in step 50 without any processing; For the vertical box number image and the vertical ISO number image, the image is rotated 90° and used as the input of the character recognition model in step 50; The horizontal box number image is processed as follows: First, character block recognition: input the trained box number character block detection model, and each horizontal box number image corresponds to the output of a prediction set containing multiple character block information. The box number character block detection model is an algorithm based on the improved EAST model, which optimizes the accuracy of character block positioning in long box number texts and complex climate environments. The EAST model includes a feature extraction module, a feature fusion module and an output module. The improvement includes constructing a ResNet50 deep neural network model to replace the feature extraction module in the original EAST, and introducing the SE channel attention module into the feature fusion module. The SE channel attention module enhances the dependency and information interaction between channels, and adaptively adjusts the weights of each channel, thereby enhancing and fusion the features of each channel, and further improving the model's detection accuracy for character blocks. The details are as follows: See also Figure 3 and Figure 4 First, a ResNet50 deep neural network model is constructed to replace the feature extraction module in the original EAST model to obtain more accurate and comprehensive image feature information. Among them, the ResNet50 of the present invention is composed of 50 layers of residual blocks, and the backbone feature network is composed of multi-level residual block (ResidualBlock) modules. Each level of residual block module contains at least three residual blocks, and the number of channels of each level of residual block module doubles step by step according to the data transmission direction. The residual block is composed of a 1×1 convolutional layer, a 3×3 convolutional layer and a 1×1 convolutional layer, which are used for dimensionality reduction, feature extraction and dimensionality increase operations, and process the input feature map respectively. Exemplarily, the network structure of ResNet50 in this embodiment includes a four-level residual block module, which is divided into four processing stages, which are specifically described as follows: Stage 1: Contains 3 residual blocks with an output channel number of 256. Each residual block is reduced in dimension by a 1×1 convolutional layer, spatial feature extraction by a 3×3 convolutional layer, and finally increased in dimension by a 1×1 convolutional layer. This stage downsamples the size of the input image.

[0025] Stage 2: Contains 4 residual blocks and has 512 output channels. Dimensionality reduction is performed using 1×1 convolution, features are extracted using 3×3 convolution, and dimensionality is increased using 1×1 convolution. The size of the feature map is further reduced in this stage.

[0026] Stage 3: Contains 6 residual blocks with an output channel number of 1024. This stage increases in depth and uses the same convolutional structure to perform deeper feature extraction and size reduction on the input feature map.

[0027] Stage 4: Contains 3 residual blocks and the number of output channels is 2048. Through the residual connection, the input feature map undergoes further convolution operations to obtain a higher-level feature representation.

[0028] After the above four stages, the input image size continues to decrease, while the number of channels gradually increases, forming the extraction of deep image features and the fusion of multi-scale information.

[0029] The steps of the feature extraction module of the box number character block detection model are as follows: the horizontal box number image passes through the convolution layer, normalization layer, activation layer, and maximum pooling layer, and then passes through a four-level residual block module for feature extraction, and finally passes through the average pooling layer and the fully connected layer to output the result to the feature fusion module.

[0030] Second, see Figure 3 and Figure 5 In the feature fusion module of the original EAST, the SE (Squeeze-and-Excitation) channel attention module is introduced as an attention mechanism to improve the feature expression ability of the network, focusing on the adaptive adjustment of the weights of feature channels. The SE module achieves feature enhancement of each channel through two key steps: Squeeze and Excitation. In the image enhancement module of the original EAST model, during the convolution process, the convolution operation of a single feature map is limited to the local spatial perception field of view, lacks information interaction between channels, and the concept of channel features is introduced only when the accumulation operation is performed at the end. The main purpose of the Squeeze operation is to break the weak correlation between channels established only by the accumulation operation, and enhance the dependency and information interaction between channels. The formula is expressed as follows: , , where H and W represent the height and width of the feature map respectively. Represents a comprehensive feature description, is the feature map output by the cth channel, * represents a 2D convolution operation, is the 2D convolution kernel, Describe the feature matrix of the input s-th channel, Represents the total number of input channels. After performing 2D convolution on all input channels, the output channels are summed up. , the feature map of each input channel will extract different spatial features through the 2D convolution kernel, and these features will be weighted and summed to obtain the comprehensive features of each output channel , the comprehensive features of each channel Compressed into a scalar, forming a 1×1×S feature output, which is the global feature description of the channel, where S represents the dimension; the connection between channels is opened up through the Squeeze compression operation, and then the feature relationship between channels is obtained by the Excitation excitation operation, which is expressed by weights. The Excitation excitation operation obtains the channel weights through two fully connected operations and the Sigmoid activation function: first, the global feature description is reduced in dimension through the first fully connected layer, and the ReLU activation function is used to increase nonlinearity. Then, the second fully connected layer restores the dimension to the original number of channels. Finally, the weights are normalized by the Sigmoid function and adjusted to the range of 0 to 1 to complete the adaptive weighting of each channel. Finally, the channel weights are applied to the feature matrix by element-by-element multiplication through the Scale function, but the spatial size of the feature matrix is ​​not changed.

[0031] Next, we build a box number character block detection model based on the EAST improvement, and create a horizontal box number image dataset, where the image samples of the training set account for 80% of the total images, and the validation set and test set each account for 10%; for each image sample, the label file contains the four coordinate points of the box number character block (upper left, lower left, upper right, lower right). And set the training hyperparameters as follows: learning rate: 0.0001, batch size (BatchSize): 16, input image size: (640,640), training batch: 500. Set the loss function to , is the score map loss (ScoreMapLoss), is the geometric graph loss (GeometryLoss), the loss of the geometry map is used to measure the difference between the geometry of the box character block predicted by the model and the geometry in the true label. is the importance coefficient that balances these two losses, which can be Set to 1. Score map loss The (ScoreMapLoss) formula is expressed as:

[0032] in, is the forward propagation prediction result, is the true value (label value), The coefficient for balancing the amplitude samples is calculated as:

[0033] Through the above steps, the training data set is input into the box number character block detection model constructed based on the EAST improvement. After forward propagation, the set loss function is used to calculate the loss value between the prediction result obtained by forward propagation and the true label. Finally, the model weight is adjusted through back propagation optimization. This operation is repeated for 500 training batches to obtain the optimized box number character block detection model.

[0034] Input the horizontal box number image into the trained box number character block detection model to obtain the character block prediction set ,in The character block information includes the bounding box coordinates , such as the upper left corner (x1, y1) and the lower right corner (x2, y2), as well as the confidence that the box contains the character block target Each character block contains one or more characters.

[0035] Then, the character blocks in the prediction set are tested using non-maximum suppression to obtain the detection set of the character blocks that are finally retained, including: Confidence ranking: confidence for each predicted character block Sort in descending order; Calculate the intersection ratio: Calculate the character block with the highest confidence The intersection and union ratio of the bounding box of the character block and the bounding boxes of other character blocks is expressed as: ,in, Indicates the character block with the highest confidence. Represents any other character block; Suppress redundant boxes: If > , then delete ,reserve , is the threshold value, which is set to 0.70 here; Iterative processing: Repeat the steps of calculating the intersection-over-union ratio and suppressing redundant boxes until all other character blocks are processed to obtain the final detection set of retained character blocks. , n≤i, n and i both represent the number of targets in the set.

[0036] Finally, if Figure 6 As shown, according to the detection set, the character block images are cropped from the horizontal box number image and spliced ​​to obtain the horizontal box number spliced ​​image. Specifically, the cropped n character block images ( ) to obtain horizontal splicing , which is set in advance The size of the image is defined as × , then we get the horizontally stitched image The size is ×( ), the splicing formula is: ,in, is the height, is the width, x represents the row coordinate of the image, and y represents the column coordinate of the image. is the sum of the widths of the first n-1 images, used to locate the horizontal position of the image after stitching. It is the pixel value of the nth character block image in the splicing.

[0037] Combination Figure 2 and Figure 6 In the unstitched image, characters may be difficult to recognize due to occlusion, distortion or various different formats, while the horizontal box number stitching image eliminates the character segmentation deviation, solves the correction angle problem, restores the complete character form, enhances the character continuity, and effectively improves the accuracy of character recognition.

[0038] Step 50: Input the horizontal box number stitching image, the horizontal ISO number image, the box door weight limit, the turned vertical ISO number image and the vertical box number image into the trained character recognition model to recognize the box number, ISO number and / or box door weight limit characters contained in each image.

[0039] The character recognition model structure adopts CRNN (Convolutional Recurrent Neural Network), the backbone network uses MobileNetV3, and the scaling factor is set to 0.5; the character recognition model training optimizer adopts Adam, and the optimization parameters are set to , 99, the learning rate adopts the Cosine decay strategy, and the initial value is set to 0.001. L2 regularization (L2: RidgeRegression, ridge regression) is used, which adds a penalty term to the loss function, that is, a factor, which encourages the value of the model parameter to be as small as possible, thereby avoiding overfitting of the model to the training data. The factor can be set to 0.00001:

[0040]

[0041]

[0042]

[0043] in, is the network weight, For The deviation correction of For The deviation correction is , t is the number of iterations, is the learning rate, is the current gradient, and is the momentum term parameter; The training loss function is set to CTC (Connectionist Temporal Classification):

[0044] Among them, y is the label sequence, x is the input image sequence, The probability of output prediction; The cropped target images in step 30 are accumulated to obtain a data set, which is divided into a training set, a test set, and a validation set. The training set samples account for 80%, and the test set and validation set each account for 10%. The number of training iterations is set to 500, and the batch size is set to 16; then the model is iteratively learned. The steps for a single iteration are as follows: Input the training set into the constructed recognition model structure network to generate a set of feature vectors X=( , ... ),in represents the features extracted by the model at time step t. Next, the model calculates the probability of the characters that may be output at each time step based on these feature vectors.

[0045] Calculate character probabilities: For each feature vector given at time step t , calculate the conditional probability distribution of the output character at this time step ,in is the character that may be output at time t. This process is similar to predicting the most likely character to be output at each moment.

[0046] Calculate path probability: Calculate the total probability of all possible paths (i.e. all character sequences) based on the character probability distribution of each time step in the previous step. Path probability formula Given the entire sequence of feature vectors Next specific path probability.

[0047] Final sequence probability: The total probability of the complete output sequence needs to consider all possible paths. By summing all possible paths, the model calculates the observed sequence The total probability of . This process uses the path summation formula: = ; Calculate loss: Calculate CTC loss , and get a value that measures the difference between the current prediction result and the target label.

[0048] Back propagation: By calculating the gradient of the loss function with respect to all parameters of the model, that is, the partial derivative of the loss function with respect to each parameter, the obtained gradient is adjusted through the Adam optimization function so that the model's prediction is closer to the target label.

[0049] In this embodiment, the character recognition model adopts a CRNN-based structure, and the backbone network uses MobileNetV3, which can reduce the computational complexity and parameter scale of the model and improve lightweight performance. During the training process, the optimizer uses Adam, and combines the Cosine attenuation strategy to adjust the learning rate so that the model can continue to optimize in the later stages of training. Furthermore, L2 regularization is used to prevent overfitting, and finally the CTC loss function is used to effectively process variable-length sequences. The accuracy and computational efficiency of character recognition are improved through model improvements.

[0050] Step 60: Since each side image contains the box number, the character recognition result of the target type of vertical box number in the target image is compared with the character recognition result of the target type of horizontal box number to determine whether the box numbers recognized in two different directions are consistent. If they are inconsistent, the box numbers with lower confidence are removed through screening by comparing the confidence of the recognition results to obtain a unique box number, thereby further improving the accuracy of the box number recognition result.

[0051] Step 70: Since the container number has a clear container number composition rule, the container number can also be verified by a container number verification algorithm, as follows: Input processing: If the input box number is empty or contains only blank characters, no verification is performed and True is returned directly, indicating that no verification is required. Otherwise, the first and last blank characters of the input box number are removed.

[0052] Length check: If the box number is 11 digits long, the next step of verification will be performed; otherwise, False will be returned directly.

[0053] Character conversion: For the first 4 letters, convert each letter into the corresponding digital code according to the character conversion rules of the international standard: If the letter is "A", the corresponding code is 10.

[0054] If the letter is between "V" and "Z", the corresponding numeric code is ord(letter)-52.

[0055] If the letter is between "L" and "U", the corresponding numeric code is ord(letter)-53.

[0056] The encoding of the remaining letters is ord(letter)-54.

[0057] The ord() function converts a character to its corresponding ASCII or Unicode code point (integer value). Digital part processing: For the 5th to 11th digits, directly convert them into integers.

[0058] Checksum calculation: Calculate the weighted sum of the first 10 characters (including converted letters and numbers) bit by bit. The weighted value is a power of 2, that is, the first bit is multiplied by 1, the second bit is multiplied by 2, and so on.

[0059] If the first 4 digits of the box number are "HLCU", subtract 2 from the weighted sum.

[0060] Check code verification: Take the weighted sum modulo 11 to get the check code. If the check code is 10, set it to 0 according to the rules.

[0061] Compare the calculated check code with the 11th digit of the box number (i.e. the entered check code): If they are equal, the verification is successful and True is returned.

[0062] If they are not equal, validation fails and False is returned.

[0063] Special processing: For box numbers with the box number prefix "HLCU", special processing is required when calculating the checksum, that is, subtract 2.

[0064] Step 80, after the verified and calculated container number is obtained in step 70, and the ISO number and load limit information are identified in step 50, the complete and highly accurate container information (container number, ISO number, load limit) is finally integrated.

[0065] The container number and box door weight limit recognition method based on deep learning of the present invention uses deep learning technology, through a multi-layer convolutional neural network and a progressive learning mechanism, to automatically recognize the container number and box door weight limit, which greatly reduces the labor burden of operators, avoids the tedious process of manual visual reading and manual recording, and improves work efficiency. At the same time, the EAST model is further improved, with good generalization ability, whether it is different lighting, climate conditions, or box numbers in multiple font formats, it can achieve a high degree of recognition accuracy.

[0066] Based on the same inventive concept, the present invention also provides a device for executing the method described in the first embodiment. Embodiment 2

[0067] A container number and door weight limit identification device based on deep learning includes a memory storing an executable program and a processor, wherein the processor runs the program to execute the method steps described in the first embodiment.

[0068] Since the device introduced in the second embodiment of the present invention is a hardware carrier for implementing the method of the first embodiment of the present invention, those skilled in the art can understand the specific implementation of the device based on the method introduced in the first embodiment of the present invention, so it is not repeated here. All methods adopted in the first embodiment of the present invention belong to the scope of protection of the present invention.

[0069] The above description is only a specific implementation mode of the present invention, and does not limit the patent scope of the present invention. Any equivalent structural transformation made by using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for identifying container numbers and door weight limits based on deep learning, characterized in that: The steps include: Collect and pre-process the side image of the container, wherein the side image of the container includes the container number, ISO number and weight limit information of the container; Input the box side image into the trained target detection model to detect the coordinates of the area where each target is located and the corresponding target type; the target includes the box number, ISO number and weight limit information, and the target type includes the horizontal box number, horizontal ISO number, vertical box number, vertical ISO number and box door weight limit; The target image is cropped from the box side image and then processed as follows: If there is an image of horizontal box number as the target type in the target image, the target image of the horizontal box number is input into the trained box number character block detection model, and each horizontal box number image outputs a prediction set containing multiple character block information, and the character blocks in the prediction set are tested by non-maximum suppression to obtain the detection set of the character blocks that are finally retained, and the character block images are cropped from the horizontal box number image according to the detection set and spliced ​​to obtain a horizontal box number spliced ​​image; the box number character block detection model is an algorithm improved based on the EAST model, and the EAST model includes a feature extraction module, a feature fusion module and an output module; the improvement includes constructing a ResNet50 deep neural network model to replace the feature extraction module in the original EAST, introducing the SE channel attention module into the feature fusion module, and the SE channel attention module enhances the dependency and information interaction between channels, and adaptively adjusts the weights of each channel, so as to enhance and fuse the features of each channel; The horizontal box number stitching image and other remaining target images are input into the trained character recognition model to recognize the box number, ISO number and box threshold weight limit characters.

2. The method for identifying container number and door weight limit based on deep learning according to claim 1 is characterized in that: The SE channel attention module first performs a Squeeze compression operation, and its expression formula is: , , where H and W represent the height, width and dimension of the feature map respectively. represents a comprehensive feature description, is the feature map output by the c-th channel, Represented as a 2D convolution operation, is the 2D convolution kernel, Describe the feature matrix of the input s-th channel, Represents the total number of input channels; the output channels are obtained by summing up all input channels after 2D convolution , the feature map of each input channel will extract different spatial features through the 2D convolution kernel, and these features will be weighted and summed to obtain the comprehensive features of each output channel The comprehensive characteristics of each channel It is compressed into a scalar to form a 1×1×S feature output, which is the global feature description of the channel, where S represents the dimension; the connection between channels is opened up through the Squeeze compression operation, and then the feature relationship between channels is obtained by the Excitation excitation operation. The Excitation excitation operation obtains the channel weight through two full connection operations and the Sigmoid activation function: first, the global feature description is reduced in dimension through the first fully connected layer, and the ReLU activation function is used to increase the nonlinearity. Then, the second fully connected layer restores the dimension to the original number of channels. Finally, the weight is normalized by the Sigmoid function and adjusted to the range of 0 to 1 to complete the adaptive weighting of each channel; finally, the channel weight is applied to the feature matrix by element-by-element multiplication through the Scale function, but the spatial size of the feature matrix is ​​not changed.

3. The method for identifying container number and door weight limit based on deep learning according to claim 1 is characterized in that: The feature extraction process of the box number character block detection model is as follows: after the horizontal box number image passes through the convolution layer, normalization layer, activation layer, and maximum pooling layer, it passes through a multi-level residual block module for feature extraction, and finally outputs the result to the feature fusion module after passing through the average pooling layer and the fully connected layer; each level of residual block module is composed of at least three residual blocks, and the number of channels of each level of residual block module is doubled step by step according to the data transmission direction; each residual block is composed of a 1×1 convolution layer, a 3×3 convolution layer, and a 1×1 convolution layer, and performs dimensionality reduction, feature extraction, and dimensionality increase operations on the input feature map. Through the gradual increase of the residual block modules at each level and the number of channels, the size of the feature map is gradually reduced, and the feature extraction depth is gradually increased, thereby realizing the extraction of deep-level features of the feature map and the fusion of multi-scale information.

4. The method for identifying container number and door weight limit based on deep learning according to claim 1 is characterized in that: The box number character block detection model outputs a character block prediction set ,in The character block information includes the bounding box coordinates and confidence ; The processing steps for predicting character blocks in a set using the non-maximum suppression test are: Confidence sorting: confidence for each character block Sort in descending order; Calculate the intersection ratio: calculate the character block with the highest confidence The intersection and union ratio of the bounding box of the character block and the bounding boxes of other character blocks is expressed as: ,in, Indicates the character block with the highest confidence. Represents any other character block; Suppress redundant boxes: If > , then delete ,reserve , is the set threshold; Iterative processing: Repeat the steps of calculating the intersection-over-union ratio and suppressing redundant boxes until all other character blocks are processed to obtain the final detection set of retained character blocks. , n≤i.

5. The method for identifying container number and door weight limit based on deep learning according to claim 4 is characterized in that: The step of cutting out character block images from the horizontal box number image according to the detection set and splicing them to obtain the horizontal box number spliced ​​image is as follows: the cut n character block images ( ) to obtain horizontal splicing , which is set in advance The size of the image is defined as × , then we get the horizontally stitched image The size is ×( ), the splicing formula is: ,in, is the height, is the width, x represents the row coordinate of the image, y represents the column coordinate of the image, is the sum of the widths of the first n-1 images, used to locate the horizontal position of the image after stitching. It is the pixel value of the nth character block image in the splicing.

6. Container number and door weight limit recognition equipment based on deep learning, characterized by: It comprises a memory storing an executable program and a processor, wherein the processor runs the program to execute the method steps described in any one of claims 1 to 5.

Images