An online identification method of slab number based on autonomous evolutionary neural network

Abstract

The present invention provides a method for online recognition of slab numbers based on an autonomous evolutionary neural network, and relates to the technical field of automatic control of hot rolling production processes in steel enterprises. The method first extracts video data of the slab traveling process, and performs data processing to obtain a slab image data set with slab numbers; then, based on the idea of ​​transfer learning, the specific position of the slab number on the slab image is located to produce a data set required for slab number recognition; based on the idea of ​​multi-objective genetic programming, an autonomous evolutionary neural network algorithm is designed to construct an integrated model for slab number recognition, and online recognition of slab numbers is realized; finally, the recognized single slab number characters are reorganized to obtain a predicted slab number. The method can realize end-to-end online recognition of slab numbers, effectively reduce labor costs, and improve the intelligent manufacturing level of hot rolling production lines.

Description

Technical Field

[0001] The present invention relates to the technical field of automatic control of hot rolling production process of steel enterprises, and in particular to an online slab number recognition method based on an autonomous evolutionary neural network. Background Art

[0002] Hot rolling production is an important production line in steel enterprises. The hot rolling workshop and the continuous casting workshop are arranged adjacent to each other. The qualified continuous casting slabs after cooling, flame treatment and marking in the continuous casting workshop, as well as the hot continuous casting slabs with qualified surface and internal quality, will be sent to the hot rolling workshop by rollers. Then the operator will identify the slab number of the slab and check it with the data in the slab warehouse. After confirmation, the subsequent operation instructions will be issued to guide the slab to enter the designated stack in the slab warehouse and wait for the completion of the established rolling plan. With the continuous deepening of intelligent manufacturing, steel enterprises have also ushered in a wave of reform. At present, the automation level of the slab storage process in the hot rolling production line is relatively weak, especially the slab number identification process. Most steel enterprises rely on operators to be on duty 24 hours a day to observe the slab number of the slab in the process of moving and check it with the data in the slab warehouse. The labor intensity is high and the work efficiency is low. At the same time, it is affected by human factors. It is easy to cause misidentification or omission, which may result in product scrapping or serious economic losses to the entire production line.

[0003] The paper "Research and Implementation of Automatic Slab Number Identification System" (Yang Liao, Xinjiang Institute of Ecology and Geography, Chinese Academy of Sciences, 2011) proposed an online slab number recognition technology based on pattern recognition; the patent "A fully automatic steel loading method for slab number detection based on machine vision" and the patent "A steel slab coding automatic recognition method and steel slab tracking system" respectively proposed two slab number recognition algorithms based on artificial neural networks. Because the design of the slab number recognition algorithm in the above methods relies on manual experience to build, the performance is greatly affected by human factors, and the algorithm has poor robustness and can only be used in specific environments. At the same time, there are also phenomena such as large recognition errors, which are difficult to meet the needs of modern intelligent manufacturing. Summary of the invention

[0004] The technical problem to be solved by the present invention is to address the deficiencies of the above-mentioned prior art and to provide an online slab number recognition method based on an autonomous evolving neural network, which automatically searches for the potential optimal neural network architecture according to the characteristics of the slab data, and realizes end-to-end online recognition of the slab number, thereby reducing labor costs while improving the intelligent manufacturing level of the hot rolling production line.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: an online slab number recognition method based on an autonomous evolutionary neural network, comprising the following steps:

[0006] Step 1: extracting video data of the slab moving process during the hot rolling production process;

[0007] Step 2: Process the video data of the slab moving process to obtain an image data set with slab numbers;

[0008] Step 2.1: Use the inter-frame difference method to compare the difference between the images of consecutive frames in the video data, extract the frames where the slab passes from the video data, and obtain the slab image;

[0009] Step 2.2: Filter the obtained slab images to select image data in which the slab numbers are clearly captured, and obtain an image data set with slab numbers;

[0010] Step 3: Based on the idea of ​​transfer learning, locate the specific position of the slab number on the slab image and create the data set required for slab number recognition. The specific method is as follows:

[0011] Step 3.1: Train the EAST scene text localization model based on the ICDAR2015 dataset;

[0012] Step 3.2: Based on the EAST scene text localization model trained in step 3.1, the image dataset with slab numbers obtained in step 2.2 is input into the trained EAST scene text localization model for testing, and the slab images whose slab numbers can be accurately located are screened out;

[0013] Step 3.3: Add the slab images that can be accurately located selected in step 3.2 as training data to the ICDAR2015 dataset, and retrain the EAST scene text localization model;

[0014] Step 3.4: Repeat steps 3.2 to 3.3 for multiple times to achieve the migration of models and data, thereby achieving accurate positioning of the slab number;

[0015] Step 3.5: For all images where the slab number position is accurately located, the image containing only the slab number is cropped from the entire slab image according to the coordinates of the positioning text box to obtain a slab number image dataset;

[0016] Step 3.6: Use computer vision and image processing methods to perform character segmentation on the slab number in the slab number image dataset. The specific method is as follows:

[0017] Step 3.6.1: Binarize the slab number image dataset by using an adaptive threshold method to convert the slab number image from a color image to a black and white image;

[0018] Step 3.6.2: Project the binary slab number image in the horizontal direction, and perform character segmentation using the trough in the histogram obtained after projection as the segmentation point to obtain a segmented slab number character data set, and name the segmented character image according to its position in the entire slab image;

[0019] Step 4: Based on the idea of ​​multi-objective genetic programming, an autonomous evolutionary neural network algorithm is designed to build an integrated model for slab number recognition to realize online recognition of slab numbers;

[0020] Step 4.1: Design an autonomous evolutionary neural network algorithm based on Cartesian genetic programming and Pareto optimal solution set ideas;

[0021] Step 4.1.1: Design a variety of convolution, pooling, residual, network series, and network parallel basic modules within the neural network. Based on Cartesian genetic programming, arrange and combine the designed basic modules to achieve variable-length neural network structure encoding;

[0022] Step 4.1.2: Taking the model parameter quantity and recognition accuracy as the two objectives, a multi-objective evolutionary strategy based on the Pareto optimal solution set is adopted to guide the autonomous evolutionary neural network algorithm to quickly search for the optimal neural network architecture with high recognition accuracy and low model complexity;

[0023] Step 4.1.3: Design genetic operators including crossover and mutation, so that the offspring solutions generated by crossover and mutation are more diverse while retaining the excellent genes of the parent solution, thereby expanding the search space of the autonomous evolutionary neural network algorithm and providing more possibilities for the algorithm to find the optimal solution;

[0024] Step 4.2: The slab number character data set obtained in step 3 is passed as training data to the autonomous evolutionary neural network algorithm designed in step 4.1, and the autonomous evolutionary neural network algorithm designed in step 4.1 is trained to obtain the final Pareto optimal frontier;

[0025] Step 4.3: Use the solution on the Pareto optimal frontier obtained in step 4.2 as the slab number recognition sub-model, and build an integrated model for online slab number recognition based on the differential evolution algorithm;

[0026] Step 4.3.1: On the premise that the total weight sum is 1, give each slab number recognition sub-model an initial weight by random initialization;

[0027] Step 4.3.2: Based on the slab number character data obtained in step 3 and each slab number recognition sub-model with initial weights obtained in step 4.3.1, the recognition accuracy of the slab number online recognition integrated model is calculated;

[0028] Step 4.3.3: Based on the DE / rand / 1 mutation strategy and adaptive crossover strategy in differential evolution, the slab number online recognition integrated model is trained to optimize the weights of each sub-model in the integrated model until the maximum number of iterations is reached or the model recognition accuracy tends to converge and no longer increases, then the training is stopped;

[0029] Step 4.3.4: Perform online recognition of the slab number characters according to the slab number online recognition integrated model obtained in step 4.3.3;

[0030] Step 5: The identified single slab number characters are reorganized according to the naming method of the character image to obtain the predicted slab number, and the predicted slab number is checked with the slab number in the slab library. At the same time, combined with the furnace parameters of the slab, automatic release, pending or alarm processing is given.

[0031] The beneficial effects of adopting the above technical solution are: the present invention provides an online slab number recognition method based on an autonomous evolutionary neural network. First, the required image data is obtained from the original industrial video data set through computer vision and basic image processing methods, and the data set on which the slab number recognition depends is constructed using the idea of ​​transfer learning, which largely solves the problem of expensive labeling in deep learning. In addition, an algorithm for online slab number recognition using an autonomous evolutionary neural network is designed, and an integrated slab number recognition model is obtained based on the model obtained by the algorithm and the idea of ​​differential evolution, which realizes the online and accurate recognition of the slab number. At the same time, combined with the relevant parameters before the slab enters the furnace, the recognition result is checked with the slab in the slab library, and automatic release, early warning or alarm processing is performed based on the check result. The method of the present invention can realize end-to-end online recognition of slab numbers, effectively reduce labor costs, and improve the intelligent manufacturing level of hot rolling production lines. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A flowchart of a method for online identification of slab numbers based on an autonomous evolving neural network provided by an embodiment of the present invention;

[0033] Figure 2 A schematic diagram of the principle of a method for online identification of slab numbers based on an autonomous evolving neural network provided by an embodiment of the present invention;

[0034] Figure 3 A schematic diagram of a slab data image processing process provided by an embodiment of the present invention;

[0035] Figure 4 A schematic diagram of gene encoding and decoding provided by an embodiment of the present invention;

[0036] Figure 5 A schematic diagram of a crossover process provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0037] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0038] This embodiment takes the hot rolling production process of a hot rolling enterprise as an example, and adopts the method of the present invention to perform online identification of slab numbers in the hot rolling production process.

[0039] In this embodiment, a video shooting device is installed on the side of the roller table to shoot video of the slab moving process. The video signal is converted into an image signal through a video processing model. The image signal is located at the position of the slab number character from the entire slab image through a character positioning operation. The entire slab number is segmented into individual character block data through an image segmentation module. The segmented data is transmitted to an autonomous evolutionary neural network algorithm designed based on the idea of ​​multi-objective genetic programming to construct a slab number recognition integrated model. After training, the model is used to identify the slab number during the moving process, and the identified results are compared with the slabs in the slab library. Release, pending or alarm processing are made based on the comparison results.

[0040] In this embodiment, a slab number online recognition method based on autonomous evolution neural network is provided. Figure 1 , 2 As shown, the following steps are included:

[0041] Step 1: extracting video data of the slab moving process during the hot rolling production process;

[0042] In this embodiment, for the video data collected during the production process of the hot rolling enterprise, the video data of the slab moving process in the last two months was extracted from the database, and a total of 6.94G of video data was obtained;

[0043] Step 2: Process the video data of the slab moving process to obtain an image data set with slab numbers;

[0044] Step 2.1: Use the inter-frame difference method to compare the difference between the images of consecutive frames in the video data, extract the frames where the slab passes from the video data, and obtain the slab image;

[0045] Step 2.2: Filter the obtained slab images to select image data in which the slab numbers are clearly captured, and obtain an image data set with slab numbers;

[0046] In this embodiment, the inter-frame difference method is first used to extract the frames in which the slab appears in the video, that is, the adjacent frame images of the video image sequence are differentially operated. When a slab is transported in the monitoring scene, the image in the video will be significantly different from the roller background. When the difference is greater than the set threshold value T=195, it is determined that the frame has a slab passing through, and the frame image is extracted and saved. In this embodiment, a total of 16573 pieces of image data are obtained;

[0047] Then, the obtained related frame images with the slab passing through are screened, and the image data with the slab number being clearly photographed are screened out, and a total of 2590 such data are obtained;

[0048] Step 3: Based on the idea of ​​transfer learning, locate the specific position of the slab number on the slab image and create the data set required for slab number recognition, such as Figure 3 As shown, the specific method is:

[0049] Step 3.1: Train the EAST scene text localization model based on the ICDAR2015 dataset;

[0050] Step 3.2: Based on the EAST scene text localization model trained in step 3.1, the image dataset with slab numbers obtained in step 2.2 is input into the trained EAST scene text localization model for testing, and the slab images whose slab numbers can be accurately located are screened out;

[0051] In this embodiment, the RGB mean of the image data set with slab numbers obtained in step 2 is calculated, and the original EAST model is fine-tuned based on the obtained RGB mean, mainly to standardize the data in the image preprocessing process. In this embodiment, the RGB mean of the image data set with slab numbers is calculated to be [65.62, 68.64, 70.47];

[0052] Step 3.3: Add the slab images that can be accurately located selected in step 3.2 as training data to the ICDAR2015 dataset, and retrain the EAST scene text localization model;

[0053] In this embodiment, the original EAST model is fine-tuned according to the mean of the slab data set calculated in 3.2, and the image data set with slab numbers obtained in 2.2 is added to the ICDAR2015 data set as training data, and the fine-tuned EAST model is tested to select images in which the slab numbers can be accurately located;

[0054] Step 3.4: Repeat steps 3.2 to 3.3 for multiple times to achieve the migration of models and data, thereby achieving accurate positioning of the slab number;

[0055] In this embodiment, steps 3.2 to 3.3 are repeated four to five times to achieve data and model migration. The variation of the slab number positioning process accuracy based on migration learning with the number of migrations is shown in Table 1, and 2071 images in which the slab number can be accurately located are obtained;

[0056] Table 1 Variation of slab number positioning accuracy with migration times

[0057] Migration times 0 1 2 3 4 5 Positioning accuracy (%) 13.27 33.03 47.92 63.00 75.56 80.83

[0058] Step 3.5: For all images where the slab number position is accurately located, the image containing only the slab number is cropped from the entire slab image according to the coordinates of the positioning text box to obtain a slab number image dataset;

[0059] In this embodiment, for all images where the positions of the slab numbers can be accurately located, images containing only the slab numbers are cropped from the entire slab image according to the coordinates of the positioning text box. More specifically, the mathematical expressions of the positioning text box that can be obtained by the EAST model include two types, namely, a rotated rectangular box (RBOX) and a quadrilateral box (QUAD), as shown in Table 2. In this embodiment, a quadrilateral box is selected, that is, the coordinates of the four corner points of the positioning text box can be obtained, such as (x1, y1), (x2, y2), (x3, y3), (x4, y4);

[0060] Table 2 Mathematical expression of the positioning text box obtained by the EAST model

[0061] Text box shape Number of channels Mathematical Expression AABB 4 G=R={d|i∈{1, 2, 3, 4}} RBOX (rotating rectangular box) 5 G={R,θ} QUAD (quadrilateral frame) 8 G=Q={(Δx,Δy)|i∈{1, 2, 3, 4}}

[0062] In view of the phenomenon that the first and last characters of the slab number are not fully selected in this embodiment, a method of correcting and expanding the positioning coordinates is adopted to further improve the accuracy of slab number positioning, which mainly includes the following steps:

[0063] (1) Text box correction: Due to the on-site shooting angle, the slab number usually has a certain tilt angle, so the positioning text box will also have a certain angle. In order to facilitate the subsequent coordinate expansion, the text box is first corrected;

[0064] More specifically, (x1, y1), (x2, y2), (x3, y3), (x4, y4) represent the coordinates of the four corner points of the positioning text box, (x0, y0) represents the geometric center of the rectangle, and the tilt angle θ can be expressed by the trigonometric relationship between the coordinates of each point of the corner text box and the coordinates of the geometric center as follows: The coordinates of the four corner points of the text box after tilt correction (x′1, y′1), (x′2, y′2), (x′3, y′3), (x′4, y′4) can be expressed as:

[0065] x′ i =(x i-x0)cosθ+(y i -y0)sin(θ)+x i , i∈[1,4]

[0066] y′ i =(y i -y0)cosθ+(x i -x0)sin(θ)+y i , i∈[1,4]

[0067] Among them, (x i ,y i ) are the coordinates of each point after rotation;

[0068] (2): Text box expansion: mainly expands the horizontal coordinates of the text box, that is, the horizontal coordinates of the upper left and lower left points of the text box are translated to the left by Δx, and the horizontal coordinates of the upper right and lower right points are translated to the right by Δx, while the vertical coordinates of each point remain unchanged. The expansion scale is L / 9, where L is the length of the text box, that is, L = (x3-x4) = (x2-x1), and the coordinates of each point in the text box before expansion are represented by (x′1, y′1), (x′2, y′2), (x′3, y′3), (x′4, y′4), then the coordinates of each point after expansion (x″1, y″1), (x″2, y″2), (x″3, y″3), (x″4, y″4) can be expressed as: (x″ i , y″ i )=(x′ i Δx,y′ i ), i = 1, 3, (x″ i , y″ i )=(x′ i +Δx,y′ i ), i=2,4;

[0069] In this embodiment, based on the slab number labeling process of the above transfer learning idea, a total of 2451 slab data of 32*400 size are obtained;

[0070] In order to reduce the impact of data scale on algorithm performance in deep learning, in this embodiment, data enhancement was performed on the 2451 slab image datasets obtained, including brightness transformation, opening and closing operations, and adding Gaussian noise and salt and pepper, a total of 6 methods, and finally an enhanced slab number dataset of 14706 images was obtained;

[0071] Step 3.6: Use computer vision and image processing methods to perform character segmentation on the slab number in the slab number image dataset. The specific method is as follows:

[0072] Step 3.6.1: Binarize the slab number image dataset by using an adaptive threshold method to convert the slab number image from a color image to a black and white image;

[0073] Step 3.6.2: Project the binary slab number image in the horizontal direction. Since there are gaps between characters, there must be troughs in the histogram after projection. Use the troughs in the histogram obtained after projection as segmentation points to perform character segmentation, and obtain the segmented slab number character data set. Name the segmented character images according to their positions in the entire slab image, such as 1-1, 1-2, ..., 1-10, where '1-1' represents the first character segmented from the first slab number, so as to facilitate the subsequent reorganization of the character recognition results into slab number sequence recognition results;

[0074] Step 4: Based on the idea of ​​multi-objective genetic programming, an autonomous evolutionary neural network algorithm is designed to build an integrated model for slab number recognition to realize online recognition of slab numbers;

[0075] Step 4.1: Design an autonomous evolutionary neural network algorithm based on Cartesian genetic programming and Pareto optimal solution set ideas;

[0076] Step 4.1.1: Design a variety of convolution, pooling, residual, network series, and network parallel basic modules within the neural network. Based on Cartesian genetic programming, arrange and combine the designed basic modules to achieve variable-length neural network structure encoding;

[0077] This embodiment designs optional basic functional modules, including convolution blocks, residual blocks, pooling blocks, and series-parallel network structures. These five types also include different sub-modules according to their specific parameters, as shown in Table 3 below. Taking the convolution module as an example, its symbol is represented by CB(C, k), where CB is the abbreviation of the convolution module, and (C, k) is the number and size of the convolution kernels in the module. Furthermore, the number of convolution kernels C∈{32,64,128} and the size of the convolution kernel k∈{1,3,5} are randomly combined, and there are 9 convolution modules to choose from. Other basic functional modules are similar, totaling 22.

[0078] In the process of encoding the network structure based on Cartesian genetic programming, 22 alternative functional modules can be randomly selected and randomly combined, and the number of selected modules is not fixed, so as to realize the network structure encoding of variable length. The schematic diagram of converting a gene code into a corresponding neural network in this embodiment is as follows: Figure 4As shown in the figure, the genotype of each network structure contains three items, namely the ID value of the node, the activation type of the node (True / False) and the specific type of the node. Only nodes with an activation type of True will be decoded into the network structure, and nodes with an activation type of False will be discarded.

[0079] Table 3 Optional neural network basic function blocks and parameter settings

[0080]

[0081] Step 4.1.2: Taking the model parameter quantity and recognition accuracy as the two objectives, a multi-objective evolutionary strategy based on the Pareto optimal solution set is adopted to guide the autonomous evolutionary neural network algorithm to quickly search for the optimal neural network architecture with high recognition accuracy and low model complexity;

[0082] In this embodiment, the accuracy of slab number recognition Acc and the complexity of the model, i.e., the number of parameters N paras Two optimization objectives are used to guide the evolution of the autonomous evolutionary neural network algorithm. This process relies on the Pareto optimal solution and eventually obtains the optimal Pareto frontier. Each solution on the frontier represents an optimal neural network architecture. These networks can be used to identify the slab number.

[0083] Step 4.1.3: Design genetic operators including crossover and mutation, so that the offspring solutions generated by crossover and mutation are more diverse while retaining the excellent genes of the parent solution, thereby expanding the search space of the autonomous evolutionary neural network algorithm and providing more possibilities for the algorithm to find the optimal solution;

[0084] In this embodiment, the crossover process is designed as a single-point crossover, that is, a crossover point is randomly selected from the two parent individuals, and the two networks of the parent generations are cross-recombined based on the crossover point. In this embodiment, the specific process of the crossover is as follows: Figure 5As shown, the first line is the genotypes of the two parent individuals before the crossover, including the activation types of their functional modules and related parameters, and the second line is the genotypes of the offspring after the crossover. According to the mutated genotypes and the encoding method of the network structure designed in this embodiment, the network architecture after the crossover can be obtained; this embodiment uses (

[300]

[1100]

[1012]

[1902]

[2043]

[700]

[056] ) and (

[1300]

[1400]

[1920]

[1511] [1431 ]

[1041]

[066] ) respectively represent the genotypes of the two parents before the crossover. When the crossover point is selected at node 3, the genotypes of the two offspring after the crossover operation are (

[300]

[1100]

[1012]

[1902]

[1431]

[1041]

[066] ) and (

[1300]

[1400]

[1920]

[1511]

[2043]

[700]

[056] ). According to the encoding and decoding strategy designed in the present invention, the corresponding network architecture can be obtained through the genotypes of the above offspring.

[0085] The design of the mutation operator in this embodiment includes two types, namely, the mutation of the module type and the mutation of the connection mode between modules. The mutation of the module type includes but is not limited to the mutation of the convolution block into a residual block, the mutation of the convolution kernel size from 3*3 to 5*5, etc. In the mutation of the connection mode, the module at the set mutation point can be randomly connected to any module located in front of it. At the same time, in order to solve the problem of mismatch in the dimensions of the two connected modules, the processing done in this embodiment is to achieve dimensional matching by using a 1*1 convolution kernel;

[0086] Step 4.2: The slab number character data set obtained in step 3 is passed as training data to the autonomous evolutionary neural network algorithm designed in step 4.1, and the autonomous evolutionary neural network algorithm designed in step 4.1 is trained to obtain the final Pareto optimal frontier;

[0087] Step 4.3: Use the solution on the Pareto optimal frontier obtained in step 4.2 as the slab number recognition sub-model, and build an integrated model for online slab number recognition based on the differential evolution algorithm;

[0088] Step 4.3.1: On the premise that the total weight sum is 1, give each slab number recognition sub-model an initial weight by random initialization;

[0089] Step 4.3.2: Based on the slab number character data obtained in step 3 and each slab number recognition sub-model with initial weights obtained in step 4.3.1, the recognition accuracy of the slab number online recognition integrated model is calculated;

[0090] Step 4.3.3: Based on the DE / rand / 1 mutation strategy and adaptive crossover strategy in differential evolution, the slab number online recognition integrated model is trained to optimize the weights of each sub-model in the integrated model until the maximum number of iterations is reached or the model recognition accuracy tends to converge and no longer increases, then the training is stopped;

[0091] Step 4.3.4: Perform online recognition of the slab number characters according to the slab number online recognition integrated model obtained in step 4.3.3;

[0092] Step 5: The identified single slab number characters are reorganized according to the naming method of the character image to obtain the predicted slab number, and the predicted slab number is compared with the slab number in the slab library. At the same time, combined with the furnace parameters such as the length, width, thickness, and weight of the slab, automatic release, pending or alarm processing is given.

[0093] In this embodiment, if the identified slab number does not exist in the slab warehouse, an automatic alarm is triggered, and it is determined that the identification result is wrong. The hot rolling production control system will issue an instruction to lock the roller and stop the forward conveyance of the slab, which must be identified by the operator; if the identified slab number exists in the slab warehouse, the hot rolling production control system retrieves the furnace parameters of the slab and compares them with the furnace parameters of the slab with the same number in the slab warehouse. If the slab in the process of moving has the same furnace parameters as the slab in the matched slab warehouse, it is released. Otherwise, the parameters are checked according to different production lines and different product specifications. If only a few parameters are unqualified, the deviation is basically small, and the production line belongs to the requirements of medium and low quality products, it is released. If the production line belongs to the production line with high quality requirements and there are deviations in individual parameters, an early warning indication is issued, and it is pending whether to proceed to the next step. The hot rolling production control system will issue an instruction to lock the roller and stop the forward conveyance of the slab, prompting manual intervention, and the operator determines whether the slab is correct. If it is correct, the early warning is manually lifted.

[0094] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. A method for online identification of slab numbers based on autonomous evolutionary neural networks, characterized in that: The following steps are involved: Step 1: extracting video data of the slab moving process during the hot rolling production process; Step 2: Process the video data of the slab moving process to obtain an image data set with slab numbers; Step 3: Based on the idea of ​​transfer learning, locate the specific position of the slab number on the slab image and create the data set required for slab number recognition; Step 4: Based on the idea of ​​multi-objective genetic programming, an autonomous evolutionary neural network algorithm is designed to build an integrated model for slab number recognition to realize online recognition of slab numbers; Step 4.1: Design an autonomous evolutionary neural network algorithm based on Cartesian genetic programming and Pareto optimal solution set ideas; Step 4.2: The slab number character data set obtained in step 3 is passed as training data to the autonomous evolutionary neural network algorithm designed in step 4.1, and the autonomous evolutionary neural network algorithm designed in step 4.1 is trained to obtain the final Pareto optimal frontier; Step 4.1.1: Design a variety of convolution, pooling, residual, network series, and network parallel basic modules within the neural network. Based on Cartesian genetic programming, arrange and combine the designed basic modules to achieve variable-length neural network structure encoding; Step 4.1.2: Taking the model parameter quantity and recognition accuracy as the two objectives, a multi-objective evolutionary strategy based on the Pareto optimal solution set is adopted to guide the autonomous evolutionary neural network algorithm to quickly search for the optimal neural network architecture with high recognition accuracy and low model complexity; Step 4.1.3: Design genetic operators including crossover and mutation, so that the offspring solutions generated by crossover and mutation are more diverse while retaining the excellent genes of the parent solutions, thereby expanding the search space of the autonomous evolutionary neural network algorithm; Step 4.3: Use the solution on the Pareto optimal frontier obtained in step 4.2 as the slab number recognition sub-model, and build an integrated model for online slab number recognition based on the differential evolution algorithm; Step 5: The identified single slab number characters are reorganized according to the naming method of the character image to obtain the predicted slab number, and the predicted slab number is checked with the slab number in the slab library. At the same time, combined with the furnace parameters of the slab, automatic release, pending or alarm processing is given.

2. The method for online identification of slab numbers based on autonomous evolving neural network according to claim 1 is characterized in that: The specific method of step 2 is: Step 2.1: Use the inter-frame difference method to compare the difference between the images of consecutive frames in the video data, extract the frames where the slab passes from the video data, and obtain the slab image; Step 2.2: Filter the obtained slab images to select image data in which the slab numbers are clearly captured, and obtain an image data set with slab numbers.

3. The method for online identification of slab numbers based on autonomous evolutionary neural network according to claim 2 is characterized in that: The specific method of step 3 is: Step 3.1: Train the EAST scene text localization model based on the ICDAR2015 dataset; Step 3.2: Based on the EAST scene text localization model trained in step 3.1, the image dataset with slab numbers obtained in step 2.2 is input into the trained EAST scene text localization model for testing, and the slab images whose slab numbers can be accurately located are screened out; Step 3.3: Add the slab images that can be accurately located selected in step 3.2 as training data to the ICDAR2015 dataset, and retrain the EAST scene text localization model; Step 3.4: Repeat steps 3.2 to 3.3 for multiple times to achieve the migration of models and data, thereby achieving accurate positioning of the slab number; Step 3.5: For all images where the slab number position is accurately located, the image containing only the slab number is cropped from the entire slab image according to the coordinates of the positioning text box to obtain a slab number image dataset; Step 3.6: Perform character segmentation on the slab number in the slab number image dataset by using computer vision and image processing methods.

4. The method for online identification of slab numbers based on autonomous evolutionary neural network according to claim 3 is characterized in that: The specific method of step 3.6 is: Step 3.6.1: Binarize the slab number image dataset by using an adaptive threshold method to convert the slab number image from a color image to a black and white image; Step 3.6.2: Project the binary slab number image in the horizontal direction, and perform character segmentation using the trough in the histogram obtained after projection as the segmentation point to obtain the segmented slab number character data set, and name the segmented character image according to its position in the entire slab image.

5. The method for online identification of slab numbers based on autonomous evolving neural network according to claim 4 is characterized in that: The specific method of step 4.3 is: Step 4.3.1: On the premise that the total weight sum is 1, give each slab number recognition sub-model an initial weight by random initialization; Step 4.3.2: Based on the slab number character data obtained in step 3 and each slab number recognition sub-model with initial weights obtained in step 4.3.1, the recognition accuracy of the slab number online recognition integrated model is calculated; Step 4.3.3: Based on the DE / rand / 1 mutation strategy and adaptive crossover strategy in differential evolution, the slab number online recognition integrated model is trained to optimize the weights of each sub-model in the integrated model until the maximum number of iterations is reached or the model recognition accuracy tends to converge and no longer increases, then the training is stopped; Step 4.3.4: Perform online recognition of the slab number characters based on the slab number online recognition integrated model obtained in step 4.3.3.

Images