Surrounding rock grade identification data preprocessing and labeling method based on tunnel face image

Abstract

The invention belongs to the technical field of data processing, and particularly relates to a surrounding rock grade identification data preprocessing and labeling method based on a tunnel face image. According to the method, high-resolution images and geological survey data are synchronously collected and subjected to standardized preprocessing, a unified data benchmark is determined, core features of rock mass, cracks and underground water are covered through a three-layer structured labeling system, a surrounding rock feature extraction system is perfected, and labeling consistency is improved by means of a unified labeling specification; by means of a deep learning model, automatic feature extraction and 2D-to-3D data fusion quantification are achieved, the labeling and identification efficiency is greatly improved, meanwhile, conversion from pixel features to real physical sizes is achieved, and accurate data support is provided for surrounding rock grade evaluation.

Description

Technical Field

[0001] This application belongs to the field of data processing technology, and more specifically, relates to a method for preprocessing and labeling surrounding rock grade identification data based on tunnel face images. Background Technology

[0002] In the field of tunnel engineering construction, accurate and rapid identification of the surrounding rock grade is the core link to ensure construction safety, optimize support design and control project costs. Traditional surrounding rock grade identification relies on on-site investigation and experience judgment by geological engineers, which has inherent defects such as strong subjectivity, different judgment standards among different personnel and low efficiency, making it difficult to adapt to the intelligent needs of large-scale tunnel construction.

[0003] With the expansion of computer vision and artificial intelligence technologies in the field of smart construction, the use of tunnel face images to achieve automated identification of surrounding rock grade has become an industry trend. However, the training of high-precision intelligent identification models relies on a large amount of standardized and accurately labeled high-quality datasets. At present, the industry lacks a systematic method for collecting, preprocessing, and labeling surrounding rock feature data. Existing labeling methods suffer from problems such as inconsistent data source quality, incomplete feature extraction systems (not covering multiple dimensions of features such as rock mass structure, fracture morphology, and groundwater status), and lack of unified labeling standards. This results in insufficient accuracy and poor consistency of labeled data, which cannot provide reliable data support for intelligent identification models and restricts the implementation and promotion of automated surrounding rock grade identification technology. Summary of the Invention

[0004] This invention provides a method for preprocessing and labeling surrounding rock grade identification data based on tunnel face images, which aims to solve the technical problems mentioned in the background section.

[0005] A method for preprocessing and labeling surrounding rock grade identification data based on tunnel face images includes the following steps: Step 1: Acquire high-resolution images of the tunnel face and synchronous geological survey data. Perform registration and stitching, region of interest extraction, image enhancement and normalization on the high-resolution images to obtain standardized images to be labeled. The synchronous geological survey data includes rock type, rock mass integrity, rock hardness and groundwater development, which serve as a true reference for image labeling. Step 2: Professional annotators, referring to the synchronous geological survey data and unified annotation specifications, perform three-layer structured annotation on the standardized image to be annotated, namely, rock mass structure and lithological area annotation, structural surface and fracture grid annotation, and groundwater exposure status annotation, to form a structured annotation dataset; Step 3: Train a deep learning model based on the structured labeled dataset to achieve automatic identification of rock mass structure and type, automatic segmentation and feature extraction of fractures, and automatic quantitative representation of groundwater distribution. Step 4: Spatial registration and mapping of the labeled 2D image data and 3D point cloud data are performed to convert image pixel features into real physical dimensions. Combined with the spatial geometric analysis results, the rock mass integrity, fracture development degree, groundwater status and rock strength index are considered, and the surrounding rock engineering grade is output through the preset evaluation model.

[0006] This invention establishes a unified data benchmark by simultaneously acquiring high-resolution images and geological exploration data and performing standardized preprocessing. It covers the core features of rock mass, fissures, and groundwater through a three-layer structured annotation system, improving the surrounding rock feature extraction system and enhancing annotation consistency by relying on unified annotation standards. By using a deep learning model to achieve automatic feature extraction and 2D to 3D data fusion and quantization, it significantly improves annotation and identification efficiency, while also realizing the conversion of pixel features to real physical dimensions, providing accurate data support for surrounding rock grade evaluation.

[0007] Preferably, the high-resolution images are acquired after blasting and muck removal at the tunnel face and before initial shotcreting, and the acquisition process uses an industrial-grade camera to take pictures from multiple angles under uniform lighting conditions.

[0008] Preferably, the image enhancement processing includes illumination compensation processing for uneven illumination, noise reduction processing for noise, and contour enhancement processing for crack details; the illumination compensation processing uses histogram equalization or adaptive gamma correction algorithm, the noise reduction processing uses Gaussian filtering algorithm, and the contour enhancement processing uses edge detection algorithm or morphological operation.

[0009] Preferably, the rock mass structure and lithological area markings are delineated using a polygon tool, and the marking categories include intact rock mass, blocky structure, fractured structure, fault fracture zone, weathered area and weak interlayer, used to distinguish areas with different geological origins or degrees of weathering.

[0010] Preferably, the structural surfaces and crack mesh annotations are drawn using an instance segmentation tool or a line segment tool. For cracks and structural surfaces whose boundaries can be clearly identified through images and whose outlines can be delineated, a polygon tool is used to delineate the complete outline. For cracks whose extension direction can only be identified and whose boundary range cannot be clearly defined, line segments are drawn along their extension direction to represent their distribution. The labeling categories are divided into transverse fractures, longitudinal fractures, oblique fractures, network fractures, and random fractures according to the fracture orientation and morphology. Each fracture instance is labeled separately.

[0011] Preferably, the groundwater exposure status is marked by using a polygon or brush tool to delineate the area, and the marking categories are divided into dry, wet, seepage, dripping and flowing water according to the severity of seepage, which are used to characterize the distribution and impact of groundwater.

[0012] Preferably, the deep learning model includes a semantic segmentation model and an instance segmentation model. The semantic segmentation model is used for rock mass structure and type identification and groundwater distribution quantification, while the instance segmentation model is used for fracture segmentation and feature extraction.

[0013] Preferably, the registration and mapping are achieved based on camera intrinsic and extrinsic parameters, converting the pixel dimensions of cracks and fracture zones in the image into their actual physical dimensions in three-dimensional space.

[0014] Preferably, the spatial geometric analysis includes the calculation of the overall flatness of the working face, the comparative analysis with the design outline, and the visualization and volume calculation of over-excavation and under-excavation areas.

[0015] Preferably, the preset evaluation model integrates various indicators through weighted calculation or neural network algorithm, and the evaluation basis conforms to the relevant rock mass classification labeling in the industry, outputting the surrounding rock engineering grade.

[0016] The beneficial effects of this invention include: This invention establishes a unified data benchmark by simultaneously acquiring high-resolution images and geological exploration data and performing standardized preprocessing. It covers the core features of rock mass, fissures, and groundwater through a three-layer structured annotation system, improving the surrounding rock feature extraction system and enhancing annotation consistency by relying on unified annotation standards. By using a deep learning model to achieve automatic feature extraction and 2D to 3D data fusion and quantization, it significantly improves annotation and identification efficiency, while also realizing the conversion of pixel features to real physical dimensions, providing accurate data support for surrounding rock grade evaluation.

[0017] This invention establishes a standardized and reproducible closed-loop system for surrounding rock data processing and annotation, effectively reducing reliance on human experience. The generated high-quality structured dataset can significantly improve the accuracy and generalization ability of the intelligent identification model, providing reliable technical support for the identification of surrounding rock grade and dynamic adjustment of construction support schemes in tunnel engineering, and promoting the application of smart construction in the field of tunnel engineering. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 The overall flowchart provided for embodiments of the present invention. Detailed Implementation

[0020] To make the technical problems, technical solutions, and beneficial effects to be solved by this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of this application.

[0021] See Figure 1 As shown, the method for preprocessing and labeling surrounding rock grade identification data based on tunnel face images includes the following steps: Step 1: Select the window period after blasting and muck removal at the working face and before initial shotcreting to acquire high-resolution images. At this time, the working face is not obscured by the support structure, which can truly reflect the original rock mass state and effectively avoid the initial shotcreting covering key features such as cracks and rock mass structure. The acquisition equipment is an industrial-grade area array camera (such as model: Basler acA2440-75um) to ensure that the image clarity meets the requirements of subsequent detail annotation.

[0022] When acquiring data using the industrial-grade area array camera, multiple angle shots are taken under uniform lighting conditions. For example, four explosion-proof LED supplementary lights are symmetrically arranged around the working face, with the distance between the supplementary lights and the working face maintained at 1.5-2m to avoid overexposure areas or shadow dead corners caused by direct strong light. At the same time, a multi-angle shooting strategy is adopted, taking pictures around the working face at eight directions: 0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315°, taking 2-3 images at each direction to ensure no shooting blind spots and provide complete data support for subsequent image stitching. During shooting, the vertical distance between the camera and the working face is controlled at about 2m, and the angle between the shooting angle and the normal of the working face does not exceed 15° to reduce image distortion.

[0023] Geological data, recorded synchronously on-site by tunnel geological survey engineers, serves as the gold standard truth reference for image annotation, addressing the subjectivity issues of manual annotation, including: Rock types: Through core observation, hardness testing and rapid mineral composition analysis, the rock types of each area (such as granite, medium-grained sandstone, and argillaceous sandstone) are determined and marked on the corresponding areas of the tunnel face; Rock mass integrity: Based on the size of the rock mass blocks and the joint spacing, it is divided into three categories: intact, relatively intact, and broken. The joint development density (strips / m²) is recorded simultaneously. Rock hardness: Tested using a rebound hammer and combined with on-site hammer impact tests, it is divided into three levels: hard (rebound value > 40), relatively hard (rebound value 25-40), and weak (rebound value < 25). Groundwater development status: Visually observe the water-bearing state at the working face, record the location, extent, and severity of seepage, and provide accurate reference for subsequent groundwater labeling.

[0024] All geological data are recorded in standardized tables, synchronously linked to the corresponding image number, shooting time, and location, ensuring a one-to-one correspondence between data and images, facilitating reference for annotators.

[0025] In this embodiment, since multi-angle images are acquired, registration and stitching are required first to obtain a complete panorama of the working face. Therefore, the SIFT (Scale Invariant Feature Transform) algorithm is used to extract feature points from each image, and the FLANN matcher is used to achieve accurate feature point matching and remove mismatched points (using the RANSAC algorithm, 1000 iterations, and an inlier threshold of 2 pixels). Then, image registration is achieved based on the homography matrix, and finally, the linear fusion algorithm is used to stitch the images, eliminate stitching gaps, and generate a complete panorama of the working face with a resolution of 8192×8192 pixels.

[0026] Furthermore, a combination of automatic extraction and manual correction is adopted. First, the threshold segmentation algorithm (Otsu algorithm) is used to automatically identify the face area and background (equipment, personnel, support structure components, etc.) to generate ROI masks. Then, the annotation personnel manually correct the mask boundaries using professional annotation software (LabelMe) to remove background interference and ensure that the ROI only includes the face rock mass area. The combination of manual correction can make up for the algorithm's error deletion defects, balance efficiency and accuracy, and avoid background interference affecting the accuracy of subsequent annotations.

[0027] Secondly, in this embodiment, an adaptive gamma correction algorithm (gamma value range 0.5-2.0, adaptively adjusted according to the image grayscale distribution) is adopted to address the difference in brightness caused by uneven local illumination at the tunnel face. This avoids excessive noise enhancement and balances the details in both bright and dark areas, making features such as cracks and rock mass boundaries clearly visible in both bright and dark areas. Then, a Gaussian filtering algorithm with 3×3 convolution kernels is used to smooth the image and eliminate Gaussian noise caused by equipment vibration and environmental interference during the shooting process. Then, the Canny edge detection algorithm (threshold range 50-150) is used, combined with dilatation-erosion morphological operations (structural element is 5×5 rectangle) to enhance the contrast of fracture outline and rock mass boundary, making fine fractures easier to identify and providing a clear basis for subsequent fracture labeling.

[0028] The enhanced ROI images are uniformly scaled to 1024×1024 pixels (to adapt to the input size of mainstream deep learning models). The scaling algorithm uses bilinear interpolation to ensure that there is no obvious distortion after the image is scaled. At the same time, the image pixel values ​​are normalized to map the pixel gray values ​​(0-255) to the [0,1] range to eliminate the difference in pixel value magnitude.

[0029] Step 2: Construct a complete labeling system through three layers of structured annotation to solve the problems of incomplete feature annotation and inconsistent standards in existing technologies, and provide structured training data for multi-task deep learning models. The specific annotations for each layer are as follows: First layer: Rock mass structure and lithological zone marking The polygon tool is used for annotation to accurately delineate the boundaries of different rock mass areas, with an annotation error of no more than 2 pixels, ensuring that the boundaries are consistent with the actual rock mass structure. The annotation categories include intact rock mass, blocky structure, fractured structure, fault fracture zone, weathered zone, and weak interlayer, etc. The exemplary definitions and annotation rules for each category are as follows: Complete rock mass: rock type is uniform, no obvious cracks, rock mass block size > 50cm, marked in light gray; Blocky structure: The rock mass is composed of large blocks of rock, with joint spacing of 0.5-2m and block size of 30-50cm, marked in gray-white; Fractured structure: The rock mass is broken into small blocks with joint spacing of 0.1-0.5m and block size of 10-30cm, marked in light brown; Fault fracture zone: There are obvious fault traces, the rock mass is extremely fractured, containing a large amount of debris and muddy filler, marked in dark brown; Weathered area: The rocks are severely weathered, lighter in color, and loose in texture, marked as pale yellow; Weak interlayers: mainly composed of mudstone and shale, with a soft texture and easy disintegration, marked in dark red.

[0030] When labeling, it is necessary to combine synchronous geological survey data. For areas with ambiguous lithological boundaries, the labeling range should be determined by referring to the test results of rock type and hardness to ensure that the labeling is consistent with the true value.

[0031] Second layer: Annotation of structural planes and fracture networks Select the corresponding annotation tool based on the crack characteristics, where: Wide cracks and structural surfaces (width ≥ 5mm, clear and identifiable boundaries): Use the instance segmentation polygon tool to outline the complete contour, and mark it in blue. Create a separate annotation instance for each crack to avoid overlapping annotations. For minute cracks (width < 5mm, only the direction of extension can be identified, but the boundary cannot be defined): use the polyline tool to draw along the direction of the crack, set the line width to 2 pixels, and mark it in cyan. Each line segment corresponds to an independent crack to ensure accurate direction.

[0032] The labeling categories are divided into transverse fractures (0°±15° angle with tunnel axis), longitudinal fractures (90°±15° angle with tunnel axis), oblique fractures (30°-60° or 120°-150° angle with tunnel axis), network fractures (multiple fractures intertwined to form a network), and random fractures (irregular orientation). Each fracture instance must be labeled with a category label simultaneously. After labeling, the system automatically calculates the labeling length and orientation angle of each fracture, providing a basis for subsequent quantitative analysis. In this embodiment, a differentiated labeling method is adopted according to the differences in fracture morphology, taking into account both the contour information of wide fractures and the orientation information of fine fractures to ensure that no fracture features are missed.

[0033] Third layer: Marking of groundwater outcrop status Use the polygon tool (for large areas of seepage) or the brush tool (for small, irregular areas of seepage) to define the water-bearing area. The brush size can be adjusted adaptively according to the seepage area (5-20 pixels). The annotation categories are as follows: Dry: The working face shows no signs of moisture and no marked areas; Moist: The surface is damp, with no obvious water marks, and is marked in light blue. Seepage: There are obvious water marks, but no water droplets form; it is marked in sky blue. Drip: Intermittent dripping, with obvious water droplet traces, marked in dark blue; Flowing water: A continuous flow of water forming a water channel, marked in dark blue.

[0034] When labeling, it is necessary to combine the location and range of seepage recorded in the on-site survey. For damp areas that are not easy to identify in the image, the difference in gray value of the image can be used to help judge (the gray value of water-bearing areas is lower than that of dry areas) to ensure that the labeled range is consistent with the actual groundwater exposure state.

[0035] After annotation is completed, two annotators first cross-check the annotations to verify the accuracy of the annotation boundaries, the accuracy of the categories, and the absence of any missing areas. Then, a senior geological engineer reviews the annotations to correct any deviations (such as misaligned boundaries or misclassified categories) to ensure that the annotation accuracy is ≥95%. The annotated data that passes the inspection is used to generate a structured annotation dataset (in JSON format, associated with the original image path, annotation coordinates, and category labels). Unqualified data is returned to the annotators for re-annotation, forming a closed-loop management system to ensure the quality of the dataset.

[0036] Step 3: Train a deep learning model based on a structured labeled dataset to achieve automated extraction and quantification of surrounding rock features, specifically including: The structured labeled dataset was divided into training, validation, and test sets in a 7:2:1 ratio. The training set was used for learning model parameters, the validation set was used for tuning model hyperparameters (such as learning rate and number of iterations), and the test set was used to evaluate the model's generalization ability. The dataset was divided using random sampling to ensure that the characteristics of each type of surrounding rock, fracture morphology, and groundwater status were evenly distributed across the three sets of data, thus avoiding data bias that could lead to a decline in model performance.

[0037] In this embodiment, a semantic segmentation model and an instance segmentation model are trained collaboratively, as detailed below: Semantic segmentation models: The U-Net model (for rock mass structure and type identification) and the DeepLabV3+ model (for groundwater distribution quantification) were selected, both trained using the PyTorch framework. The U-Net model features an Encoder-Decoder structure, accurately capturing rock mass boundaries and is suitable for segmenting macroscopic geological units at the working face. The DeepLabV3+ model incorporates dilated convolutions, effectively expanding the receptive field and adapting to the segmentation of irregular groundwater distributions. For example, the training parameters are set as follows: The learning rate is 0.001, the number of iterations is 100, the batch size is 8, the optimizer is Adam, and the loss function is the cross-entropy loss function.

[0038] Instance segmentation model: The Mask R-CNN model is selected for crack segmentation and feature extraction. This model can simultaneously detect crack targets and segment instances, accurately distinguishing different crack instances, and outputting crack contour masks and location information to support subsequent crack quantization (length, width). For example, the training parameters are set as follows: The learning rate is 0.002, the number of iterations is 80, the batch size is 4, the optimizer is SGD, and the loss function is a weighted sum of classification loss, bounding box regression loss, and mask loss.

[0039] During training, the model performance (accuracy, recall, intersection-over-union ratio, IoU) is monitored in real time using the validation set. When the validation set metrics do not improve for 10 consecutive rounds, an early stopping strategy is adopted to prevent the model from overfitting, and the model weights with the best performance are ultimately retained.

[0040] The trained model is evaluated using a test set. If it meets the preset requirements, the model is applied; otherwise, it is retrained.

[0041] Step 4: By fusing 2D images with 3D point cloud data, the pixel features are converted into actual physical dimensions. Combined with spatial geometric analysis, quantitative indicators are provided for the evaluation of surrounding rock grade. This specifically includes the following steps: First, camera intrinsic parameters (focal length, principal point coordinates, distortion coefficient) and extrinsic parameters (camera attitude, position) are obtained through camera calibration (using Zhang Zhengyou calibration method). Then, 3D point cloud data (point cloud density 100 points / cm²) of the tunnel face is collected using a 3D laser scanner. The point cloud data and 2D image data are then unified into the same coordinate system (tunnel construction coordinate system).

[0042] Based on camera intrinsic and extrinsic parameters and coordinate system transformation matrix, a mapping relationship between 2D image pixel coordinates and 3D point cloud spatial coordinates is established to realize the conversion from pixel size to real physical size: the pixel length of the fracture and the pixel area of ​​the fracture zone marked on the image are converted into real length (meters) and real area (square meters) through the mapping relationship; for the fracture width, the spatial distance between the rock masses on both sides of the fracture is calculated by combining the depth information of the point cloud data to obtain the real width (millimeters); by fusing camera parameters and point cloud data, the deficiency of 2D image lack of depth information is made up for, and the quantitative characterization of surrounding rock features is realized.

[0043] Overall flatness calculation of the tunnel face: The least squares method is used to fit the three-dimensional point cloud data of the tunnel face to obtain an ideal plane. The distance deviation of each point cloud to the ideal plane is calculated. The flatness of the tunnel face is characterized by the mean deviation (≤5cm is flat, 5-10cm is relatively flat, and >10cm is uneven). The flatness directly reflects the blasting effect and rock mass stability. Over-excavation and under-excavation analysis: The 3D point cloud of the tunnel face is compared with the tunnel design outline (generated based on construction drawings). Through spatial distance calculation, over-excavation areas (point cloud located outside the design outline) and under-excavation areas (point cloud located inside the design outline) are identified. Visual color marking is used (over-excavation areas are marked in red, and under-excavation areas are marked in yellow). At the same time, the over-excavation and under-excavation volumes are calculated through spatial integration algorithm to provide data for subsequent support adjustment and engineering quantity calculation. The spatial geometric analysis results are synchronously linked to the 2D labeled image, realizing the linkage between 2D features and 3D spatial information, and providing comprehensive quantitative indicators for the comprehensive evaluation of the surrounding rock grade.

[0044] Then, by integrating multi-dimensional quantitative indicators through a pre-set evaluation model, the engineering grade of the surrounding rock is output, achieving standardized and intelligent identification. Specifically: Extract 4 core quantitative indicators: Rock mass integrity: Calculated by the ratio of the area of ​​intact rock mass to the total area of ​​the working face, denoted as Kv (Kv = area of ​​intact rock mass / total area of ​​working face); Fracture development degree: The fracture development index F is obtained by weighted summation of fracture density (total fracture length / face area, unit: m / m²), average length, and dominant orientation (the orientation of the fracture with the highest proportion) (weight allocation: density 0.4, average length 0.3, dominant orientation 0.3). Groundwater status: Calculated by weighting the area proportion of each level of seepage zone, denoted as W (dry 0 points, damp 1 point, seepage 2 points, dripping 3 points, flowing water 4 points, W=Σ(area proportion of each level × corresponding score)). Rock strength: Based on the test results of the field rebound hammer, the average value of the rock rebound at the working face is taken and denoted as Rc.

[0045] The preset evaluation model is constructed using a neural network algorithm (e.g., 4 neurons in the input layer, 16 neurons in each of the 2 hidden layers, and 5 neurons in the output layer corresponding to Class IV surrounding rock). The model training data uses historical surrounding rock data (content indicators and corresponding grades, a total of 500 sets) of the tunnel and similar projects. After training, the model parameters are fixed.

[0046] The extracted four indicators, Kv, F, W, and Rc, are input into a preset evaluation model. The model outputs a probability distribution of the surrounding rock grade through nonlinear mapping, and the grade with the highest probability is taken as the final identification result. At the same time, a weighted calculation backup scheme is set up (the weights are determined based on the "Engineering Rock Mass Classification Standard" GB / T 50218-2014, for example: the weight of Kv is 0.3, the weight of F is 0.3, the weight of W is 0.2, and the weight of Rc is 0.2). When the probability output by the neural network model is lower than 80%, the weighted calculation result is used for calibration to ensure the accuracy of the identification.

[0047] Based on this, a geological unfolding map of the tunnel face (marking the rock mass structure, fissure distribution, groundwater area and actual size) and a surrounding rock grade report (including various quantitative indicators, identification grade and stability evaluation) are generated and pushed to the tunnel construction management platform through the data interface.

[0048] The above are merely preferred embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for preprocessing and labeling surrounding rock grade identification data based on tunnel face images, characterized in that, Includes the following steps: Step 1: Acquire high-resolution images of the tunnel face and synchronous geological survey data. Perform registration and stitching, region of interest extraction, image enhancement and normalization on the high-resolution images to obtain standardized images to be labeled. The synchronous geological survey data includes rock type, rock mass integrity, rock hardness and groundwater development, which serve as a true reference for image labeling. Step 2: Professional annotators, referring to the synchronous geological survey data and unified annotation specifications, perform three-layer structured annotation on the standardized image to be annotated, namely, rock mass structure and lithological area annotation, structural surface and fracture grid annotation, and groundwater exposure status annotation, to form a structured annotation dataset; Step 3: Train a deep learning model based on the structured labeled dataset to achieve automatic identification of rock mass structure and type, automatic segmentation and feature extraction of fractures, and automatic quantitative representation of groundwater distribution. Step 4: Spatial registration and mapping of the labeled 2D image data and 3D point cloud data are performed to convert image pixel features into real physical dimensions. Combined with the spatial geometric analysis results, the rock mass integrity, fracture development degree, groundwater status and rock strength index are considered, and the surrounding rock engineering grade is output through the preset evaluation model.

2. The method for preprocessing and labeling surrounding rock grade identification data based on tunnel face images according to claim 1, characterized in that, The high-resolution images were acquired after blasting and muck removal at the tunnel face and before initial shotcreting. The acquisition process used an industrial-grade camera to take multi-angle shots under uniform lighting conditions.

3. The method for preprocessing and labeling surrounding rock grade identification data based on tunnel face images according to claim 1, characterized in that, The image enhancement processing includes illumination compensation for uneven illumination, noise reduction for noise, and contour enhancement for crack details. The illumination compensation uses histogram equalization or adaptive gamma correction algorithms, the noise reduction uses Gaussian filtering algorithms, and the contour enhancement uses edge detection algorithms or morphological operations.

4. The method for preprocessing and labeling surrounding rock grade identification data based on tunnel face images according to claim 1, characterized in that, The rock mass structure and lithological area markings use polygon tools to delineate boundaries. The marking categories include intact rock mass, blocky structure, fractured structure, fault fracture zone, weathered area, and weak interlayer, which are used to distinguish areas with different geological origins or degrees of weathering.

5. The method for preprocessing and labeling surrounding rock grade identification data based on tunnel face images according to claim 1, characterized in that, The structural surfaces and crack mesh annotations are drawn using an instance segmentation tool or a line segment tool. For cracks and structural surfaces whose boundaries can be clearly identified by images and whose contours can be outlined, a polygon tool is used to outline the complete contours. For cracks whose extension direction can only be identified and whose boundary range cannot be clearly defined, line segments are drawn along their extension direction to represent their distribution. The labeling categories are divided into transverse fractures, longitudinal fractures, oblique fractures, network fractures, and random fractures according to the fracture orientation and morphology. Each fracture instance is labeled separately.

6. The method for preprocessing and labeling surrounding rock grade identification data based on tunnel face images according to claim 1, characterized in that, The groundwater exposure status is marked by using polygons or brush tools to delineate the area. The marking categories are divided into dry, wet, seepage, dripping, and flowing water according to the severity of seepage, which are used to characterize the distribution and impact of groundwater.

7. The method for preprocessing and labeling surrounding rock grade identification data based on tunnel face images according to claim 1, characterized in that, The deep learning model includes a semantic segmentation model and an instance segmentation model. The semantic segmentation model is used for rock mass structure and type identification and groundwater distribution quantification, while the instance segmentation model is used for fracture segmentation and feature extraction.

8. The method for preprocessing and labeling surrounding rock grade identification data based on tunnel face images according to claim 1, characterized in that, The registration and mapping are achieved based on the camera's intrinsic and extrinsic parameters, converting the pixel dimensions of cracks and fracture zones in the image into their actual physical dimensions in three-dimensional space.

9. The method for preprocessing and labeling surrounding rock grade identification data based on tunnel face images according to claim 1, characterized in that, The spatial geometric analysis includes the calculation of the overall flatness of the working face, the comparative analysis with the design outline, and the visualization and volume calculation of over-excavation and under-excavation areas.

10. The method for preprocessing and labeling surrounding rock grade identification data based on tunnel face images according to claim 1, characterized in that, The preset evaluation model integrates various indicators through weighted calculation or neural network algorithm, and the evaluation criteria conform to the relevant rock mass classification labeling in the industry, outputting the surrounding rock engineering grade.

Images