Tunnel face structural surface automatic identification method

Abstract

The invention relates to a tunnel face structural surface automatic identification method based on digital photography and deep learning, which comprises the following steps: firstly, carrying out image acquisition on a tunnel face by adopting high-definition shooting equipment, and establishing a high-definition image data set of the tunnel face by means of illumination processing, data enhancement and the like; secondly, tunnel face contour extraction is carried out based on an SAM neural network model, the model is optimized through supervised training, segmentation precision is improved through multiple mask outputs, and model precision is evaluated through an intersection-to-union ratio; and finally, establishing a YOLOv8 neural network model, carrying out adaptive size splitting on the tunnel face image, extracting structural face features, and combining prediction results to form a complete tunnel face fissure image. According to the method, the tunnel face structural surface information can be extracted based on the digital photography technology and the deep learning algorithm, and automatic and efficient recognition of the tunnel face rock mass structural surface can be achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of data acquisition, and particularly relates to an automatic recognition method for tunnel face structural planes based on digital photography and deep learning. Background Art

[0002] During the construction process of tunnel engineering, due to the complex and variable geological conditions through which the tunnel passes, it is often impossible to comprehensively and accurately explore the development of the rock mass in the tunnel body through traditional means such as drilling and geophysical prospecting in the reconnaissance and design stage. The rock mass where the tunnel is located may have irregular structures such as faults, joints, and fissures, and these factors may change during the tunnel construction process, thereby affecting the construction safety and progress of the tunnel. Therefore, as the tunnel construction progresses, especially in the face area, the rock mass conditions need to be monitored and updated in real time, and particular attention should be paid to the changes in the rock mass structural planes at the face. The face is a key area in tunnel construction, and the development of its rock mass structural planes is crucial for the stability of the tunnel surrounding rock, the long-term safety of the tunnel, and the selection of construction methods. Accurately identifying and positioning the rock mass structural planes quickly and in a timely manner can provide real-time feedback on the rock mass conditions to the construction team, so as to dynamically adjust and correct unreasonable construction parameters, and effectively avoid risks or quality problems caused by changes in geological conditions during the construction process.

[0003] At the present stage, traditional methods such as the artificial traverse method and the window measurement method are commonly used to extract rock mass structural plane information, and each structural plane is measured manually. The measurement method is simple and the results are accurate, but there are problems of time-consuming, laborious, and incomplete information extraction in the extraction of traditional core structural plane information, which affects the construction progress; the drilling method is to determine the strike and dip of the structural plane of the core through drilling, but this method has high requirements for the quality of the hole wall, and can only reflect the information around the drilling hole, and it is also necessary to combine geophysical prospecting and other means for interpretation and speculation, with poor accuracy. The three-dimensional laser scanning method has the advantages of automation, high density, and high accuracy, but this method has a large amount of point cloud data, and has the disadvantages of low automation and accuracy, and the laser scanning equipment is expensive and has high requirements for line of sight visibility, which is restricted in many engineering applications. With the development of digital photogrammetry technology, the interpretation of the geometric information of rock mass structural planes by combining digital photogrammetry and image processing technology has been studied, but its efficiency and accuracy still need to be improved. With the rapid development of computer vision technology and the continuous upgrading and renewal of image acquisition equipment, it has become a new idea to quickly obtain the rock mass structural plane information on the surface of the tunnel face through digital photography and machine learning methods. Summary of the Invention

[0004] To solve the above problems, the present invention provides an automatic recognition method for structural planes of tunnel faces based on digital photography and deep learning, which can extract the information of structural planes of tunnel faces based on digital photography technology and deep learning algorithms (SAM neural network and Yolov8 neural network).

[0005] To achieve the above object, the technical solution of the present invention is as follows:

[0006] An automatic recognition method for structural planes of tunnel faces adopts the following steps:

[0007] S1, on-site collection of tunnel face images;

[0008] S2, perform data processing on the on-site collected face images to improve the clarity of the images;

[0009] S3, perform data augmentation on the image set after data processing to improve the generalization ability of the model;

[0010] S4, use the label studio tool to manually annotate the face contours and rock mass fractures in the data set;

[0011] S5, use the image data set labeled in step S4 as the training set and put it into the SAM model for training;

[0012] S6, use a more advanced multi-scale deformable attention transformer as the default pixel decoder, adopt a Transformer decoder, use binary cross-entropy loss and stochastic loss as the mask loss, use the Adam optimizer and step learning rate scheduling, and debug the model to the highest accuracy;

[0013] S7, evaluate the accuracy of the SAM model after debugging by using intersection over union;

[0014] S8, split the face contour images extracted by the SAM model to ensure that the sizes after splitting are consistent;

[0015] S9, establish a YOLOv8 neural network model for automatic recognition of structural planes;

[0016] S10, merge the photos predicted by the YOLOv8 neural network model in step S9 into a complete face fracture map according to the topological relationship.

[0017] Preferably, in step S1, a photographing device is used to photograph the face. The pixel of the photographing device is greater than 20 million, the photographing device is placed 5-10 meters in front of the center of the face, and the photographing is completed before drilling.

[0018] Preferably, in the step S1, a projection lamp is used to improve the light condition in the tunnel. The projection lamp is placed 5 meters away from the heading face, and the lateral position of the projection lamp is adjusted according to the actual light condition in the tunnel.

[0019] Preferably, in the step S2, an automatic gamma correction algorithm is used to adaptively correct the image, and the light clarity is improved by adjusting the light of the image.

[0020] Preferably, in the step S3, the processed high-definition image of the heading face is randomly flipped and rotated to obtain an enhanced image dataset, where the training set: test set: validation set is randomly divided in the ratio of 8:1:1.

[0021] Preferably, in the step S5, the full-batch algorithm is used to solve the energy function of the SAM model during the training process, and the optimal image segmentation result is obtained; when a fuzzy prompt is given, the model outputs multiple valid masks on average, and three masks are used for output during the training to improve the accuracy of image segmentation.

[0022] Preferably, in the step S5, during the training process, the prompt encoder and the mask decoder are prompted to run in a web browser, and a linear combination of the focal loss and the dice loss is used to supervise the mask prediction.

[0023] Preferably, in the step S6, during the model debugging process, 6 MSDeformAttn layers are applied to the feature maps with resolutions of 1 / 8, 1 / 16, and 1 / 32, and a simple upsampling layer with lateral connections is used on the final 1 / 8 feature map to generate a feature map with a resolution of 1 / 4 as the per-pixel embedding.

[0024] Preferably, in the step S6, during the model debugging process, an initial learning rate of 0.0001 and a weight decay of 0.05 are used. The learning rate multiplier 0.1 is applied to the backbone, and the learning rate is decayed to 0.9 and 0.95 fractions of the total number of training steps, multiplied by 10.

[0025] Preferably, in the step S7, the threshold of the intersection over union is 0.8. When it is greater than 0.8, it is considered that the heading face contour predicted by the SAM model is accurate.

[0026] Compared with the prior art, the advantages of the present invention are as follows:

[0027] The operation steps of the automatic recognition method for tunnel face structural planes provided by the present invention are clear, with high systematicness and operability. First, through clear steps such as image acquisition, data processing, model training, and evaluation, the present invention can efficiently and accurately extract the characteristics of rock mass structural planes on the tunnel face. Second, the machine learning method adopted by the present invention, especially the combination based on the SAM neural network and the YOLOv8 neural network, has strong generality and adaptability. This method can adjust the model according to the specific conditions of different tunnel projects, so as to be applicable to various types of tunnel construction environments and has good flexibility. In addition, through the automated feature extraction process, the present invention significantly reduces the time of traditional manual geological sketching. Compared with the manual measurement method, the present invention can quickly obtain accurate information on the face structural planes, thus greatly improving the efficiency of tunnel construction. Especially in the key link of tunnel construction - the face monitoring stage, quickly identifying and positioning changes in rock mass structural planes helps to timely adjust construction parameters, reduce construction risks and potential safety hazards caused by unclear geology, and ensure construction quality. Finally, since this method uses digital photography and deep learning technologies, the equipment cost is relatively low, the operation is simple, and the adaptability is strong, which is convenient for on-site operation and popularization. Through this efficient, accurate, and easy-to-deploy method, the monitoring and management of changes in rock mass structural planes during tunnel construction will be more convenient, providing strong support for the safe and efficient construction of tunnel projects. Description of the Drawings

[0028] Figure 1 is the flow chart of the automatic recognition of the tunnel face structural plane in the embodiment of the present invention;

[0029] Figure 2 is the schematic diagram of on-site acquisition of tunnel face images in the embodiment of the present invention.

[0030] Description of the Reference Numerals:

[0031] 1 - Tunnel face; 2 - Rock mass fracture; 3 - Photographing device; 4 - Projection lamp. Detailed Embodiment

[0032] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art belong to the present invention.

[0033] In addition, the described features, structures, or characteristics may be combined in one or more embodiments in any suitable manner. In the following description, numerous specific details are provided to give a thorough understanding of the embodiments of the present application. However, those skilled in the art will realize that the technical solutions of the present application can be practiced without one or more of the specific details, or other methods, components, devices, steps, etc. may be adopted. In other cases, well-known methods, devices, implementations, or operations are not shown or described in detail to avoid obscuring aspects of the present application.

[0034] Referring to Figure 1-2 , in this embodiment, the method for collecting the tunnel face image is first determined. In view of the characteristics of insufficient light and uneven light in the tunnel, lighting treatment measures are taken for the tunnel face to improve the clarity of the photos. The shooting distance from the tunnel face is determined according to the shooting ability and height of the camera, and the generalization ability of the model is improved through data augmentation methods, thereby establishing a tunnel face image dataset. Then, a SAM neural network model is established, and the model is supervised and trained. By adjusting parameters such as the decoder, loss weights, and hyperparameter settings, the extraction of the tunnel face contour in the digital image is realized. It should be noted that: in view of the different standard cross-sections of tunnels in different highway, railway, and other underground engineering projects, the network model can be classified and trained. Finally, a YOLOv8 neural network model for automatic identification of structural planes is established. In view of the large amount of training data and poor recognition effect when directly training the overall image of the tunnel face, the present invention adaptively splits the tunnel face image into appropriate sizes, and the split images are used as the dataset of the YOLOv8 model. The predicted photos are merged and spliced according to the topological relationship to form a complete tunnel face fracture map.

[0035] The method for automatic identification of structural planes on the tunnel face based on digital photography and deep learning in this embodiment specifically adopts the following steps:

[0036] S1, on-site collection of tunnel face images.

[0037] In this step, referring to Figure 2 , Figure 2 is a schematic diagram of on-site collection of tunnel face images. The tunnel face 1 is photographed using the photographing device 3. The pixel of the photographing device is greater than 20 million. To ensure that the photographing angle and range can cover the entire tunnel face 1 and the angle is directly facing, the photographing device 3 is placed 5 - 10 meters in front of the center of the tunnel face 1. The photographing is completed before drilling. Construction in the tunnel will cause difficulties in image collection, and the photographing should be carried out during construction breaks. The projection lamp 4 is used to improve the light conditions in the tunnel. The projection lamp 4 is placed 5 meters or approximately 5 meters away from the tunnel face, and the lateral position of the projection lamp 4 is adjusted according to the actual light conditions in the tunnel.

[0038] S2, perform data processing on the on-site collected tunnel face images to improve the clarity of the images.

[0039] In this step, the Auto Gamma Correction algorithm can be used to adaptively correct the image, and by adjusting the light of the image, the light clarity can be improved.

[0040] S3. Perform data augmentation on the image set after data processing to improve the generalization ability of the model.

[0041] In this step, the processed high-definition face image of the tunnel face is randomly flipped and rotated to obtain an enhanced image data set, in which the training set: test set: validation set is randomly divided in the ratio of 8:1:1.

[0042] S4. Use the label studio tool to manually annotate the tunnel face contour and rock mass fracture 2 in the data set.

[0043] In this step, the Label Studio tool is used to manually annotate the tunnel face contour and rock mass fracture 2 in the collected tunnel face image data set. By using this tool, the annotator can accurately mark the contour line of the tunnel face and the distribution of fractures in the image, providing high-quality annotation data for the subsequent deep learning model training. Manual annotation is a key step in training a deep learning model, and the accuracy of the annotation data directly determines the effect and accuracy of model training.

[0044] S5. Use the image data set annotated in step S4 as the training set and put it into the SAM model for training.

[0045] In this step, the full-batch algorithm is used during training to solve the energy function of the SAM model and obtain the optimal image segmentation result.

[0046] Specifically, for an output result, if a vague hint is given, the model will average multiple valid masks. To solve this problem, the present invention modifies the model and uses 3 mask outputs during model training.

[0047] Specifically, to improve the driving efficiency of the model, given the pre-computed image embedding, the prompt encoder and the mask decoder can run in a web browser, and the running time on the CPU is about 50 milliseconds.

[0048] Specifically, a linear combination of focal loss and dice loss is used to supervise mask prediction, and the training uses a hybrid prompt segmentation task with geometric prompts. By randomly sampling 11 rounds of prompts in each mask to simulate an interactive setting, the SAM can be seamlessly integrated into the data engine.

[0049] S6. Use the more advanced multi-scale deformable attention transformer MSDeformAttn as the default pixel decoder, adopt the Transformer decoder, use binary cross-entropy loss and stochastic loss as the mask loss, use the Adam optimizer and the step learning rate scheduler, and debug the model to the highest accuracy;

[0050] In this step, according to the debugging process, 6 MSDeformAttn layers are used to apply to the feature maps with resolutions of 1 / 8, 1 / 16, and 1 / 32, and a simple upsampling layer with lateral connections is used on the final 1 / 8 feature map to generate a feature map with a resolution of 1 / 4 as the per-pixel embedding.

[0051] Specifically, according to the debugging process, an initial learning rate of 0.0001 and a weight decay of 0.05 are used. The learning rate multiplier 0.1 is applied to the backbone, and the learning rate is decayed to 0.9 and 0.95 fractions of the total number of training steps, multiplied by 10.

[0052] S7. Use the intersection over union to evaluate the accuracy of the debugged SAM model; considering the difficulty of joint fracture identification, the present invention selects the threshold of the intersection over union (IoU) to be 0.8. When it is greater than 0.8, it is considered that the tunnel face contour predicted by the SAM model is accurate.

[0053] S8. Split the tunnel face contour image extracted by the SAM model to ensure that the sizes after splitting are consistent, so as to ensure the quality of the training set of the YOLOv8 model.

[0054] S9. Establish the YOLOv8 neural network model for automatic structural plane recognition; considering the similarity of neural network model training, the process of model training, debugging, and evaluation will not be elaborated too much in this step. For specific content, refer to steps S5 to S7.

[0055] S10. Merge the photos predicted by the YOLOv8 neural network model in step S9 into a complete tunnel face fracture map according to the topological relationship. Based on this, a high-definition tunnel face image with fracture markings is obtained.

[0056] In summary, the overall steps of the present invention can be divided into three main stages: The first stage is the method for collecting the high-definition image dataset of the tunnel face, including steps S1, S2, and S3. In this stage, high-definition images of the tunnel face are collected by a high-resolution imaging device, and the images are processed by an automatic gamma correction algorithm to improve the image clarity. At the same time, data augmentation methods are adopted to enhance the generalization ability of the model, providing high-quality data support for subsequent model training. The second stage is the establishment of the tunnel face contour extraction model based on the SAM neural network, including steps S4, S5, S6, and S7. In this stage, the Label Studio tool is used to manually annotate the tunnel face contour and rock mass fractures, and the annotated image dataset is used to train the SAM neural network model. Through optimization algorithms and loss functions, the extraction accuracy of the model for the tunnel face contour is improved. Finally, the intersection over union is used to evaluate the model accuracy to ensure the accuracy of the extraction results. The third stage is the establishment of the YOLOv8 neural network model for automatic identification of structural planes, including steps S8, S9, and S10. In this stage, the tunnel face contour images extracted by the SAM model are split to ensure the consistency of the image size for use in YOLOv8 model training, thereby realizing the automatic identification of rock mass structural planes. Finally, the photos predicted by the YOLOv8 model are merged according to the topological relationship to obtain a complete tunnel face fracture map. Through the collaborative work of these three main stages, the present invention can efficiently and accurately achieve the automatic extraction and identification of structural plane information of the tunnel face.

[0057] The present invention provides a method for automatic identification of structural planes of tunnel faces based on digital photography and deep learning. Through a series of steps such as image acquisition, data processing, manual annotation, neural network model training and optimization, etc., the automatic extraction of information on structural planes of tunnel faces (such as rock mass fractures, joints, etc.) is realized. Specifically, by using a high-resolution imaging device to collect images of the tunnel face and combining projection lamps to optimize the lighting in the tunnel, it is ensured that the collected images are clear and comprehensive. The tunnel face contour is extracted by the SAM neural network model, and the YOLOv8 model is used to automatically identify rock mass fractures. The core of this method lies in the precise analysis and processing of tunnel face images through deep learning models, so as to quickly and accurately identify rock mass structural planes and improve the efficiency of obtaining geological information during the construction process. In practical engineering applications, the steps and deep learning models provided by the present invention can be programmed into software programs and installed in a computer network to support real-time data processing and decision-making support at the tunnel construction site, having strong promotion and application value.

[0058] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; under the idea of the present invention, the technical features in the above embodiments or different embodiments can also be combined, and the steps can be implemented in any order, and there are many other variations in different aspects of the present invention as described above. For the sake of brevity, they are not provided in detail; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for automatically identifying a tunnel face structure surface, characterized in that: Use the following steps: S1, on-site acquisition of tunnel face images; S2, data processing of the tunnel face images collected on site to improve the image clarity; S3, performs data enhancement on the image set after data processing to improve the generalization ability of the model; S4, label studio tool was used to manually label the tunnel face contours and rock mass fractures in the dataset; S5, putting the image dataset annotated in step S4 into the SAM model as a training set for training; S6, uses a more advanced multi-scale deformable attention transformer MSDeformAttn as the default pixel decoder, adopts the Transformer decoder, uses binary cross entropy loss and random loss as mask loss, uses the Adam optimizer and step learning rate scheduling, and debugs the model to the highest accuracy; S7, using intersection and union comparison to evaluate the accuracy of the debugged SAM model; S8, splitting the tunnel face contour image extracted by the SAM model to ensure that the sizes of the split images are consistent; S9, establish a YOLOv8 neural network model for automatic structural surface recognition; S10, merging the photos predicted by the YOLOv8 neural network model in step S9 into a complete tunnel face crack map according to the topological relationship.

2. The method for automatic identification of tunnel face structure according to claim 1, characterized in that: In step S1, the tunnel face is photographed by a photographing device, the pixel of the photographing device is greater than 20 million, the photographing device is placed 5-10 meters in front of the center of the tunnel face, and the photographing time is completed before drilling.

3. The method for automatic identification of tunnel face structure according to claim 2, characterized in that: In the step S1, a projection lamp is used to improve the lighting condition in the tunnel. The projection lamp is placed 5 meters away from the tunnel face, and the lateral position of the projection lamp is adjusted according to the actual lighting condition in the tunnel.

4. The method for automatic identification of tunnel face structure according to claim 1, characterized in that: In step S2, an automatic gamma correction algorithm is used to perform adaptive correction on the image, and the light clarity of the image is improved by adjusting the light of the image.

5. The method for automatic identification of tunnel face structure according to claim 1, characterized in that: In step S3, the processed high-definition image of the tunnel face is randomly flipped and rotated to obtain an enhanced image data set, wherein the training set: the test set: the validation set is randomly divided in the manner of 8:1:

1.

6. The method for automatic identification of tunnel face structure according to claim 1, characterized in that: In step S5, a full-batch algorithm is used during the training process to solve the energy function of the SAM model and obtain the optimal image segmentation result; when a fuzzy prompt is given, the model outputs multiple valid masks on average, and uses three masks for output during training to improve the accuracy of image segmentation.

7. The method for automatic identification of tunnel face structure according to claim 6, characterized in that: In step S5, during the training process, the prompt encoder and the mask decoder are run in a web browser, and a linear combination of focal loss and dice loss is used to supervise the mask prediction.

8. The method for automatic identification of tunnel face structure according to claim 1, characterized in that: In step S6, during the model debugging process, 6 MSDeformAttn layers are applied to feature maps with resolutions of 1 / 8, 1 / 16 and 1 / 32, and a simple upsampling layer with lateral connections is used on the final 1 / 8 feature map to generate a feature map with a resolution of 1 / 4 as a per-pixel embedding.

9. The method for automatic identification of tunnel face structure according to claim 8, characterized in that: In step S6, during the model debugging process, an initial learning rate of 0.0001 and a weight decay of 0.05 are used, a learning rate multiplier of 0.1 is applied to the backbone, and the learning rate is decayed to 0.9 and 0.95 fractions of the total number of training steps, multiplied by 10.

10. The method for automatic identification of tunnel face structure according to claim 1, characterized in that: In step S7, the threshold of the intersection-to-union ratio is 0.

8. When it is greater than 0.8, it is considered that the tunnel face profile predicted by the SAM model is accurate.

Images