A tunnel face analysis method based on image data

Abstract

The present invention relates to a tunnel face analysis method based on image data, including quality identification of collected face images, and three-dimensional reconstruction of qualified face images to obtain a three-dimensional point cloud model; obtaining the face area and two-dimensional joint data in the two-dimensional image; extracting the structural surface features of the image using a neural network model, and processing the structural surface feature image to extract the structural surface line data; combining the structural surface projected onto the two-dimensional image by the three-dimensional analysis and the structure extracted by the neural network, and obtaining all the identified face joints through clustering grouping, and drawing a face sketch using a computer. The advantages of the present invention are: geological analysis based on the geometric structure of the face, and ultra-short distance prediction for the face ahead. Combined with three-dimensional geological information, with the help of deep learning and image recognition technology, the structural surface is identified, reducing the difficulty of sketching and the operation time.

Description

Technical Field

[0001] The invention relates to the technical field of tunnel geological exploration, and in particular to a tunnel face analysis method based on image data. Background Art

[0002] "Tunnel excavation, geology first", the primary task of tunnel safety control is to quickly, accurately, comprehensively and timely grasp the tunnel geology. Relevant specifications require that during the tunnel construction process, geological surveys, advanced geological forecasts and face geological sketches should be carried out in sequence until the tunnel is through. However, due to the limitations of survey technology and funds, the geological survey work is relatively rough, and geological omissions occur from time to time, especially in long and large tunnels. The face geological information revealed by tunnel excavation is the most accurate and intuitive, and can be used to make a near-term comprehensive forecast of fault fracture zones, large-scale landslides, coal and gas outbursts, and can also be used to carry out geological sketches, identify the integrity of the rock mass and surrounding rock classification, etc., to provide basic data for the realization of information-based dynamic construction.

[0003] Geological sketching is highly professional and requires tunnel geological engineers to comprehensively identify the geological conditions revealed by the face, geological survey data for tunnel design, on-site supplementary surveys, and geostress test results. However, most sites do not meet their requirements, resulting in incorrect identification results or errors. Face geological work is very time-sensitive and needs to be accurately and quickly recorded after the face slag is completed. Otherwise, the geological information revealed by the face excavation will disappear with the next cycle or the next process. Moreover, on-site face sketching is currently mainly completed through naked eye observation, manual sketching, and paper records, which has low work efficiency, poor analysis accuracy, limited information, and long feedback time.

[0004] At present, the technology for face analysis based on images generally extracts joint features through traditional image processing methods such as image preprocessing, filtering and binarization, which results in many misidentifications and missed identifications, and the effect is not good. In addition, the three-dimensional geometric information and two-dimensional image information are not combined for analysis, and the image information is not fully utilized for combined identification. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art, provide a tunnel face analysis method based on image data, and solve the problems existing in the current tunnel face analysis.

[0006] The object of the present invention is achieved by the following technical solution: a tunnel face analysis method based on image data, the tunnel face analysis method comprising:

[0007] The quality of the acquired tunnel face images is judged, and the qualified tunnel face images are reconstructed into three-dimensional point cloud models;

[0008] The tunnel face model is extracted from the segmentation of the 3D point cloud model data, and the tunnel face excavation contour is projected onto the 2D tunnel face image to obtain the tunnel face area in the 2D image;

[0009] Identify, segment and project the structural surface of the 3D point cloud model to obtain 2D joint data;

[0010] The neural network model is used to extract the structural surface features of the image, and the structural surface feature image is processed to extract the structural surface line data;

[0011] The structural surface projected onto the two-dimensional image by the three-dimensional analysis and the structure extracted by the neural network are analyzed together, and all the identified joints of the tunnel face are obtained through clustering grouping, and a sketch of the tunnel face is drawn by computer.

[0012] Furthermore, the quality identification of the acquired tunnel face image includes:

[0013] Based on the collected tunnel face images at different positions and angles, a mean low-pass filter is used to obtain reference images in the horizontal and vertical directions respectively;

[0014] Obtaining the adjacent pixel value difference images in the vertical and horizontal directions of the image to be evaluated, and taking their absolute values, obtaining the adjacent pixel value difference images in the vertical and horizontal directions of the reference image, and taking their absolute values;

[0015] Compare the difference images between the image to be evaluated and the reference image in the horizontal and vertical directions, and take the absolute values;

[0016] Calculating the sum of the differences between adjacent pixels of the image to be evaluated in the horizontal and vertical directions respectively, and calculating the sum of the differences between the image to be evaluated and the reference image in the horizontal and vertical directions;

[0017] Then, the clarity evaluation parameters of the image to be evaluated in the horizontal and vertical directions are calculated, and images with clarity evaluation parameters higher than a threshold are eliminated.

[0018] Furthermore, the three-dimensional reconstruction of the qualified tunnel face image to obtain a three-dimensional point cloud model includes:

[0019] Extract the feature points in the qualified tunnel face image for pairwise matching, and calculate the scene information of the camera position based on the matching structure and projection theorem;

[0020] The scene information is combined with the original image to obtain the three-dimensional point cloud of the object in the image, and the three-dimensional point cloud is connected into a surface through the Poisson surface reconstruction algorithm to construct a three-dimensional point cloud model.

[0021] Furthermore, the segmentation and extraction of the tunnel face model from the three-dimensional point cloud model data, and the projection of the tunnel face excavation contour onto the two-dimensional tunnel face image, and the acquisition of the tunnel face area in the two-dimensional image include:

[0022] The mileage value of each point is calculated based on the mileage value of the tunnel face combined with the 3D point cloud model and the design line data, and the 3D point cloud model is divided into the primary support part and the side wall tunnel face part according to the mileage value of the primary support and the excavation boundary;

[0023] According to the perpendicular relationship between the normal vectors of the triangular surfaces of the tunnel face and the side wall, the tunnel face part of the side wall is divided into the side wall and tunnel face models;

[0024] The segmented tunnel face model is projected onto a plane perpendicular to the design line data to obtain the analysis contour of the two-dimensional tunnel face, and the contour is refined through spline curves to obtain a smooth outer contour of the tunnel face excavation surface.

[0025] Furthermore, the identification, segmentation and projection of the structural surface of the three-dimensional point cloud model to obtain two-dimensional joint data includes:

[0026] According to the differences in the normal vector orientations of different structural surfaces in the face model, a clustering algorithm is used to segment the structural surface into small closed areas, each with a different normal vector orientation.

[0027] The segmented structural surface area is projected onto a plane perpendicular to the design line data to obtain the joint data of the two-dimensional structural surface contour.

[0028] Furthermore, the extracting of structural surface features of an image using a neural network model includes:

[0029] The collected tunnel face images are annotated with the same identification standard, and the annotated structural surface pixel data is recorded in the form of text. The annotated images, text data and original images are formed into a data set, and the data set expansion method is used to expand the data set.

[0030] An end-to-end neural network is used to train the expanded data set, and the network model obtained after user training is used to predict and extract features from the newly input tunnel face image.

[0031] Furthermore, the step of processing the structural surface feature image to extract structural surface line data comprises:

[0032] According to the characteristic image obtained by the neural network model, a closed area is formed by using the analysis contour of the two-dimensional tunnel face for mask processing, and the image pixel values ​​of the non-analysis area are cleared;

[0033] The preprocessed image is labeled by the conditional random field algorithm, and the energy of the image label is judged by threshold, and the pixel points with probability greater than the preset threshold are retained to obtain the final threshold label image;

[0034] Connected regions are extracted from the threshold label image, and line thinning is performed on the connected regions extracted separately to obtain irregular lines with a single pixel width;

[0035] All pixel intersections are calculated using convolution to connect the separated small pixel connection areas;

[0036] After obtaining the connection result of a single connected area, the connected area extraction step and the pixel intersection calculation step are repeated to obtain the final result of all connected areas after local connection.

[0037] Furthermore, the image marking of the preprocessed image by the conditional random field algorithm includes:

[0038] Set the random variable x i ∈L={l1,l2,…,l L} represents the label of pixel i, and according to x1,x2,…,x N Form a random vector X, N represents the number of pixels in the image;

[0039] Assume that image G = (V, E), where V = {x1, x2, ..., x N}, the global observation is I, and Gibbs distribution is used to transform G = (V, E) through the conditional random field model to

[0040] In the fully connected conditional random field model, the label x i The energy is expressed as in, is a unary energy term representing the classification of pixel i into label x i The energy It is to split the pixels i and j into x i , x j energy.

[0041] Furthermore, the structural surface projected onto the two-dimensional image by the three-dimensional analysis and the structure extracted by the neural network are analyzed together, and all the identified tunnel face joints are obtained by clustering and grouping, and the tunnel face sketch is drawn by computer, including:

[0042] The analysis contour of the two-dimensional tunnel face and the two-dimensional structural surface contour are judged for coincidence, and the analysis contour of the two-dimensional tunnel face and the two-dimensional structural surface contour whose coincidence is less than a preset value are eliminated;

[0043] Map the angles of the lines on the final structural surface to the circumference to obtain the coordinate points of the line positions;

[0044] The coordinate points are input as data and clustered using the Meanshift algorithm. The top three groups with the largest number are taken as retained structural surfaces, and the remaining groups are taken as messy structural surfaces. The joint data of all tunnel faces are obtained, and a sketch of the tunnel face is drawn by computer.

[0045] Furthermore, the tunnel face analysis method further comprises the step of acquiring tunnel face images at different positions and angles by means of a three-dimensional digital scanner before performing quality assessment on the tunnel face images.

[0046] The present invention has the following advantages:

[0047] 1. Fast data collection, one-click shooting, intelligent data collection. It is more convenient, comprehensive and efficient than traditional mobile phone and camera data collection. One data collection only takes two minutes, reducing construction interference and field time.

[0048] 2. Equipped with a special lightweight fill light, which can be remotely operated, one person can complete all field work. It is light, durable, and easy to carry, making field collection simple and safe.

[0049] 3. Rapidly identify the image quality and then perform geological 3D reconstruction. The tunnel face is like the real scene again, and the joints and fissures are clearly visible. Laser ranging is used to obtain the real scale and highly restore the real construction site.

[0050] 4. Perform geological analysis based on the geometric structure of the tunnel face and make ultra-short-distance predictions for the tunnel face ahead. Combined with 3D geological information, deep learning and image recognition technology are used to identify the structural surface, reducing the difficulty of sketching and the time required for operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 is a flow chart of the method of the present invention;

[0052] Figure 2 Schematic diagram of face collection Figure 1 ;

[0053] Figure 3 Schematic diagram of face collection Figure 2 ;

[0054] Figure 4 To show the case to the boss Figure 3

[0055] Figure 5 Comparison chart for annotated tunnel face images;

[0056] Figure 6 This is the effect diagram of the tunnel face feature extraction;

[0057] Figure 7 This is a schematic diagram of the threshold label image of the tunnel face;

[0058] Figure 8 This is a schematic diagram of the tunnel face connected domain results;

[0059] Fig. 9 Schematic diagram of the mapping method for clustering grouping;

[0060] Fig.10 This is the final recognition effect diagram after clustering and grouping. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of protection of the application claimed for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present application. The present invention is further described below in conjunction with the drawings.

[0062] like Figure 1 As shown, the present invention relates to a tunnel face analysis method based on image data, which comprises the following specific steps:

[0063] S1. Collection of face data

[0064] The best time to shoot is after the slag is discharged from the face and before the steel arch is erected. A cycle of construction procedures on site is blasting, slag discharge, mechanical troubleshooting, trolley forward, manual troubleshooting, steel arch support, anchor bolt application, net hanging, and concrete spraying. The advantage of shooting after blasting and slag discharge is that the entire face can be photographed without the steel arch blocking and construction workers interfering. Moreover, after the excavator and manual troubleshooting, the front of the face is relatively safe, and the dust near the face has been discharged through the ventilator, the air quality is good, and the picture has less noise. Considering the safety of the photographer, the quality of the photos, and the time margin, it is most appropriate to collect data after the slag is discharged from the face and before the steel arch is erected; the photographer sets up the equipment on a tripod and keeps the equipment basically level.

[0065] like Figure 2 As shown, fill-in lights are placed 2 to 3 meters away from the face of the palm, and can be placed 1.5 to 3 meters away from the side wall. Adjust the irradiation angle of the fill-in lights to evenly illuminate the face of the palm. During shooting, flashlights or external strong light should be avoided from directly shining on the face of the palm, which will affect the fill-in light effect.

[0066] like Figure 3 As shown, after ensuring that there are no people, shadows, or equipment blocking the tunnel face to the equipment, data collection on the tunnel face is performed. First, set up the equipment on the center line of the tunnel (position 1 in the figure), and let the equipment face the tunnel face to collect data (in order to ensure the integrity of the data, it is recommended to collect more than 3 images from different angles at one location. The equipment control module can control automatic collection), and complete the data collection of one location point. Then collect data from the remaining two locations on the left and right sides, respectively, and the distance between each point should be greater than 3 meters. The distance between the collection equipment and the tunnel face is based on the equipment's ability to collect the entire tunnel face, and it can be moved back and forth according to the size of the tunnel section.

[0067] like Figure 4 As shown, in order to collect data on the excavation side wall, move the device to just below the excavation side wall of the face and collect data in two times. As marked in the figure above, positions 1 and 2 are on the left and right sides of the tunnel centerline, respectively, and the two positions are 1.5 meters to 2.5 meters apart. Place the device and the tripod at 90 degrees, with the camera facing the left or right side wall. The device control module automatically controls the device to shoot, rotate, and measure distance, and records all intermediate data (the automatic shooting of the side wall will face the excavation side wall, rotate and shoot 6-8 times). There must be no obstructions or strong light exposure during the entire shooting process, and it is necessary to avoid equipment shaking and blurring caused by moving without completing the shooting.

[0068] S2. Face image quality assessment

[0069] Based on the collected tunnel face images at different positions and angles, a mean low-pass filter is used to obtain reference images in the horizontal and vertical directions respectively;

[0070] Obtaining the adjacent pixel value difference images in the vertical and horizontal directions of the image to be evaluated, and taking their absolute values, obtaining the adjacent pixel value difference images in the vertical and horizontal directions of the reference image, and taking their absolute values;

[0071] Compare the difference images between the image to be evaluated and the reference image in the horizontal and vertical directions, and take the absolute values;

[0072] Calculating the sum of the differences between adjacent pixels of the image to be evaluated in the horizontal and vertical directions respectively, and calculating the sum of the differences between the image to be evaluated and the reference image in the horizontal and vertical directions;

[0073] Then, the clarity evaluation parameters of the image to be evaluated in the horizontal and vertical directions are calculated, and images with clarity evaluation parameters higher than a threshold are eliminated.

[0074] The value range of the clarity evaluation parameter is 0 to 1. The lower the value, the clearer it is. The maximum value of the clarity evaluation parameter of the image to be evaluated in the horizontal and vertical directions is selected as the clarity evaluation parameter of the image to be evaluated. The threshold value can be set by the user according to the specific situation, generally between 0.3 and 0.5.

[0075] S3. 3D reconstruction of tunnel face images

[0076] Take out the set of tunnel face photos that have been judged as qualified by the quality identification module, and process them through the following steps of the 3D reconstruction module to obtain the 3D point cloud model of the tunnel face:

[0077] S31. Find the feature points in each photo and match them two by two. This step requires precise identification of the local features of the object and fast and accurate matching. In actual shooting, the object may be rotated, scaled, or have brightness changes, so such scenes need to be processed in the algorithm.

[0078] S32. Based on the matching results, the projection theorem is used to calculate the camera position and other scene information; this step is also called Structure from Motion. The main criterion for measuring the results is accuracy. We use the Bundler method based on the Levenberg-Marquardt algorithm to process the face photos. This step can be called Sparse Reconstruction.

[0079] S33. Combine the scene information with the original photo to obtain the 3D point cloud of the object in the photo. With the scene information, multi-view stereo reconstruction can be performed. Since the image processing accuracy is usually high, the amount of calculation in this step is large, and the execution efficiency algorithm is the first priority. In addition to efficiency, the accuracy and completeness of the reconstruction also need to be considered, because these factors determine the quality of the point cloud. In order to indicate the difference from sparse reconstruction, this step is also called dense reconstruction.

[0080] S34. Construct a 3D model based on the 3D point cloud. After obtaining the 3D point cloud of the object surface, use the Poisson Surface Reconstruction algorithm to connect these points into a surface before it can be displayed in the subsequent 3D scene.

[0081] S4. Segmentation of the tunnel face and excavation contour

[0082] Due to the limited viewing angle of the camera used for shooting, it is necessary to shoot from a distance away from the tunnel face to capture the full view of the tunnel face. In this case, the image information of the primary support position will inevitably be captured. At the same time, in order to use the three-dimensional information of the side wall, it is necessary to place the device under the excavation face and rotate it 180° to collect photos of the side wall. The model data obtained after processing by the three-dimensional reconstruction module consists of three parts: primary support, side wall and tunnel face. In order to facilitate data processing, these three parts need to be separated separately through algorithms.

[0083] Before the construction of the tunnel face, accurate design data for the initial support and excavation sections will be given, and the mileage value at the boundary between the initial support and the excavation will also have accurate design data. After 3D reconstruction, the reconstructed 3D point cloud model is placed on the design line data according to the mileage value of the face, and the mileage value of each point can be calculated. The 3D model can be divided into the initial support part and the side wall face part according to the mileage value of the initial support and the excavation boundary.

[0084] Since the normal vector of the triangular surface of the tunnel face is basically parallel to the design line, and the normal vector of the triangular surface of the side wall part is basically perpendicular to the design line, based on this idea, the side wall tunnel face part can be further divided into the side wall and tunnel face models.

[0085] S5. Identify the two-dimensional analysis contour

[0086] The segmented tunnel face model is projected onto a plane perpendicular to the design line data to obtain a two-dimensional analysis contour of the tunnel face. The preliminary contour obtained can be refined using a spline curve to obtain a smoother outer contour of the excavation surface.

[0087] S6. Identify structural surfaces based on 3D models

[0088] The structural surface of the 3D tunnel face point cloud model is identified, and the normal vector orientations of different structural surfaces are quite different. Therefore, the structural surface can be segmented based on the clustering algorithm. The segmentation result is a series of closed small areas, and each small area has a relatively large difference in the approximate normal vector with the surrounding small areas.

[0089] S7. Project the three-dimensional structural surface onto the two-dimensional tunnel face image to obtain the two-dimensional joints of the tunnel face.

[0090] Since the result data provided by the equipment to the user is mainly two-dimensional joint data, it is necessary to project the structural surface area obtained by three-dimensional space identification onto a two-dimensional plane. The two-dimensional structural surface contour can be obtained by projecting small areas of the three-dimensional structural surface onto a plane perpendicular to the line data.

[0091] S8. Extracting structural features of the tunnel face using neural networks

[0092] like Figure 5 As shown in the figure, the collected tunnel face images are annotated. The geological engineer annotates the collected tunnel face images with structural surfaces. The annotation effect is shown in the figure below. The left side is the original tunnel face image, and the right side is the annotated structural surface. The annotated structural surface pixel data is recorded in text form. All tunnel face image annotations need to be based on the same recognition standard. Data sets annotated with different recognition standards will make subsequent model training difficult.

[0093] The annotated images and text data are aggregated and organized into folders together with the original images to form a data set. Then the data set is expanded using the data set expansion method. The specific data expansion method is as follows:

[0094] A1. Randomly crop a part of the original image. The cropping size should not be too small. After the acquisition is completed, manual screening is required to remove the tunnel face images without structural surface markings.

[0095] A2. The image is flipped horizontally, vertically or mirrored;

[0096] A3. Set the floating angle range within 30 degrees to randomly rotate the image;

[0097] A4. Adjust the brightness or contrast of the image, make it brighter or darker, increase or decrease the contrast; adjust the chroma, change the ratio of the R, G, and B color components; adjust the saturation of the image.

[0098] The above steps can be used to perform any permutation and combination, and the original image can be expanded in a random order. The labeled image data also needs to be processed together during the expansion process to maintain consistency.

[0099] The expanded dataset is trained using an end-to-end neural network. In this method, the classic model of image segmentation is used for training. The image is scaled to 800×800 pixels and input into the model. The convolutional neural network and supervised learning are used. Through back propagation, the stochastic gradient descent algorithm is used to minimize the cost function.

[0100] like Figure 6 As shown, the trained network model is used to predict and extract features from the newly input tunnel face image.

[0101] S9. Perform subsequent processing on the structural surface feature image, extract the structural surface line data as the main direction and connect them

[0102] S91, feature image preprocessing, performing mask processing on the feature image obtained in step S8 by using the analysis contour obtained in step S5 to form a closed area, and clearing the image pixel values ​​in the non-analysis area;

[0103] S92. Use a conditional random field (CRF) algorithm to label the image. The specific labeling method is as follows:

[0104] Set the random variable x i ∈L={l1,l2,…,l L} represents the label of pixel i, and according to x1,x2,…,x N Form a random vector X, N represents the number of pixels in the image;

[0105] Assume that image G = (V, E), where V = {x1, x2, ..., x N}, the global observation is I, and Gibbs distribution is used to transform G = (V, E) through the conditional random field model to

[0106] In the fully connected conditional random field model, the label x i The energy is expressed as in, is a unary energy term representing the classification of pixel i into label x i The energy It is to split the pixels i and j into x i , x j energy.

[0107] A threshold is determined for the obtained E(x), and pixels with a probability greater than 0.5 are retained, and the following is obtained: Figure 7 The final thresholded label image is shown.

[0108] S93, extracting connected regions from the threshold label image, and then performing line thinning processing on the connected regions extracted separately to obtain irregular lines with a single pixel width;

[0109] S94. Calculate all pixel intersections using convolution. Use matrix convolution to calculate the refined image. Set a 3×3 convolution kernel with all values ​​set to 1. Perform convolution on the refined image to obtain a processed image. Then find the position where the sum of the products is greater than 3. This position is the intersection of the lines. This method has greater advantages in operating efficiency than the original traversal query method and can save computing resources more effectively.

[0110] S95, connect the separated small pixel connection areas. All the intersections of the connected area lines have been obtained in step S93, and it is easy to disconnect from the intersection positions to finally obtain all the sub-connection area lines. This step needs to refer to the overall angle of the structural surface obtained in step S7, and use this angle to guide the connection.

[0111] S96, after obtaining the connection result of a single connected area, repeat steps S93 to S95, and finally obtain the following Figure 8 The final result of all connected domains after local connection is shown.

[0112] S10: Eliminate close structural surfaces and perform closure identification

[0113] The degree of overlap between the structural surface obtained in step S7 and the structural surface obtained by the two-dimensional image analysis is judged; specifically, the minimum closed rectangle overlap rate of the two is calculated, and combined with the absolute difference between the two angles, it can be determined that the two overlap, and the areas with smaller closed areas or less than a preset threshold are eliminated.

[0114] S11. Provide the identified joint data of the tunnel face and draw a sketch of the tunnel face using computer graphics

[0115] like Fig. 9 As shown in Figure 1, the Meanshift algorithm is used to group the final structural surface lines obtained by S10. Here, the angle of the line needs to be mapped to the perimeter of the circle to solve the problem that 0 degrees and 360 degrees overlap and cannot be simply clustered using Euclidean distance.

[0116] The angle based on the unified reference coordinate axis is calculated using the formula θ = arctan[(y2-y0) / (x2-x0)]-arctan[(y1-y0) / (x1-x0)], and then the angle is calculated using the formula P(x,y) = (sin(θ 2 ),cos(θ 2 )) Map the angle to the coordinate point on the circle to get the position;

[0117] The coordinate points are used as data input, and the Meanshift algorithm is used for clustering. The top three groups with the largest number are taken as the retained structural surfaces, and the remaining groups are taken as the messy structural surfaces. After grouping, the most red recognition effect is obtained as follows Fig.10 As shown, a sketch of the tunnel face is drawn by computer.

[0118] The above is only a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications and environments, and can be modified within the scope of the concept described herein through the above teachings or the technology or knowledge of the relevant field. The changes and modifications made by those skilled in the art shall not deviate from the spirit and scope of the present invention, and shall be within the scope of protection of the claims attached to the present invention.

Claims

1. A tunnel face analysis method based on image data, characterized in that: The tunnel face analysis method comprises: The quality of the acquired tunnel face images is judged, and the qualified tunnel face images are reconstructed into three-dimensional point cloud models; The tunnel face model is extracted from the 3D point cloud model data segmentation, and the tunnel face excavation contour is projected onto the 2D tunnel face image to obtain the tunnel face area in the 2D tunnel face image; Identify, segment and project the structural surface of the 3D point cloud model to obtain 2D joint data of the tunnel face area; The neural network model is used to extract the structural surface features of the tunnel face image, and the structural surface feature image is processed to extract the structural surface line data; The structural surface projected onto the two-dimensional tunnel face image by three-dimensional analysis and the structure extracted by the neural network are combined for analysis, and all the tunnel face joints are identified through clustering grouping, and a sketch of the tunnel face is drawn by computer.

2. The method for analyzing a tunnel face based on image data according to claim 1, characterized in that: The quality identification of the acquired tunnel face image includes: Based on the collected tunnel face images at different positions and angles, a mean low-pass filter is used to obtain reference images in the horizontal and vertical directions respectively; Obtaining the adjacent pixel value difference images in the vertical and horizontal directions of the image to be evaluated, and taking their absolute values, obtaining the adjacent pixel value difference images in the vertical and horizontal directions of the reference image, and taking their absolute values; Compare the difference images between the image to be evaluated and the reference image in the horizontal and vertical directions, and take the absolute values; Calculating the sum of the differences between adjacent pixels of the image to be evaluated in the horizontal and vertical directions respectively, and calculating the sum of the differences between the image to be evaluated and the reference image in the horizontal and vertical directions; Then, the clarity evaluation parameters of the image to be evaluated in the horizontal and vertical directions are calculated, and images with clarity evaluation parameters higher than a threshold are eliminated.

3. The method for analyzing a tunnel face based on image data according to claim 1, characterized in that: The three-dimensional reconstruction of the qualified tunnel face image to obtain the three-dimensional point cloud model comprises: Extract the feature points in the qualified tunnel face image for pairwise matching, and calculate the scene information of the camera position based on the matching structure and projection theorem; The scene information is combined with the original image to obtain the three-dimensional point cloud of the object in the image, and the three-dimensional point cloud is connected into a surface through the Poisson surface reconstruction algorithm to construct a three-dimensional point cloud model.

4. The method for analyzing a tunnel face based on image data according to claim 1, characterized in that: The step of segmenting and extracting the tunnel face model from the three-dimensional point cloud model data, and projecting the tunnel face excavation contour onto the two-dimensional tunnel face image, and obtaining the tunnel face area in the two-dimensional tunnel face image comprises: The mileage value of each point is calculated based on the mileage value of the tunnel face combined with the 3D point cloud model and the design line data, and the 3D point cloud model is divided into the primary support part and the side wall tunnel face part according to the mileage value of the primary support and the excavation boundary; According to the perpendicular relationship between the normal vectors of the triangular surfaces of the tunnel face and the side wall, the tunnel face part of the side wall is divided into the side wall and tunnel face models; The segmented tunnel face model is projected onto a plane perpendicular to the design line data to obtain the analysis contour of the two-dimensional tunnel face, and the contour is refined through spline curves to obtain a smooth outer contour of the tunnel face excavation surface.

5. The method for analyzing a tunnel face based on image data according to claim 1, characterized in that: The method of identifying, segmenting and projecting the structural surface of the three-dimensional point cloud model to obtain two-dimensional joint data of the tunnel face area includes: According to the differences in the normal vector orientations of different structural surfaces in the face model, a clustering algorithm is used to segment the structural surface into small closed areas, each with a different normal vector orientation. The segmented structural surface area is projected onto a plane perpendicular to the design line data to obtain the joint data of the two-dimensional structural surface contour.

6. The method for analyzing a tunnel face based on image data according to claim 1, characterized in that: The method of extracting structural surface features of the tunnel face image by using a neural network model includes: The collected tunnel face images are annotated with the same identification standard, and the annotated structural surface pixel data is recorded in the form of text. The annotated images, text data and original images are formed into a data set, and the data set expansion method is used to expand the data set. An end-to-end neural network is used to train the expanded data set, and the network model obtained after user training is used to predict and extract features from the newly input tunnel face image.

7. The method for analyzing a tunnel face based on image data according to claim 6, characterized in that: The processing of the structural surface feature image to extract structural surface line data comprises: According to the feature image obtained by the neural network model, a closed area is formed by analyzing the feature image using the contour, and the image pixel values ​​of the non-analysis area are cleared; The preprocessed image is labeled by the conditional random field algorithm, and the energy of the image label is judged by threshold, and the pixel points with probability greater than the preset threshold are retained to obtain the final threshold label image; Connected regions are extracted from the threshold label image, and line thinning is performed on the connected regions extracted separately to obtain irregular lines with a single pixel width; All pixel intersections are calculated using convolution to connect the separated small pixel connection areas; After obtaining the connection result of a single connected area, the connected area extraction step and the pixel intersection calculation step are repeated to obtain the final result of all connected areas after local connection.

8. The method for analyzing a tunnel face based on image data according to claim 7, characterized in that: The image marking of the preprocessed image by the conditional random field algorithm comprises: Set the random variable x i ∈L={l1,l2,…,l L } represents the label of pixel i, and according to x1,x2,…,x N Form a random vector X, N represents the number of pixels in the image; Assume that image G = (V, E), where V = {x1, x2, ..., x N }, the global observation is I, and G = (V, E) is transformed through the conditional random field model using Gibbs distribution; In the fully connected conditional random field model, the label x i The energy is expressed as in, is a unary energy term representing the classification of pixel i into label x i The energy It is to split the pixels i and j into x i , x j energy.

9. The method for analyzing a tunnel face based on image data according to claim 5, characterized in that: The method of combining and analyzing the structural surface projected onto the two-dimensional image by the three-dimensional analysis and the structure extracted by the neural network, and obtaining all the identified joints of the tunnel face by clustering and grouping, and drawing a sketch of the tunnel face by computer includes: The overlap between the analysis contour of the two-dimensional tunnel face and the two-dimensional structural surface contour is judged, and the analysis contour of the two-dimensional tunnel face and the two-dimensional structural surface contour whose overlap is less than a preset value are eliminated; Map the angles of the lines on the final structural surface to the circumference to obtain the coordinate points of the line positions; The coordinate points are input as data and clustered using the Meanshift algorithm. The top three groups with the largest number are taken as retained structural surfaces, and the remaining groups are taken as messy structural surfaces. The joint data of all tunnel faces are obtained, and a sketch of the tunnel face is drawn by computer.

10. A tunnel face analysis method based on image data according to any one of claims 1 to 9, characterized in that: The tunnel face analysis method further comprises the step of collecting tunnel face images at different positions and angles by means of a three-dimensional digital scanner before performing quality identification on the tunnel face images.

Images