A method for recognizing working tooth number of a heading machine based on improved YOLO V5

Abstract

The application discloses an improved YOLOV5-based working tooth number recognition method for a heading machine, relates to the technical field of working tooth number recognition of a heading machine, and comprises the following steps: collecting an actually-shot image of the working of a heading machine in a mine, and performing image enhancement, classification and labeling; and inputting the image into an improved YOLOV5 model to obtain a non-working pick detection model; the model improvement method is that a Focus layer in a backbone layer is replaced by a convolution layer with k=6, and an attention mechanism module is added; an ASFF adaptive feature fusion network is used in a Neck layer; and a small target detection layer is added to a prediction layer. According to the non-working pick detection model, the number of non-working picks in a to-be-detected image can be recognized, and then the working tooth number of the heading machine is obtained. Through the improvement of the YOLOV5 model, the application improves the interference of factors such as poor illumination and much dust in the mine on the pick recognition, simplifies the network model, and realizes the rapid and accurate recognition of the working tooth number of the heading machine.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of tunneling machine working tooth number recognition, and more particularly to a tunneling machine working tooth number recognition method based on improved YOLOV5. BACKGROUND

[0002] The development of the coal industry has an important influence on the supply of social production energy. In recent years, with the frequent occurrence of coal mine safety accidents, all sectors of society have shown high concern for coal mining work. To realize the unmanned and autonomous development of coal mines, it is very important to realize the intelligent and stable control of the tunneling machine.

[0003] The environment of the tunneling machine is complex, the working load is large, and the construction environment is poor. During the cutting process, the load changes dramatically, and the cutting mechanism is subjected to frequent alternating impact loads during work, which can easily cause the cutting teeth to break and fall off, seriously affecting the cutting efficiency and coal mining efficiency. During the operation of the tunneling machine, the main cause of cutting tooth damage is that the working load is too large, exceeding the maximum stress of the cutting tooth, thereby causing the cutting tooth to be damaged and resulting in shutdown. Therefore, to ensure that the cutting tooth can work normally for a long time, it is necessary to analyze the stress of each working cutting tooth to ensure that the stress of the working cutting tooth is less than the maximum value. Since the cutting head changes with the working load and the cutting depth during the cutting process, the number of working cutting teeth and the stress of each cutting tooth are different. Under the conditions of obtaining the cutting hardness, cutting head vibration, cylinder stress, and motor power, if the number of working cutting teeth can be known, the stress of each cutting tooth can be analyzed. By adjusting the cutting speed in real time, the stress of each cutting tooth can be controlled within the allowable range of the rated value, ensuring that the cutting tooth is not damaged and works normally, and realizing the intelligent and stable control of the tunneling machine. However, the existing tunneling machine working tooth number recognition method has insufficient accuracy. With the continuous improvement of camera accuracy and the widespread application of image processing technology, target detection technology can be used to recognize the cutting teeth of the tunneling machine.

[0004] The existing target detection method is mainly divided into two-stage detection (Two Stage) and one-stage detection (One Stage). Two-stage detection such as R-CNN (Region-CNN) first generates regions, and then classifies samples through a convolutional neural network, which has high detection accuracy. One-stage detection such as the YOLO (You Only Look Once) series directly extracts features in the network to predict object classification and position, without the need for region generation, and has fast detection speed. YOLOV5 in the YOLO series improves the detection accuracy while maintaining fast detection speed, making it more suitable for detecting cutting teeth of the tunneling machine.

[0005] The target detection of the cutting pick in view of the special environment of the tunneling face faces the following three problems. First, the tunneling face environment is poor, the image acquisition is greatly affected by factors such as light intensity and dust, and the collected pictures need to be preprocessed; second, the tunneling machine cuts coal at a high speed, and the number of working teeth changes greatly, so the network inference speed needs to be improved to realize real-time detection; third, the camera is placed behind the spraying system of the tunneling machine, and the part of the cutting pick photographed is small, so the ability of the identification model to detect small targets is required to be higher.

[0006] Therefore, how to overcome the problems of great detection difficulty and insufficient real-time performance caused by factors such as light intensity, dust influence, coal cutting speed, and shooting position in the process of detecting the cutting pick of the tunneling machine is a problem that needs to be solved by those skilled in the art. SUMMARY

[0007] Therefore, how to overcome the problems of great detection difficulty and insufficient real-time performance caused by factors such as light intensity, dust influence, coal cutting speed, and shooting position in the process of detecting the cutting pick of the tunneling machine is a problem that needs to be solved by those skilled in the art.

[0008] In order to achieve the above-mentioned purpose, the present application provides the following technical scheme:

[0009] A tunneling machine working tooth number identification method based on an improved YOLO V5, comprising the following steps:

[0010] Step 1, collecting underground tunneling machine working real image to construct a data set;

[0011] Step 2, image enhancement is performed on the underground tunneling machine working real image in the data set to obtain an enhanced working image;

[0012] Step 3, classifying and labeling the enhanced working image to obtain a tunneling machine working image data set with class labels;

[0013] Step 4, inputting the tunneling machine working image data set with class labels into the improved YOLO V5 model for training to obtain a non-working pick detection model;

[0014] Step 5, inputting the working image to be detected into the non-working pick detection model to identify the number of non-working picks, and obtaining the number of working picks of the tunneling machine according to the total number of picks and the number of non-working picks.

[0015] Optionally, in step 2, the image enhancement is performed using a multi-scale Retinex model, wherein the single scale filter of the Retinex model is replaced by three Gaussian filters of different scales to perform dust removal and brightness enhancement on the image, thereby enhancing the effect of the original algorithm in highlighting the working details of the roadheader and dust removal.

[0016] Optionally, in step 3, the classification and labeling of the enhanced working image refers to classifying and labeling the enhanced working image according to the categories of the cutting teeth and the cutting head.

[0017] Optionally, in step 4, the improved YOLO V5 model refers to improvements made to the backbone layer, the Neck layer, and the prediction layer based on the YOLO V5 model.

[0018] The Focus layer in the backbone layer is replaced by a convolution layer with k = 6, which reduces the parameter quantity and computational complexity while ensuring the same detection effect of the cutting teeth, improves the operation speed, and uses a larger convolution kernel to increase the receptive field, thereby improving the performance and further adding an SE attention mechanism module to increase the weight of the key features of the cutting teeth and improve the detection effect.

[0019] The original feature fusion network PANet in the Neck layer is replaced by an ASFF adaptive feature fusion network, which suppresses inconsistent features by filtering conflict information in the spatial domain, fully utilizes the shallow layer features of the network for segmentation, and improves the feature fusion capability of the network for cutting teeth of different scales.

[0020] The large target detection anchor box in the prediction layer is deleted, and a small target detection anchor box is added. The large target detection head is deleted, and a small target detection layer is added.

[0021] Optionally, in step 4, the specific process of inputting the roadheader working image dataset with class labels into the improved YOLOV5 model is as follows:

[0022] The roadheader working image dataset with class labels is input into the preprocessing layer of the improved YOLO V5 model for image preprocessing, and the preprocessed dataset is obtained.

[0023] The preprocessed dataset is input into the backbone layer of the improved YOLO V5 model to obtain cutting tooth feature maps of different scales.

[0024] The cutting tooth feature maps of different scales are input into the Neck layer of the improved YOLO V5 model for sampling and feature fusion, and cutting tooth feature fusion maps of different scales are obtained.

[0025] The different scale roadheader pick tooth feature fusion graph is input into the prediction layer of the improved YOLO V5 model, and a non-working pick tooth number prediction result is obtained.

[0026] Optionally, the pre-processing layer performs the following process on the image:

[0027] First step, data enhancement: image splicing, affine transformation, random scaling;

[0028] Second step, adaptive anchor box calculation: output the prediction box based on the initial anchor box, compare with the real box, calculate the difference between the prediction box and the real box, and then update the network parameters in reverse;

[0029] Third step, adaptive image scaling: uniformly adjust the image to a preset size.

[0030] Optionally, the backbone layer processes the image in the pre-processed data set in the following process:

[0031] The image in the pre-processed data set is sequentially input into the convolution layer of k=6, the CBL module, the BottleneckCSP module, the spatial pyramid pooling module and the SE attention mechanism module, to obtain different scale roadheader working image feature maps with weights.

[0032] Optionally, the Neck layer processes the different scale roadheader working image feature maps in the following process:

[0033] The different scale roadheader working image feature maps are first sampled by the FPN feature pyramid to obtain high-level semantic information and low-level positioning information; and then the extracted high-level semantic information and low-level positioning information are fused at different scales by the ASFF adaptive feature fusion network to obtain different scale roadheader pick tooth feature fusion graphs.

[0034] Optionally, the method further comprises step 6 of evaluating the roadheader working tooth number recognition result from three aspects of detection accuracy, generalized intersection over union and detection speed, in particular:

[0035] The generalized intersection over union is a bounding box regression loss function, and the calculation formula is as follows:

[0036]

[0037]

[0038] Wherein, GIOU represents the generalized intersection over union, IOU represents the intersection over union, A represents the real box, B represents the target prediction box, C represents the minimum frame containing A and B at the same time, A∩B represents the area of A and B, A∪B represents the area of (A+B-A∩B).

[0039] The detection accuracy takes average accuracy as an index, the average accuracy is obtained through a PR curve diagram, the precision is the accuracy rate, as the horizontal coordinate, and the recall is the recall rate, as the vertical coordinate;

[0040] The detection speed uses the number of detected images in one second as an index.

[0041] Through the above technical solution, the present application provides a tunneling machine working tooth number recognition method based on improved YOLO V5, which has the following beneficial effects compared with the prior art:

[0042] The present application improves the interference of factors such as poor underground lighting and much dust on the recognition of cutting teeth by improving the YOLO V5 model, simplifies the network model, optimizes the model's detection and recognition ability of cutting teeth, realizes the rapid and accurate recognition of the working tooth number of the tunneling machine, provides an important parameter for realizing the intelligent and stable control of the tunneling machine, can effectively reduce the coal production reduction caused by cutting tooth damage, and at the same time improve the service life of the tunneling machine. BRIEF DESCRIPTION OF DRAWINGS

[0043] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced below, and obviously, the drawings in the following description are only embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of the provided drawings.

[0044] Figure 1 YOLO V5 model improvement schematic diagram for the present application;

[0045] Figure 2 Generalized intersection-over-union value comparison diagram of YOLO V5 model before and after improvement;

[0046] Figure 3 PR curve comparison diagram of YOLO V5 model before and after improvement;

[0047] Figure 4 Average accuracy value comparison diagram of YOLO V5 model before and after improvement;

[0048] Figure 5 Step diagram of tunneling machine working tooth number recognition method of the present application. DETAILED DESCRIPTION

[0049] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0050] Embodiment 1

[0051] The embodiments of the present application disclose a kind of based on improved YOLO V5's heading machine working tooth number identification method, see Figure 5 , comprising the following steps:

[0052] Step 1, collect the working image of underground heading machine to build dataset.

[0053] In the specific implementation process, two explosion-proof cameras can be respectively erected at the top end and the bottom end of the rear part of the heading machine spray system to realize real-time shooting of 180° up and down, and save the shot images, a total of 1500 underground heading machine working real shot pictures are obtained.

[0054] Step 2, using multi-scale Retinex model to enhance the image of underground heading machine working real shot image in data set, get enhanced working image;The multi-scale Retinex model refers to replacing the single scale filter of Retinex model with three different scale Gaussian filters to carry out dust removal and brightness enhancement on the image.

[0055] Step 3, the enhanced working image is classified and labeled according to the pick and cutting head, and the heading machine working image dataset with class label is obtained, and the training set, test set and validation set are divided according to the ratio of 8:1:1, of which 1200 are as training set, 150 are as test set, and 150 are as validation set.

[0056] Step 4, the heading machine working image dataset with class label is input into the improved YOLO V5 model for training, and a non-working pick detection model is obtained.

[0057] The improved YOLO V5 model refers to improving backbone layer, Neck layer and prediction layer based on YOLO V5 model, and the specific method is:

[0058] (1) replace the Focus layer in backbone layer with convolution layer with k=6, and add SE attention mechanism module;

[0059] (2) change the original feature fusion network PANet of Neck layer to ASFF adaptive feature fusion network;

[0060] (3) Delete large target detection anchor boxes in the prediction layer, add small target detection anchor boxes, delete large target detection heads, and add small target detection layers.

[0061] According to the improved YOLO V5 model, the specific process of inputting the tunneling machine working image data set with class labels into the improved YOLO V5 model is as follows:

[0062] (1) The tunneling machine working image data set with class labels is input into the pre-processing layer of the improved YOLO V5 model for image pre-processing to obtain a pre-processed data set, specifically:

[0063] First step, data augmentation: perform operations such as image splicing, affine transformation, and random scaling to expand and enhance the training effect of the model;

[0064] Second step, adaptive anchor box calculation: output the predicted box based on the initial anchor box, compare it with the real box, calculate the difference between the predicted box and the real box, and then update the network parameters in reverse;

[0065] Third step, adaptive image scaling: uniformly adjust the image to a preset size (640, 640) to more effectively utilize the information of the receptive field.

[0066] (2) The pre-processed data set is input into the backbone layer of the improved YOLO V5 model to obtain tunneling machine working image feature maps of different scales, specifically:

[0067] The images in the pre-processed data set are input into the convolution layer with k=6 for convolution;

[0068] Input into the CBL module for convolution: the CBL module is composed of a standard two-dimensional convolution layer, a normalization processing layer, and an activation function layer;

[0069] Input into the BottleneckCSP module to fuse feature maps of different levels, retain and accumulate more features of different receptive fields to obtain information-rich feature maps;

[0070] Input into the spatial pyramid pooling module: first reduce the input channel by half through a standard convolution module, then perform maximum pooling with convolution sizes of 5, 9, and 13, and concatenate the results of the three maximum pooling operations with the data that has not been subjected to the pooling operation;

[0071] Input into the SE attention mechanism module to increase the weight of the key features of the cutting teeth and improve the detection effect:

[0072] The first step is to perform global average pooling on the tunneling machine's working feature map, compressing a feature map with c channels and a width and height of H×W into a feature map with c channels and a width and height of 1×1. The resulting image represents global information, and the specific relationships are as follows:

[0073]

[0074] Among them, u c This represents the feature map of channel c before pooling, with dimensions H×W, and F. sq Indicates global average pooling, z c This represents a feature map with a width and height of 1×1 for the c-channel after pooling.

[0075] The second step involves establishing connections between channels using two fully connected layers and a non-linear activation function, which facilitates the stitching of feature maps. The specific relationships are as follows:

[0076]

[0077] Where z represents the feature map before processing. T1 represents the processed feature map, T1 represents the first fully connected layer, ReLU represents the non-linear activation function, and T2 represents the second fully connected layer.

[0078] Then, the sigmoid activation function is used to obtain normalized weights, and finally, the weights are multiplied and applied to each channel of the tunnel boring machine working feature map. This completes the channel attention recalibration of the cutting tooth features, resulting in tunnel boring machine working image feature maps of different scales with weights.

[0079] (3) Input the feature maps of tunnel boring machine working images at different scales into the Neck layer of the improved YOLO V5 model, perform sampling and feature fusion, and obtain the feature fusion maps of tunnel boring machine cutting teeth at different scales, specifically:

[0080] The feature maps of tunnel boring machine working images at different scales are first downsampled from top to bottom by FPN feature pyramid to extract the semantic information of the cutting teeth at the high level, and then upsampled from bottom to top to extract the localization information of the cutting teeth at the low level, thus obtaining the semantic information of the high level and the localization information of the low level. Then, the ASFF adaptive feature fusion network is used to fuse the extracted high-level semantic information and low-level localization information at different scales to obtain the feature fusion map of the tunnel boring machine cutting teeth at different scales.

[0081] (4) The improved YOLO V5 model is inputted with the prediction layer of the feature fusion image of the different sizes of the heading machine pick, the prediction layer is composed of a convolution layer, an initial anchor frame and three detection heads, the adaptive anchor frame is adjusted, the anchor frame for large target detection is deleted, the anchor frame for small target detection is added, the adjusted anchor frames are [5, 6, 8, 14, 15, 11], [10, 13, 16, 30, 33, 23] and [30, 61, 62, 45, 59, 119], the detection head for large target is deleted, the sizes of the adjusted target detection heads are 38*38*255 and 19*19*255, the small target detection layer is added, the feature map of the size of 160*160 of the heading machine pick obtained is concatenated with the feature map of the second layer of the backbone pick to obtain a larger feature map of the heading machine pick for small pick detection, and a non-working pick number prediction result is obtained.

[0082] Step 5, inputting the to-be-detected working image into the non-working pick detection model, since the working pick is located in the coal seam, the non-working pick number is recognized in the image, and thus the total pick number of the heading machine is subtracted from the non-working pick number, so that the working pick number of the heading machine is obtained.

[0083] The method can further include step 6, evaluating the working pick number recognition result of the heading machine from three aspects of detection accuracy, generalized intersection over union and detection speed, and specifically:

[0084] The generalized intersection over union is a boundary box regression loss function, and the calculation formula is as follows:

[0085]

[0086]

[0087] Wherein, GIOU represents the generalized intersection over union, IOU represents the intersection over union, A represents a real frame, B represents a target prediction frame, C represents a minimum frame containing A and B at the same time, A∩B represents the area of A and B, and A∪B represents the area of (A+B-A∩B);

[0088] The detection accuracy takes the average precision as an index, the average precision is obtained through a PR curve diagram, the precision is an accuracy rate, serving as an abscissa, and the recall is a recall rate, serving as an ordinate;

[0089] The detection speed uses the number of images detected per second as an index.

[0090] Embodiment 2

[0091] The application further provides a heading machine working pick number recognition system based on the improved YOLO V5, comprising:

[0092] An image acquisition module is configured to acquire an underground tunneling machine working real image to construct a dataset.

[0093] An image enhancement module is configured to perform image enhancement on the underground tunneling machine working real image in the dataset to obtain an enhanced working image.

[0094] An image classification module is configured to classify and label the enhanced working image to obtain a tunneling machine working image dataset with a category label.

[0095] A model training module is configured to input the tunneling machine working image dataset with a category label into an improved YOLOV5 model for training to obtain a non-working pick detection model.

[0096] A pick identification module is configured to input a to-be-detected working image into the non-working pick detection model to identify a number of non-working picks, and to obtain a working pick number of the tunneling machine according to a total pick number of the tunneling machine and the number of non-working picks.

[0097] Embodiment 3

[0098] According to the simulation experiment of the scheme of the present application, the beneficial effects of the present application are verified.

[0099] Hardware resources: the operating system of the experimental platform is 11th Gen Intel(R) Core(TM) i5-11260H, the model training is performed under a NVIDIA 3050 GPU card, the software version is PyCharm Community Edition 2022.2, and the experimental framework is CUDA11.2. The experimental operation steps are as follows:

[0100] (1) Constructing a dataset: in this embodiment, a self-built dataset is used, the data source is an underground tunneling working face real image, a total of 1500 pictures are collected, of which 1200 pictures are used as a training set, 150 pictures are used as a test set, and 150 pictures are used as a verification set, and the experimental object categories are picks and cutting heads.

[0101] (2) Optimization strategy: refer to Figure 1 On the basis of the original YOLO V5 model, the Focus layer in the backbone layer is replaced by a convolution layer with k=6, and an SE attention mechanism module is added; the original feature fusion network PANet in the Neck layer is replaced by an ASFF adaptive feature fusion network; the large target detection anchor frame in the prediction layer is deleted, a small target detection anchor frame is added, the large target detection head is deleted, a small target detection layer is added, and the accuracy of detecting small picks is improved.

[0102] (3) Parameter setting: the training parameters of the improved YOLO V5 model are as follows: the training pre-weight is a yolov5m.pt model, the training round is 200 rounds, 4 pictures are input each time, and the picture size is 640*640.

[0103] (4) Result analysis: the training results of the improved YOLO V5 model are evaluated from three aspects of generalized intersection over union, detection accuracy and detection speed. The smaller the value of generalized intersection over union, the closer the target prediction box is to the real box, and the higher the prediction accuracy; the average precision is used as the index of model detection accuracy, the average precision is obtained through the PR curve diagram, and the larger the area surrounded by the curve and the coordinate axis, the higher the model detection accuracy; the faster the detection speed, the better the real-time performance of the model prediction. Figure 2 The generalized intersection over union diagram of the improved YOLO V5 model is shown in the figure, the horizontal axis is the training round, the vertical axis is the generalized intersection over union value, the left side is the original model, and the right side is the improved model. After 300 iterations, the ordinary YOLO V5 model is 0.092, and the improved YOLO V5 model is 0.055. The improved YOLO V5 model has higher target box positioning accuracy. Figure 3 The PR diagram of the improved YOLO V5 model is shown in the figure, the horizontal coordinate is the recall rate (Racall), the vertical coordinate is the accuracy (Precision), the lower curve represents the cutting tooth detection effect, the upper curve represents the cutting head detection effect, and the middle curve represents the comprehensive detection effect. It can be seen that the improved PR diagram has a larger area surrounded by the coordinate axis, and the detection accuracy is higher, Figure 4 The average precision (MeanAverage Precision map) value diagram of the improved YOLO V5 model is shown in the figure, the horizontal coordinate is the training round, and the vertical coordinate is the map value. The average precision value is improved from 86.5 to 88.7, and the improved YOLO V5 model has higher detection accuracy. The detection speed is improved from 121 pictures per second to 156 pictures per second, which is increased by 29%, and the improved YOLO V5 model has faster detection speed.

[0104] From the above data, it can be concluded that the positioning accuracy, accuracy and detection speed of the cutting tooth of the tunneling machine are all better than those of the original model, and the non-working cutting tooth of the tunneling machine can be quickly and accurately identified, and the actual working tooth number of the tunneling machine can be calculated.

[0105] Each embodiment in the specification is described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same or similar parts of each embodiment can be referred to each other.

[0106] The foregoing description of the disclosed embodiments enables a person skilled in the art to make or use the application. Modifications of these embodiments will occur to persons of skill in the art, and that the appended claims are intended to cover all such modifications that do not depart from the true spirit and scope of the application. Therefore, the application is not limited to the embodiments shown but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for identifying the number of working teeth of a heading machine based on improved YOLO V5, characterized in that, The method comprises the following steps: Step 1, collecting downhole tunneling machine working real image construction dataset; Step 2, image enhancement is performed on the downhole tunneling machine working real image in the dataset to obtain an enhanced working image; Step 3, classifying and labeling the enhanced working image to obtain a tunneling machine working image dataset with class labels; Step 4, inputting the tunneling machine working image dataset with class labels into an improved YOLO V5 model for training to obtain a non-working pick detection model; Step 5, inputting a to-be-detected working image into the non-working pick detection model to identify the number of non-working picks, and obtaining the number of working picks of the tunneling machine according to the total number of picks and the number of non-working picks. In step 4, the improved YOLO V5 model refers to improvements made to the backbone layer, the neck layer and the prediction layer of the YOLO V5 model, and the specific method is as follows: Replace the Focus layer in the backbone layer with a convolution layer with k=6 and add an SE attention mechanism module; Replace the original feature fusion network PANet in the neck layer with an ASFF adaptive feature fusion network; Delete the large target detection anchor box in the prediction layer, add a small target detection anchor box, delete the large target detection head, and add a small target detection layer.

2. The improved YOLO V5-based working tooth number recognition method for a heading machine according to claim 1, characterized in that, In step 2, a multi-scale Retinex model is used for image enhancement, wherein the multi-scale Retinex model refers to replacing the single-scale filter of the Retinex model with three Gaussian filters of different scales to perform dust removal and brightness enhancement on the image.

3. The improved YOLO V5-based working tooth number recognition method for a heading machine according to claim 1, characterized in that, In step 3, classifying and labeling the enhanced working image refers to classifying and labeling the enhanced working image according to the picks and cutting heads.

4. The improved YOLO V5-based working tooth number recognition method for a heading machine according to claim 1, characterized in that, In step 4, the specific process of inputting the tunneling machine working image dataset with class labels into the improved YOLO V5 model is as follows: Input the tunneling machine working image dataset with class labels into the preprocessing layer of the improved YOLO V5 model to perform image preprocessing and obtain a preprocessed dataset; Input the preprocessed dataset into the backbone layer of the improved YOLO V5 model to obtain tunneling machine working image feature maps of different scales; Input the tunneling machine working image feature maps of different scales into the neck layer of the improved YOLO V5 model to perform sampling and feature fusion and obtain tunneling machine pick feature fusion maps of different scales; Input the tunneling machine pick feature fusion maps of different scales into the prediction layer of the improved YOLO V5 model to obtain a non-working pick number prediction result.

5. The improved YOLO V5-based working tooth number recognition method for a heading machine according to claim 4, characterized in that, The preprocessing process of the preprocessing layer is as follows: First step, data enhancement: image splicing, affine transformation, random scaling; Second step, adaptive anchor box calculation: output the prediction box based on the initial anchor box, compare with the real box, calculate the difference between the prediction box and the real box, and then update the network parameters in reverse; Third step, adaptive image scaling: uniformly adjust the image to a preset size.

6. The improved YOLO V5-based working tooth number recognition method for a heading machine according to claim 4, characterized in that, The process of processing the image in the preprocessed dataset by the backbone layer is as follows: The images in the preprocessed data set are sequentially input into a convolutional layer with k=6, a CBL module, a BottleneckCSP module, a spatial pyramid pooling module, and an SE attention mechanism module to obtain different scale roadheader working image feature maps with weights.

7. The improved YOLO V5-based working tooth number recognition method for a heading machine according to claim 4, characterized in that, The process of the Neck layer processing different scale roadheader working image feature maps is as follows: The different scale roadheader working image feature maps are first sampled by the FPN feature pyramid to obtain high-level semantic information and low-level positioning information; then the extracted high-level semantic information and low-level positioning information are fused at different scales by the ASFF adaptive feature fusion network to obtain different scale roadheader pick tooth feature fusion maps.

8. The improved YOLO V5-based working tooth number recognition method for a heading machine according to claim 1, characterized in that, It also includes step 6, which evaluates the roadheader working tooth number recognition result from three aspects of detection accuracy, generalized intersection over union, and detection speed, and the specific steps are as follows: The generalized intersection over union is a bounding box regression loss function, and the calculation formula is as follows: ； ； Wherein, GIOU represents the generalized intersection over union, IOU represents the intersection over union, A represents the real box, B represents the target prediction box, C represents the minimum frame containing A and B at the same time, A∩B represents the area of A and B, and A∪B represents the area of (A+B-A∩B); The detection accuracy takes the average precision as the index, and the average precision is obtained by the PR curve graph, where precision is the accuracy rate as the horizontal coordinate, and recall is the recall rate as the vertical coordinate; The detection speed uses the number of detected images per second as an index.

Images