A round billet end face billet number direction classification method and device based on deep learning

Abstract

The application discloses a kind of based on deep learning's round billet end face billet number direction classification scheme, belong to intelligent metallurgical technology field, the method includes: establishing round billet end face billet number data set;Round billet end face billet number data set is divided into training data set and test data set;Each image in training data set is preprocessed;Based on the image after pre-processing, improved AlexNet network is trained, and classification model is obtained;Classification model is tested based on test data set, in the case where test result is classification model meets preset condition, it is determined that the classification model training is completed;Based on the classification model of training completion and text recognition unit, to be tested image is handled, to determine the class that round billet end face billet number direction in to-be-tested image belongs to.The based on deep learning's round billet end face billet number direction classification scheme provided in application can efficiently and accurately identify round billet end face billet number direction.

Description

Technical Field

[0001] This invention relates to the field of intelligent metallurgical technology, and in particular to a method and apparatus for classifying the billet number direction on the end face of a round billet based on deep learning. Background Technology

[0002] In the production of round billets, specific numbers, letters, or symbols are sprayed onto the billets as important information to identify their batches. As a key raw material in steel production, the quality of steel billets is paramount, and the billet number becomes a crucial index for quality traceability. Confirmation of these markings before rolling is a critical step in ensuring product quality, avoiding loading errors and production quality accidents. Simultaneously, the marking information serves as data for inventory management, ensuring the automation of billet outbound and inbound tracking, scheduling, and management. With the widespread adoption of automated production lines, various billet number character recognition systems exist in the market. Mainstream solutions often employ visual processing technology, combining traditional methods with deep learning algorithms to achieve more efficient and accurate billet number recognition. The detection and recognition of end-face markings generally involves the following steps: acquiring an image, determining whether a billet number exists on the end face of the billet and whether the billet number is complete, coarsely locating the character area, correcting the angle of the billet number character line, and performing single-character or character line detection and recognition.

[0003] During on-site image acquisition, due to differences in roller conveyor speed and camera sensitivity limitations, the billet positions in the original images vary, resulting in instances where the end-face billet number is missing or incomplete, and the direction of the billet number also differs, especially for round billets. Therefore, before detecting and recognizing the end-face billet number of round billets, it is necessary to first check whether the billet number in the image is complete. Furthermore, it is necessary to identify and correct the angle of the text line of the billet number, correcting the billet number in the image containing the complete billet number to a horizontal position.

[0004] Therefore, how to accurately and efficiently detect the direction of the billet number in an image is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a method and apparatus for classifying the billet number direction on the end face of a round billet based on deep learning, which can solve the problem in the prior art that the billet number direction cannot be accurately and efficiently detected in an image.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a deep learning-based method for classifying the billet number orientation on the end face of a round billet, wherein the method includes:

[0008] Establish a dataset of blank numbers on the end face of round billets, wherein the dataset contains multiple images of blank numbers on the end face of round billets;

[0009] The blank number dataset on the end face of the round blank is divided into a training dataset and a test dataset;

[0010] Preprocess each image in the training dataset;

[0011] The improved AlexNet network is trained based on the preprocessed images to obtain a classification model;

[0012] The classification model is tested based on the test dataset. If the test results show that the classification model meets the preset conditions, the training of the classification model is determined to be complete. The improved AlexNet network includes a blank number text direction classification function.

[0013] Based on the trained classification model and text recognition unit, the test image is processed to determine the category to which the billet number direction of the end face of the billet in the test image belongs.

[0014] Optionally, the step of establishing the billet number dataset for the end face of the round billet includes:

[0015] Using an industrial camera to capture images of the billet number on the end face of a round billet;

[0016] For each of the blank end face blank number images, the images are subjected to preset processing to establish a blank end face blank number dataset; wherein, the preset processing includes: data augmentation processing and manual annotation, wherein the data augmentation processing is to rotate the blank end face blank number image of the blank multiple times according to a preset rotation angle, and to keep the rotated blank end face blank number image after each rotation.

[0017] Optionally, the step of preprocessing each image in the training dataset includes:

[0018] For each image in the training dataset, grayscale processing, filtering, and data augmentation are performed sequentially.

[0019] Optionally, the improved AlexNet network includes: a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a first fully connected layer, a second fully connected layer, and a third fully connected layer;

[0020] The first convolutional layer, the second convolutional layer, and the fourth convolutional layer each include a convolutional layer, an activation function ReLU, and a pooling layer arranged sequentially; wherein, a local response normalization layer (LRN) is added between the convolutional layer and the pooling layer contained in the first convolutional layer and the second convolutional layer.

[0021] The third convolutional layer includes a convolutional layer and an activation function ReLU arranged sequentially.

[0022] Both the first fully connected layer and the second fully connected layer include a fully connected layer, ReLU, and a random deactivation Dropout layer arranged sequentially.

[0023] The third fully connected layer consists of a fully connected layer and a softmax function arranged sequentially.

[0024] Optionally, the step of training the improved AlexNet network based on the preprocessed image to obtain the classification model includes:

[0025] In network training, the improved AlexNet was selected as the blank number orientation classification model, and the AdamW optimizer was selected as the optimizer for training the blank number orientation classification and recognition model; the cross-entropy loss function was selected as the loss function for training the blank number orientation classification and recognition model.

[0026] Optionally, the step of processing the test image based on the trained classification model and text recognition unit to determine the category to which the billet number direction of the round billet end face in the test image belongs includes:

[0027] The test image with the billet number on the end face of the round billet is input into the trained classification model for orientation determination, and the orientation classification result of the billet number on the end face of the round billet is obtained.

[0028] Based on the classification results, the image to be tested is rotated and corrected, and the corrected image to be tested is input into the text recognition unit for recognition to obtain the billet number recognition result of the end face of the billet.

[0029] This invention also provides a deep learning-based orientation classification device for the end face number of a round billet, wherein the device includes:

[0030] A module is established to create a dataset of blank numbers on the end faces of round billets, wherein the dataset contains multiple images of blank numbers on the end faces of round billets.

[0031] A partitioning module is used to divide the round billet end face billet number dataset into a training dataset and a test dataset;

[0032] The preprocessing module is used to preprocess each image in the training dataset;

[0033] The training module is used to train the improved AlexNet network based on the preprocessed images to obtain a classification model;

[0034] The testing module is used to test the classification model based on the test dataset. If the test result shows that the classification model meets the preset conditions, it is determined that the classification model training is complete. The improved AlexNet network includes a blank number text direction classification function.

[0035] The prediction module is used to process the test image based on the trained classification model and text recognition unit to determine the category to which the billet number direction of the end face of the billet in the test image belongs.

[0036] Optionally, the establishment module includes:

[0037] The first submodule is used to capture images of the billet number on the end face of the round billet using an industrial camera;

[0038] The second submodule is used to perform preset processing on each of the round billet end face billet number images to establish a round billet end face billet number dataset; wherein, the preset processing includes: data augmentation processing and manual annotation, wherein the data augmentation processing is to rotate the round billet end face billet number image multiple times according to a preset rotation angle, and to keep the rotated round billet end face billet number image after each rotation.

[0039] Optionally, the preprocessing module is specifically used for:

[0040] For each image in the training dataset, grayscale processing, filtering, and data augmentation are performed sequentially.

[0041] Optionally, the improved AlexNet network includes: a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a first fully connected layer, a second fully connected layer, and a third fully connected layer;

[0042] The first convolutional layer, the second convolutional layer, and the fourth convolutional layer each include a convolutional layer, an activation function ReLU, and a pooling layer arranged sequentially; wherein, a local response normalization layer (LRN) is added between the convolutional layer and the pooling layer contained in the first convolutional layer and the second convolutional layer.

[0043] The third convolutional layer includes a convolutional layer and an activation function ReLU arranged sequentially.

[0044] Both the first fully connected layer and the second fully connected layer include a fully connected layer, ReLU, and a random deactivation Dropout layer arranged sequentially.

[0045] The third fully connected layer consists of a fully connected layer and a softmax function arranged sequentially.

[0046] Optionally, the training module is specifically used for:

[0047] In network training, the improved AlexNet was selected as the blank number orientation classification model, and the AdamW optimizer was selected as the optimizer for training the blank number orientation classification and recognition model; the cross-entropy loss function was selected as the loss function for training the blank number orientation classification and recognition model.

[0048] Optionally, the prediction module includes:

[0049] The third submodule is used to input the test image with the billet number on the end face of the billet into the trained classification model for orientation determination, and the orientation classification result of the billet number on the end face of the billet.

[0050] The fourth submodule is used to perform angular rotation correction on the image to be tested based on the classification result, and input the corrected image to be tested into the text recognition unit for recognition to obtain the billet end face number recognition result.

[0051] This invention provides an electronic device, which includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor. When the program or instructions are executed by the processor, they implement the steps of any of the above-described deep learning-based methods for classifying the billet number direction of the end face of a round billet.

[0052] The deep learning-based classification scheme for billet end face number orientation provided in this application involves: establishing a billet end face number dataset; dividing the billet end face number dataset into a training dataset and a test dataset; preprocessing each image in the training dataset; training an improved AlexNet network based on the preprocessed images to obtain a classification model; testing the classification model based on the test dataset; and determining that the classification model training is complete when the test results show that the classification model meets the preset conditions.

[0053] Based on the trained classification model and text recognition unit, the test image is processed to determine the category to which the billet number direction of the round billet end face belongs. The scheme provided in this application extracts features from the original acquired image through an improved AlexNet multi-layer convolution. During the training of the classification model, fully connected layers and Softmax are used to extract global features to classify the acquired image. Finally, a classification model with preset conditions is used to classify the billet number direction of the round billet end face based on the trained classification model, which can efficiently and accurately identify the billet number direction of the round billet end face. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating the steps of a deep learning-based method for classifying the billet number direction on the end face of a round billet, according to an embodiment of this application.

[0055] Figure 2 This is a structural block diagram illustrating a deep learning-based round billet end face billet number orientation classification device according to an embodiment of this application. Detailed Implementation

[0056] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0057] The following description, in conjunction with the accompanying drawings, details the deep learning-based billet number orientation classification scheme for the end face of round billets provided in this application, through specific embodiments and application scenarios.

[0058] As attached Figure 1 As shown, the deep learning-based method for classifying the billet number orientation on the end face of a round billet according to an embodiment of this application includes the following steps:

[0059] Step 101: Establish a dataset of blank numbers for the end face of the round blank.

[0060] The billet end face number dataset contains multiple images of billet end face numbers, and each image contains a billet number.

[0061] An optional method for creating a dataset of billet end face numbers for round billets is as follows:

[0062] First, use an industrial camera to capture images of the billet number on the end face of the round billet;

[0063] Secondly, for each round billet end face billet number image, the image is pre-processed to establish a round billet end face billet number dataset.

[0064] The preset processing includes data augmentation and manual annotation. Data augmentation involves rotating the billet end face image multiple times at a preset rotation angle, maintaining the rotated billet end face image after each rotation. The preset rotation angle can be flexibly set by those skilled in the art, and this embodiment does not impose specific limitations on it; for example, it can be set to 30 degrees, 60 degrees, or 45 degrees. Manual annotation can be divided into 13 folders according to the direction of the billet number text lines, including 12 directional classes and one class for billet number images not present on the end face.

[0065] Taking the acquisition of images of round billets on a production line as an example, the acquisition process uses a fixed-point shooting method where the camera is stationary and the billet moves on the conveyor rollers. Due to uneven lighting in the environment, a lighting supplement unit is included in the acquisition device. Although the camera initially filters out images that do not contain billets, it is inevitable that some images will not contain billet information. On-site, 400 images containing the billet end face number were acquired using an industrial camera. The image size is 1600x1264, RGB format. The billet end face number contains uppercase letters A and Z, numbers 0-9, and the symbol -.

[0066] Step 102: Divide the round billet end face billet number dataset into a training dataset and a test dataset.

[0067] The generated dataset is divided into training and testing datasets proportionally, with the ratio of training to testing sets ranging from 7:3 to 9:1. One feasible approach is to crop the original image set to 800×800 pixels before creating the dataset, using the area where the end face blank number appears as the cropping criterion.

[0068] Step 103: Preprocess each image in the training dataset.

[0069] An optional way to preprocess the images in the training dataset is as follows:

[0070] For each image in the training dataset, grayscale processing, filtering, and data augmentation are performed sequentially.

[0071] Median filtering was used for filtering, and gamma transform was employed for image enhancement.

[0072] Gamma transform is mainly used for image correction, adjusting images with excessively high or low gray levels to enhance contrast. The transform formula involves multiplying the values ​​of each pixel in the original image. , (Representing input and output grayscale respectively)

[0073]

[0074] Step 104: Train the improved AlexNet network based on the preprocessed image to obtain the classification model.

[0075] An improved AlexNet network is used to train an image classification model. The network improvement is based on a text direction classification task, and the main improvements are: reducing the number of convolutional kernels from 5 to 4. The original AlexNet network used two GPUs for training; however, a single GPU can easily handle the current data volume, so the improved AlexNet network is trained using a single GPU. Given the small data volume and no more than 15 classifications in the current task, reducing the number of convolutional kernels reduces model complexity, lowers computational resource requirements, and accelerates training while maintaining performance. Therefore, in this example, the number of convolutional kernels is uniformly reduced to half that of the original AlexNet network, reducing the dimensionality of feature extraction; and convolutional layers C3 and C5 in the original AlexNet network are connected.

[0076] The improved AlexNet network includes: first convolutional layer C1, second convolutional layer C2, third convolutional layer C3, fourth convolutional layer C4, first fully connected layer FC5, second fully connected layer FC6, and third fully connected layer FC7.

[0077] The first, second, and fourth convolutional layers each include a convolutional layer, a ReLU activation function, and a pooling layer arranged sequentially. A Local Response Normalization (LRN) layer is added between the convolutional layers and pooling layers in the first and second convolutional layers. The third convolutional layer includes a convolutional layer and a ReLU activation function arranged sequentially. The first and second fully connected layers each include a fully connected layer, a ReLU activation function, and a Dropout layer arranged sequentially. The third fully connected layer includes a fully connected layer and a softmax function arranged sequentially.

[0078] A more specific improved AlexNet network structure is shown below, including the following convolutional layers and fully connected layers, as well as the data input and output information for each layer:

[0079] Convolutional layer C1 (convolution-ReLU-pooling): Convolution: Input 227x227x3, 48 11x11x3 convolutional kernels with no edge padding = 0, stride = 4, output feature map is 55x55x48; activation function is ReLU; pooling: pooling kernel is 3x3.

[0080] Convolutional layer C2 (convolution-ReLU-pooling): Convolution: Input 27x27x48, 128 convolutional kernels of size 5x5x96, padding = 2, stride = 1, output feature map is 27x27x128; activation function is ReLU; Pooling: pooling kernel size 3 × 3, no padding = 0, stride = 2, therefore its feature map output is 13×13×128.

[0081] Convolutional layer C3 (Convolution-ReLU): Convolution: Input is 13x13x128, 192 convolutional kernels are 3x3x256, padding = 1, stride = 1, output feature map is 13x13x192; activation function is ReLU.

[0082] Convolutional layer C4 (convolution-ReLU-pooling): Convolution: Input is 13x13x192, 128 convolutional kernels are 3x3x384, padding = 1, stride = 1, output feature map is 13x13x128; activation function is ReLU; pooling: pooling kernel size is 3 × 3, no padding = 0, stride = 2, therefore its feature map output is 6×6×128.

[0083] The fully connected layer FC5 (fully connected – ReLU – Dropout) has the following characteristics: fully connected: input is 6x6x128, 4096 convolutional kernels of 6x6x128, padding = 0, stride = 1, output feature map is 1x1x4096; activation function is ReLU; Dropout: randomly removes some neural nodes to prevent overfitting, output is 1x1x4096.

[0084] Fully connected layer FC6 (fully connected – ReLU – Dropout): Fully connected: input is 1x1x4096, activation function is ReLU; Dropout: randomly removes some neural nodes to prevent overfitting, output is 1x1x4096.

[0085] The fully connected layer FC7 (fully connected – softmax) has a 1x1x4096 input, a softmax value of 5, and an output of 1x1x5. In layers C1 and C2, LRN (Local Response Normalization) is added between the convolutional and pooling layers to prevent overfitting.

[0086] Local response normalization specifically refers to normalizing a feature map corresponding to each convolutional kernel in convolutional layers C1 and C2. More specifically, for each location on each feature map, the sum of squares of the pixels surrounding that location is calculated, and then the pixel value at the current location is divided by the sum. The calculation process can be expressed by the following formula:

[0087] .

[0088] in, Indicates the first Each feature map at location The normalized value, Indicates the first Each feature map at location The original value, Indicates the size of the domain. Indicates the number of feature maps. , and These are hyperparameters, in the embodiments of this application. , , .

[0089] The ReLU function is a common activation function, and its expression is as follows: The softmax layer transforms the output of neurons into probabilities, and determines which category the direction of the blank belongs to based on the magnitude of the probabilities.

[0090] An optional method for training the improved AlexNet network based on the preprocessed image to obtain the classification model is as follows: In network training, the improved AlexNet is selected as the blank number orientation classification model, the AdamW optimizer is selected as the optimizer for training the blank number orientation classification and recognition model, and the cross-entropy loss function is selected as the loss function for training the blank number orientation classification and recognition model.

[0091] Step 105: Test the classification model based on the test dataset. If the test results show that the classification model meets the preset conditions, the training of the classification model is considered complete.

[0092] The improved AlexNet network includes a function for classifying the orientation of the blank number text.

[0093] In practical implementation, the AdamW optimizer can be used more specifically as the optimizer for training the billet number orientation classification and recognition model; the cross-entropy loss function is used as the loss function for training the billet number orientation classification and recognition model; the learning parameters include the base learning rate, batch size, Adam optimizer parameters, and maximum number of iterations. AdamW adds weight decay processing to Adam. Weight decay is a regularization technique used to prevent model overfitting. During updates, the first moment estimate m and the second moment estimate v of the gradient are calculated, and then weight decay processing is performed. In AdamW, weight decay is directly applied to the weight parameters, and the weight parameters are adjusted according to the updated m and v, as well as the learning rate and weight decay rate. The learning parameters of the recognition model are set, and iterative learning is performed according to the set parameters until the loss function value decreases very slowly. During training, the best weights are saved when the classification accuracy is above 99%. New sample images are used for testing, and the model is saved when the test accuracy reaches above 98.5%.

[0094] It should be noted that the above is only an example of using a test accuracy rate of 98.5% or higher as a criterion for determining whether a classification model meets the preset conditions. In actual implementation, the preset conditions are not limited to this and can be flexibly set by those skilled in the art.

[0095] Step 106: Based on the trained classification model and text recognition unit, process the image to be tested to determine the category to which the billet number direction of the end face of the billet in the image to be tested belongs.

[0096] An optional method for processing the test image based on a trained classification model and text recognition unit to determine the category to which the billet number direction on the end face of the billet in the test image belongs can be as follows:

[0097] The test image with the billet end face number is input into the trained classification model for orientation determination, and the orientation classification result of the billet end face number is obtained. Based on the classification result, the test image is rotated and corrected, and the corrected test image is input into the text recognition unit for recognition to obtain the billet end face number recognition result.

[0098] In actual implementation, after the image is fed into the classification model for classification, it needs to be angle-corrected according to the classification result until the text is in a normally readable horizontal state range, and then fed into the text recognition unit to finally obtain the recognition result.

[0099] The classification model achieved an accuracy of 99.5% on the training set and 99.1% on the test set. Adding the model to the existing billet number recognition model on the rolling line significantly improved the overall recognition accuracy from 71.5% to 80.2%. This result demonstrates that the proposed method possesses high classification accuracy and stability, effectively addressing the problem of classifying billet number directions on the end face of round billets.

[0100] The deep learning-based method for classifying the billet end face number orientation of round billets provided in this application involves: establishing a round billet end face number dataset; dividing the round billet end face number dataset into a training dataset and a test dataset; preprocessing each image in the training dataset; training an improved AlexNet network based on the preprocessed images to obtain a classification model; testing the classification model based on the test dataset, and determining that the classification model training is complete when the test results show that the classification model meets preset conditions; and processing the test image based on the trained classification model and a text recognition unit to determine the category to which the round billet end face number orientation belongs in the test image. This method, through the improved AlexNet multi-layer convolution, extracts features from the original acquired images, and during the training of the classification model, utilizes fully connected layers and Softmax to extract global features to classify the acquired images, ultimately setting a classification model based on preset conditions. Based on the trained classification model, the method classifies the round billet end face number orientation, enabling efficient and accurate identification of the round billet end face number orientation.

[0101] Figure 2 The structural block diagram of a deep learning-based circular billet end face orientation classification device is shown in the embodiment of this application.

[0102] The orientation classification device for round billet end face numbers based on deep learning provided in this application includes the following functional modules:

[0103] Module 201 is used to create a dataset of billet end face numbers for round billets, wherein the dataset contains multiple images of billet end face numbers for round billets.

[0104] The partitioning module 202 is used to partition the round billet end face billet number dataset into a training dataset and a test dataset;

[0105] Preprocessing module 203 is used to preprocess each image in the training dataset;

[0106] Training module 204 is used to train the improved AlexNet network based on the preprocessed images to obtain a classification model;

[0107] Test module 205 is used to test the classification model based on the test dataset, and determine that the classification model training is complete when the test result shows that the classification model meets the preset conditions; wherein, the improved AlexNet network includes a blank number text direction classification function;

[0108] The prediction module 206 is used to process the test image based on the trained classification model and text recognition unit to determine the category to which the blank number direction of the end face of the blank in the test image belongs.

[0109] Optionally, the establishment module includes:

[0110] The first submodule is used to capture images of the billet number on the end face of the round billet using an industrial camera;

[0111] The second submodule is used to perform preset processing on each of the round billet end face billet number images to establish a round billet end face billet number dataset; wherein, the preset processing includes: data augmentation processing and manual annotation, wherein the data augmentation processing is to rotate the round billet end face billet number image multiple times according to a preset rotation angle, and to keep the rotated round billet end face billet number image after each rotation.

[0112] Optionally, the preprocessing module is specifically used for:

[0113] For each image in the training dataset, grayscale processing, filtering, and data augmentation are performed sequentially.

[0114] Optionally, the improved AlexNet network includes: a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a first fully connected layer, a second fully connected layer, and a third fully connected layer;

[0115] The first convolutional layer, the second convolutional layer, and the fourth convolutional layer each include a convolutional layer, an activation function ReLU, and a pooling layer arranged sequentially; wherein, a local response normalization layer (LRN) is added between the convolutional layer and the pooling layer contained in the first convolutional layer and the second convolutional layer.

[0116] The third convolutional layer includes a convolutional layer and an activation function ReLU arranged sequentially.

[0117] Both the first fully connected layer and the second fully connected layer include a fully connected layer, ReLU, and a random deactivation Dropout layer arranged sequentially.

[0118] The third fully connected layer consists of a fully connected layer and a softmax function arranged sequentially.

[0119] Optionally, the training module is specifically used for:

[0120] In network training, the improved AlexNet was selected as the blank number orientation classification model, and the AdamW optimizer was selected as the optimizer for training the blank number orientation classification and recognition model; the cross-entropy loss function was selected as the loss function for training the blank number orientation classification and recognition model.

[0121] Optionally, the prediction module includes:

[0122] The third submodule is used to input the test image with the billet number on the end face of the billet into the trained classification model for orientation determination, and the orientation classification result of the billet number on the end face of the billet.

[0123] The fourth submodule is used to perform angular rotation correction on the image to be tested based on the classification result, and input the corrected image to be tested into the text recognition unit for recognition to obtain the billet end face number recognition result.

[0124] The deep learning-based billet end face number orientation classification device provided in this application establishes a billet end face number dataset; divides the billet end face number dataset into a training dataset and a test dataset; preprocesses each image in the training dataset; trains an improved AlexNet network based on the preprocessed images to obtain a classification model; tests the classification model based on the test dataset, and determines that the classification model training is complete when the test result shows that the classification model meets preset conditions; and processes the test image based on the trained classification model and a text recognition unit to determine the category to which the billet end face number orientation belongs in the test image. The device provided in this application extracts features from the original acquired images through improved AlexNet multi-layer convolution, and uses fully connected layers and Softmax to extract global features to classify the acquired images during the training of the classification model, ultimately obtaining a classification model with preset conditions. Based on the trained classification model, it classifies the billet end face number orientation, enabling efficient and accurate identification of the billet end face number orientation.

[0125] In the embodiments of this application Figure 2The deep learning-based orientation classification device for the end face number of round billets shown is installed in an electronic device or server. The electronic device or server equipped with this device can be a device with an operating system. This operating system can be Android, iOS, or other possible operating systems; this application embodiment does not specifically limit it.

[0126] The embodiments provided in this application Figure 2 The deep learning-based orientation classification device for the end face of a round billet, as shown, can achieve... Figure 1 The various processes implemented in the method implementation examples will not be described again here to avoid repetition.

[0127] Optionally, embodiments of this application also provide an electronic device, including a processor, a memory, and a program or instructions stored in the memory and executable on the processor. When the program or instructions are executed by the processor, they implement the processes performed by the aforementioned deep learning-based circular billet end face billet number orientation classification device and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0128] It should be noted that the electronic device in this application embodiment includes the server described above.

[0129] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0130] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0131] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A deep learning-based method for classifying the billet number orientation on the end face of a round billet, characterized in that, include: Establish a dataset of blank numbers on the end face of round billets, wherein the dataset contains multiple images of blank numbers on the end face of round billets; The blank number dataset on the end face of the round blank is divided into a training dataset and a test dataset; Preprocess each image in the training dataset; The improved AlexNet network is trained based on the preprocessed images to obtain a classification model; The classification model is tested based on the test dataset. If the test results show that the classification model meets the preset conditions, the training of the classification model is determined to be complete. The improved AlexNet network includes a blank number text direction classification function. Based on the trained classification model and text recognition unit, the test image is processed to determine the category to which the billet number direction of the round billet end face in the test image belongs; The improved AlexNet network includes: a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a first fully connected layer, a second fully connected layer, and a third fully connected layer. The first convolutional layer, the second convolutional layer, and the fourth convolutional layer each include a convolutional layer, an activation function ReLU, and a pooling layer arranged sequentially; wherein, a local response normalization layer (LRN) is added between the convolutional layer and the pooling layer contained in the first convolutional layer and the second convolutional layer. The third convolutional layer includes a convolutional layer and an activation function ReLU arranged sequentially. Both the first fully connected layer and the second fully connected layer include a fully connected layer, ReLU, and a random deactivation Dropout layer arranged sequentially. The third fully connected layer consists of a fully connected layer and a softmax function arranged sequentially.

2. The method according to claim 1, characterized in that, The step of establishing the billet number dataset for the end face of the round billet includes: Using an industrial camera to capture images of the billet number on the end face of a round billet; For each of the blank end face blank number images, the images are subjected to preset processing to establish a blank end face blank number dataset; wherein, the preset processing includes: data augmentation processing and manual annotation, wherein the data augmentation processing is to rotate the blank end face blank number image of the blank multiple times according to a preset rotation angle, and to keep the rotated blank end face blank number image after each rotation.

3. The method according to claim 2, characterized in that, The steps for preprocessing each image in the training dataset include: For each image in the training dataset, grayscale processing, filtering, and data augmentation are performed sequentially.

4. The method according to claim 1, characterized in that, The steps for training the improved AlexNet network based on the preprocessed images to obtain the classification model include: In network training, the improved AlexNet was selected as the blank number orientation classification model, and the AdamW optimizer was selected as the optimizer for training the blank number orientation classification and recognition model; the cross-entropy loss function was selected as the loss function for training the blank number orientation classification and recognition model.

5. The method according to claim 1, characterized in that, Based on the trained classification model and text recognition unit, the steps for processing the test image to determine the category to which the billet number direction on the end face of the billet in the test image belongs include: The test image with the billet number on the end face of the round billet is input into the trained classification model for orientation determination, and the orientation classification result of the billet number on the end face of the round billet is obtained. Based on the classification results, the image to be tested is rotated and corrected, and the corrected image to be tested is input into the text recognition unit for recognition to obtain the billet number recognition result of the end face of the billet.

6. A deep learning-based orientation classification device for the end face number of a round billet, characterized in that, include: A module is established to create a dataset of blank numbers on the end faces of round billets, wherein the dataset contains multiple images of blank numbers on the end faces of round billets. A partitioning module is used to divide the round billet end face billet number dataset into a training dataset and a test dataset; The preprocessing module is used to preprocess each image in the training dataset; The training module is used to train the improved AlexNet network based on the preprocessed images to obtain a classification model; The testing module is used to test the classification model based on the test dataset. If the test result shows that the classification model meets the preset conditions, it is determined that the classification model training is complete. The improved AlexNet network includes a blank number text direction classification function. The prediction module is used to process the test image based on the trained classification model and text recognition unit to determine the category to which the billet number direction of the end face of the billet in the test image belongs. The improved AlexNet network includes: a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a first fully connected layer, a second fully connected layer, and a third fully connected layer. The first convolutional layer, the second convolutional layer, and the fourth convolutional layer each include a convolutional layer, an activation function ReLU, and a pooling layer arranged sequentially; wherein, a local response normalization layer (LRN) is added between the convolutional layer and the pooling layer contained in the first convolutional layer and the second convolutional layer. The third convolutional layer includes a convolutional layer and an activation function ReLU arranged sequentially. Both the first fully connected layer and the second fully connected layer include a fully connected layer, ReLU, and a random deactivation Dropout layer arranged sequentially. The third fully connected layer consists of a fully connected layer and a softmax function arranged sequentially.

7. The apparatus according to claim 6, characterized in that, The establishment module includes: The first submodule is used to capture images of the billet number on the end face of the round billet using an industrial camera; The second submodule is used to perform preset processing on each of the round billet end face billet number images to establish a round billet end face billet number dataset; wherein, the preset processing includes: data augmentation processing and manual annotation, wherein the data augmentation processing is to rotate the round billet end face billet number image multiple times according to a preset rotation angle, and to keep the rotated round billet end face billet number image after each rotation.

8. The apparatus according to claim 7, characterized in that, The preprocessing module is specifically used for: For each image in the training dataset, grayscale processing, filtering, and data augmentation are performed sequentially.

Images