Tunnel face geological information acquisition and image processing method based on machine vision

Abstract

The invention discloses a tunnel face geological information acquisition and image processing method based on machine vision, which belongs to the technical field of image processing, and comprises the following steps: resolving the pose of a terminal in real time through a visual inertial odometer, generating augmented reality guide information, and guiding the terminal to move to a standard acquisition point; at each acquisition point, performing multi-stage verification of scene semantic compliance, motion blur and defocus and illumination uniformity on the image flow to ensure the acquisition quality; for tunnel high-dynamic illumination, adopting an adaptive brightness segmentation and multi-resolution fusion technology based on an improved Otsu algorithm to perform high-dynamic range reconstruction on a single-frame image so as to recover details; and finally, performing clustering optimization based on the feature matching similarity, and screening out an optimal image subset for three-dimensional reconstruction. According to the method, the problems of subjective and random acquisition process, many image quality defects, detail loss under extreme illumination and large image data redundancy are solved.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more particularly to a method for acquiring and processing geological information at tunnel faces based on machine vision. Background Technology

[0002] In modern tunnel and underground engineering construction, accurate and timely acquisition and analysis of geological information at the excavation face (working face) is crucial for ensuring construction safety, optimizing support design, and achieving information-based construction. Traditional methods mainly rely on manual on-site sketches and experience-based judgment by geological engineers, which suffers from drawbacks such as strong subjectivity, low efficiency, and inability to meet the requirements of rapid construction. Therefore, digital geological information acquisition technology based on machine vision has emerged. By acquiring high-resolution images of the working face and applying computer vision algorithms, it aims to achieve objective and quantitative analysis of geological parameters such as the attitude, lithology, and weathering degree of structural surfaces, providing data support for engineering decision-making.

[0003] Currently, mainstream technical solutions in this field typically follow a standardized post-processing workflow. First, during construction breaks, on-site personnel use high-definition cameras to acquire image sequences of the tunnel face from multiple directions and angles. After acquisition, the image data is taken out of the tunnel for further processing, including manual selection of valid images and preprocessing such as illumination correction. Then, photogrammetric algorithms such as structure-of-motion reconstructing (SCOR) are used to reconstruct a 3D point cloud or mesh model of the tunnel face. Finally, intelligent algorithms are developed and applied based on the 3D model to automatically identify and extract geological feature parameters.

[0004] However, the existing technological system suffers from a systemic contradiction: a severe disconnect exists between advanced backend analysis algorithms and the frontend data acquisition stage, which lacks quality control. Research typically focuses on optimizing backend algorithms such as 3D reconstruction and geological identification, while assuming the raw image data is idealized. Tunnels suffer from poor lighting conditions, relying on a few high-powered lights to create an extremely high dynamic range of brightness at the tunnel face, resulting in both overexposure in highlights and loss of detail in shadows. Traditional global image preprocessing methods are limited in their effectiveness in such localized, extreme lighting unevenness, and may even introduce noise or damage image realism due to over-enhancement. The physical loss of information during the optical imaging stage is difficult for post-processing algorithms to compensate for. Summary of the Invention

[0005] This invention overcomes the shortcomings of the prior art and provides a machine vision-based method for geological information acquisition and image processing at the tunnel face.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: a method for acquiring and processing geological information at a tunnel face based on machine vision, comprising:

[0007] S1. By fusing visual and inertial measurement data, the spatial pose of the mobile terminal is calculated in real time, and dynamic guidance information is generated to move it to the preset standard acquisition point.

[0008] S2. After reaching each standard acquisition point, perform multi-level quality checks on the real-time image stream and output qualified original images. The multi-level quality checks include: scene semantic compliance check, motion blur and defocus check, and illumination uniformity assessment.

[0009] S3. For qualified original images, perform single-frame high dynamic range reconstruction through adaptive brightness segmentation and multi-resolution fusion processing to output an enhanced image sequence.

[0010] S4. Clustering algorithm based on feature matching similarity is used to optimize the enhanced image sequence and select image subsets;

[0011] S5. The selected subset of images is used as input data for the subsequent 3D reconstruction algorithm of the face of the tunnel.

[0012] In a preferred embodiment of the present invention, in step S1, the real-time calculation of the spatial pose of the mobile terminal adopts the visual inertial odometry method, which integrates visual image features and inertial measurement data, and calculates the six-degree-of-freedom pose of the mobile terminal in real time through a tightly coupled factor graph optimization method; and based on the real-time pose, a virtual guide box is dynamically generated on the mobile terminal display interface to guide the operator to move to the preset standard acquisition point.

[0013] In a preferred embodiment of the present invention, step S1 also includes monitoring the face contour, using a lightweight convolutional neural network model to detect and locate predefined control points on the face contour in real time, and prompting operators to adjust the terminal posture through voice alarms and screen pop-ups to ensure complete coverage of the face.

[0014] In a preferred embodiment of the present invention, in step S2, the scene semantic compliance verification uses a lightweight convolutional neural network model to perform semantic analysis on the real-time image stream, determine whether the image has the face of the working face as the core subject, whether there is large-area occlusion or scene deviation, and determine compliance through continuous frame confidence evaluation.

[0015] In a preferred embodiment of the present invention, in step S2, the motion blur and defocus verification evaluates motion blur by calculating the variance of the Laplacian operator response image of the grayscale image, and evaluates defocus by calculating the gradient energy of the image, and locks the acquisition function based on a preset threshold until the image clarity meets the standard.

[0016] In a preferred embodiment of the present invention, in step S2, the illumination uniformity assessment is carried out by dividing the image brightness channel into local blocks, calculating the average brightness of each block, and calculating the ratio of underexposed and overexposed pixels. Based on the ratio results, the operator is prompted to adjust the lighting equipment.

[0017] In a preferred embodiment of the present invention, step S3, the single-frame high dynamic range reconstruction includes: converting the qualified original image to the YUV color space; performing adaptive luminance segmentation on the luminance channel to generate multiple luminance sub-intervals; calculating adaptive enhancement parameters for each luminance sub-interval and generating a pseudo-exposure image; and fusing the pseudo-exposure image into a high dynamic range luminance image using a multi-resolution fusion algorithm.

[0018] In a preferred embodiment of the present invention, the adaptive brightness segmentation employs an improved Otsu algorithm, which introduces an adaptive recursive termination mechanism based on geological texture complexity and a regional contrast weight factor to dynamically determine the segmentation depth and interval merging.

[0019] In a preferred embodiment of the present invention, in step S4, the clustering algorithm based on feature matching similarity includes: extracting scale-invariant feature transformation features from the image; calculating the similarity between images through feature matching and random sampling consensus algorithms; adopting an adaptive cluster center initialization strategy and dynamically determining the number of clusters; and selecting a representative image subset based on a multi-objective optimization method.

[0020] In a preferred embodiment of the present invention, the multi-objective optimization method adopts the Pareto optimality concept and constructs a multi-dimensional quantitative evaluation system, including geological feature saliency, texture complexity, sharpness, noise level, matching stability and key point spatial distribution indicators, and selects representative images through Pareto front and hypervolume contribution.

[0021] This invention addresses the shortcomings of the prior art and has the following beneficial effects:

[0022] (1) This invention constructs a fully automated spatial positioning system for data acquisition by deeply integrating visual inertial odometry with augmented reality guidance, achieving real-time tracking of terminal pose and automatic locking of standard acquisition points in complex tunnel environments. Based on the tightly coupled factor graph optimization method, high-frequency IMU data and visual features are fused and calculated, replacing the subjective judgment of traditional manual visual framing, and realizing the mathematical standardization of the spatial layout of acquisition points. Compared with the existing experience-based operation mode, this method enables the image sequence to have the optimal spatial geometric relationship, further providing strong constraints for three-dimensional reconstruction.

[0023] (2) This invention integrates semantic understanding, physical imaging analysis and illumination assessment into a unified real-time quality control system by constructing a multimodal quality verification model; wherein a lightweight convolutional neural network realizes pixel-level scene semantic segmentation, and combined with the dual sharpness detection of Laplacian operator and gradient energy, the multi-dimensional parallel verification mechanism can instantly identify composite defects such as dust occlusion and motion blur, solving the quality control problem caused by dynamic changes in the working face. Compared with traditional post-processing schemes, this method can complete quality control on the acquisition site, further avoiding the chain effect of defective images on subsequent processing.

[0024] (3) This invention proposes a texture-aware single-frame high dynamic range reconstruction algorithm. It introduces geological feature complexity as a recursive termination condition through an improved Otsu segmentation method, combined with multi-resolution fusion technology. Targeting the characteristics of tunnel lighting, it achieves a balance between shadow detail enhancement and highlight suppression while maintaining computational efficiency. This overcomes the insufficient adaptability of traditional global enhancement methods in local high dynamic range scenes. Compared with multi-frame HDR technology, this method eliminates ghosting hazards and is more suitable for construction efficiency requirements, ensuring the complete reproduction of subtle features such as rock joints and fissures.

[0025] (4) This invention uses an adaptive single-frame image enhancement algorithm based on a priori illumination model to convert qualified original images to YUV space and perform improved Otsu brightness segmentation and multi-resolution fusion. It introduces texture complexity and contrast weight to dynamically generate pseudo-exposure images and restores high dynamic range details through Laplacian pyramid fusion. This overcomes the technical difficulties of tunnel highlight overexposure and shadow texture compression. Compared with traditional HDR technology, which requires multiple frames to capture and is prone to ghosting and is inefficient, this method achieves efficient detail enhancement on the basis of a single frame and further preserves the original texture of geological features.

[0026] (5) This invention combines the advantages of the aforementioned standardized acquisition, real-time quality verification and single-frame enhancement, and optimizes the image sequence based on the clustering algorithm of feature matching similarity. It redefines the similarity metric based on the actual number of matching points and multi-objective optimization, ensuring that the smallest but most information-rich representative image subset is selected, thereby improving the computational efficiency and accuracy of 3D reconstruction. Compared with traditional clustering methods that rely on pixel-level differences, this method is more in line with the geometric relationship of the tunnel scene. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, 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 recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a perspective structural diagram of a preferred embodiment of the present invention;

[0029] Figure 2 These are examples of tunnel face image captures according to a preferred embodiment of the present invention;

[0030] Figure 3 This is a roadmap of high dynamic range image generation technology according to a preferred embodiment of the present invention;

[0031] Figure 4 This is a pseudo-exposure image of a preferred embodiment of the present invention;

[0032] Figure 5 This is a high dynamic range image according to a preferred embodiment of the present invention. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein. Therefore, the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0035] Application Overview:

[0036] This application focuses on the acquisition and image processing of geological information at tunnel faces based on machine vision. Although existing technologies have achieved the process from two-dimensional images to three-dimensional models and then to geological information interpretation, there are defects in image quality during the acquisition stage. On the one hand, the high dynamic range illumination of tunnels causes irreversible information loss, with overexposure of highlights and compression of shadow textures coexisting, directly destroying the integrity of geological features. On the other hand, open-loop acquisition lacks real-time quality feedback, and defects such as motion blur, defocus, and dust obstruction can only be detected offline outside the tunnel. However, the tunnel face changes rapidly as construction progresses, making re-shooting infeasible.

[0037] The limitations of existing technologies lie in their reliance on conventional post-processing approaches and their attempt to solve the unique challenges of tunnels using general algorithms. From an algorithmic perspective, traditional image preprocessing techniques such as gamma correction and global histogram equalization are primarily designed for general scenes with overall lighting that is either too dark or too bright. Their global pixel adjustment logic cannot adapt to the high dynamic range environment of tunnels, where highlights and shadows coexist. This makes it difficult to recover geological textures in shadowed areas and may even introduce noise due to over-enhancement, violating the requirements for geological logging authenticity. Secondly, when dealing with high dynamic range lighting environments, conventional HDR technology, through exposure bracketing, can extend the dynamic range, but its multi-frame acquisition requirements contradict the extreme efficiency demands of tunnel construction. Furthermore, the synthesis algorithm is prone to producing ghosting in moving scenes, disrupting the continuity of geological textures.

[0038] To address the aforementioned issues, this application employs a multi-dimensional image quality real-time evaluation model to quantitatively analyze key indicators such as sharpness, illumination uniformity, and coverage integrity online. It also combines this with an augmented reality-based real-time human-computer interaction guidance mechanism to provide operators with immediate adjustment instructions. Furthermore, for the high dynamic range lighting environment of tunnels, an innovative single-frame image enhancement algorithm based on a priori illumination model and adaptive tone mapping has been developed, effectively restoring shadow details while maintaining efficiency.

[0039] Exemplary method:

[0040] like Figure 1 As shown, a method for acquiring geological information and processing images at a tunnel face based on machine vision includes the following steps:

[0041] S1. By fusing visual and inertial measurement data, the spatial pose of the mobile terminal is calculated in real time, and dynamic guidance information is generated to move it to the preset standard acquisition point.

[0042] S2. After reaching each standard acquisition point, perform multi-level quality checks on the real-time image stream and output qualified original images. The multi-level quality checks include: scene semantic compliance check, motion blur and defocus check, and illumination uniformity assessment.

[0043] S3. For qualified original images, perform single-frame high dynamic range reconstruction through adaptive brightness segmentation and multi-resolution fusion processing to output an enhanced image sequence.

[0044] S4. Clustering algorithm based on feature matching similarity is used to optimize the enhanced image sequence and select image subsets;

[0045] S5. Use the selected subset of images as input data and input it into the 3D reconstruction algorithm of the tunnel face.

[0046] In the entire process of geological information acquisition and image processing at the tunnel face, step S1 is used to address the subjectivity and uncertainty of manual operation in the existing acquisition process. By using technical means, the acquisition behavior is standardized to ensure that the image sequence has a geometric configuration that meets the requirements of three-dimensional reconstruction.

[0047] In step S1, environmental perception is performed using the inertial measurement unit built into the mobile smart terminal. The inertial measurement unit includes a three-axis accelerometer and a three-axis gyroscope. The sampling frequency is not less than 200 Hz. The inertial measurement unit can continuously measure the motion state of the device and output the angular velocity and linear acceleration data of the mobile smart terminal at a high frequency. In addition, the mobile smart terminal is also equipped with a rear high-resolution image sensor responsible for capturing visual information, with a resolution of not less than 48 million pixels.

[0048] Collecting geological information at the tunnel face faces three challenges: extremely uneven lighting conditions, with areas of strong direct sunlight and dark shadows existing simultaneously; severe dust interference, affecting image quality and the stability of feature extraction; and complex spatial structure, requiring high-precision six-degree-of-freedom pose information.

[0049] Relying solely on visual positioning has significant limitations in tunnel environments: in areas with sparse rock texture, insufficient visual feature point extraction leads to decreased positioning accuracy; during rapid movement, image blurring increases the failure rate of feature tracking. While inertial measurement units (IMUs) can provide high-frequency motion information, they suffer from cumulative errors, resulting in severe pose drift when used alone.

[0050] The vision-inertial tight coupling scheme selected in this invention is an optimized combination designed based on the special requirements of the tunnel environment. Visual data provides an absolute position reference, correcting the accumulated error of the IMU; while IMU data provides stable motion estimation in high-dynamic scenes, compensating for the information loss within the visual sampling interval.

[0051] In areas with rich texture at the tunnel face, visual positioning is the primary method, with the IMU providing motion priors to improve feature matching efficiency. When there is dust interference or drastic changes in lighting, the IMU maintains short-term pose estimation to avoid complete positioning failure. Through tight coupling optimization, the positioning error is controlled within 3cm to meet the accuracy requirements of geological logging.

[0052] During the data collection task, the system will guide the operator to align the mobile smart terminal's camera with the intersection of the left edge of the tunnel face and the tunnel invert arch through both voice prompts and on-screen text guidance. After ensuring that at least 1 / 3 of the effective area of ​​the tunnel face is included in the viewfinder, the first reference image will be taken. The purpose of this reference image is to establish an initial reference coordinate system for subsequent spatial positioning.

[0053] After successful data acquisition, the system uses visual inertial odometry to achieve precise positioning, rather than simply superimposing visual and inertial data. Visual inertial odometry is a method that calculates the position and attitude of the mobile terminal in three-dimensional space in real time by fusing visual image features and inertial measurement data.

[0054] The visual inertial odometry calculation method employs a tightly coupled factor graph optimization approach. It pre-integrates the high-frequency angular velocity and linear acceleration data output by the inertial measurement unit (IMU) to effectively compensate for errors caused by IMU hardware bias, thus forming inertial measurement constraints. The factor graph optimization uses a sliding window size of 10 frames to balance long-term pose constraints with real-time computational load. The IMU pre-integration error threshold is set to 0.05 to ensure that the pose drift rate is below 3 cm / s.

[0055] Meanwhile, visual feature points with corner response values ​​greater than a preset threshold are extracted from the image sequence. The KLT sparse optical flow method is used to track the feature points across frames. When the number of tracked feature points is lower than the threshold or the distribution is uneven, the selection of a new key frame is automatically triggered to construct visual reprojection error constraints.

[0056] By unifying the visual constraints and IMU pre-integration constraints of all keyframes within the sliding window into the factor graph model, and using a nonlinear optimization library to perform batch optimization solutions for the terminal pose and IMU bias within the window, the six-degree-of-freedom pose of the terminal in the tunnel space can be calculated in real time with high accuracy and low latency. This includes three-dimensional position and three-dimensional rotation attitude, namely displacement in the forward, backward, left, right, and up / down directions and rotation in the pitch, yaw, and roll angles. This provides a comprehensive description of the state of the mobile terminal in space.

[0057] Based on the accurate pose acquired in real time, the system dynamically generates and renders a virtual guide box on the terminal display interface. The position, size, and perspective of the guide box are adjusted synchronously according to the real-time changes in the terminal's viewing angle, precisely corresponding to the projection of the next preset standard acquisition point under the current viewing angle.

[0058] The standard acquisition points are systematically laid out along the horizontal direction of the working face with a fixed translation interval. The translation interval is set at 500 mm with a tolerance range of ±50 mm to ensure that the overlap rate between adjacent images is between 60% and 75%.

[0059] When the operator moves the handheld terminal horizontally along the working face, the interface will render the relative deviation between the current pose and the target pose in real time in the form of multi-dimensional feedback, and guide the operator to move accurately through graphical instructions and numerical prompts.

[0060] The graphical instructions use color gradient logic. When the three-dimensional spatial deviation is large, the guide box displays red; when the deviation is moderate, it displays yellow; and when the preset accuracy requirement is met, it displays green. At the same time, the screen will accurately display the remaining translation distance and the deviation values ​​of the three attitude angles in digital form, intuitively guiding the operator to accurately adjust the terminal position and attitude.

[0061] When the Euclidean distance between the real-time pose of the mobile terminal and the target pose is less than 50 mm and the attitude angle deviation is less than 3 degrees, the system determines that the standard acquisition point has been reached, locks the acquisition function, and triggers the subsequent image quality verification process.

[0062] Throughout the guided movement, the system also monitors the tunnel face contour. Employing a lightweight convolutional neural network model specifically trained for tunnel scenarios, the system enables efficient inference on mobile devices. It detects and locates at least three predefined control points on the tunnel face contour in real time: the intersections of the tunnel face with the left, right, and top arch lines of the tunnel. By continuously tracking the pixel coordinates of these three control points, the system ensures that these key points remain within the imaging field of view. If any control point is detected about to move out of the field of view, the system immediately alerts the operator via voice alarm and screen pop-up to adjust the terminal's orientation. This fundamentally ensures that the acquired images completely cover the entire tunnel face area, avoiding the omission of edge geological information.

[0063] The training of the YOLOv5-nano lightweight convolutional neural network model first constructs a dedicated dataset containing tens of thousands of tunnel site images. Each image is precisely labeled with the coordinates of the intersection points of the tunnel face contour with the tunnel invert and crown lines, serving as the detection targets. During training, uniformly scaled RGB images are input into the network, sequentially passing through a Focus slicing layer, a CSPNet cross-stage local network, and an SPPF spatial pyramid pooling layer for feature extraction. Multi-scale feature fusion is achieved through a PANet path aggregation network, ultimately outputting a prediction result containing the target location and category. The training process employs an adaptive momentum optimizer, calculating the CIoU loss between the predicted and ground truth boxes and the classification loss for backpropagation, and applying mosaic data augmentation to improve model robustness.

[0064] like Figure 2 As shown in the figure, an example of a tunnel face image was captured.

[0065] Step S1 transforms the previously experience-dependent and uncertain acquisition process into a standardized and quantifiable deterministic operation. Through high-precision pose calculation using visual inertial odometry, combined with real-time guidance from a virtual bounding box, the acquired image sequence possesses a spatially uniform preset baseline length, providing optimal geometric constraints for subsequent motion reconstruction algorithms. Simultaneously, the facet contour monitoring mechanism avoids omissions or duplications in the acquisition range, ensuring that each image contributes effective information to 3D reconstruction.

[0066] Step S2 serves as a quality control node. Through a multi-level quality verification system, defective images such as semantic anomalies, blurring, out-of-focus images, and lighting imbalances are directly screened and removed at the acquisition site. At the same time, real-time interactive feedback guides operators to make on-site adjustments, solving the problem of the infeasibility of reshooting caused by dynamic changes in the working face.

[0067] Scene semantic compliance verification is a content validity screening mechanism designed for tunnel acquisition scenarios. Through image semantic analysis, it determines whether the acquired image focuses on the tunnel face and whether there are serious interfering factors that affect the identification of geological features. Specifically, it includes whether the proportion of the tunnel face area meets the standard, whether there is large-area dust obstruction or construction tool obstruction, and whether the shooting scene deviates from the tunnel face range. Essentially, it ensures that the image content meets the core requirements of geological information acquisition and avoids invalid scene data from flowing into subsequent processing.

[0068] Motion blur and defocus checks are dual, precise detection methods for image sharpness. Motion blur specifically refers to dynamic motion blur caused by the terminal being unstable during acquisition, while defocus is caused by the camera not focusing on the working face, resulting in global edge blur. The two checks target different types of sharpness defects and use quantitative indicators to evaluate the recognizability of geological textures in the image, avoiding the impact of detail loss on the accuracy of subsequent feature matching.

[0069] Illumination uniformity assessment is a specialized verification for adapting to the high dynamic range lighting environment of tunnels. It analyzes the local and global brightness distribution characteristics of the image to determine whether there is overexposure of highlights leading to the annihilation of details or dark shadows leading to texture compression. This ensures that the image lighting conditions meet the adaptation requirements of subsequent enhancement processing, while avoiding misjudgment of geological features caused by lighting defects.

[0070] The system employs a lightweight convolutional neural network model for scene semantic compliance verification.

[0071] In a preferred implementation, the MobileNetV2 architecture is used, where the top fully connected classification layer is replaced and retrained using a dataset containing more than 20,000 tunnel site images. The dataset covers scene samples with different dust concentrations and occlusion types, enabling rapid pixel-level semantic segmentation of the tunnel face region, background region, and occluded region.

[0072] The images in the dataset are labeled into three categories: valid working face, people or equipment occlusion, and non-working face environment. During the verification process, the video stream captured by the image sensor is fed into the model for real-time inference at a rate of 15 frames per second. The model outputs a confidence score for each frame of the image, indicating whether it belongs to one of the three categories.

[0073] The system determines that the scene is compliant and proceeds to the next verification stage only when the confidence level of the valid face category in ten consecutive frames of images is higher than the threshold. If occlusion or environmental error is detected, the system immediately guides the operator to adjust the framing through voice and visual prompts.

[0074] After passing the scene semantic compliance check, the system immediately performs a quantitative evaluation of the optical sharpness of the image frame, including motion blur and defocus checks.

[0075] The motion blur verification quantization metric is the variance of the Laplacian operator response of the grayscale image. Specifically, a 3x3 Laplacian kernel is used to convolve the image. The resulting image reflects the high-frequency edge information in the original image. The variance of all pixel values ​​in the resulting image is calculated. This variance effectively measures the overall sharpness of the image. A sharp image, due to its rich edge details, will have a larger Laplacian response variance. When the variance value is lower than a preset blur threshold, the system determines that the current image has motion blur caused by hand tremors or terminal movement.

[0076] The defocus verification quantization index is the gradient energy of the image. The Sobel operator is used to calculate the first-order gradient of the image in the horizontal and vertical directions, and then the sum of squares of the gradient magnitude of each pixel is calculated. The sum of squares represents the overall gradient energy of the image. The gradient energy of a sharp image is significantly higher than that of a blurry image. When the sum of squares is lower than the preset focus threshold, the image is determined to be out of focus. The system will lock the acquisition function until the Laplacian variance of 5 consecutive frames is greater than the blur threshold and the gradient energy is greater than the focus threshold.

[0077] Laplacian variance is sensitive to changes in the second derivative of an image, effectively capturing edge blurring caused by motion. Choosing Laplacian variance as the motion blur assessment metric is based on a dual consideration of tunnel image characteristics and geological feature protection. The Laplacian operator, a second-order differential operator, is particularly sensitive to edge and detail information in images, and the rock mass joints and fissures relied upon for geological logging are precisely reflected through edge information. Gradient energy, by calculating the overall gradient magnitude of the image, reflects the richness of detail, avoiding the loss of rock mass texture due to defocus. The combination of these two methods covers both dynamic blur and optical defocus defects, while meeting real-time computational efficiency requirements.

[0078] Finally, an assessment of illumination uniformity was conducted. Since tunnel interior lighting typically uses a few high-power searchlights, it is very easy to cause extremely uneven brightness distribution on the tunnel face, that is, there are large areas of dead black areas and areas of overflowing highlights at the same time. Both of these situations will lead to the permanent loss of geological texture information.

[0079] The assessment of illumination uniformity in tunnel lighting environments employs a block-based statistical method, designed based on the actual distribution characteristics of the tunnel face lighting.

[0080] Step S2 uses the logic of local region analysis rather than global evaluation;

[0081] First, we make a preliminary assessment by analyzing the histogram of the image's luminance channel. Specifically, we convert the image to the HSV color space to extract the luminance channel, then divide it into 16×16 local blocks, calculate the average luminance of each block, and count the proportion of pixels with luminance values ​​below 30 to the total number of pixels, as well as the proportion of pixels with luminance values ​​above 225 to the total number of pixels.

[0082] When the proportion of pixels with a brightness value lower than 30 is higher than the threshold, the image is judged to have severe underexposure and loss of detail in dark areas; when the proportion of pixels with a brightness value higher than 225 is higher than the threshold, the image is judged to have severe overexposure and loss of detail in bright areas.

[0083] The average brightness statistics of each local block can identify local overexposed or underexposed areas, while the global underexposed / overexposed pixel ratio calculation ensures the rationality of the overall lighting conditions.

[0084] In any case, the system will determine that the current lighting conditions are extremely poor and will prompt the operator to turn on the mobile terminal's built-in fill light or adjust the position of external lighting equipment to improve the uniformity of lighting.

[0085] Once the image passes all three levels of verification, the system selects the frame with the highest comprehensive score among the three frames as the qualified original image, marks it, and stores it in the local cache. If any verification fails in three consecutive acquisitions, the system will fine-tune the pose parameters of the current standard acquisition point, restart the S1 guidance process, and try to acquire the image again.

[0086] Only when an image frame has passed all the checks of the three verification modules—scene semantic compliance, motion blur and defocus, and illumination uniformity—in sequence without interruption, does the system finally confirm that the optical imaging conditions at that moment have reached the optimal standard. At this point, the system will automatically execute a single image capture command and store this high-quality original image, which has undergone rigorous screening and verification, in the device's local cache, awaiting further processing.

[0087] Step S3, based on the successful selection of qualified original images in Step S2, constructs a single-image high dynamic range information recovery method based on adaptive brightness region segmentation and multi-resolution fusion, specifically addressing the unique high dynamic range lighting environment of tunnel construction sites. Through computational photography, it recovers and enhances geological texture details lost due to extreme lighting conditions from a single standard dynamic range image, particularly solving the technical challenges of detail annihilation in shadow areas and overexposure in highlight areas.

[0088] In step S3, the qualified original image output in step S2 is converted from the RGB color space to the YUV color space.

[0089] The RGB color space is an additive color model based on the three primary colors of red, green, and blue, while the YUV color space separates luminance information from chrominance information. The Y component represents luminance information, while the U and V components represent blue and red color differences, respectively. The purpose of this color space conversion is to focus subsequent processing on the luminance channel, which carries the main detail information, avoiding unnecessary interference with the chrominance channel and thus maintaining the natural color balance of the image.

[0090] The conversion formula is based on international standards and maps RGB values ​​to YUV space. Given the RGB value (R, G, B) of a pixel, its YUV components are calculated as follows: The coefficients are weighted based on the human eye's sensitivity to brightness and color. After conversion, the system only processes the Y channel, while the U and V channels are temporarily stored for later inverse conversion. This conversion not only reduces computational complexity but also ensures color fidelity, laying the foundation for brightness zoning and enhancement.

[0091] like Figure 3 As shown, after completing the color space conversion, the system performs adaptive segmentation of the luminance channel Y using an iterative luminance region based on the maximum inter-class variance.

[0092] The Otsu algorithm, also known as the maximum inter-class variance method, is an automatic threshold selection method based on the histogram distribution of an image. Its principle is to find a threshold to divide image pixels into two classes, thereby maximizing the variance between the two classes of pixels and achieving the best separation effect.

[0093] The formula for calculating between-class variance is: ,in, For candidate thresholds, These represent the weights of the two types of pixels, i.e., the percentage of pixels in the total number of pixels. The average brightness of the two types of pixels is maximized. The optimal threshold can then be obtained.

[0094] While the traditional Otsu's algorithm can automatically determine the segmentation threshold, its core principle of maximizing inter-class variance faces technical challenges in tunnel face image processing. The non-uniformity of tunnel illumination leads to a significant multimodal characteristic in image brightness distribution, making it impossible for a single global threshold to effectively distinguish all important brightness ranges. Secondly, the face region often contains large areas of rock masses with similar brightness, making the traditional Otsu's algorithm prone to bias, confusing important geological feature areas with the background. Most importantly, traditional recursive segmentation methods lack specific consideration for geological texture features, potentially resulting in insufficient segmentation in areas rich in critical geological information.

[0095] This method improves upon the traditional Otsu algorithm;

[0096] Based on the traditional Otsu algorithm, an adaptive recursive termination mechanism based on geological texture complexity is introduced. The system evaluates the texture richness index within the newly generated sub-intervals after each segmentation. Texture richness is quantified by calculating the local variance and entropy value of the image within the sub-interval. The local variance reflects the dispersion of pixel values, while the entropy value measures information content based on histogram distribution. Specifically, for a sub-interval image patch, its local variance is calculated as follows: ,in, This is the pixel brightness value. The average brightness within the block. This represents the number of pixels.

[0097] Entropy calculation is based on the brightness histogram: ,in, Let be the probability of brightness level j. This represents the brightness level. When the entropy or local variance of a sub-interval exceeds a threshold, segmentation will continue even if the brightness range of that interval is relatively wide; otherwise, segmentation will terminate. This ensures that limited computational resources are prioritized for regions with rich textures.

[0098] Furthermore, a regional contrast weighting factor was incorporated into the threshold calculation process. The traditional Otsu algorithm only considers inter-class variance, while the improved algorithm also considers local contrast features within sub-intervals.

[0099] For each candidate threshold T, the system calculates the average contrast weight of the two sub-intervals. The contrast is calculated by extracting the edge intensity map using the Sobel operator, and then calculating the average edge amplitude within the sub-interval.

[0100] The weighting factor is defined as: ,in, The adjustment factor is used, and EdgeMap is the edge magnitude map. The modified inter-class variance calculation is as follows: .

[0101] For areas containing high-frequency details such as rock joints and fissures, the system lowers the segmentation threshold, allowing these important geological features to be divided into independent processing intervals.

[0102] Due to the localized nature of face lighting, traditional recursive segmentation can generate too many small and similar brightness ranges. After completing the basic segmentation, the system intelligently merges these ranges based on brightness differences and texture similarity, avoiding redundant processing of similar areas while ensuring that different geological features receive appropriate enhancement.

[0103] The merging criteria include: the brightness difference is less than a preset threshold, and the texture similarity is measured by the structural similarity index.

[0104] The structural similarity index is calculated as follows: ,in, The average brightness over the interval Standard deviation, For covariance, A constant. When the SSIM value is higher than the threshold, the intervals are merged to avoid duplicate processing.

[0105] In practice, the system first linearly normalizes the pixel values ​​of the luminance channel and calculates its luminance histogram. For the first time, an improved Otsu algorithm is applied to find an optimal segmentation threshold, dividing the entire luminance range into two sub-ranges. Unlike traditional methods, the system evaluates the geological information content of the two sub-ranges while determining the segmentation threshold. If a sub-range is found to contain particularly rich geological features, it will be prioritized for deeper segmentation.

[0106] Subsequently, the system dynamically determines the further segmentation strategy for each sub-region based on preset information value assessment indicators. The content-value-based segmentation method ensures that limited computing resources are prioritized for processing the image regions with the greatest geological value; the entire segmentation process is recursively performed a maximum of three times, but the actual segmentation depth is adaptively adjusted according to the image content, ultimately generating six to ten brightness sub-regions.

[0107] After completing the brightness partitioning, the system calculates adaptive enhancement parameters for each brightness sub-interval. The enhancement intensity is based not only on the interval average brightness but also on texture complexity and edge density. Texture complexity is jointly evaluated using local variance and entropy, while edge density is calculated using the gradient magnitude map generated by the Sobel operator.

[0108] Considering the unique lighting environment and geological features of tunnels, the enhancement intensity is determined not only based on the average brightness of the interval but also by comprehensively taking into account geological feature indicators such as texture complexity and edge density within the interval. For areas containing important geological structures, appropriate enhancement processing is applied even if the brightness is moderate, ensuring that all valuable geological information is fully displayed.

[0109] To avoid overprocessing pure black areas with no information and blown-out highlights with no detail in the image, the system actively discards the darkest and brightest luminance sub-regions in subsequent processing, focusing the processing core on the target luminance sub-region in the middle.

[0110] Adaptive gamma correction values ​​are calculated for each target brightness sub-region. Gamma correction is a non-linear operation used to adjust image brightness; its mathematical form is that the output equals the input raised to the power of gamma. Traditional gamma correction typically uses a fixed gamma value, while this method dynamically calculates the corresponding gamma value based on the average brightness of each brightness sub-region.

[0111] The determination of the gamma value is linearly related to the average brightness of the corresponding sub-interval, ensuring that dark areas with lower average brightness receive a larger gamma value, resulting in significant brightening and contrast enhancement, while bright areas with higher average brightness receive a smaller gamma value, achieving highlight compression and detail protection.

[0112] like Figure 4 As shown, using different calculated gamma values, the system performs independent gamma transformation operations on the original brightness channel to generate multiple pseudo-exposure images. Each pseudo-exposure image corresponds to the optimal exposure version of a brightness sub-range, essentially a virtual multi-exposure image derived from a single original image. These images reveal the geological texture details within their respective ranges in different ways, providing a foundation for subsequent fusion.

[0113] like Figure 5 As shown, finally, the system executes a multi-resolution fusion algorithm based on the Laplacian pyramid to seamlessly fuse the pseudo-exposure images into a final high dynamic range brightness image.

[0114] Traditional methods use weighted graphs for fusion, but this method employs a multi-scale fusion strategy based on feature selection, avoiding the complexity and subjectivity of weight calculation. The core of the fusion is the Laplacian pyramid, a multi-scale image representation method that effectively captures detailed features at different scales through the difference of Gaussian pyramids.

[0115] Laplace Pyramid Fusion targets the macro- and micro-joint features of tunnel rock masses by separating multi-scale details of the image: enhancing the macro-fracture outline at the bottom of the pyramid and preserving fine texture information at the top, avoiding the smoothing of details caused by traditional weighted fusion; it is particularly suitable for enhancing and reproducing multi-scale fractures in sedimentary rock strata.

[0116] For each pseudo-exposure image, a Gaussian pyramid is first constructed. The Gaussian pyramid is generated by iteratively applying Gaussian blur and downsampling.

[0117] Gaussian blur uses a Gaussian kernel convolution with a standard deviation and a downsampling factor of 2. The number of Gaussian pyramid layers L is dynamically determined based on the image resolution, typically set to 4 to 5 layers.

[0118] In one specific embodiment, the number of pyramid layers is set to 5, which is based on the multi-scale characteristics of the rock mass: the first and second layers deal with the macroscopic joint system with a spacing of >10cm, the third layer deals with the mesoscale fissures with a spacing of 2-10cm, and the fourth and fifth layers protect the microscopic texture with a spacing of <2cm.

[0119] The Laplace Pyramid was obtained by differentiating adjacent layers of the Gaussian Pyramid.

[0120] The upsampling operation uses bilinear interpolation to double the image size. The Laplacian image captures high-frequency detail information at each scale.

[0121] During the fusion process, the system does not use a weight map, but instead selects based on the absolute value of the Laplacian coefficient. For each layer of the pyramid and each pixel location, the fused Laplacian coefficient is obtained by selecting the value with the largest absolute value of the Laplacian coefficient at that location among all pseudo-exposure images; ensuring that the most significant detail features are preserved at each scale and location, avoiding detail smoothing.

[0122] For the lowest layer of the Gaussian pyramid, i.e. the highest resolution layer, the system adopts a direct averaging strategy to maintain the consistency of the basic brightness.

[0123] After fusion, a high dynamic range (HDR) luminance channel is obtained through inverse reconstruction using the Laplacian pyramid. The reconstruction process starts from the coarsest layer, upsampling layer by layer and adding to the next finer Laplacian layer, ultimately yielding the fused luminance image. This multi-scale reconstruction ensures seamless integration of details while maintaining overall image consistency.

[0124] The fused high dynamic range luminance channel is re-merged with the U and V chrominance channels of the original image, and the final RGB format image is obtained through inverse color space conversion.

[0125] After inverse conversion, the system outputs a high dynamic range image, which retains rich geological texture details in both shadow and highlight areas, meeting the analysis needs of tunnel construction sites.

[0126] Step S3 achieves high dynamic range information recovery of a single image through color space conversion, adaptive brightness segmentation, enhanced parameter calculation, pseudo-exposure image generation, and multi-resolution fusion. The method improves upon the traditional Otsu algorithm by introducing texture complexity and contrast weights, and employs a maximum selection fusion rule based on the Laplacian pyramid, thus avoiding the limitations of weighted methods.

[0127] In the 3D reconstruction process, the quantity and quality of the image sequence directly affect computational efficiency and reconstruction accuracy. Step S4, based on the enhanced image sequence generated in step S3, constructs an adaptive clustering optimization method based on improved feature matching similarity.

[0128] The core objective of this method is to select the smallest but most informative and geometrically optimal representative subset of images from a preprocessed image sequence. By redefining the similarity metric between images, the adaptability of traditional clustering algorithms is improved, significantly enhancing subsequent processing efficiency while maintaining reconstruction quality.

[0129] Step S4 redefines the similarity metric between images and improves the adaptability of traditional clustering algorithms. Feature matching similarity no longer depends on simple pixel-level differences or global feature distance, but is based on the number of actual matchable feature points between images. This metric better reflects the geometric relationship of images in three-dimensional space.

[0130] The scale-invariant feature transform (SMT) algorithm is used to extract key points and their descriptors from images. SMT is a local feature description algorithm that is invariant to changes in image scaling, rotation, and brightness. It detects extreme points by constructing a Gaussian pyramid and generates a 128-dimensional descriptive vector by calculating the gradient direction histogram of the region around the key point.

[0131] In tunnel face images, this algorithm can effectively identify stable features such as rock joints and fracture edges, providing a reliable foundation for subsequent matching. Traditional clustering methods have significant limitations in image selection. The K-means algorithm typically groups images based on Euclidean distance, but this metric cannot accurately reflect the visual overlap and geometric consistency between images. In tunnel scenes, image content is highly similar, and simple feature distance can easily lead to clustering results deviating from the actual 3D reconstruction requirements.

[0132] Step S4 improves the clustering process, specifically including:

[0133] Introduce a similarity metric based on matching stability;

[0134] When calculating the similarity between two images, a distance-based feature point matching algorithm is used to establish the initial feature point correspondence. This algorithm systematically calculates the similarity between feature descriptors to provide reliable matching candidates for subsequent geometric verification, and then applies a random sampling consensus algorithm to estimate the fundamental matrix.

[0135] In one specific embodiment, the mechanism for dynamically determining the number of clusters in the feature matching similarity clustering algorithm employs an improved silhouette coefficient method. The average silhouette coefficient is calculated when the number of clusters k ranges from 2 to 10, and the k value that maximizes the silhouette coefficient is selected as the optimal number of clusters.

[0136] Scale-Invariant Feature Transform (SIFT) descriptors are extracted for each image. These descriptors are 128-dimensional vectors that characterize the gradient distribution around keypoints. The matching process evaluates similarity by calculating the Euclidean distance between descriptors. For each descriptor in the reference image, its Euclidean distance to each descriptor in the target image is calculated.

[0137] The nearest neighbor distance ratio (NNDR) criterion is used for matching. For each feature point to be matched, its two nearest neighbor feature points in the target image are found, and the ratio of the nearest neighbor distance to the second nearest neighbor distance is calculated. When this ratio is lower than a preset threshold, the nearest neighbor match is accepted. This mechanism ensures that the matching pair has sufficient discriminative power in the feature space and avoids fuzzy matching.

[0138] Random sampling consistency is a robust parameter estimation method that eliminates false matches through iterative random sampling and model validation. The number of geometrically validated inlier matching pairs is defined as image similarity, and this metric ensures that similarity calculations are based on real three-dimensional geometric constraints.

[0139] Design an adaptive cluster center initialization strategy; traditional methods randomly select initial centers, which can easily lead to clustering results getting trapped in local optima.

[0140] This method analyzes the spatial distribution characteristics of image sequences and prioritizes images with high matching degree with most images as initial centers. The specific process is as follows: calculate the sum of similarity between each image and all other images, and select the top k images with the highest sum as initial cluster centers.

[0141] The optimal number of clusters is dynamically determined. Traditional methods require pre-setting the number of clusters, while this method automatically determines the optimal number based on the matching characteristics of image sequences.

[0142] The system first calculates the average matching density of the entire sequence, and then finds the minimum number of clusters that can retain most of the matching information by analyzing the eigenvalue distribution of the matching matrix. The adaptive mechanism ensures that the clustering results can cover all important perspectives and avoid unnecessary redundancy.

[0143] In the specific implementation process, the input enhanced image sequence is first subjected to feature extraction; each image generates hundreds of key points and their descriptors through the scale-invariant feature transformation algorithm; in order to improve the stability of the feature points, the system performs principal component analysis to reduce the dimensionality of the descriptors, retaining the first 64 principal components, reducing the amount of computation while maintaining the discriminative power; the dimensionality-reduced descriptors are more suitable for real-time processing on mobile devices.

[0144] Next, a matching matrix between images is constructed; for each pair of images in the sequence, a detailed matching analysis is performed; the matching process adopts a two-way consistency check, requiring that the matching of feature points in both directions be consistent; after estimating the fundamental matrix through a random sampling consistency algorithm, the number of interior points that satisfy the geometric constraints is counted; this number not only reflects the visual similarity between images, but also reflects their relative positional relationship in three-dimensional space.

[0145] The clustering process employs an improved iterative optimization algorithm. After the initial cluster centers are determined using the aforementioned adaptive strategy, each image is assigned to the cluster to which the cluster center with the largest number of interior point matches belongs. In each iteration, the cluster center update no longer simply calculates the mean, but instead selects the image within the cluster with the highest average similarity to all other members. This center update strategy based on actual matching degree ensures that the cluster centers are always the most representative images.

[0146] In the image quality assessment stage after cluster convergence, this method adopts a Pareto optimal multi-objective optimization method to avoid the subjectivity of traditional weighted fusion.

[0147] In the image quality assessment stage after cluster convergence, a Pareto-optimal multi-objective optimization framework is used for representative image selection. This method achieves objective image quality evaluation and selection by establishing a multi-dimensional quantitative evaluation system and combining it with mathematical optimization theory.

[0148] The evaluation system includes six core dimensions, which are deeply linked to the core requirements of geological information collection at the tunnel face, forming a comprehensive control over image quality.

[0149] First, a Gaussian pyramid is constructed for the image, consisting of three scale layers: the original image, half-downsampled image, and quarter-downsampled image.

[0150] Calculations were performed at three different scales:

[0151] Macroscale (original) Figure 1 / 4 downsampling), to assess the continuity and clarity of the main joint system;

[0152] Mesoscale (original) Figure 1 / 2 downsampling), to analyze the integrity of the fracture network;

[0153] Microscopic scale (original image) to detect the texture features of rock mineral particles.

[0154] Gradient magnitude maps are calculated at each scale level based on the partial derivatives of the image in the horizontal and vertical directions. For the linear characteristics of rock mass joints, a directional consistency index is calculated by analyzing the histogram of gradient directions, with particular attention paid to the gradient direction range of 45 degrees to 135 degrees, which corresponds to the main strike of the rock mass joints. The directional consistency index is the average ratio of the histogram values ​​for this directional range to the histogram values ​​for all directions across the three scale levels.

[0155] Texture complexity analysis employs the Local Binary Pattern Variance (LCVV) method. The image is divided into 16x16 pixel local regions, and a rotation-invariant uniform LCV is calculated for each region—a texture descriptor robust to rotational changes. The LCVV variance for each local region is then calculated to represent texture richness. The texture complexity metric is defined as the proportion of local regions whose variance exceeds a specific threshold.

[0156] The basic image quality assessment adopts a blind image quality assessment method based on natural scene statistics, which includes three core indicators.

[0157] The sharpness index is calculated by analyzing the energy distribution of an image in the complex wavelet domain and comparing the average ratio of high-frequency coefficients to low-frequency coefficients; it directly reflects the clarity of subtle features such as rock joints and striations; insufficient sharpness will lead to increased feature point matching errors, while high-sharpness images can provide more stable feature constraints for 3D reconstruction.

[0158] The noise level index is estimated using the wavelet contraction method and calculated using the median absolute deviation of the HH subband coefficients of the first-level wavelet decomposition; this avoids noise introduced by tunnel dust and equipment vibration from obscuring the true geological texture, for example, preventing noise from being misjudged as micro-fractures;

[0159] Artifact detection is achieved by analyzing the difference between the gradient statistical distribution of image patches and the prior distribution of natural images. The Kullback-Leibler divergence is used to measure the difference between the distribution of each image patch and the prior distribution, and the average value is used as the artifact index. It can effectively identify false textures introduced by uneven lighting or algorithm processing, ensuring the authenticity of geological logging.

[0160] Feature matching stability assessment is achieved through historical matching data analysis. A matching inlier rate sequence is defined as the sequence of ratios of the number of inlier features to the total number of features in each match. The robust coefficient of variation of the inlier rate is calculated using the interquartile range divided by the median; this statistic is insensitive to outliers.

[0161] Furthermore, the evaluation of matching stability also incorporates temporal consistency and spatial distribution quality metrics. Temporal consistency is evaluated through the smoothness of the matching point trajectories between consecutive frames;

[0162] The Voronoi diagram method is used to evaluate the spatial distribution of key points. First, a Voronoi diagram of the feature points is constructed, and the area of ​​each Voronoi cell is calculated. Then, the coefficient of variation of the area distribution is calculated, which is the ratio of the standard deviation to the mean. A low coefficient of variation indicates a more uniform distribution of feature points.

[0163] Traditional weighted methods require manual setting of weight coefficients, which introduces subjectivity. This method employs the Pareto optimality concept of multi-objective optimization to avoid weight setting. For each image within a cluster, a six-dimensional objective vector is constructed, including a geological feature saliency index, a texture complexity index, a sharpness index, a negative noise level index, a negative matching stability coefficient of variation index, and a negative distribution uniformity index. Among these, noise level, matching stability coefficient of variation, and distribution uniformity are negative indices, and their negative signs transform the problem into a maximization problem.

[0164] Pareto dominance is used to compare the quality of images. For two images, if one image is no less than the other in all objective metrics and is strictly superior to the other in at least one metric, then the image dominates the other. If no image dominates a given image, then that image is a Pareto optimal solution. The Pareto front is the set of all Pareto optimal solutions, constructed using the following algorithm: Initialize an empty set as the Pareto front; for each image within a cluster, if no image in the current Pareto front dominates that image, add that image to the Pareto front and remove all images dominated by it; finally, output the Pareto front.

[0165] When the Pareto front contains multiple non-dominated solutions, a method based on hypervolume contribution is used to select the final representative. Hypervolume is a metric for the quality of a solution set in multi-objective optimization, defined as the volume enclosed by the solution set and a reference point in the objective space. The reference point is defined by the minimum or maximum value of each objective. The hypervolume contribution of each Pareto solution is calculated, i.e., the reduction in hypervolume after removing that solution from the Pareto front. The image with the largest hypervolume contribution is selected as the final representative.

[0166] Finally, the selected subset of images is used as input data and fed into the 3D reconstruction algorithm of the tunnel face.

[0167] In one specific embodiment, to verify the effectiveness of the present invention, systematic tests were carried out in three typical tunnel projects, and the traditional manual data acquisition method, basic visual positioning method and the method of the present invention were compared in all aspects.

[0168] Traditional manual data acquisition methods rely on geological engineers operating professional cameras, and basic visual positioning methods depend solely on visual odometry, without incorporating IMU data or a quality verification step. The method of this invention, however, fully includes the entire process of visual-inertial fusion, multi-level quality verification, and HDR reconstruction.

[0169] Table 1 Systematic Test Results

[0170] Comparison indicators Traditional manual collection methods Basic visual positioning methods Method of the present invention Average acquisition time / min 45 32 18 Qualified image ratio 68% 75% 92% Point cloud reconstruction error / cm 3.5±0.8 4.2±1.2 2.1±0.3 Joint surface dip angle measurement error 5.2° 4.8° 1.8° Joint surface orientation measurement error 7.5° 6.3° 2.3°

[0171] The joint surface dip angle measurement error refers to the deviation between the maximum acute angle between the rock mass fracture surface (joint surface) and the horizontal plane measured by the three-dimensional reconstruction model and the actual dip angle of the joint surface in the real geological environment; this error reflects the accuracy of the technology in revealing key parameters of rock mass stability and potential landslide risk.

[0172] The joint surface orientation measurement error refers to the deviation between the azimuth angle of the intersection line (i.e., the orientation line) of the joint surface measured by the model and its actual geographical azimuth angle; this error affects the accuracy of the analysis of the overall structural morphology and geological structure of the surrounding rock of underground engineering.

[0173] A qualified image refers to an original image that, during the on-site acquisition process at the tunnel face, passes real-time multi-level quality verification and is determined by the system to simultaneously meet strict requirements such as scene semantic compliance, image clarity, and acceptable illumination distribution. The ratio of qualified images to the total number of images successfully stored in a single acquisition task to the total number of all shooting attempts is the qualified image ratio. This indicator measures the intelligence level of the acquisition system and the initial quality reliability of the output data.

[0174] Test results show that, in terms of acquisition efficiency, this invention reduces the average acquisition time from 45 minutes in the traditional method to 18 minutes, improving efficiency by 150%; in terms of image quality, the proportion of qualified images increases from 68% to 92%, and the number of defective images is significantly reduced by 75%; the advantage in 3D reconstruction accuracy is even more significant, with the point cloud reconstruction error of this invention being only 2.1±0.3cm, far superior to the 3.5±0.8cm of the traditional method and the 4.2±1.2cm of the basic visual method. In the critical joint surface attitude measurement, the dip angle measurement error is reduced from 5.2° to 1.8°, and the strike error is reduced from 7.5° to 2.3°, meeting the high-precision requirements of geological logging.

[0175] Based on the preferred embodiments of the present invention described above, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A method for acquiring geological information and processing images at a tunnel face based on machine vision, characterized in that, Including the following steps: S1. By fusing visual and inertial measurement data, the spatial pose of the mobile terminal is calculated in real time, and dynamic guidance information is generated to move it to the preset standard acquisition point. S2. After reaching each standard acquisition point, perform multi-level quality checks on the real-time image stream and output qualified original images. The multi-level quality checks include: scene semantic compliance check, motion blur and defocus check, and illumination uniformity assessment. S3. For qualified original images, perform single-frame high dynamic range reconstruction through adaptive brightness segmentation and multi-resolution fusion processing to output an enhanced image sequence. S4. Clustering algorithm based on feature matching similarity is used to optimize the enhanced image sequence and select image subsets; S5. The selected subset of images is used as input data for the subsequent 3D reconstruction algorithm of the face of the tunnel.

2. The method for acquiring and processing geological information at a tunnel face based on machine vision according to claim 1, characterized in that: In step S1, the real-time calculation of the spatial pose of the mobile terminal adopts the visual inertial odometry method, which integrates visual image features and inertial measurement data, and calculates the six-degree-of-freedom pose of the mobile terminal in real time through a tightly coupled factor graph optimization method; and based on the real-time pose, a virtual guide box is dynamically generated on the mobile terminal display interface to guide the mobile terminal to move to the preset standard acquisition point.

3. The method for acquiring and processing geological information at a tunnel face based on machine vision according to claim 1, characterized in that: In step S1, the face contour is also monitored. A lightweight convolutional neural network model is used to detect and locate predefined control points on the face contour in real time. Voice alarms and screen pop-ups prompt operators to adjust the terminal posture to ensure complete coverage of the face.

4. The method for acquiring and processing geological information at a tunnel face based on machine vision according to claim 1, characterized in that: In step S2, the scene semantic compliance verification uses a lightweight convolutional neural network model to perform semantic analysis on the real-time image stream, determine whether the image has the working face as the core subject, whether there is large-area occlusion or scene deviation, and determine compliance through continuous frame confidence assessment.

5. The method for acquiring and processing geological information at a tunnel face based on machine vision according to claim 1, characterized in that: In step S2, the motion blur and defocus verification evaluates motion blur by calculating the variance of the Laplacian operator response image of the grayscale image and evaluates defocus by calculating the gradient energy of the image. The acquisition function is locked based on a preset threshold until the image sharpness meets the standard.

6. The method for acquiring and processing geological information at a tunnel face based on machine vision according to claim 1, characterized in that: In step S2, the illumination uniformity assessment divides the image brightness channel into local blocks, calculates the average brightness of each block, and calculates the ratio of underexposed and overexposed pixels. Based on the ratio results, the operator is prompted to adjust the lighting equipment.

7. The method for acquiring and processing geological information at a tunnel face based on machine vision according to claim 1, characterized in that: In step S3, the single-frame high dynamic range reconstruction includes: converting the qualified original image to the YUV color space; performing adaptive luminance segmentation on the luminance channel to generate multiple luminance sub-intervals; calculating adaptive enhancement parameters for each luminance sub-interval and generating a pseudo-exposure image; and fusing the pseudo-exposure image into a high dynamic range luminance image using a multi-resolution fusion algorithm.

8. The method for acquiring and processing geological information at a tunnel face based on machine vision according to claim 7, characterized in that: The adaptive brightness segmentation employs an improved Otsu algorithm, introducing an adaptive recursive termination mechanism based on geological texture complexity and a regional contrast weight factor to dynamically determine the segmentation depth and interval merging.

9. The method for acquiring and processing geological information at a tunnel face based on machine vision according to claim 1, characterized in that: In step S4, the clustering algorithm based on feature matching similarity includes: extracting scale-invariant feature transformation features from the image; calculating the similarity between images through feature matching and random sampling consensus algorithms; adopting an adaptive cluster center initialization strategy and dynamically determining the number of clusters; and selecting a representative image subset based on a multi-objective optimization method.

10. The method for acquiring and processing geological information at a tunnel face based on machine vision according to claim 9, characterized in that: The multi-objective optimization method adopts the Pareto optimality concept and constructs a multi-dimensional quantitative evaluation system, including geological feature saliency, texture complexity, sharpness, noise level, matching stability, and key point spatial distribution indicators. Representative images are selected through Pareto front and hypervolume contribution.

Images