CN115311456A

CN115311456A - Tuyere coke segmentation method based on improved DeepLabv3+

Info

Publication number: CN115311456A
Application number: CN202210959538.XA
Authority: CN
Inventors: 李豪; 王世友; 张颖伟; 冯琳
Original assignee: Northeastern University China; Bengang Steel Plates Co Ltd
Current assignee: Northeastern University China; Bengang Steel Plates Co Ltd
Priority date: 2022-08-11
Filing date: 2022-08-11
Publication date: 2022-11-08

Abstract

The invention provides a tuyere coke segmentation method based on improved deep Labv3+, and relates to the technical field of coke identification. Firstly, acquiring coke video data of a blast furnace tuyere raceway, converting the acquired coke video data of the blast furnace tuyere raceway into picture data, and then labeling coke particles in the picture data to obtain a tuyere coke data set; then constructing a Deeplabv3+ model fused with a coordinate attention mechanism as a tuyere coke segmentation model, and training the tuyere coke segmentation model by using a tuyere coke data set; and finally, segmenting the coke particles in the tuyere coke image to be segmented by using the trained tuyere coke segmentation model. According to the method, a CoordinateAttention network is introduced into the existing Deeplabv3+ model, so that the recognition and perception of the model on the coke particles are enhanced, and the risk of misclassification is reduced; and four parallel cavity convolution extraction features are introduced into the ASPP module, so that the expression capability of the feature map is continuously enhanced, the prediction precision is improved, and the generalization capability of the model is also enhanced.

Description

Tuyere coke segmentation method based on improved DeepLabv3+

Technical Field

The invention relates to the technical field of coke identification, in particular to a tuyere coke segmentation method based on improved deep Labv3 +.

Background

In the blast furnace ironmaking, coke is required to be added besides the iron-containing raw material, and the coke provides 78% of energy for the whole blast furnace ironmaking; secondly, the coke also acts as a reducing agent for reducing the blast furnace gas; moreover, the main component of the dead material column at the bottom of the furnace hearth is coke, so that the function of supporting furnace burden is achieved on one hand, and the air permeability in the furnace is guaranteed to meet certain standards on the other hand. Therefore, coke plays an important role in the iron-making process. After the coke is put in from the top of the blast furnace, the coke is subjected to collision, friction, extrusion, carbon dissolution reaction, gasification and other processes in the blast furnace from the top, so that the strength of the coke is reduced, the particle size of the coke is gradually reduced, and after the coke reaches a tuyere raceway, the coke is subjected to aggravated abrasion under the impact of high-speed airflow and violent combustion, and the particle size is rapidly reduced. For the change of the coke granularity of the blast furnace, the change is only known to be a process which is continuously reduced, but for the numerical value of the coke granularity of a specific part, no reliable method can be used for accurately solving the numerical value. Therefore, the development of the partition work of the coke at the tuyere plays a key role in the subsequent solution of the coke size.

The image segmentation refers to a process of dividing a region of interest in a picture according to the difference of pixels to form a foreground and a background. At present, the segmentation field is divided into a traditional segmentation method and a deep learning segmentation method. In the conventional segmentation techniques, threshold segmentation, region segmentation, edge segmentation, and the like are applied more frequently. Because the traditional method has poor segmentation precision and is influenced by the pixels of the original input image and the segmentation conditions, the segmentation method based on deep learning is the mainstream in the market at present.

Disclosure of Invention

The technical problem to be solved by the invention is to provide a tuyere coke cutting method based on improved DeepLabv3+ aiming at the defects of the prior art, so as to realize the cutting of the tuyere coke.

In order to solve the technical problems, the technical scheme adopted by the invention is as follows: a tuyere coke segmentation method based on improved DeepLabv3+ is characterized in that a tuyere coke data set is constructed through a coke image of a blast furnace tuyere convolution area with coke labels;

constructing a Deeplabv3+ model fused with a coordinate attention machine as a tuyere coke segmentation model;

training a tuyere coke segmentation model by using a tuyere coke data set;

segmenting coke in a tuyere coke image to be segmented by using the trained tuyere coke segmentation model;

the method specifically comprises the following steps:

step 1: collecting coke video data of a blast furnace tuyere raceway;

erecting a CCD industrial camera at the small blast furnace tuyere mirror to acquire data in real time, so as to obtain coke video data of a rotary area of the blast furnace tuyere;

step 2: converting the acquired coke video data of the blast furnace tuyere raceway into picture data, and labeling coke particles in the picture data to obtain a tuyere coke data set;

labeling coke particles in a tuyere raceway in the picture data by using a labelme labeling tool;

and step 3: taking 80% of tuyere coke data in the tuyere coke data set as a training set, and taking 20% of tuyere coke data as a testing set;

and 4, step 4: constructing a Deeplabv3+ model fused with a coordinate attention mechanism as a tuyere coke segmentation model, and training the tuyere coke segmentation model by using a training set;

the Deeplabv3+ model fused with the Coordinate Attention machine system selects an Aligned Xconcentration network as a trunk feature extraction network of an Encoder module on the basis of an original Deeplabv3+ model, a coding attachment network is added to process a feature map extracted by the trunk feature extraction network, an ASPP (automatic sequence protocol) module in the original Deeplabv3+ model is improved, and the rest structures of the Deeplabv3+ model fused with the Coordinate Attention machine system are consistent with the Deeplabv3+ model; the improved ASPP module consists of a 1 multiplied by 1 convolution layer, four parallel hollow convolution layers and a pooling layer;

after the input picture is processed by the Aligned Xcenter network, two characteristic graphs are obtained, wherein one characteristic graph is a shallow characteristic, the other characteristic graph is a deep characteristic, and the two characteristic graphs are input into a Coordinate Attention network for processing and then follow-up operation of an original Deeplabv3+ model is executed; the shallow feature processed by the CoordinateAttention network is directly transmitted into the Decoder module; inputting the deep features processed by the CoordinateAttention network into an improved ASPP module, and extracting semantic information of different scales; inputting the deep features processed by the improved ASPP module into a Decode module, fusing the deep features and the shallow features by the Decode module to obtain a final feature map, performing up-sampling by adopting bilinear interpolation, and recovering the features extracted by the Encoder module to the same size as the original picture to obtain a segmentation result map of the tuyere coke;

and 5: evaluating the generalization ability of the trained tuyere coke segmentation model by using the test set to obtain an optimal tuyere coke segmentation model;

and 6: and acquiring coke video data of a blast furnace tuyere raceway in real time, converting the coke video data into image data, and then segmenting the coke particles by using an optimal tuyere coke segmentation model.

Adopt the produced beneficial effect of above-mentioned technical scheme to lie in: aiming at the problems that the existing Deeplabv3+ model is inaccurate in image target edge segmentation, slow in image feature fitting and incapable of effectively utilizing attention information, a CoordinateAttention network is introduced behind a deep convolution neural network module, so that the identification and perception of the model on coke particles are enhanced, and the risk of misclassification is reduced; and simultaneously, the ASPP module network structure is optimally designed, four parallel cavity convolutions are introduced to extract features, and the expression capability of the feature map is continuously enhanced. The improved Deeplabv3+ model is verified through the constructed tuyere coke data set, and the result shows that the improved model can effectively improve the defects of the original model, can more finely divide the target, better solves the problems of rough dividing boundary and the like, and enhances the generalization capability of the model while improving the prediction precision.

Drawings

FIG. 1 is a flowchart of a method for tuyere coke division based on the modified DeepLabv3+ provided by an embodiment of the present invention;

FIG. 2 is a tuyere image collected from a tuyere mini-mirror provided in an embodiment of the present invention;

FIG. 3 is a diagram illustrating the labeling result of coke particles by using labelme according to an embodiment of the present invention;

FIG. 4 is a diagram illustrating the results of an original tuyere image and a label provided in an embodiment of the present invention;

fig. 5 is a diagram of a deplabv 3+ model structure of a fusion coordinate attention mechanism according to an embodiment of the present invention;

fig. 6 is an original deepabv 3+ model architecture diagram provided in an embodiment of the present invention;

FIG. 7 is a diagram of an Aligned Xreception network structure according to an embodiment of the present invention;

fig. 8 is a schematic structural diagram of an improved ASPP module provided in an embodiment of the present invention;

fig. 9 is a schematic diagram of a CoordinateAttention network structure according to an embodiment of the present invention;

FIG. 10 is a graph showing the results of coke particle segmentation in different models according to the image data of a portion of the tuyere provided in the embodiment of the present invention.

Detailed Description

The following detailed description of embodiments of the present invention is provided in connection with the accompanying drawings and examples. The following examples are intended to illustrate the invention but are not intended to limit the scope of the invention.

This example uses a steel from 2500m ³ Taking data of the actual smelting process of the blast furnace as an example, the method for dividing the coke at the tuyere based on the improved DeepLabv3+ is adopted to realize the blast furnaceAnd (4) dividing coke in a tuyere raceway.

In this embodiment, a tuyere coke cutting method based on improved deep lab v3+, as shown in fig. 1, includes the following steps:

step 1: collecting coke video data of a blast furnace tuyere raceway;

in the embodiment, the CCD industrial camera is erected at the small mirror of the blast furnace tuyere to acquire the coke video data of the rotary area of the blast furnace tuyere in real time. The collected tuyere images are respectively shown in fig. 2, the black rod-shaped object at the upper right corner of the image is a coal injection pipe, and the area with darker middle color and approximate to an ellipse is coke particles. Because the view of the small air port mirror is limited, the image data collected by the embodiment is carried out during the coal stopping of the blast furnace, and the integral stable operation of the blast furnace cannot be influenced by the short coal stopping.

And 2, step: converting the acquired coke video data of the blast furnace tuyere raceway into picture data, and labeling coke particles in the picture data to obtain a tuyere coke data set;

the tuyere image data acquired from the tuyere small mirror needs to be further processed, and the specific operation is to convert the acquired video data into picture data and then finish labeling coke particles in the picture data. The method for marking the original pictures is various, often adopts labelme software, is simple and convenient to operate and easy to operate, has low requirements on hardware equipment, and can be installed in any equipment and finish corresponding marking work. In this example, labeling of coke particles on the tuyere raceway picture data was accomplished using a labelme labeling tool.

Importing all pictures to be marked into labelme software, then carrying out point tracing operation on a target area, and independently dividing the target area, namely coke particles in the images, wherein the marked area is represented by a specific name and aims to well distinguish objects to be identified in subsequent semantic segmentation codes, clicking a storage button to convert the format of the objects to json after marking is finished, and forming a label picture to be put into a label data set. The labelme labeling result is shown in fig. 3, wherein the closed area formed by connecting small circular points in sequence is the position of the tuyere coke. And after completing the data annotation work of the collected tuyere image by using labelme software, obtaining a tuyere raceway coke data set, wherein the original tuyere image and the label result are displayed as shown in FIG. 4, wherein the left side is a real tuyere image, and the right side is an artificial label.

And step 3: taking 80% of tuyere coke data in the tuyere coke data set as a training set and 20% of tuyere coke data as a test set;

according to the Deeplabv3+ model fusing the coordinate attention mechanism, on the basis of the original Deeplabv3+ model, an align Xconcentration network is selected as a trunk feature extraction network of an Encoder module, a CoordinateAttention network is added to process a feature map extracted by the trunk feature extraction network, an ASPP module in the original Deeplabv3+ model is improved, and the rest of the structure of the Deeplabv3+ model fusing the coordinate attention mechanism is consistent with that of the Deeplabv3+ model, as shown in FIG. 5; the improved ASPP module consists of a 1 multiplied by 1 convolution layer, four parallel hollow convolution layers and a pooling layer;

after the input picture is processed by the Aligned Xcenter network, two characteristic graphs are obtained, wherein one characteristic graph is a shallow characteristic, the other characteristic graph is a deep characteristic, and the two characteristic graphs are input into a Coordinate Attention network for processing and then follow-up operation of an original Deeplabv3+ model is executed; the shallow feature processed by the CoordinateAttention network is directly transmitted into the Decoder module; inputting the deep features processed by the CoordinateAttention network into an improved ASPP module, and extracting semantic information of different scales; inputting the deep layer characteristics processed by the improved ASPP module into a Decode module, fusing the deep layer characteristics and the shallow layer characteristics by the Decode module to obtain a final characteristic diagram, performing up-sampling by adopting bilinear interpolation, and recovering the characteristics extracted by the Encoder module to the same size as the original image to obtain a segmentation result diagram of the tuyere coke;

the architecture of the Deeplabv3+ model is shown in FIG. 6, and the overall network structure of the model is composed of an Encoder module and a Decode module, wherein the Encoder module is mainly responsible for processing image semantic features, and the Decode module is mainly responsible for restoring feature information. The Encoder module comprises a Deep Convolutional Neural Network (DCNN) and a cavity space pyramid pooling module (ASPP); the Decode module carries out subsection processing on the extracted features and then executes fusion processing of deep features and shallow features, so that the extracted features are comprehensive, and processing of edge detail information can be well enhanced. And finally, completing segmentation operation by adopting a bilinear interpolation technical means to obtain a segmentation result.

In the Encoder module, the DCNN is used to extract detail features of a specific object or a target scene in an image to be segmented. The network often used includes VGGNet, google lenet, resNet, mobileNet, xception, etc., and the backbone network used in this embodiment is an Aligned Xception network, which is an enhanced version of the Xception network. The Xception network reduces the complexity of a model by introducing deep separable convolution, accelerates the processing time of the model, and simultaneously, a residual mechanism further promotes the network convergence speed. Compared with the original Xpacket network, the Aligned Xpacket network has the following improvements. Firstly, the Aligned Xcenter network has a deeper network structure, a midle flow module in the original network structure is iterated for only 8 times, and the fine adjustment of the Aligned Xcenter is changed into iteration for 16 times; meanwhile, the depth separable convolution is used for replacing the maximum pooling operation in the original network, and preparation is made for extracting feature maps with any size by adopting the void convolution in the subsequent process; and finally, adding a batch normalization link and a ReLU activation link after each deep convolution process, and further improving the accuracy of feature collection. The network structure is shown in fig. 7.

ASPP (advanced Spatial gradient Pooling) completes corresponding convolution operation by adding convolution kernels with different sampling rates and obtains information of different scales, so that the extraction of the feature information of the target image is more comprehensive, and the improvement of the image segmentation precision is facilitated. A simple understanding is a distinguished pooling layer, whose purpose is to extract features as much as possible, consistent with a common pooling layer. The structure of the improved ASPP module is shown in fig. 8, which consists of a 1 × 1 convolution (left-most) + Pooling pyramid (three in the middle) + Pooling (right-most). The method is characterized in that cavity convolution is adopted for each layer of the pooling pyramid to replace common convolution, the cavity convolution is applied to the image segmentation field at the earliest time, the problem can be faced in the research of the segmentation field at first, in the feature extraction stage, repeated convolution and pooling can increase the size of a receptive field, but the size of a feature map can be reduced correspondingly, and when the image is finally restored, namely the size of the feature map is restored by adopting up-sampling, the problem that the size of the feature map is small and the size of the image is large is easy to generate, so that certain loss of precision can be generated. After the hole convolution occurs, it becomes easy to solve the above problem.

The hole convolution is different from the conventional convolution in that a hole binary word is provided with a key parameter called a relation rate in the hole binary word, generally called a sampling rate, the number of zero elements inserted into a convolution kernel in the hole convolution is determined by the size of the sampling rate, the size of the whole convolution kernel is increased by adding the zero elements, but the parameter number of the convolution kernel is kept unchanged, so that the receptive field can be expanded while the calculation rate is ensured, and more context information can be acquired;

the combination of the void ratios of the original deep labv3+ model ASPP module is [6,12,18], the embodiment adjusts the network structure of the ASPP to a certain extent, and when the combination of the void ratios is [3,6,12,18], the overall performance index of the model is optimal.

The Decoder module mainly comprises two upsampling processes, and bilinear interpolation is adopted to restore the image size. The decoding process comprises the following steps that firstly, output dimensionality is adjusted by utilizing 1 x 1 convolution on shallow layer characteristics, the dimensionality can be kept consistent with the upsampled high-layer characteristics, accordingly, subsequent concatenate operation can be executed, the concatenate processing process is not simple addition of two characteristics, a characteristic is obtained after stacking processing of one characteristic, and a segmentation graph with the same size as an input image can be obtained after the characteristic graph is subjected to the upsampling operation.

The CoordinateAttention (CA) network can capture cross-channel information, and can also capture direction perception and position perception information, which is helpful for more accurate positioning and identification of a specific target by a model; second, the CA network is flexible and lightweight, and can be easily inserted into classical modules to enhance features by a method of enforcing information representation, the network structure of which is shown in fig. 9.

The overall structure of the CA network is relatively simple and easy to understand, and the process of introducing attention can be divided into two processes, wherein the first process is a coordinate information embedding stage, and the second process is a coordinate attention generating stage. In the first process, the input features are aggregated into two separate direction-aware feature maps in the vertical and horizontal directions, respectively, by using two one-dimensional globopooling operations. The CA network has the main advantages that the problem of losing position information in the prior channel attention is solved, direction and position information are merged into the channel attention on the premise of keeping the channel attention, and the attention module is prompted to capture remote dependence by accurate position information. In a comparative experiment with other attention mechanism performances, the CA network is excellent in performance, and is greatly promoted in task processing of specific scenes such as image recognition, target detection and semantic segmentation.

And 5: evaluating the generalization ability of the trained tuyere coke segmentation model by using a test set to obtain an optimal tuyere coke segmentation model, and evaluating the optimal tuyere coke segmentation model by using two evaluation indexes of MIoU and MPA;

the performance of the semantic segmentation model is judged, and besides the segmentation effect accident of directly observing the segmentation graph, the performance can be realized by using some evaluation indexes, wherein the commonly used evaluation indexes comprise MIoU, MPA and the like. In order to facilitate understanding of specific meanings of these evaluation indexes, the present embodiment needs to implement the following assumptions:

(1) The pixels in the image belong to different categories, wherein the total category is k +1 and comprises a background.

(2)p _ij Indicating the number of pixels that are originally of class i but were misjudged as class j. When i = j, the prediction is correct, otherwise the prediction is wrong.

MIoU (mean intersectionohunion): calculating the average value of the ratio of the intersection and the union of the two sets of the true value and the predicted value for the common evaluation standard of semantic segmentation, wherein the formula is as follows:

MPA (Mean PixelAccuracy): calculating the proportion of the predicted correct number to the total number of the pixels in each category, and averaging the calculation results of all the categories, as shown in the following formula:

step 6: and acquiring coke video data of a blast furnace tuyere raceway in real time, converting the coke video data into image data, and then segmenting the coke particles by using an optimal tuyere coke segmentation model.

In the embodiment, a migration learning method is used for initializing the tuyere coke segmentation model, the calculation time of the model is reduced, and parameter fine adjustment is performed according to model loss, wherein the total amount of tuyere coke data sets in an experiment is 1000, 80% of the tuyere coke data sets are used for training, and 20% of the tuyere coke data sets are used for testing. The relevant parameters are shown in table 1.

Table 1 network parameters and related configurations

Model training uses a pixel cross entropy loss function, and obtains a loss value by comparing each pixel one by one, wherein the expression is as follows:

where k represents the category, y _i Representing a one-bit thermal coding vector, and taking the value of 1 if the category is the same as the label category, and taking the value of 0 if the category is different from the label category; p is a radical of _i Representing the probability that the prediction sample belongs to i.

In order to verify the reliability and availability of the deplab v3+ model of the fusion attention mechanism, the FCN model with the backbone network of VGGnet and the deplab v3+ model with the DCNN of Xception are adopted in the embodiment to compare with the improved deplab v3+ model of the invention. The results of the models on the test set can reflect the integral segmentation performance of the models, the coke data set of the tuyere raceway which is manufactured in advance is transmitted into the Deeplabv3+ model improved by the method and other models to perform the segmentation work of the tuyere coke, and the performance of each model is evaluated by counting the experimental results. And evaluating different models by using the evaluation indexes MIoU and MPA, wherein the evaluation indexes of different models are shown in a table 2.

TABLE 2 comparison of segmentation results for different models

Segmentation model	MIoU	MPA
			FCN-8s	74.25％	76.34％
DeepLabv3+	84.81％	87.92％
			CA_DeepLabv3+	87.68％	94.39％

From experimental results, the FCN model performs the worst in the tuyere coke segmentation task, and the MIoU is only 74.25%, because the FCN does not process the detail information of the input image, the relation among the pixels is not fully considered, and the space regularization step used in the segmentation method based on pixel classification is omitted, resulting in lack of overall space; the DeepLabv3+ model is excellent in performance by depending on a special coding-decoding structure of the model; the DeepLabv3+ model which is the fusion coordinate attention mechanism has the best performance, wherein the MIoU reaches 87.68 percent, is improved by 13.43 percent compared with the FCN model and is improved by 2.87 percent compared with the original DeepLabv3+ model, and the integral DeepLabv3+ model has larger performance improvement after the CA network is introduced.

The ASPP module extracts semantic information of different scales in an original image by introducing hole convolution of different hole rates, so that the extraction of characteristic information of a target image is more comprehensive. Therefore, it is very important to design a network structure capable of capturing multi-scale semantic information and obtaining a large enough receiving domain. Three parallel void convolutions are introduced into the ASPP module in the original deep Labv3+ model, the void ratios are [6,12,18], but different networks formed by different research objects often need deep research on a specific void ratio combination to obtain an optimal result, in order to obtain an optimal segmentation model, the embodiment conducts deep research on the ASPP module void ratio combination on the basis of the CA _ deep Labv3+ model, and designs the void ratios of different combinations to perform experiments, wherein the experimental results are shown in Table 3.

TABLE 3 comparison of the segmentation results of the CA _ DeepLabv3+ model under different sampling rate combinations

Sampling rate	MIoU	MPA
			[3,6,12]	86.46％	91.83％
[6,12,18]	87.68％	94.39％
			[3,6,12,18]	88.62％	95.01％
[6,12,18,24]	87.45％	93.56％

From the results in the table, the void combination rate [3,6,12] is inferior to the original [6,12,18] in performance, because the void ratio is small, the receptive field of the feature map is also small, and the overall extraction of the features is not facilitated; the combination rate of the holes is [6,12,18,24], the structure of ASPP is changed, four parallel hole convolutions are introduced, but the performance result is still not the same as that of the original model, because when the hole rate is designed to be too large, the phenomena of not dense sampling and losing a large amount of information are easily caused, and the whole characteristic extraction process is influenced; when the hole combination rate is [3,6,12,18], all performance indexes of the model reach the optimum,

the MIoU index is 88.62%, which is 0.94% higher than the CA _ DeepLabv3+ model. Finally, the improved DeepLabv3+ model which integrates an attention mechanism and optimizes an ASPP structure is obtained, and the MIoU of the model is improved by 14.37% compared with the FCN model and is improved by 3.81% compared with the original DeepLabv3+ model.

In order to more clearly and intuitively feel the segmentation effect of each model, fig. 10 shows the segmentation result of part of the tuyere image data under different models, and the segmentation result sequentially includes an original image, an artificial label, an FCN segmentation result, a deep labv3+ segmentation result and an improved deep labv3+ model segmentation result from left to right in the segmentation graph.

From the segmentation result, the FCN model can roughly segment the coke particles, but the segmentation effect is not ideal, the coke edge is discontinuous, and meanwhile, some misclassification phenomena exist. The discontinuity of the coke edge causes that the subsequent coke particle size calculation work cannot be completed by the segmentation result; compared with the FCN model, the original DeepLabv3+ model is improved to a certain extent, the coke edge can be continuously outlined, but small particles are mistakenly identified, and the phenomenon of model misclassification is caused only when the feature extraction of the tuyere image is not complete; the improved DeepLabv3+ model provided by the invention has a better segmentation effect than an FCN model and an original DeepLabv3+ model, and can clearly see that the segmentation result of the improved DeepLabv3+ model can better identify and segment coke particles from the comparison of the segmentation result and an artificial label, and meanwhile, the control treatment is also compared in the details of the coke edge, so that the improved DeepLabv3+ model can be used for completing the segmentation work of the coke at the tuyere, and a solid foundation is laid for the subsequent solving work of the coke particle size.

Finally, it should be noted that: the above examples are only intended to illustrate the technical solution of the present invention, and not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some or all of the technical features may be equivalently replaced; such modifications and substitutions do not depart from the spirit of the corresponding technical solutions and scope of the present invention as defined in the appended claims.

Claims

1. A tuyere coke cutting method based on improved DeepLabv3+ is characterized in that:

constructing a tuyere coke data set through a coke image of a blast furnace tuyere raceway with coke labels;

training a tuyere coke cutting model by using a tuyere coke data set;

and segmenting the coke in the tuyere coke image to be segmented by using the trained tuyere coke segmentation model.

2. A method for tuyere coke partitioning based on deplab v3+ as claimed in claim 1, wherein: the method comprises the following steps:

step 1: acquiring coke video data of a blast furnace tuyere raceway;

3. A method for tuyere coke cutting based on deplab v3+ improvement as claimed in claim 2, wherein: the specific method of the step 1 comprises the following steps:

a CCD industrial camera is erected at the small mirror of the blast furnace tuyere to acquire data in real time, so that the coke video data of the blast furnace tuyere raceway is obtained.

4. A method for tuyere coke cutting based on deplab v3+ improvement as claimed in claim 2, wherein: and 2, labeling the coke particles in the tuyere raceway in the picture data by using a labelme labeling tool.

5. A method for tuyere coke cutting based on deplab v3+ improvement as claimed in claim 2, wherein: the Deeplabv3+ model fused with the Coordinate Attention machine system selects an Aligned Xconcentration network as a trunk feature extraction network of an Encoder module on the basis of an original Deeplabv3+ model, a coding attachment network is added to process a feature map extracted by the trunk feature extraction network, an ASPP (automatic sequence protocol) module in the original Deeplabv3+ model is improved, and the rest structure of the Deeplabv3+ model fused with the Coordinate Attention machine system is consistent with the Deeplabv3+ model.

6. A method for tuyere coke partitioning based on DeepLabv3+ improvement as set forth in claim 5, wherein: the improved ASPP module consists of a 1 × 1 convolutional layer, four parallel hole convolutional and pooling layers.

7. A method for tuyere coke partitioning based on DeepLabv3+ improvement as set forth in claim 6, wherein: the specific method for performing coke segmentation on the tuyere coke segmentation model comprises the following steps:

after the input picture is processed by the Aligned Xconcept network, two characteristic graphs are obtained, wherein one characteristic graph is a shallow characteristic, the other characteristic graph is a deep characteristic, and the two characteristic graphs are input into a coding attachment network for processing and then the subsequent operation of the original Deeplabv3+ model is executed; the shallow feature processed by the CoordinateAttention network is directly transmitted into the Decoder module; inputting deep features processed by the CoordinateAttention network into an improved ASPP module, and extracting semantic information with different scales; and finally, inputting the deep features processed by the improved ASPP module into a Decode module, fusing the deep features and the shallow features by the Decode module to obtain a final feature map, performing up-sampling by adopting bilinear interpolation, and recovering the features extracted by the Encoder module to the same size as the original picture to obtain a segmentation result map of the tuyere coke.