A method and device for identifying the main structural surface of tunnel rock mass based on image recognition

Abstract

The present invention relates to the technical field of rock mass structural surface recognition, and in particular to a method and device for identifying the main structural surface of a tunnel rock mass based on image recognition, comprising: obtaining an excavation surface image and slicing it, performing recognition based on a pre-trained rock mass structural surface trace recognition model, obtaining an image with a predicted trace, and then performing cluster fitting processing to obtain a structural surface trace. The present invention can effectively reduce the requirements for the original image during the recognition process, and can improve the capture accuracy of the predicted trace, avoid overfitting, and greatly eliminate noise points on the predicted trace, so that the obtained structural surface trace can be closer to the real structural surface trace, thereby improving the accuracy of trace recognition.

Description

Technical Field

[0001] The invention relates to the technical field of rock mass structural surface recognition, and in particular to a method and device for identifying the main structural surface of a tunnel rock mass based on image recognition. Background Art

[0002] Tunnel engineering is an important part of urban construction and transportation infrastructure construction. The structural surface in the tunnel rock mass is one of the key factors to be considered in evaluating the stability and grade of the surrounding rock. The traditional structural surface identification method mainly relies on the experience of geological personnel and equipment-side identification, which has problems such as low identification efficiency and large errors, and cannot meet actual needs. At the moment when the upstream management end is becoming more and more refined, there is a huge efficiency gap between the manual identification method and the management needs. Automated and intelligent on-site construction methods are gradually becoming popular, but the site is restricted by the construction environment and it is difficult to deploy large and medium-sized equipment. Taking the rock structure surface identification equipment of the tunnel excavation face as an example, there are already a variety of special equipment based on SLR cameras and laser radars on the market, which have supporting algorithms for automatic modeling and intelligent identification. However, such equipment and algorithms need to be deployed and consume a lot of time for on-site construction personnel to adjust the angle, shoot and model. Not only does the equipment hinder the construction channel and bring greater safety hazards, but it also wastes the time of on-site personnel and seriously delays the construction progress.

[0003] For example, a Chinese patent application with the publication number CN115731390A discloses a method for identifying the structural surface of a limestone tunnel rock mass. The method mainly obtains the structural image of the limestone tunnel face and obtains the skeleton line under the structural image according to an algorithm, and then linearizes and performs pixel-level statistics on the skeleton line to obtain the length and apparent inclination of each skeleton line. In obtaining the skeleton line, the method corrodes the structural surface trace and performs an open operation to obtain the skeleton line of the structural surface. The simple corrosion and open operation process cannot accurately identify and eliminate redundant noise points, and may even perform erroneous processing on the original skeleton line, resulting in inaccurate skeleton lines. At the same time, the method has certain quality requirements for the acquired image, which will bring great inconvenience to the acquisition of images during the construction process.

[0004] Therefore, it is urgent to propose a method that can complete structural surface identification under ordinary shooting conditions without wasting too much time of construction workers, or even only using the tunnel face image, while improving the efficiency and accuracy of structural surface identification, saving costs, and providing favorable support for advanced geological forecasting, comprehensive geological analysis and other processes carried out in the same period. Summary of the invention

[0005] The purpose of the present invention is to overcome the problem that the structural surface trace recognition method of the existing rock structure surface recognition technology is greatly affected by the quality of the acquired structural surface image and the accuracy is not high. A tunnel rock main structural surface recognition method based on image recognition is provided, which can recognize the ordinary image of the rock excavation surface, and at the same time use the noise reduction technology to process the image, and cluster and straight line fit the trace discrete point set of the trace. The method improves the accuracy of trace recognition by performing noise reduction processing on the image and based on the configured algorithm, has low requirements on the quality of the original image, and has high engineering practicality.

[0006] In order to achieve the above-mentioned object of the invention, the present invention provides the following technical solutions:

[0007] A method for identifying the main structural surface of a tunnel rock mass based on image recognition comprises the following steps:

[0008] S1, acquiring an excavation surface image as a first image, and slicing and segmenting the first image to form a plurality of second images;

[0009] S2. Identify the plurality of second images based on a pre-trained rock mass structural surface trace recognition model to form a plurality of third images, wherein the plurality of third images are binary images and have predicted traces;

[0010] S3, stitching and restoring the plurality of third images to form a fourth image, wherein the fourth image is equal in size to the first image, is a binary image, and has a prediction trace;

[0011] S4. Perform cluster fitting processing on the predicted traces in the fourth image to obtain structural surface traces, wherein the cluster fitting processing includes density-based spatial clustering algorithm processing and Hough space-based straight line fitting.

[0012] Preferably, in the above-mentioned method for identifying the main structural surface of a tunnel rock mass based on image recognition, S2 specifically includes:

[0013] S21, inputting the plurality of second images in a round-robin manner, calling the pre-trained rock mass structural surface trace recognition model for prediction on each second image, and outputting a corresponding predicted image;

[0014] S22, processing the predicted image based on average value pooling, and then binarizing the processed predicted image to form a trace prediction image;

[0015] S23, after the round robin is completed, all the trace prediction images corresponding to each second image are obtained, and all the trace prediction images are combined to obtain the third image.

[0016] Preferably, in the above-mentioned tunnel rock main structural surface identification method based on image recognition, the pre-trained rock structural surface trace identification model is constructed by the following steps, including:

[0017] Get the original images as training sets and manually annotate them;

[0018] Cross entropy is used as the loss function to calculate the loss of each layer and modify the weight;

[0019] The initial model for identifying the rock mass structural surface trace is trained based on the training set to obtain a trained rock mass structural surface trace identification model.

[0020] Preferably, in the above-mentioned tunnel rock main structural surface identification method based on image recognition, the pre-trained rock structural surface trace identification model includes 5 3×3 convolutional layers, corresponding 5 Relu function activation layers, corresponding 5 deconvolution layers, 1 fully connected layer, 1 1×1 convolutional layer and 1 maximum pooling layer.

[0021] Preferably, in the above-mentioned method for identifying the main structural surface of a tunnel rock mass based on image recognition, S4 specifically includes:

[0022] S41, setting density-based spatial clustering algorithm parameters, and inputting the fourth image;

[0023] S42, clustering points belonging to the same trace in combination with the signal area in the fourth image to obtain a classification space composed of several classifications, wherein the centroid point of the classification space is the trace point;

[0024] S43, cycling through the classification spaces of each classification based on the Hough space to obtain fitting straight lines for all classifications;

[0025] S44, performing noise filtering on the fitting straight lines of all the classifications to obtain structural surface traces.

[0026] Preferably, in the above-mentioned method for identifying the main structural surface of a tunnel rock mass based on image recognition, the setting of density-based spatial clustering algorithm parameters specifically includes the domain distance of samples and the number of samples within the domain distance.

[0027] Preferably, in the above-mentioned method for identifying the main structural surface of a tunnel rock mass based on image recognition, the S43 specifically includes:

[0028] S431, performing coordinate transformation on the classification space composed of several classifications to transform it into a polar coordinate Hough space;

[0029] S432: a straight line is drawn using the centroid of the classification space under a certain classification, and is rotated at a fixed angle, and at the same time, the distance from the centroid of other classification spaces in the classification to the straight line is taken;

[0030] S433, when the distance does not exceed the set threshold, the number of point sets on the straight line is increased by 1, and when the maximum number of rotations is reached, a straight line with the largest number of point sets is obtained as the fitting straight line for the classification;

[0031] S434. Repeat steps S432-S433 until fitting straight lines of all categories are obtained.

[0032] Preferably, in the above-mentioned tunnel rock main structural surface identification method based on image recognition, the slicing and segmenting of the first image specifically comprises: obtaining the length and width of the first image, and setting the number of slices, and performing slicing and segmenting with a ratio of the width to the number of slices;

[0033] In S3, the plurality of third images are stitched together and restored to form a fourth image according to the number of slices and the ratio.

[0034] Preferably, in the above-mentioned method for identifying the main structural surface of a tunnel rock mass based on image recognition, in the binary image, 0 represents a non-trace and 1 represents a trace.

[0035] On the other hand, the present invention also provides a device for identifying the main structural surface of a tunnel rock mass based on image recognition, characterized in that it includes at least one processor and a memory communicatively connected to the at least one processor; the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute any of the above-mentioned methods for identifying the main structural surface of a tunnel rock mass based on image recognition.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] The present invention provides a method for identifying the main structural surfaces of a tunnel rock mass based on image recognition. By slicing and segmenting the acquired image, the recognition error caused by the excessive size of the original image is reduced, thereby reducing the quality requirements for the image to be identified and improving the recognition range. On the basis of slicing and segmenting, the rock mass structural surface trace identification model that has been pre-trained is used for identification, thereby improving the capture accuracy in obtaining the predicted trace of the structural surface and avoiding overfitting. Then, a density-based spatial clustering algorithm is used to greatly eliminate noise points on the predicted trace, so that the acquired structural surface trace can be closer to the real structural surface trace, thereby improving the accuracy of trace identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 A flowchart of the present invention in one embodiment;

[0039] Figure 2 A schematic diagram of image segmentation in one embodiment of the present invention;

[0040] Figure 3 It is a schematic diagram of the structure of a rock mass structural surface trace identification model in one embodiment of the present invention;

[0041] Figure 4 A predicted output diagram of the present invention in one embodiment;

[0042] Figure 5 A binary map of the predicted output in one embodiment of the present invention;

[0043] Figure 6 A binary spliced ​​image in one embodiment of the present invention;

[0044] Figure 7 A straight line fitting flow chart of an embodiment of the present invention;

[0045] Figure 8 A schematic diagram of structural surface trace identification results in one embodiment of the present invention;

[0046] Fig. 9 Schematic diagram of trace labeling software in one embodiment of the present invention. DETAILED DESCRIPTION

[0047] The present invention is further described in detail below in conjunction with test examples and specific implementation methods. However, this should not be understood as the scope of the above subject matter of the present invention being limited to the following embodiments, and all technologies realized based on the content of the present invention belong to the scope of the present invention.

[0048] Example 1

[0049] Figure 1 A method for identifying the main structural surface of a tunnel rock mass based on image recognition according to an exemplary embodiment of the present invention is shown, comprising the following steps:

[0050] S1, obtaining an excavation surface image as a first image, slicing and segmenting the first image to form a plurality of second images, thereby reducing recognition errors caused by the original image being too large, thereby reducing image requirements and increasing the recognition range;

[0051] S2. Identify the plurality of second images based on a pre-trained rock mass structural surface trace recognition model to obtain a plurality of third images, wherein the plurality of third images are binary images with predicted traces, thereby improving capture accuracy and avoiding overfitting;

[0052] S3, stitching and restoring the plurality of third images to form a fourth image, wherein the fourth image is equal in size to the first image, is a binary image, and has a prediction trace;

[0053] S4. Perform cluster fitting processing on the predicted traces in the fourth image to obtain structural surface traces, wherein the cluster fitting processing includes density-based spatial clustering algorithm processing and Hough space-based straight line fitting.

[0054] It is understandable that the identification of the rock structure surface of the tunnel excavation face needs to be carried out during the tunnel excavation process in order to evaluate the stability of the rock mass and the grade of the surrounding rock. One of the most important indicators for evaluating the stability and grade of the rock mass is the integrity of the rock mass. According to the "Technical Regulations for Advanced Geological Forecasting of Railway Tunnels" and other relevant specifications, the integrity of the rock mass mainly depends on the development characteristics of the structural surface. At present, it mainly relies on on-site geologists to conduct manual identification and record on site.

[0055] Therefore, in order to achieve the purpose of identifying the main structural surface of the rock mass through the image of the newly exposed excavation surface, the present embodiment provides a method for identifying the main structural surface of the tunnel rock mass based on image recognition. The method obtains the image by slicing segmentation, thereby reducing the recognition error caused by the original image being too large, thereby reducing the image requirements and improving the recognition range; and recognizes through a pre-trained rock mass structural surface trace identification model, thereby improving the capture accuracy in obtaining the predicted trace of the structural surface and avoiding overfitting; and then using a density-based spatial clustering algorithm to greatly eliminate the noise points on the predicted trace, so that the obtained structural surface trace can be closer to the real structural surface trace, thereby improving the accuracy of trace identification.

[0056] Example 2

[0057] In a possible implementation, the above S1 specifically includes:

[0058] S11. Receive an image uploaded on-site, which is usually taken with an ordinary mobile phone and recorded as the first image.

[0059] S12, further, segmenting and slicing the first image;

[0060] Specifically, since the image received by the deep learning model is (224×224), if compression is used for the first image, semantics will be lost as the image becomes smaller, and keeping the original size will make it too large to be processed. Therefore, in this embodiment, the first image is segmented and sliced ​​to ensure that semantic information is not lost.

[0061] like Figure 3 As shown, the above S12 specifically includes:

[0062] S121, image slicing. Obtain the length and width (h×w) of the original image. According to the configuration, the image is divided into 4 blocks according to the narrow side (w), that is, kernel = (w / 4, w / 4), and the missing part of the image is padded with (0,0,0).

[0063] S122, image compression. The kernel size after slicing still exceeds (224×224) and must be compressed. Use the OpenCV library to compress the image and save the compressed slices in order from top to bottom and from left to right, which are recorded as the second image. In order to facilitate the subsequent re-stitching of the image, the original size (h×w) and the slice size (kernel) need to be cached and returned together.

[0064] In a possible implementation, the above S2 includes: rock mass trace identification based on deep learning. Among them, the rock mass structural surface trace identification corresponds to the edge detection algorithm in the machine vision algorithm. Traditional algorithms include Canny operator, support vector machine and other methods. The present invention adopts a neural network based on holistically-nested edge detection (HED) to complete the identification of rock mass structural surface traces. The specific steps are as follows:

[0065] S.21 Network design and loss function. The technical principle of HED is mainly to use VGG-16 as the backbone network. VGG-16 is a deep convolutional neural network (CNN) proposed by Karen Simonyan and Andrew Zisserman of the Visual Geometry Group of the University of Oxford in 2014. The network has a depth of 16 layers and uses a relatively small 3×3 convolution kernel, so it is named VGG-16. VGG-16 is one of the best performing models in the ImageNet image classification competition. The network uses very small filters during training while increasing the depth of the network, which enables it to better capture complex features in the image. VGG-16 also uses a maximum pooling layer to reduce the number of parameters in the network to avoid overfitting. In VGG-16, the structures of the convolutional layer and the fully connected layer are the same, except that their depths are different. The first convolutional layer of the network uses 64 3×3 convolution kernels, the second convolutional layer uses 128 3×3 convolution kernels, the third convolutional layer uses 256 3×3 convolution kernels, the fourth convolutional layer uses 512 3×3 convolution kernels, and the fifth convolutional layer also uses 512 3×3 convolution kernels. Each convolutional layer is followed by a maximum pooling layer to reduce the size of the feature map. After the convolutional layer, VGG-16 has three fully connected layers. The first fully connected layer has 4096 neurons, the second fully connected layer also has 4096 neurons, and the last fully connected layer has 1000 neurons, corresponding to the 1000 categories in the ImageNet dataset. VGG-16 has become one of the classic deep learning models in image classification and has been widely used in image recognition, object detection, and image segmentation. The network design uses 5 convolutional layers, each with a different number of (3×3) convolution kernels, and uses ReLU as the activation function. Slightly different from VGG-16, each convolutional layer is followed by a deconvolution, and then the maximum pooling layer is used to achieve downsampling and enter the next layer. The deconvolved image will be saved and passed to the fully connected layer as features of different dimensions. Finally, a (1×1) convolution kernel is used to reduce the 5 channels of the image pixels to 1 channel. During the training phase, HED uses cross entropy as the loss function, which can calculate the loss and modify the weights at each layer. Overall, the architecture of HED is not complicated, but due to the weights of the 5 layers in the middle, the training time is long, and the resulting model takes up more space than other models. The HED network and optimized structure are as follows: Figure 3 shown.

[0066] S.22 trace annotation. From the perspective of the HED network, it returns one-dimensional data. Therefore, in the training set annotation stage, the rock image trace annotation results need to be normalized. The simplest way is to generate a black and white binary image, with black pixels representing non-traces and white pixels representing traces. The present invention annotated a total of 2,000 images in the training stage, and used data enhancement, such as modifying lighting, magnification, and screenshots, to annotate the remaining 2,000 images, with a total of nearly 5,000 annotated images.

[0067] S.23 Training: Based on a large number of training sets, the rock mass trace identification model is retrained and persisted as an H5 model file.

[0068] S.24 prediction. Input the second image, loop through the second images, call the prediction function for each second image, pass in the model file, and output the predicted image, such as Figure 4 shown.

[0069] S.25 Pooling. Since there are many noise points in the predicted image, it is necessary to continue average pooling to remove the noise points and then binarize the processed image. The trace prediction image is formed, and the effect is as follows Figure 5 After the round robin is completed, all trace prediction images form the third image.

[0070] In a possible implementation, the above S3, image stitching and restoration, specifically includes: the third image generated by the rock mass trace identification step only represents a small area, and all small images should be re-stitched and restored to the size of the first image:

[0071] S31, image stitching: Reassemble the third image according to the logic of aspect ratio and narrow edge four images as a group.

[0072] S32, image restoration. According to the original size (h×w) of the cache and the slice size (kernel), the ratio is converted, and the assembled image is first cropped and then stretched to form a binary image of the original image size, which is recorded as the fourth image. In this binary image, 0 still represents non-trace and 1 represents trace. The final binary image is as follows: Figure 6 shown.

[0073] In a possible implementation, the above S4 specifically includes: a density-based spatial clustering algorithm and a Hough space-based straight line fitting. Cluster analysis is an extension of classification statistics. Based on the similarity of objects, sample points of similar classification in sample data are grouped together. Common clustering algorithms include K-means, BIRCH, DBSCAN and other algorithms. Different algorithms have different effects in dealing with different clustering requirements and sample conditions, and it is necessary to select the most suitable algorithm according to the specific situation. Taking DBSCAN as an example, it is a density-based clustering algorithm. The core idea is to divide all data points into three types: core points (Core Point), border points (BorderPoint) and noise points (NoisePoint). First, randomly select an unvisited point, mark it as visited, and then find all points in its eps neighborhood. If the point is a core point, it is classified into the same category as all the points in its eps neighborhood, and recursively find the eps neighborhood of each point. If the point is a border point, it is classified into the same category as the core point. Finally, the unvisited point is marked as a noise point. Line fitting algorithm is a commonly used algorithm in the field of computer vision. It can automatically identify straight lines in images, thereby achieving the purpose of image processing and analysis. Commonly used line fitting algorithms include: least squares method, RANSAC, Hough line. Line fitting algorithms are widely used in the field of computer vision, such as edge detection, target tracking, image stitching, etc. Taking Hough Line as an example, its core idea is Hough transform, which transforms rectangular coordinates into polar coordinates: yi = θxi + c, θ and c are variables, there are countless straight lines through a point (x1, y1), that is, there are countless θ and c, and the θ and c of another point (x2, y2) passing through the collinear line must be the same as (x1, y1), so the line fitting problem becomes a problem of finding the most collinear points in the slope and intercept space.

[0074] Specifically, as described in S4 above, the result after trace recognition cannot directly form a straight line. On the contrary, in the previous binarization process, points with strong signals and general signals are considered to be trace signals, resulting in a large number of discrete points, causing the trace in the first image to appear as a large white area in the fourth image. It is necessary to cluster such areas to group points that may belong to the same trace.

[0075] 4.1 Distance and density. Combined with the signal area of ​​the binary image (the fourth image), DBSCAN is the best clustering method. The reason is that the areas of the same trace are relatively independent, and the points in the area and along the trace are densely arranged. The density-based algorithm can well discover and summarize these discrete points. Secondly, DBSCAN does not need to specify the number of clusters in advance, but automatically determines the number of clusters based on the distribution characteristics of the data itself, which is conducive to discovering the maximum number of trace areas. The density of DBSCAN is calculated based on the neighborhood distance. Specifically, it means that all points within the neighborhood distance eps are marked as the same category, and then recursion continues between these points to find other discrete points within the neighborhood distance eps until all points are classified or marked as noise points.

[0076] 4.2 Clustering Implementation Based on multiple experiments and tests, the present invention selects eps=10, min_samples=5 as the parameters of DBSCAN, where eps is the domain distance and min_sampleswei is the number of samples within the domain distance, and implements DBSCAN clustering on the binary image (the fourth image) to obtain a classification space T composed of several classifications.

[0077] 4.3 Hough transform. In the rectangular coordinate system (x, y) space, each straight line y = θx + c can be confirmed by two parameters, the slope θ and the intercept c. If θ and c are regarded as independent variables, c = -xθ + y, and this (θ, c) space is the Hough space. The conversion from the rectangular coordinate system to the polar coordinate Hough space is called the Hough transform.

[0078] 4.4 Line fitting to achieve trace. In the classification space T, loop through each classification, take the centroid point of the classification, draw a straight line with the point as the center and rotate it counterclockwise from 0 to 180°, with a step size of 1°. For every 1° rotation, take other points in the classification and calculate the distance d from the straight line. When d is within the threshold range, the point is considered to be on the straight line, and the number of points on the straight line is increased by 1. Finally, the straight line with the largest number of points in 180 rotations is counted as the fitting straight line of the classification. Continue to the next classification until the classification is completed. The process is as follows Figure 7 shown.

[0079] 4.5 Main structural surface filtering. Remove the straight line segments that are too short. The filtered straight line segments can be considered as the traces of the main structural surface on the excavation surface. The effect is as follows: Figure 8 shown.

[0080] Example 3

[0081] On the other hand, the present invention also provides a tunnel rock main structural surface identification device based on image recognition, characterized in that it includes at least one processor and a memory communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute a tunnel rock main structural surface identification method based on image recognition as described in any one of the above items.

[0082] Based on the above technical solution, in this embodiment, using a preset processor, the WinForm program can implement a simple rock mass trace marking software, such as Fig. 9 At the same time, in this embodiment, the final structural surface trace obtained can return the result. In order to combine various on-site applications, such as the advanced geological prediction system, this embodiment encapsulates the prediction process as an HTTP interface call, passes in the newly exposed excavation surface image through the POST method, and returns the straight line vertex coordinates of the structural surface on the original image, which are drawn by the relevant system or used for other services.

[0083] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for identifying the main structural surface of a tunnel rock mass based on image recognition, characterized in that: The following steps are involved: S1, obtaining an excavation surface image, recorded as a first image, and slicing and segmenting the first image to form a plurality of second images; S2. Identify the plurality of second images based on a pre-trained rock mass structural surface trace recognition model to form a plurality of third images, wherein the plurality of third images are binary images and have predicted traces; S3, stitching and restoring the plurality of third images to form a fourth image, wherein the fourth image is equal in size to the first image, is a binary image, and has a prediction trace; S4. Perform cluster fitting processing on the predicted traces in the fourth image to obtain structural surface traces, wherein the cluster fitting processing includes density-based spatial clustering algorithm processing and Hough space-based straight line fitting.

2. The method for identifying the main structural surface of a tunnel rock mass based on image recognition according to claim 1, characterized in that: The S2 specifically includes: S21, inputting the plurality of second images in a round-robin manner, calling the pre-trained rock mass structural surface trace recognition model for prediction on each second image, and outputting a corresponding predicted image; S22, processing the predicted image based on average value pooling, and then binarizing the processed predicted image to form a trace prediction image; S23, after the round robin is completed, all the trace prediction images corresponding to each second image are obtained, and all the trace prediction images are combined to obtain the third image.

3. The method for identifying the main structural surface of a tunnel rock mass based on image recognition according to claim 1, characterized in that: The pre-trained rock mass structural surface trace identification model is constructed by the following steps, including: obtaining original images as training sets and manually annotating them; Cross entropy is used as the loss function to calculate the loss of each layer and modify the weight; The initial model for identifying the rock mass structural surface trace is trained based on the training set to obtain a trained rock mass structural surface trace identification model.

4. The method for identifying the main structural surface of a tunnel rock mass based on image recognition according to claim 1, characterized in that: The pre-trained rock structure surface trace identification model includes 5 3×3 convolutional layers, 5 corresponding Relu function activation layers, 5 corresponding deconvolutional layers, 1 fully connected layer, 1 1×1 convolutional layer and 1 maximum pooling layer.

5. The method for identifying the main structural surface of a tunnel rock mass based on image recognition according to any one of claims 1 to 4, characterized in that: The S4 specifically includes: S41, setting density-based spatial clustering algorithm parameters, and inputting the fourth image; S42, clustering points belonging to the same trace in combination with the signal area in the fourth image to obtain a classification space composed of several classifications, wherein the centroid point of the classification space is the trace point; S43, cycling through the classification spaces of each classification based on the Hough space to obtain fitting straight lines for all classifications; S44, performing noise filtering on the fitting straight lines of all the classifications to obtain structural surface traces.

6. The method for identifying the main structural surface of a tunnel rock mass based on image recognition according to claim 5, characterized in that: The density-based spatial clustering algorithm parameters include: the domain distance of samples, and the number of samples within the domain distance.

7. The method for identifying the main structural surface of a tunnel rock mass based on image recognition according to claim 5, characterized in that: The S43 specifically includes: S431, performing coordinate transformation on the classification space composed of several classifications to transform it into a polar coordinate Hough space; S432, draw a straight line with the centroid of the classification space under a certain classification, rotate it at a fixed angle, and at the same time take the distance from the centroid of other classification spaces in the classification to the straight line; S433, when the distance does not exceed the set threshold, the number of point sets on the straight line is increased by 1, and when the maximum number of rotations is reached, a straight line with the largest number of point sets is obtained as the fitting straight line for the classification; S434. Repeat steps S432-S433 until fitting straight lines of all categories are obtained.

8. The method for identifying the main structural surface of a tunnel rock mass based on image recognition according to claim 1, characterized in that: The slicing and segmenting the first image specifically includes: obtaining the length and width of the first image, setting the number of slices, and performing slicing and segmenting according to the ratio of the width to the number of slices; In S3, the plurality of third images are stitched together and restored to form a fourth image according to the number of slices and the ratio.

9. The method for identifying the main structural surface of a tunnel rock mass based on image recognition according to claim 1, characterized in that: In the binary image, 0 represents a non-trace, and 1 represents a trace.

10. A device for identifying the main structural surface of a tunnel rock mass based on image recognition, characterized in that: It includes at least one processor and a memory communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the method described in any one of claims 1 to 9.

Images