Fully-mechanized face coal-rock interface recognition method based on semantic segmentation model

Abstract

The invention discloses a semantic segmentation model-based fully-mechanized face coal-rock interface recognition method, which comprises the following steps of: inputting a coal-rock image into a trained improved SegNeXt semantic segmentation model, and outputting the classification and the position of the coal-rock image; the step of training the improved SegNeXt semantic segmentation model comprises the steps of collecting coal rock images and classification and positions of the coal rock images, and preprocessing the coal rock images to obtain a coal rock data set; expanding a coal rock data set through a data enhancement method, enhancing coal rock feature expression, and dividing the coal rock data set to obtain a training set and a verification set; a cross-space efficient multi-scale attention module is introduced into an encoder in the SegNeXt semantic segmentation model, and a depth strip convolution in an original multi-scale convolution attention module is replaced with an equal-difference height-increasing separable strip convolution; and adopting a cosine annealing scheduling and data enhancement strategy, training the improved SegNeXt semantic segmentation model by using the data set, and verifying the improved SegNeXt semantic segmentation model by using the verification set.

Description

Technical Field

[0001] The present invention relates to a method for identifying coal-rock interface in a fully mechanized mining face based on a semantic segmentation model, and belongs to the technical field of coal-rock interface identification methods. Background Art

[0002] The development of intelligent coal mines relies on highly reliable and adaptable intelligent mining equipment. As the core equipment in fully mechanized mining faces, the level of intelligence of shearers impacts mining efficiency and safety. During the mining process, shearers frequently switch between coal seams and hard rock, reducing equipment life and increasing maintenance costs. Mixed coal and rock cutting also results in 15% to 20% of gangue mixed with the raw coal, affecting coal quality, reducing its calorific value, and increasing harmful gas emissions.

[0003] Coal-rock identification technology is one of the core technologies for the intelligent development of coal mines. Through coal-rock identification technology, coal mining machines can accurately identify the coal-rock interface and adaptively adjust the cutting height and traction rate of the drum, which can not only reduce the wear of the drum, but also reduce the gangue content of coal and improve the quality of coal.

[0004] Deep learning image processing technology has high-precision feature extraction and automatic learning capabilities. It can adapt to the complex environment of underground coal mines and can accurately identify coal and rock characteristics and distribution in real time, providing technical support for research in this field.

[0005] Although there are significant differences between coal and rock in terms of color, texture, and distribution, the geological conditions underground in coal mines are complex. The large amount of dust and water mist generated during the mining process of coal mining machines, as well as the noise and vibration generated during the operation of the equipment, cause the properties and distribution of coal and rock to change dynamically. This has led to many existing coal and rock identification methods not being widely used in actual working conditions. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to overcome the defects of the existing technology and provide a method for identifying the coal-rock interface in a fully mechanized mining face based on a semantic segmentation model. This method realizes a new type of accurate identification model of the coal-rock interface in a fully mechanized mining face by improving the SegNeXt semantic segmentation model. Compared with other lightweight semantic segmentation models, it achieves very high segmentation accuracy and speed; this coal-rock interface recognition model achieves higher accuracy in coal-rock interface recognition while only increasing the number of parameters by a small amount, achieving a better balance between the segmentation accuracy and speed of the model.

[0007] Preferably, the present invention provides a method for identifying coal-rock interface in a fully mechanized mining face based on a semantic segmentation model, comprising:

[0008] Input the coal and rock images to be identified into the trained improved SegNeXt semantic segmentation model, and use the trained improved SegNeXt semantic segmentation model to predict and output the coal and rock image classification and coal and rock image location;

[0009] Among them, the training completed the improved SegNeXt semantic segmentation model, including:

[0010] S100: Collect coal and rock images, classify coal and rock images, and locate coal and rock images, and pre-process the coal and rock images to obtain a coal and rock dataset;

[0011] S200: Expand the coal and rock dataset through data augmentation methods, enhance the coal and rock feature expression, and divide the coal and rock dataset into training set and validation set;

[0012] S300: Introducing a cross-space efficient multi-scale attention module into the encoder of the SegNeXt semantic segmentation model. The cross-space efficient multi-scale attention module includes a first cross-space efficient multi-scale attention module, a second cross-space efficient multi-scale attention module, and a third cross-space efficient multi-scale attention module. The encoder of the SegNeXt semantic segmentation model includes a stage1 stage, a stage2 stage, a first cross-space efficient multi-scale attention module, a stage3 stage, a second cross-space efficient multi-scale attention module, a stage4 stage, and a third cross-space efficient multi-scale attention module connected in sequence. The depthwise strip convolution in the original multi-scale convolutional attention module in the encoder of the SegNeXt semantic segmentation model is replaced by an arbitrarily increased depthwise separable strip convolution to obtain an improved SegNeXt semantic segmentation model.

[0013] S400: Using cosine annealing scheduling and data augmentation strategy, the improved SegNeXt semantic segmentation model is trained using the dataset and verified using the validation set. If the accuracy of the coal and rock image classification output by the SegNeXt semantic segmentation model is higher than the preset accuracy threshold, the SegNeXt semantic segmentation model is judged to be qualified. Preferably, in step S100, the coal and rock images are preprocessed, including:

[0014] Use cameras to randomly capture coal and rock images in different angles and magnification combinations;

[0015] The height and width of the collected coal rock image are obtained, and the size of the coal rock image is adjusted to a preset size.

[0016] Preferably, in step S200, the coal and rock dataset is expanded by a data enhancement method, the coal and rock feature expression is enhanced, and the coal and rock dataset is divided into a training set and a validation set, including:

[0017] Methods including image color and brightness change, image fogging, image size scaling, rotation, pixel translation, contrast change, perspective transformation, and noise addition are used to expand the coal-rock dataset and enhance the coal-rock interface feature expression. Labelme image annotation software is used to annotate the coal-rock images, and the coal-rock dataset is divided into a training set and a validation set.

[0018] Preferably, in step S300, a cross-space efficient multi-scale attention module is introduced into the encoder of the SegNeXt semantic segmentation model, and the depthwise strip convolution in the original multi-scale convolutional attention module in the encoder of the SegNeXt semantic segmentation model is replaced with equidistantly increased depthwise separable strip convolution to obtain an improved SegNeXt semantic segmentation model, including:

[0019] The depthwise separable strip convolution with kernel sizes of 7, 11, and 21 in the multi-scale convolutional attention module is replaced by arbitrarily increased depthwise separable strip convolution with kernel sizes of 7, 11, and 15 respectively;

[0020] The equidistantly increased convolution kernel is used to gradually increase the receptive field size, extract the local boundary information of the coal rock image, and accumulate and capture the contextual relationship layer by layer.

[0021] Preferably, the first cross-space efficient multi-scale attention module, the second cross-space efficient multi-scale attention module and the third cross-space efficient multi-scale attention module include an input layer Input, a segmentation layer, a one-dimensional horizontal global pool X AvgPool, a one-dimensional vertical global pool Y Avg Pool, a first average pooling layer Avg Pool, a second average pooling layer Avg Pool, a connection layer Concat, a first convolution layer Conv, a second convolution layer Conv, a first activation function Sigmoid, a second activation function Sigmoid, a third activation function Sigmoid, a group normalization layer Group Norm, a first classification layer Softmax, a second classification layer Softmax, a first matrix multiplication Matmal, a second matrix multiplication Matmal and an output layer Output.

[0022] The input layer Input, the segmentation layer, the one-dimensional horizontal global pool X Avg Pool, the connection layer Concat, the second convolution layer Conv, the first activation function Sigmoid, the group normalization layer Group Norm, the second classification layer Softmax, the second matrix multiplication Matmal and the output layer Output are connected in sequence. The segmentation layer is used to divide the feature map into G small sub-feature maps;

[0023] The one-dimensional vertical global pool Y Avg Pool, the connection layer Concat, the second convolution layer Conv, the second activation function Sigmoid, the group normalization layer Group Norm, the second average pooling layer Avg Pool and the first matrix multiplication Matmal are connected in sequence.

[0024] The first convolutional layer Conv, the first classification layer Softmax, and the first matrix multiplication Matmal are connected in sequence. The first convolutional layer Conv, the first average pooling layer Avg Pool, and the second matrix multiplication Matmal are connected in sequence.

[0025] Preferably, step S300 includes:

[0026] The segmentation layer divides the feature map into G small sub-feature maps;

[0027] The input sub-feature map is processed by the first convolutional layer Conv and then input into the first classification layer Softmax and the first average pooling layer Avg Pool. The output result of the first classification layer Softmax is input into the first matrix multiplication Matmal, and the output result of the first average pooling layer Avg Pool is input into the first matrix multiplication Matmal;

[0028] For the input sub-feature map, the one-dimensional horizontal global pool X Avg Pool is used to perform the average pooling operation in the X direction and then input into the connection layer Concat. The one-dimensional vertical global pool Y Avg Pool is used to perform the average pooling operation in the Y direction and then input into the connection layer Concat. The output result of the connection layer Concat is input into the second convolutional layer Conv. Based on the output result of the second convolutional layer Conv, the first activation function Sigmoid and the second activation function Sigmoid are used to activate and generate the channel attention map, and the group normalization layer Group Norm is used for group normalization. The output result of the group normalization layer Group Norm is processed by the second classification layer Softmax and the second matrix multiplication Matmal. The output result of the group normalization layer Group Norm is average pooled by the second average pooling layer Avg Pool and then input into the first matrix multiplication Matmal. The output result of the first matrix multiplication Matmal and the output result of the second matrix multiplication Matmal are fused and input into the third activation function Sigmoid. The output result of the third activation function Sigmoid is input into the output layer Output.

[0029] Preferably, in step S400, the improved SegNeXt semantic segmentation model is trained using the dataset using a cosine annealing schedule and data augmentation strategy, including setting the number of training generations, the initial learning rate, random flipping, random rotation, and color jittering. Preferably, the improved multi-scale convolution attention module includes 5×5 depth convolution, 1×7 depth strip convolution, 7×1 depth strip convolution, 1×11 depth strip convolution, 11×1 depth strip convolution, 1×15 depth strip convolution, 15×1 depth strip convolution, a first cross-space efficient multi-scale attention module, a second cross-space efficient multi-scale attention module, and a third cross-space efficient multi-scale attention module.

[0030] The 5×5 depth convolution is connected to the first branch 1×7 depth strip convolution, the second branch 1×11 depth strip convolution and the third branch 1×15 depth strip convolution. The first branch 1×7 depth strip convolution is connected to the 7×1 depth strip convolution; the second branch 1×11 depth strip convolution is connected to the second 11×1 depth strip convolution, the third branch 1×15 depth strip convolution is connected to the 15×1 depth strip convolution, the first branch 7×1 depth strip convolution is connected to the first cross-space efficient multi-scale attention module, the second 11×1 depth strip convolution is connected to the second cross-space efficient multi-scale attention module, the 15×1 depth strip convolution is connected to the third cross-space efficient multi-scale attention module, the first cross-space efficient multi-scale attention module, the second cross-space efficient multi-scale attention module and the third cross-space efficient multi-scale attention module are connected with 1×1 convolution.

[0031] Preferably, the present invention provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any one of the methods described in the first aspect when executing the program.

[0032] Preferably, the present invention provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of any one of the methods described in the first aspect when executed by a processor.

[0033] The beneficial effects achieved by the present invention are:

[0034] This invention improves the SegNeXt semantic segmentation model architecture and replaces the original large convolution kernel with an equidistantly increased depth-separable strip convolution. This effectively solves the problem of coal rock detail loss caused by the boundary smoothing effect in traditional methods. It also combines a cross-space efficient multi-scale attention module to enhance cross-space feature fusion capabilities and improve pixel-level classification accuracy in complex texture areas. After several rounds of transfer learning optimization, this invention uses a data enhancement strategy that combines physical simulation with geometric transformations such as image resizing and rotation by changing brightness, adding fog, and adding noise to simulate high-dust, low-light environments underground. This allows the model to achieve significantly improved segmentation accuracy compared to traditional methods while only slightly increasing the number of parameters. This provides a guarantee for the real-time cutting needs of coal mining machines, effectively reduces the waste rock mixed mining rate and the annual maintenance cost of equipment, and provides a high-precision, low-latency visual perception solution for intelligent coal mining. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solution of the present application, the following is a brief introduction to the drawings required for use in the embodiments. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0036] Figure 1 A flowchart provided for an embodiment of the present invention;

[0037] Figure 2 A structural diagram of a cross-space efficient multi-scale attention module provided by an embodiment of the present invention;

[0038] Figure 3 A structural diagram of the improved multi-scale convolutional attention module provided by an embodiment of the present invention;

[0039] Figure 4 This is a structural diagram of the improved SegNeXt semantic segmentation model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0040] See also Figure 1 The present application discloses a method for identifying coal-rock interface in a fully mechanized mining face based on a semantic segmentation model, which is characterized by comprising:

[0041] Input the coal and rock images into the trained improved SegNeXt semantic segmentation model, and use the trained improved SegNeXt semantic segmentation model to predict the output coal and rock image classification and location;

[0042] Among them, the training completed the improved SegNeXt semantic segmentation model, including:

[0043] S100: Collect coal and rock images, classify and locate coal and rock images in the working face and similar simulation experiments, pre-process the coal and rock images, and obtain a coal and rock dataset;

[0044] S200: Expand the coal and rock dataset through data augmentation methods, enhance the coal and rock feature expression, and divide the coal and rock dataset into training and validation sets according to an 8:1:1 ratio.

[0045] S300: Introducing a cross-space efficient multi-scale attention module in the encoder of the SegNeXt semantic segmentation model, and using arbitrarily increased depthwise separable strip convolution to replace the depthwise strip convolution in the original multi-scale convolutional attention module in the encoder of the SegNeXt semantic segmentation model. Specifically, the depthwise strip convolution with kernel sizes of 7, 11, and 21 in the multi-scale convolutional attention module is replaced by arbitrarily increased depthwise separable strip convolution with kernel sizes of 7, 11, and 15, respectively, to obtain the improved SegNeXt semantic segmentation model;

[0046] S400: Using cosine annealing scheduling and data enhancement strategy, the improved SegNeXt semantic segmentation model is trained using the data set, and the improved SegNeXt semantic segmentation model is verified using the validation set. If the accuracy of the coal and rock image classification predicted by the SegNeXt semantic segmentation model is higher than the preset accuracy threshold, the SegNeXt semantic segmentation model is judged to be qualified.

[0047] In step S100, the coal rock image is preprocessed, including:

[0048] Use high-resolution Hikvision industrial cameras to randomly capture coal and rock images at different angles and magnifications.

[0049] The height and width of the collected coal and rock images are obtained, and the size of the coal and rock images is adjusted to a preset size: 512 pixels × 512 pixels.

[0050] In step S200, the coal rock dataset is expanded by a data enhancement method, the coal rock feature expression is enhanced, and the coal rock dataset is divided into a training set and a validation set, including:

[0051] Methods including image color and brightness change, image fogging, image size scaling, rotation, pixel translation, contrast change, perspective transformation, and noise addition are used to expand the coal-rock dataset and enhance the coal-rock interface feature expression. The coal-rock images are annotated using Labelme image annotation software, and the coal-rock dataset is divided into training and validation sets according to a preset ratio of 8:1:1.

[0052] In step S300, a cross-space efficient multi-scale attention module is introduced into the encoder of the SegNeXt semantic segmentation model, and the depthwise strip convolution in the original multi-scale convolutional attention module in the encoder of the SegNeXt semantic segmentation model is replaced with equidistantly increased depthwise separable strip convolution to obtain an improved SegNeXt semantic segmentation model, including:

[0053] Large convolution kernels will cause the features of the boundary area to be over-smoothed, weakening the ability to capture detailed information;

[0054] The depthwise separable strip convolution with kernel sizes of 7, 11, and 15 is used to replace the depthwise strip convolution with kernel sizes of 7, 11, and 21 in the original multi-scale convolutional attention module;

[0055] The equidistantly increased convolution kernel is used to gradually increase the receptive field size, extract the local boundary information of the coal rock image, accumulate and capture the contextual relationship layer by layer, and more flexibly combine information of different scales.

[0056] Step S300 includes:

[0057] The cross-space efficient multi-scale attention module can efficiently learn features across multiple scales and locations, extracting semantic information more accurately at different scales and positions, and effectively improving the SegNeXt semantic segmentation model's ability to extract complex features.

[0058] The encoder in the SegNeXt semantic segmentation model includes a stage1 stage, a stage2 stage, a first cross-space efficient multi-scale attention module, a stage3 stage, a second cross-space efficient multi-scale attention module, a stage4 stage and a third cross-space efficient multi-scale attention module connected in sequence. The stage1 stage in the encoder in the SegNeXt semantic segmentation model is used to extract low-level features of coal rock texture and shape, the stage2 stage in the encoder in the SegNeXt semantic segmentation model is used to extract mid-level features of coal rock interface contour and distribution information, the stage3 stage in the encoder in the SegNeXt semantic segmentation model is used to extract large-area distribution of coal rock interface, and the features of the stage4 stage in the encoder in the SegNeXt semantic segmentation model are used to extract high-level semantic features of global context information;

[0059] Use the features of stage2, stage3 and stage4 for splicing;

[0060] After the stage 2, stage 3 and stage 4 of encoder feature extraction, efficient cross-space multi-scale attention modules are added to perform spatial and channel feature selection, realizing multi-scale cross-space extraction and fusion of coal rock images from low-level features to high-level semantic features.

[0061] In step S400, the cosine annealing scheduling and data enhancement strategy are adopted to train the improved SegNeXt semantic segmentation model using the dataset, including setting the training generation, initial learning rate, random flipping, random rotation, and color jittering.

[0062] like Figure 3 As shown in the figure, the MSCA module (multi-scale convolutional attention module) before the improvement includes 5×5 depth convolution, 1×7 depth strip convolution, 7×1 depth strip convolution, 1×11 depth strip convolution, 11×1 depth strip convolution, 1×21 depth strip convolution, 21×1 depth strip convolution and 1×1 convolution.

[0063] The 5×5 depth convolution is connected to the first branch 1×7 depth strip convolution and 7×1 depth strip convolution; the 5×5 depth convolution is connected to the second branch 1×11 depth strip convolution and 11×1 depth strip convolution; the 5×5 depth convolution is connected to the third branch 1×21 depth strip convolution and 21×1 depth strip convolution; finally, the three branches are connected with 1×1 convolution.

[0064] The improved multi-scale convolution attention module includes 5×5 depth convolution, 1×7 depth strip convolution, 7×1 depth strip convolution, 1×11 depth strip convolution, 11×1 depth strip convolution, 1×15 depth strip convolution, 15×1 depth strip convolution, the first cross-space efficient multi-scale attention module, the second cross-space efficient multi-scale attention module and the third cross-space efficient multi-scale attention module.

[0065] The 5×5 depth convolution is connected to the first branch 1×7 depth strip convolution, the second branch 1×11 depth strip convolution and the third branch 1×15 depth strip convolution. The first branch 1×7 depth strip convolution is connected to the 7×1 depth strip convolution; the second branch 1×11 depth strip convolution is connected to the second 11×1 depth strip convolution, the third branch 1×15 depth strip convolution is connected to the 15×1 depth strip convolution, the first branch 7×1 depth strip convolution is connected to the first cross-space efficient multi-scale attention module, the second 11×1 depth strip convolution is connected to the second cross-space efficient multi-scale attention module, the 15×1 depth strip convolution is connected to the third cross-space efficient multi-scale attention module, the first cross-space efficient multi-scale attention module, the second cross-space efficient multi-scale attention module and the third cross-space efficient multi-scale attention module are connected with 1×1 convolution.

[0066] like Figure 4 As shown in the figure, the input of the SegNeXt semantic segmentation model is a coal rock map of size 512×512. The encoder in the SegNeXt semantic segmentation model includes the stage1 stage, stage2 stage, first EMA module, stage3 stage, second EMA module, stage4 stage and third EMA module connected in sequence; the decoder in the SegNeXt semantic segmentation model includes the connection layer Concat, the head layer Head and the multi-layer perceptron layer MLP connected in sequence.

[0067] In an embodiment of the present application, the present invention provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any of the above methods when executing the program.

[0068] In an embodiment of the present application, the present invention provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of any of the above methods when executed by a processor.

[0069] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments.

[0070] Those skilled in the art will readily appreciate other embodiments of the present invention after considering the specification and practicing the invention as disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not invented herein, and the description and examples are to be considered merely as exemplary.

[0071] The above specific implementation methods further illustrate the purpose, technical solutions and beneficial effects of this application in detail. It should be understood that the above are only specific implementation methods of this application and are not intended to limit the scope of protection of this application. Any modifications, equivalent replacements, improvements, etc. made on the basis of the technical solutions of this application should be included in the scope of protection of this application.

Claims

1. A method for identifying coal-rock interface in fully mechanized mining face based on a semantic segmentation model, characterized by: include: Input the coal and rock images to be identified into the trained improved SegNeXt semantic segmentation model, and use the trained improved SegNeXt semantic segmentation model to predict and output the coal and rock image classification and coal and rock image location; Among them, the training completed the improved SegNeXt semantic segmentation model, including: S100: Collect coal and rock images, classify coal and rock images, and locate coal and rock images, and pre-process the coal and rock images to obtain a coal and rock dataset; S200: Expand the coal and rock dataset through data augmentation methods, enhance the coal and rock feature expression, and divide the coal and rock dataset into training set and validation set; S300: Introducing a cross-space efficient multi-scale attention module into the encoder of the SegNeXt semantic segmentation model. The cross-space efficient multi-scale attention module includes a first cross-space efficient multi-scale attention module, a second cross-space efficient multi-scale attention module, and a third cross-space efficient multi-scale attention module. The encoder of the SegNeXt semantic segmentation model includes a stage1 stage, a stage2 stage, a first cross-space efficient multi-scale attention module, a stage3 stage, a second cross-space efficient multi-scale attention module, a stage4 stage, and a third cross-space efficient multi-scale attention module connected in sequence. The depthwise strip convolution in the multi-scale convolutional attention module in the encoder of the SegNeXt semantic segmentation model is replaced by an arbitrarily increased depthwise separable strip convolution to obtain an improved multi-scale convolutional attention module and an improved SegNeXt semantic segmentation model. S400: Using cosine annealing scheduling and data enhancement strategy, the improved SegNeXt semantic segmentation model is trained using the data set, and the improved SegNeXt semantic segmentation model is verified using the validation set. If the accuracy of the coal and rock image classification predicted by the SegNeXt semantic segmentation model is higher than the preset accuracy threshold, the SegNeXt semantic segmentation model is judged to be qualified.

2. The method for identifying coal-rock interface in a fully mechanized mining face based on a semantic segmentation model according to claim 1, characterized in that: In step S100, the coal rock image is preprocessed, including: Use cameras to randomly capture coal and rock images in different angles and magnification combinations; The height and width of the collected coal rock image are obtained, and the size of the coal rock image is adjusted to a preset size.

3. The method for identifying coal-rock interface in a fully mechanized mining face based on a semantic segmentation model according to claim 1, characterized in that: In step S200, the coal rock dataset is expanded by a data enhancement method, the coal rock feature expression is enhanced, and the coal rock dataset is divided into a training set and a validation set, including: Methods including image color and brightness change, image fogging, image size scaling, rotation, pixel translation, contrast change, perspective transformation, and noise addition are used to expand the coal-rock dataset and enhance the coal-rock interface feature expression. Labelme image annotation software is used to annotate the coal-rock images, and the coal-rock dataset is divided into a training set and a validation set.

4. The method for identifying coal-rock interface in a fully mechanized mining face based on a semantic segmentation model according to claim 1, characterized in that: In step S300, a cross-space efficient multi-scale attention module is introduced into the encoder of the SegNeXt semantic segmentation model, and the depthwise strip convolution in the multi-scale convolutional attention module in the encoder of the SegNeXt semantic segmentation model is replaced with arithmetically increased depthwise separable strip convolution to obtain an improved SegNeXt semantic segmentation model, including: The depthwise separable strip convolution with kernel sizes of 7, 11, and 21 in the multi-scale convolutional attention module is replaced by arbitrarily increased depthwise separable strip convolution with kernel sizes of 7, 11, and 15 respectively; The equidistantly increased convolution kernel is used to gradually increase the receptive field size, extract the local boundary information of the coal rock image, and accumulate and capture the contextual relationship layer by layer.

5. The method for identifying coal-rock interface in a fully mechanized mining face based on a semantic segmentation model according to claim 1, characterized in that: The first cross-space efficient multi-scale attention module, the second cross-space efficient multi-scale attention module and the third cross-space efficient multi-scale attention module include input layer Input, segmentation layer, one-dimensional horizontal global pool X Avg Pool, one-dimensional vertical global pool Y Avg Pool, first average pooling layer Avg Pool, second average pooling layer Avg Pool, connection layer Concat, first convolution layer Conv, second convolution layer Conv, first activation function Sigmoid, second activation function Sigmoid, third activation function Sigmoid, group normalization layer Group Norm, first classification layer Softmax, second classification layer Softmax, first matrix multiplication Matmal, second matrix multiplication Matmal and output layer Output. The input layer Input, the segmentation layer, the one-dimensional horizontal global pool X Avg Pool, the connection layer Concat, the second convolution layer Conv, the first activation function Sigmoid, the group normalization layer Group Norm, the second classification layer Softmax, the second matrix multiplication Matmal and the output layer Output are connected in sequence. The segmentation layer is used to divide the feature map into G small sub-feature maps; The one-dimensional vertical global pool Y Avg Pool, the connection layer Concat, the second convolution layer Conv, the second activation function Sigmoid, the group normalization layer Group Norm, the second average pooling layer Avg Pool and the first matrix multiplication Matmal are connected in sequence. The first convolutional layer Conv, the first classification layer Softmax, and the first matrix multiplication Matmal are connected in sequence. The first convolutional layer Conv, the first average pooling layer Avg Pool, and the second matrix multiplication Matmal are connected in sequence.

6. The method for identifying coal-rock interface in a fully mechanized mining face based on a semantic segmentation model according to claim 5, characterized in that: Step S300 includes: The segmentation layer divides the feature map into G small sub-feature maps; The input sub-feature map is processed by the first convolutional layer Conv and then input into the first classification layer Softmax and the first average pooling layer Avg Pool. The output result of the first classification layer Softmax is input into the first matrix multiplication Matmal, and the output result of the first average pooling layer Avg Pool is input into the first matrix multiplication Matmal; For the input sub-feature map, the one-dimensional horizontal global pool X Avg Pool is used to perform the average pooling operation in the X direction and then input into the connection layer Concat. The one-dimensional vertical global pool Y Avg Pool is used to perform the average pooling operation in the Y direction and then input into the connection layer Concat. The output result of the connection layer Concat is input into the second convolutional layer Conv. Based on the output result of the second convolutional layer Conv, the first activation function Sigmoid and the second activation function Sigmoid are used to activate and generate the channel attention map, and the group normalization layer Group Norm is used for group normalization. The output result of the group normalization layer Group Norm is processed by the second classification layer Softmax and the second matrix multiplication Matmal. The output result of the group normalization layer Group Norm is average pooled by the second average pooling layer Avg Pool and then input into the first matrix multiplication Matmal. The output result of the first matrix multiplication Matmal and the output result of the second matrix multiplication Matmal are fused and input into the third activation function Sigmoid. The output result of the third activation function Sigmoid is input into the output layer Output.

7. The method for identifying coal-rock interface in a fully mechanized mining face based on a semantic segmentation model according to claim 1, characterized in that: The improved multi-scale convolution attention module includes 5×5 depth convolution, 1×7 depth strip convolution, 7×1 depth strip convolution, 1×11 depth strip convolution, 11×1 depth strip convolution, 1×15 depth strip convolution, 15×1 depth strip convolution, the first cross-space efficient multi-scale attention module, the second cross-space efficient multi-scale attention module and the third cross-space efficient multi-scale attention module. The 5×5 depth convolution is connected to the first branch 1×7 depth strip convolution, the second branch 1×11 depth strip convolution and the third branch 1×15 depth strip convolution. The first branch 1×7 depth strip convolution is connected to the 7×1 depth strip convolution; the second branch 1×11 depth strip convolution is connected to the second 11×1 depth strip convolution, the third branch 1×15 depth strip convolution is connected to the 15×1 depth strip convolution, the first branch 7×1 depth strip convolution is connected to the first cross-space efficient multi-scale attention module, the second 11×1 depth strip convolution is connected to the second cross-space efficient multi-scale attention module, the 15×1 depth strip convolution is connected to the third cross-space efficient multi-scale attention module, the first cross-space efficient multi-scale attention module, the second cross-space efficient multi-scale attention module and the third cross-space efficient multi-scale attention module are connected with 1×1 convolution.

8. The method for identifying coal-rock interface in a fully mechanized mining face based on a semantic segmentation model according to claim 1, characterized in that: In step S400, the improved SegNeXt semantic segmentation model is trained using the dataset using cosine annealing scheduling and data augmentation strategy, including setting training generations, initial learning rate, random flipping, random rotation and color jittering.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the method according to any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images