A coal and rock identification method for a coal mining face based on deep learning

Abstract

This invention provides a deep learning-based method for coal and rock identification in coal mine faces, aiming to solve the problems of low accuracy and slow processing efficiency in traditional methods. The method first acquires original coal and rock images of the coal mine face using a high-resolution camera and applies image enhancement techniques based on RSTF and DA to optimize image quality. Then, a deep learning model based on YOLOv5s and the lightweight semantic segmentation network BiSeNet is used for coal and rock detection and semantic segmentation, effectively extracting the location and boundary information of the coal and rock. Finally, the generalization ability and recognition accuracy of the model are evaluated using a test set, verifying the effectiveness and reliability of the method in a real coal mine environment. The advantages of this invention are that it achieves efficient and accurate coal and rock identification in coal mine faces through deep learning technology, enabling rapid and accurate completion of coal and rock identification tasks in complex coal mine environments, thereby improving mine production efficiency and safety.

Description

Technical Field

[0001] This invention relates to the field of coal mining, and in particular to a method for coal and rock identification in coal mining faces based on deep learning. Background Technology

[0002] In coal mining, accurately identifying the distribution of coal and rock on the face is a key technology for improving mining efficiency and ensuring safe production. Traditional coal and rock identification methods suffer from low efficiency and accuracy. With the development of deep learning technology, deep learning-based image recognition methods have made significant progress in many fields, providing new solutions for coal and rock identification.

[0003] Currently, common deep learning image recognition algorithms include Mask R-CNN, SegNet, and U-Net. Mask R-CNN can perform object detection and instance segmentation simultaneously, but its model complexity is high, and it consumes a lot of computational resources, making it unsuitable for deployment on low-configuration devices. SegNet's main advantages are its high segmentation accuracy and good edge detection capabilities, but its computational complexity and memory consumption are relatively high. U-Net performs well in semantic segmentation tasks, with high segmentation accuracy and fast inference speed, but it may have certain limitations when processing complex images.

[0004] Therefore, the problem of existing technologies failing to simultaneously address recognition efficiency, accuracy, and computational resource requirements needs to be solved. A balance needs to be struck between accuracy, computational complexity, and inference speed to meet the needs of coal and rock identification in practical applications. Summary of the Invention

[0005] To improve the accuracy and efficiency of coal and rock identification, this invention proposes a deep learning-based method for coal and rock identification in coal mine faces.

[0006] The above-mentioned technical problems of the present invention are mainly solved by the following technical solutions:

[0007] In a first aspect, the present invention provides a method for coal and rock identification in coal mine faces based on deep learning, comprising the following steps:

[0008] Obtain the original coal and rock image, perform preprocessing, and obtain the preprocessed image;

[0009] The preprocessed image is enhanced using the RSTF method combined with the DA optimization algorithm to obtain the enhanced image.

[0010] A deep learning model is constructed using the YOLOv5s network as the base model and adding a segmentation head by combining the BiSeNet segmentation network.

[0011] Deep learning model training;

[0012] The final deep learning model was obtained and used for coal and rock identification in the mine face.

[0013] One possible implementation method includes: labeling the acquired coal and rock images using the Labelme open-source annotation tool.

[0014] One possible implementation method for image enhancement based on the RSTF method and combined with the DA optimization algorithm includes: labeling the acquired coal and rock images using the Labelme open-source annotation tool.

[0015] Initial image analysis: Analyze the input preprocessed image, calculate the maximum value and standard deviation of its pixel values, and use similarity calculation methods to find reference images with similar features to the input image to obtain the analysis results;

[0016] Parameter optimization: Based on the analysis results, evaluate the current state of the input image and the required enhancement direction and magnitude, and determine the image enhancement transform coefficients of RSTF;

[0017] DA optimization algorithm: The initial image enhancement transform coefficients are set as the initial values ​​of the DA group optimization algorithm, and the parameters are iteratively optimized to obtain the best image enhancement effect;

[0018] Apply optimized parameters: Use the optimal parameters determined by the DA optimization algorithm as the image enhancement transform coefficients.

[0019] Furthermore, the transformation function of the RSTF method is:

[0020]

[0021] Where f(i,j) and g(i,j) are the pixel values ​​of the images before and after transformation at point (i,j), respectively, S(i,j) is the pixel value of the similar image at point (i,j), Max, denoted as the maximum pixel value and standard deviation; a, b, c, and k are the image enhancement transform coefficients of RSTF.

[0022] One possible implementation is that the deep learning network is designed based on YOLOv5s, wherein the input layer is reduced from the original RGB three-channel to grayscale single-channel for receiving enhanced images; the backbone and neck networks are inherited from YOLOv5s and are used to extract image features, including a series of convolutional layers, pooling layers and residual modules, to gradually extract high-level features.

[0023] Furthermore, the segmentation head is constructed using the BiSeNet segmentation network to perform semantic segmentation on the output of the neck part in YOLOv5s, including an RFB+ layer, a stacking layer, a feature fusion module, a random deactivation layer, a convolutional layer, and an upsampling layer.

[0024] Two RFB+ layers are stacked in parallel, and then connected sequentially to a feature fusion module, a random deactivation layer, a convolutional layer, and an upsampling layer to output the segmentation result.

[0025] Secondly, the present invention provides a coal and rock identification device for coal mine faces based on deep learning, comprising:

[0026] The image acquisition and preprocessing module is used to acquire raw coal and rock images, perform preprocessing, and obtain preprocessed images.

[0027] The image enhancement module is used to enhance the preprocessed image based on the RSTF method and combined with the DA optimization algorithm to obtain the enhanced image;

[0028] The model building module is used to build a deep learning model that uses the YOLOv5s network as the base model and adds a segmentation head by combining the BiSeNet segmentation network.

[0029] The model training module is used for training deep learning models.

[0030] The final model acquisition module is used to obtain the final deep learning model.

[0031] Thirdly, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the deep learning-based coal and rock identification method for coal mine faces as described above.

[0032] Fourthly, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the deep learning-based coal and rock identification method for coal mine faces as described above.

[0033] Fifthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the deep learning-based coal and rock identification method for coal mine faces as described above.

[0034] The beneficial effects of this invention are as follows:

[0035] 1. This invention employs an image enhancement method based on region similarity transformation and determines enhancement parameters through a dragonfly swarm optimization algorithm, resulting in more prominent enhanced image features and improved model recognition capabilities. This process enhances the model's adaptability to different lighting, noise, and environmental changes, thereby improving its robustness.

[0036] 2. This invention converts the three-channel color image of the input layer into a single-channel grayscale image. This conversion significantly reduces computation while retaining sufficient image information for coal and rock identification. Compared to processing color images, processing grayscale images not only reduces the number of model parameters and computational complexity but also accelerates model training and inference, improving overall efficiency.

[0037] 3. This invention uses YOLOv5s as the basic detection model, balancing accuracy and efficiency. YOLOv5s performs excellently in target detection tasks, and its efficient structural design enables the model to achieve fast inference while maintaining high accuracy. This is particularly important for coal and rock identification tasks that require real-time processing.

[0038] 4. This invention introduces the BiSeNet network structure based on YOLOv5s, adding a new segmentation head to achieve efficient semantic segmentation. BiSeNet, through its unique spatial and contextual path design, can capture rich semantic information while maintaining high resolution. Its lightweight design ensures efficient semantic segmentation without adding excessive computational burden.

[0039] 5. The improved model described in this invention fully considers practical application scenarios, especially the computing power of low-configuration devices. By inputting grayscale images and employing a lightweight segmentation network, the model ensures efficient operation on resource-constrained devices (low-configuration devices), achieving real-time coal and rock identification in the field. This makes the method widely applicable and practically valuable in actual coal mine face environments. Attached Figure Description

[0040] Figure 1 This is a flowchart of the coal and rock image enhancement process using RSTF and DA according to the present invention.

[0041] Figure 2 This is a structural diagram of the coal and rock deep learning model based on YOLOv5s and the lightweight semantic segmentation network BiSeNet, which is the basis of this invention.

[0042] Figure 3 This is a diagram showing the results of identifying actual coal mine faces using the present invention.

[0043] Figure 4 This is a diagram showing the coal and rock identification results of this invention and several existing advanced semantic segmentation models.

[0044] Figure 5 This figure shows the comparison results of the real-time inference performance between the present invention and several existing advanced semantic segmentation models.

[0045] Figure 6 This is a schematic diagram of the structure of the deep learning-based coal and rock identification device for coal mine faces provided by the present invention.

[0046] Figure 7 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0047] The technical solution of the present invention will be further described in detail below through embodiments and in conjunction with the accompanying drawings.

[0048] A deep learning-based method for coal and rock identification in coal mine faces includes the following steps:

[0049] Step 100: Obtain the original coal and rock image, perform preprocessing, and obtain the preprocessed image.

[0050] As a specific embodiment, the method for acquiring raw coal and rock images is as follows: Raw coal and rock images of the coal mine face are acquired using a high-resolution camera or other image sensing equipment. During acquisition, the optimal acquisition position and angle are determined based on the specific conditions of the coal mine face to ensure coverage of the entire face area and reduce blind spots. Lighting conditions and reflection are also considered to avoid strong light and shadows affecting image quality.

[0051] As a specific implementation, preprocessing includes: labeling the acquired coal and rock images using the Labelme open-source annotation tool. During annotation, semantic segmentation and annotation are performed separately for coal and rock. The annotation file records the location and category information of each object (coal is labeled "Coal", and rock is labeled "Rock").

[0052] Step 200: Enhance the preprocessed image using the RSTF method combined with the DA optimization algorithm to obtain the enhanced image.

[0053] As a specific embodiment, a method for image enhancement based on the RSTF method and combined with the DA optimization algorithm for preprocessed images includes:

[0054] Step 210: First, the input preprocessed image is analyzed, calculating its maximum pixel value and standard deviation. Furthermore, a similarity calculation method is used to find reference images with features similar to the input image. Similar images are calculated using the following formula:

[0055]

[0056] Where S(i,j) is the pixel value of the similar image at point (i,j), S klThe similarity score is calculated using the following formula:

[0057]

[0058] Where, d kl The distance between two pixels is denoted as , and D is the normalization coefficient.

[0059] Step 220: Based on the initial image analysis results from Step 210, including the standard deviation and maximum value of pixel values, and the characteristics of the reference image used for similarity calculation, evaluate the current state of the input preprocessed image and the direction and magnitude of the required enhancement. Next, the image enhancement transformation coefficients a, b, c, and k of the RSTF need to be determined. Based on the initial calculation results, to ensure the optimization process starts from a reasonable starting point, the initial image enhancement transformation coefficients are set to [1.5, 0.5, 1, 1.5].

[0060] The transformation function of the RSTF method is:

[0061]

[0062] Where f(i,j) and g(i,j) are the pixel values ​​of the images before and after transformation at point (i,j), respectively, S(i,j) is the pixel value of the similar image at point (i,j), Max, denoted as the maximum pixel value and standard deviation; a, b, c, and k are the image enhancement transform coefficients of RSTF.

[0063] Step 230: Set the initial transformation coefficients to the initial values ​​for the DA swarm optimization algorithm. Then, through iterative calculation, update the position and velocity of individual dragonflies each time to optimize these parameters. The DA optimization algorithm performs a global optimization search by simulating five key behaviors of the dragonfly swarm. These key behaviors include separation (S), alignment (A), aggregation (C), foraging (F), and predator avoidance (E), and the update calculation method for these behaviors is as follows:

[0064]

[0065] F i =X + -X

[0066] E i =X - -X

[0067] Where X is the current position of the dragonfly, X j It is the position of the j-th neighboring dragonfly, N is the number of neighboring dragonflies, and V is the position of the j-th neighboring dragonfly. j X is the speed of the j-th neighboring dragonfly. + It is the location of the food source, X - This is the enemy's location, S iLet A be the separation behavior component of the i-th dragonfly individual. i For the alignment behavior components of the i-th dragonfly individual, C i Let F be the aggregated behavioral component for the i-th dragonfly individual. i Let E be the foraging behavior component of the i-th dragonfly individual. i Let X be the enemy avoidance behavior component of the i-th dragonfly individual. To effectively update the dragonfly's position, two key vectors ΔX and X' are needed, and the update calculation method is as follows:

[0068] ΔX t+1 =(sS) i +aA i +cC i +fF i +eE i )+wΔX t

[0069] X t+1 =X t +ΔX t+1

[0070] Where s, a, c, f, and e are the weights for separation, alignment, aggregation, foraging, and enemy avoidance, respectively; w is the inertia weight; and ΔX... t Let X be the change in the dragonfly's displacement after the t-th iteration. t Let be the displacement of the dragonfly after the t-th iteration.

[0071] The following is a detailed execution process:

[0072] First, calculate the object value of all dragonfly individuals based on the initial a, b, c, and k, and randomly initialize the positions ΔX and X for each dragonfly individual;

[0073] Then, update the food source location based on the best value of the current dragonfly individual, and update the enemy location based on the worst value of the current dragonfly individual;

[0074] Next, update the weight coefficients s, a, c, f, e, and w, calculate the behavior vectors S, A, C, F, and E, adjust the proximity radius of the individual dragonfly to determine its neighboring individuals, and update its velocity and position if an individual dragonfly has at least one neighboring dragonfly.

[0075] Finally, check whether the current optimization results meet the preset criteria. If they do not meet the criteria, repeat the above steps.

[0076] The standard uses the fitness function F(x) to evaluate the image enhancement effect. It quantifies the quality of image enhancement by combining three performance indicators, including entropy value, number of edge pixels and Sobel value of the image.

[0077] The calculation method is as follows:

[0078]

[0079] Where: I(x) is the grayscale image to be evaluated, H(I(x)) is the image entropy value, and n edge (I(x)) is the number of pixels at the edge of the image, E(I(x)) is the Sobel value of the image, M and N are the width and height of the image, and x is the pixel gray value.

[0080] Step 2.4: The optimal parameters determined by the DA optimization algorithm are applied as image enhancement transform coefficients to the original coal and rock image. After enhancement processing, a coal and rock image with high clarity, moderate contrast, and prominent features is generated, providing high-quality input data for the subsequent training and prediction of the deep learning model.

[0081] In step 200, an RSTF-based method is applied to enhance the preprocessed coal and rock image. The DA optimization algorithm is used to optimize the image enhancement parameters to achieve the best enhancement effect, thereby improving the image's feature representation capability.

[0082] Step 300, Deep learning model construction: Using the YOLOv5s network as the base model, a new segmentation head is added based on the BiSeNet segmentation network, enabling the model to perform coal and rock detection and semantic segmentation simultaneously.

[0083] As a specific embodiment, the deep learning model is designed based on YOLOv5s. The input layer is reduced from the original RGB three-channel to grayscale single-channel to receive the coal and rock image after step 200 enhancement. The backbone and neck networks are inherited from YOLOv5s and are used to extract image features, including a series of convolutional layers, pooling layers and residual modules, to gradually extract high-level features.

[0084] Specifically, the deep learning model structure design is as follows:

[0085] Focus layer → First convolutional layer → First C3 layer → Second convolutional layer → Second C3 layer → Third convolutional layer → Third C3 layer → Fourth convolutional layer → SPP layer → First C3_False layer → Fifth convolutional layer → First upsampling layer → First fusion layer → Second C3_False layer → Sixth convolutional layer → Second upsampling layer → Second fusion layer → Third C3_False layer → Seventh convolutional layer → Third fusion layer → Fourth C3_False layer → Eighth convolutional layer → Fourth fusion layer → Fifth C3_False layer; The third, fourth, and fifth C3_False layers are connected to the detection head.

[0086] Building upon existing detection heads, this invention adds a segmentation head to the backend of the neck network. Specifically, the third fusion layer and the fifth C3_False layer are connected to the segmentation head to achieve efficient coal and rock detection and semantic segmentation. The segmentation head design incorporates the advantages of the BiSeNet segmentation network, including spatial path, context path, and feature fusion modules, to comprehensively extract spatial and semantic information from the image and generate accurate segmentation results. The segmentation head combines an RFB+ layer and a feature fusion module (FFM, derived from BiSeNet), as well as convolutional layers, upsampling layers, and random deactivation layers to efficiently achieve the semantic segmentation task of coal and rock. The specific structural design is as follows:

[0087] First RFB+ layer → First convolutional layer → First upsampling layer → Stacking layer → FFM layer → Random deactivation layer → Second convolutional layer → Second upsampling layer; First RFB+ layer is connected to the third fusion layer;

[0088] The second branch, connected to the stacking layer, is the second RFB+ layer, which in turn connects to the fifth C3_False layer. The RFB+ layer is a further improvement on the RFB layer, employing a cascaded convolutional layer structure to enhance feature representation and image detail capture. Its structure is: first convolutional layer → second convolutional layer → fusion layer. The other branch connected to the fusion layer consists of three parallel convolutional layers.

[0089] Specifically, the PyTorch deep learning framework is used to build deep learning models.

[0090] Step 400, Model Training: Input the enhanced coal and rock image from Step 2 into the improved YOLOv5s model for training.

[0091] As a specific implementation, the enhanced coal and rock images are divided into training, validation, and test sets in a 6:2:2 ratio for model training and evaluation.

[0092] The specific training process is as follows: PyTorch is used to load the pre-defined training and validation sets and perform batch processing. Combining detection and segmentation tasks, a loss function (IoU) is defined with a learning rate of 0.001 and a training epoch of 200. Batch gradient descent is used to train the model, with a batch size of 32. After each training epoch, the model's performance is evaluated on the validation set, including detection accuracy, segmentation accuracy, and computation speed. Based on the validation results, model parameters and training strategies are adjusted to ensure gradual convergence and improved performance. The model is then saved.

[0093] Deep learning model testing:

[0094] Test the model on the test set to verify its performance in real-world applications and ensure that the model has good generalization ability. Figure 3 The model described in this embodiment is demonstrated to perform well in guessing actual surfaces, and the model has good recognition and segmentation performance.

[0095] To comprehensively evaluate the advantages of the model in this invention, its performance in real-time inference is compared with that of several existing advanced semantic segmentation models. The selected comparison models include Mask R-CNN, SegNet, and U-Net, and the evaluation metric is the mean accuracy of prediction (mAP). Figure 3 This is a diagram showing the results of identification at an actual coal mine face. Figure 4 The performance comparison of several models is shown, with the first row displaying different input images, and rows 2-4 showing the recognition results of U-Net, SegNet, Mask R-CNN, and this embodiment, respectively. Figure 4 As can be seen, the model described in this embodiment performs excellently on segmentation tasks. While U-Net is effective in segmentation, its encoder-decoder structure contains skip connections, resulting in high computational cost and slow speed during inference, especially on high-resolution images. SegNet is a lightweight segmentation network with fewer parameters and shorter inference time, but its accuracy is lower. Mask R-CNN demonstrates good accuracy with low mAP. 50 and mAP [50,95] The values ​​were 84.6 and 76.3, respectively. However, the model's inference time for a single coal and rock image was relatively long. Mask-RCNN is a two-stage method that first generates region proposals and then performs pixel-by-pixel segmentation, resulting in higher computational complexity and slower inference speed. In contrast, the model described in this embodiment achieves a better balance between prediction accuracy and inference speed. The combination of YOLOv5 and BiSeNet architectures, along with optimized processing techniques, achieves fast inference speed. These design choices ensure that the model maintains high accuracy while significantly reducing computational load and inference time. By integrating the lightweight semantic segmentation network BiSeNet, although mAP 50 It only increased slightly by 6.6%, but the reasoning speed doubled.

[0096] The model described in this embodiment is designed for deployment on low-configuration devices, addressing the problem of limited computing resources in many coal mines. By optimizing the model structure and algorithm, this invention can reduce computing resource requirements while maintaining high-accuracy recognition, thus adapting to the needs of practical application environments.

[0097] The following describes the deep learning-based coal and rock identification device for coal mine faces provided by the present invention. The deep learning-based coal and rock identification device for coal mine faces described below can be referred to in correspondence with the deep learning-based coal and rock identification method for coal mine faces described above.

[0098] Figure 6 This is a schematic diagram of the structure of the coal and rock identification device for coal mine faces provided in an embodiment of the present invention, as shown below. Figure 6 As shown, it includes: an image acquisition and preprocessing module 61, an image enhancement module 62, a model building module 63, a model training module 64, and a final model acquisition module 65, wherein:

[0099] Image acquisition and preprocessing module 61 is used to acquire the original coal and rock image, perform preprocessing, and obtain a preprocessed image;

[0100] Image enhancement module 62 is used to enhance the preprocessed image based on the RSTF method and combined with the DA optimization algorithm to obtain an enhanced image;

[0101] Model building module 63 is used to build a deep learning model with YOLOv5s network as the base model and BiSeNet segmentation network to add a segmentation head.

[0102] Model training module 64 is used for training deep learning models;

[0103] The final model acquisition module 65 is used to obtain the final deep learning model.

[0104] Figure 7 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 7 As shown, the electronic device may include: a processor 710, a communication interface 720, a memory 730, and a communication bus 740, wherein the processor 710, the communication interface 720, and the memory 730 communicate with each other through the communication bus 740. The processor 710 can call logical instructions in the memory 730 to execute a deep learning-based coal face coal and rock identification method, which includes:

[0105] Obtain the original coal and rock image, perform preprocessing, and obtain the preprocessed image;

[0106] The preprocessed image is enhanced using the RSTF method combined with the DA optimization algorithm to obtain the enhanced image.

[0107] A deep learning model is constructed using the YOLOv5s network as the base model and adding a segmentation head by combining the BiSeNet segmentation network.

[0108] Deep learning model training;

[0109] The final deep learning model was obtained and used for coal and rock identification in the mine face.

[0110] Furthermore, the logical instructions in the aforementioned memory 730 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0111] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the deep learning-based coal and rock identification method for coal mining faces provided by the above methods.

[0112] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the deep learning-based coal and rock identification method for coal mining faces provided by the methods described above.

[0113] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0114] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0115] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A deep learning-based method for coal and rock identification in coal mine faces, characterized in that, Includes the following steps: Obtain the original coal and rock image, perform preprocessing, and obtain the preprocessed image; The preprocessed image is enhanced using the RSTF method combined with the DA optimization algorithm to obtain the enhanced image, including: Initial image analysis: Analyze the input preprocessed image, calculate the maximum value and standard deviation of its pixel values, and use similarity calculation methods to find reference images with similar features to the input image to obtain the analysis results; Parameter optimization: Based on the analysis results, evaluate the current state of the input image and the required enhancement direction and magnitude, and determine the image enhancement transform coefficients of RSTF; The transformation function of the RSTF method is: in, f(i, j) , g(i, j) The images before and after the transformation are respectively in ( i , j ) pixel value, S(i, j) For similar images (i, j) pixel value of a point Max , φ ( i , j ) represents the maximum pixel value and standard deviation; a , b , c and k These are the image enhancement transform coefficients for RSTF; DA optimization algorithm: The initial image enhancement transform coefficients are set as the initial values ​​of the DA group optimization algorithm, and the parameters are iteratively optimized to obtain the best image enhancement effect; Apply optimized parameters: Use the optimal parameters determined by the DA optimization algorithm as the image enhancement transform coefficients; A deep learning model is constructed using the YOLOv5s network as the base model and adding a segmentation head to the BiSeNet segmentation network. The segmentation head, built based on the BiSeNet segmentation network, is used to perform semantic segmentation on the output of the neck part in YOLOv5s. It includes an RFB+ layer, a stacking layer, a feature fusion module, a random deactivation layer, a convolutional layer, and an upsampling layer. Two RFB+ layers are stacked in parallel and then connected to the feature fusion module, the random deactivation layer, the convolutional layer, and the upsampling layer in sequence to output the segmentation result. Deep learning model training; the deep learning network is designed based on YOLOv5s, wherein the input layer is reduced from the original RGB three-channel to grayscale single-channel for receiving enhanced images; the backbone and neck networks are inherited from YOLOv5s and are used to extract image features, including a series of convolutional layers, pooling layers and residual modules, to gradually extract high-level features; The final deep learning model was obtained and used for coal and rock identification in the mine face.

2. The deep learning-based coal and rock identification method for coal mine faces according to claim 1, characterized in that, The preprocessing method includes: using the Labelme open-source annotation tool to annotate the acquired coal and rock images.

3. A coal and rock identification device for coal mine faces based on deep learning, characterized in that, To implement the method of claim 1 or 2, it includes: The image acquisition and preprocessing module is used to acquire raw coal and rock images, perform preprocessing, and obtain preprocessed images. The image enhancement module is used to enhance the preprocessed image based on the RSTF method and combined with the DA optimization algorithm to obtain the enhanced image; The model building module is used to build a deep learning model that uses the YOLOv5s network as the base model and adds a segmentation head by combining the BiSeNet segmentation network. The model training module is used for training deep learning models. The final model acquisition module is used to obtain the final deep learning model.

4. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the deep learning-based coal and rock identification method for coal mining faces as described in any one of claims 1 to 2.

5. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the deep learning-based coal and rock identification method for coal mining faces as described in any one of claims 1 to 2.

6. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the deep learning-based coal and rock identification method for coal mining faces as described in any one of claims 1 to 2.

Images