Karst tunnel lining quality intelligent detection and image analysis system

Abstract

The application discloses a karst tunnel lining quality intelligent detection and image analysis system, relates to the technical field of image processing and defect identification, and comprises a multispectral image acquisition module, a lining defect automatic identification engine and a defect quantitative evaluation platform.The multispectral image acquisition module acquires visible light and near-infrared images and performs fusion enhancement, the lining defect automatic identification engine adopts a double-branch attention feature extraction network to identify multiple types of defects, and the defect quantitative evaluation platform calculates defect geometric features and evaluates the severity, so that the whole process of detection, identification and evaluation decision is automatically realized.The detection precision of the application reaches a sub-millimeter level, the identification accuracy is more than 95%, and the detection efficiency is more than 10 times that of the artificial method.

Description

Technical Field

[0001] This invention relates to the field of tunnel inspection technology, specifically to image processing and defect recognition technology, and particularly to intelligent detection and quantitative assessment of surface and internal defects in tunnel linings in karst areas. Background Technology

[0002] Tunnels in karst areas are prone to various defects during operation due to complex geological conditions, groundwater erosion, and variations in construction quality. These defects include cracks, honeycombing, voids, and water seepage. If these defects are not detected and addressed promptly, they will seriously threaten the structural safety of the tunnel and the safety of traffic. Traditional tunnel lining inspection mainly relies on manual visual inspection and simple instrument measurements, which suffers from low efficiency, strong subjectivity, insufficient accuracy, and difficulty in quantitative assessment, failing to meet the needs of modern tunnel operation and maintenance.

[0003] In recent years, with the development of computer vision and deep learning technologies, image recognition-based tunnel defect detection methods have gradually become a research hotspot. However, existing tunnel lining image processing techniques have significant shortcomings in the following aspects:

[0004] In terms of image acquisition and preprocessing, existing technologies mostly employ single-spectral imaging, which is insufficiently adaptable to the complex lighting environments and diverse types of defects within tunnels. Multispectral image registration and fusion algorithms have limited accuracy, making pixel-level precise alignment difficult. Image enhancement processing typically uses globally uniform parameters, failing to adaptively adjust based on the contrast characteristics of local image regions. This results in insufficient enhancement of low-contrast areas and excessive enhancement of high-contrast areas, impacting the accuracy of subsequent defect identification.

[0005] In terms of feature extraction and fusion, existing deep learning methods mostly employ a single network architecture for feature extraction, making it difficult to simultaneously capture both global semantic information and local detailed features. For different types of defects such as cracks, honeycombs, and cavities, their feature distributions differ significantly. Fixed-weight feature fusion strategies cannot adaptively adjust the fusion ratio of global and local features according to the defect type, limiting further improvements in recognition accuracy.

[0006] In terms of defect geometric feature extraction, existing technologies have limited accuracy in detecting defect edges and extracting contours, making it difficult to achieve sub-millimeter-level geometric dimension measurements. The assessment of defect severity lacks a scientific quantitative indicator system, resulting in highly subjective assessment results. Furthermore, defect identification and quantitative assessment are independent of each other, lacking an effective information feedback mechanism and failing to dynamically optimize the discrimination threshold of the identification model based on assessment results.

[0007] Chinese patent CN118505690A discloses a deep learning-based method for detecting cracks in tunnel lining. This method uses the YOLOv8 algorithm to identify crack locations and the DeepCrack model to identify crack sizes, achieving automatic crack detection and size measurement. However, this method has the following main shortcomings: First, image preprocessing only uses histogram equalization, which is insufficient for the complex lighting and diverse defect types in karst tunnels. Second, crack identification and size measurement are two independent processes, lacking overall coherence and coordination, resulting in low detection efficiency. Third, it only focuses on cracks as a single defect type, lacking the ability to identify defects unique to karst areas such as honeycomb and cavities. Fourth, it lacks a quantitative assessment function for the severity of defects.

[0008] Therefore, it is necessary to develop an image processing and intelligent recognition system for karst tunnel lining to solve image processing technology problems such as multispectral image adaptive fusion enhancement, bi-branch feature adaptive weight fusion, accurate extraction of defect geometric features, and closed-loop feedback for recognition and evaluation, so as to achieve high-precision recognition and quantitative evaluation of multiple types of defects. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of the prior art and provide an intelligent detection and image analysis system for the quality of karst tunnel lining. Through the deep coupling and synergy of multispectral image acquisition, deep learning defect identification and quantitative geometric evaluation, it can achieve high-precision automatic detection and quantitative analysis of surface and internal defects of karst tunnel lining.

[0010] The intelligent detection and image analysis system for karst tunnel lining quality provided by this invention includes a multispectral image acquisition module, an automatic lining defect identification engine, a defect quantitative assessment platform, and a maintenance plan generator.

[0011] The multispectral image acquisition module is used to acquire visible light and near-infrared images of the tunnel lining. The acquired images are registered and fused. Adaptive enhancement parameters are determined based on the local contrast characteristics of the images. The fused images are then subjected to adaptive enhancement processing to generate preprocessed image data, which is then transmitted to the lining defect automatic recognition engine.

[0012] The automatic lining defect identification engine is connected to the multispectral image acquisition module to receive preprocessed image data, construct a dual-branch attention feature extraction network to extract multi-scale features from the preprocessed image, adaptively weight and fuse the extracted global semantic features and local detail features, identify the defect type and locate the defect area through the defect classification head, generate a defect identification result containing defect type, location coordinates and confidence level, and transmit the defect identification result to the defect quantitative assessment platform.

[0013] The defect quantitative assessment platform is connected to the lining defect automatic identification engine to receive defect identification results, perform edge detection and contour extraction on the identified defect areas, calculate the geometric feature parameters of the defects based on the defect contours, determine the severity score of the defects according to the defect type, geometric feature parameters and distribution density, classify and assess the health status of the lining according to the score results, and generate a defect quantitative assessment report.

[0014] Compared with the prior art, the present invention has the following beneficial effects:

[0015] This invention enhances the ability to identify different types of defects by fusing visible and near-infrared images through a multispectral image acquisition module, and improves image quality under complex lighting conditions through adaptive image enhancement processing. By establishing an automatic lining defect identification engine, employing a dual-branch attention feature extraction network and an adaptive weight fusion mechanism, the invention improves the identification accuracy and location accuracy of various defect types such as cracks, honeycombing, and voids. Furthermore, by setting up a quantitative defect assessment platform, the invention enables precise calculation of defect geometric features and scientific classification of lining health status, providing a quantitative basis for maintenance decisions.

[0016] This invention feeds back statistical information on defect features to the automatic lining defect identification engine through a quantitative defect assessment platform, optimizing the discrimination threshold and forming a closed-loop collaborative mechanism for detection, identification, and evaluation feedback. This enables the system to have adaptive learning and continuous optimization capabilities. Deep coupling of data flows between modules achieves full automation from image acquisition, defect identification, to quantitative assessment. Compared to existing technologies that focus only on a single detection step, this collaborative system achieves a non-linear synergy greater than the sum of its parts, achieving sub-millimeter level detection accuracy, over 95% recognition accuracy, and more than 10 times the detection efficiency of manual methods, significantly improving the intelligent level of karst tunnel lining quality inspection. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall structure of the intelligent detection and image analysis system for karst tunnel lining quality of the present invention;

[0018] Figure 2 This is a schematic diagram illustrating the workflow of the multispectral image acquisition module of the present invention;

[0019] Figure 3 This is a schematic diagram of the network architecture of the automatic lining defect identification engine of the present invention;

[0020] Figure 4 This is a schematic diagram of the evaluation process of the defect quantitative evaluation platform of the present invention;

[0021] Figure 5 This is a schematic diagram of the decision-making process of the maintenance solution generator of the present invention. Detailed Implementation

[0022] Please refer to the attached document. Figures 1-5 The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0023] like Figure 1 As shown, the intelligent detection and image analysis system for karst tunnel lining quality provided by the present invention includes a multispectral image acquisition module 1, an automatic lining defect identification engine 2, a defect quantitative assessment platform 3, and a maintenance plan generator 4.

[0024] like Figure 2 As shown, the multispectral image acquisition module 1 is used to acquire multispectral images of the tunnel lining and perform preprocessing to provide high-quality input data for subsequent defect identification.

[0025] The multispectral image acquisition module 1 includes a spectral acquisition unit and an image preprocessing unit. The spectral acquisition unit is equipped with a visible light camera and a near-infrared camera. Both cameras are mounted on a moving inspection platform and scan the tunnel lining surface using a synchronous triggering method. The visible light camera acquires visible light images in the wavelength range of 400nm to 700nm, clearly reflecting surface defects such as cracks, spalling, and water seepage. The near-infrared camera acquires near-infrared images in the wavelength range of 700nm to 1400nm, providing better penetration and sensitivity to hidden defects such as voids and honeycombing within the lining. Both cameras have a resolution of at least 4096×3000 pixels and a frame rate of at least 30fps, ensuring clear and continuous image sequences can be acquired when the inspection vehicle is traveling at a speed of 5m / s to 15m / s.

[0026] The image preprocessing unit first performs geometric registration on the acquired visible light and near-infrared images. It uses a feature point matching method to calculate the spatial transformation relationship between the two images, registering the near-infrared image to the coordinate system of the visible light image to ensure that corresponding pixels represent the same physical location. After registration, the two images are fused using a weighted fusion strategy. This strategy dynamically determines the fusion weights based on the response intensity of each pixel in the visible and near-infrared bands, generating a fused image.

[0027] To adapt to the complex and variable lighting conditions within karst tunnels, the image preprocessing unit employs an adaptive enhancement strategy to enhance the fused image. Specifically, the fused image is divided into several local regions, and a contrast feature value is calculated for each region. Enhancement parameters for that region are determined based on these contrast feature values. The enhancement intensity is increased for regions with low contrast and decreased for regions with high contrast, avoiding over-enhancement that could lead to detail loss. The enhancement process utilizes an adaptive histogram equalization method, which can enhance the overall image contrast while preserving local details, effectively mitigating image quality degradation caused by uneven lighting, shadow occlusion, and other factors.

[0028] In a preferred embodiment of the present invention, the innovative algorithm of the adaptive enhancement strategy is as follows. Let the contrast feature value of a certain local region in the fused image be... The adaptive enhancement factor in this region The calculation formula is:

[0029] ,

[0030] in, As the benchmark enhancement factor, The target contrast threshold, The adjustment parameters are used to control the sensitivity of the enhancement factor to changes in contrast. A minimum enhancement factor is used to prevent the enhancement factor from becoming too small. In a preferred embodiment, A value of 2.5 can provide sufficient enhancement for low-contrast areas; The value is 0.6, which was determined through statistical analysis of a large number of karst tunnel images and represents a suitable contrast level. A value of 8.0 ensures that the enhancement factor decays rapidly when the contrast deviates from the target value, avoiding over-enhancement of high-contrast areas. A value of 0.8 ensures that all areas receive a basic enhancement effect. This formula achieves non-linear adaptive adjustment through an exponential function, allowing the enhancement factor to dynamically change according to local contrast characteristics. Near target contrast At that time, the enhancement factor reaches its maximum value. Provides the strongest enhancement effect; when contrast... keep away At that time, the enhancement factor gradually decayed to Avoid over-enhancing.

[0031] After adaptive enhancement, the image undergoes denoising to eliminate sensor noise and environmental interference. Denoising is achieved using a bilateral filter, which effectively preserves edge information and avoids blurring of defect boundaries while smoothing noise. After denoising, the image is normalized, mapping pixel values ​​to the range of 0 to 1, generating preprocessed image data.

[0032] The multispectral image acquisition module 1 transmits the generated preprocessed image data to the lining defect automatic identification engine 2, providing high-quality input for subsequent defect identification.

[0033] like Figure 3 As shown, the automatic lining defect recognition engine 2 is used to perform deep learning analysis on the preprocessed image to identify the defect type and accurately locate the defect area.

[0034] The automatic lining defect identification engine 2 employs an improved deep learning network architecture, which includes a dual-branch attention feature extraction network and a defect classification and localization head.

[0035] The dual-branch attention feature extraction network comprises two parallel feature extraction paths: a global semantic branch and a local detail branch. The global semantic branch employs a Transformer encoder structure, capturing long-range dependencies and global semantic information of the image through a self-attention mechanism. This branch is capable of understanding the positional relationship and contextual information of defects within the overall lining. The local detail branch uses a convolutional neural network structure, extracting local texture features and edge details of the image through multi-layer convolution and pooling operations. This branch has a strong expressive ability for the boundary morphology and surface texture of defects.

[0036] In a preferred embodiment of the present invention, the dual-branch feature fusion employs an adaptive weight fusion mechanism. Let the feature map extracted by the global semantic branch be... The feature map extracted from the local detail branch is The fused feature map The calculation formula is:

[0037] ,

[0038] in, For adaptive weights of global semantic features, The two weights are adaptive weights for local detail features, and they satisfy the normalization condition. The adaptive weights are calculated based on the information content of the feature map, and the specific calculation formula is as follows:

[0039] ,

[0040] ,

[0041] in, and These represent the information entropy of the global semantic feature map and the local detail feature map, respectively. A temperature coefficient is used to control the smoothness of the weight distribution. Information entropy. The calculation employs the Shannon entropy formula, which involves normalizing all pixel values ​​in the feature map and then calculating the negative log-weighted sum of their probability distributions. In a preferred embodiment, The value is set to 1.5, which was determined through experimental verification on defect identification tasks and enables a reasonable weight allocation between global and local features. This adaptive weight fusion mechanism allows the network to dynamically adjust the fusion weights according to the feature distribution of different defect types. For defect types that require global context information, such as cellular defects and voids, the network automatically increases the weights. The value of is increased to enhance the contribution of global semantic features; for defect types with clear boundaries such as cracks and peeling, the network automatically increases . The value of highlights the local details.

[0042] The fused feature map is input into the defect classification and localization head for defect identification and localization. The defect classification and localization head consists of two parallel output branches: a classification sub-network and a localization sub-network. The classification sub-network uses fully connected layers and a Softmax activation function to output the probability distribution of defect types, including categories such as cracks, honeycombing, voids, water seepage, spalling, and no defects. The localization sub-network uses fully connected layers to output the bounding box coordinates of the defect region, including the coordinates of the bounding box's center point, width, and height.

[0043] To improve the accuracy of defect identification, the lining defect automatic identification engine 2 adopts a multi-scale feature pyramid structure. This structure extracts features at different levels of the dual-branch feature extraction network to construct a feature pyramid. Shallow features retain rich spatial details and are suitable for detecting small-sized defects, while deep features have stronger semantic expressive power and are suitable for detecting large-sized defects. Through the feature pyramid structure, the network can simultaneously detect defects of different scales, improving its ability to identify defects of various sizes.

[0044] The automatic tunnel lining defect recognition engine 2 was trained using supervised learning. The training dataset contained a large number of tunnel lining images labeled with defect types and locations, collected from on-site tunnel observations in multiple karst areas. The network was trained using a stochastic gradient descent optimization algorithm, updating network parameters through backpropagation to minimize the weighted sum of classification and localization losses. Cross-entropy loss was used for classification, and smoothed L1 loss was used for bounding box regression. Data augmentation techniques, including random cropping, rotation, flipping, and brightness adjustment, were employed during training to increase the diversity of training samples and improve the model's generalization ability.

[0045] The automatic lining defect identification engine 2 generates defect identification results including defect type, location coordinates, and confidence level, and transmits these results to the defect quantitative assessment platform 3. Simultaneously, the automatic lining defect identification engine 2 receives defect feature statistical information from the defect quantitative assessment platform 3, optimizes the defect classification threshold based on the statistical information, forming a closed-loop feedback mechanism that enables the identification engine to continuously learn and adaptively optimize.

[0046] like Figure 4 As shown, the defect quantitative assessment platform 3 is used to perform quantitative analysis on the identified defects, calculate the geometric characteristic parameters of the defects, assess the severity of the defects, and classify the health status of the lining.

[0047] The defect quantitative assessment platform 3 includes a geometric feature extraction unit, a severity scoring unit, and a health grading unit.

[0048] The geometric feature extraction unit receives the defect recognition results from the lining defect automatic recognition engine 2 and extracts a sub-image of the defect region from the preprocessed image based on the defect's location coordinates. Edge detection is performed on the sub-image, and the Canny edge detection algorithm is used to extract the defect's boundary contour. After edge detection, morphological processing is performed on the edge image, including operations such as dilation, erosion, and closing, to fill broken edge lines and remove isolated noise points, resulting in a complete and continuous defect contour.

[0049] Geometric feature parameters are calculated based on defect contours. For crack-type defects, the extracted geometric feature parameters include crack length, average width, maximum width, and crack direction angle. Crack length is obtained by calculating the total length of the contour skeleton. A skeleton extraction algorithm is used to refine the crack contour into skeleton lines of single-pixel width. The number of pixels on the skeleton lines is counted and multiplied by the pixel resolution to obtain the actual length. Crack width is obtained by measuring the width of the contour perpendicular to the skeleton lines. The width is measured at regular intervals along the skeleton lines, and the average of all measured points is calculated to obtain the average width. The maximum measured value is taken as the maximum width. For honeycomb-type defects, the extracted geometric feature parameters include defect area, perimeter, roundness, and texture roughness. Defect area is obtained by counting the number of pixels within the contour and multiplying by the pixel area resolution. Perimeter is obtained by calculating the total length of the contour boundary. Roundness is defined as 4π multiplied by the area divided by the square of the perimeter, reflecting the regularity of the defect shape. Texture roughness is obtained by calculating the standard deviation of grayscale variation within the defect area, reflecting the roughness of the honeycomb surface. For void-type defects, the extracted geometric feature parameters include void area, depth estimate, and shape irregularity. The calculation method for the cavity area is the same as that for the honeycomb area. The depth estimate is obtained by analyzing the shadow distribution of the cavity edge, and a shape shadow recovery algorithm is used to estimate the three-dimensional depth information of the cavity from the two-dimensional image. The shape irregularity is obtained by calculating the ratio of the area of ​​the contour to its smallest circumscribed ellipse; this parameter reflects the complexity of the cavity shape.

[0050] The severity scoring unit determines the severity score of a defect based on its type, geometric features, and distribution density. A weighted scoring model is used, with different weighting coefficients corresponding to different defect types and features. For crack defects, the severity score primarily considers crack length and maximum width; longer and wider cracks result in higher scores, indicating more severe defects. For honeycomb defects, the severity score primarily considers defect area and texture roughness; larger areas and higher roughness result in higher scores. For void defects, the severity score primarily considers void area and estimated depth; larger areas and deeper voids result in higher scores. Furthermore, the scoring model also considers the spatial distribution density of defects. If a region has a large number of defects or a denser defect distribution, the severity score for that region is increased, as densely distributed defects pose a greater threat to the integrity of the lining structure.

[0051] The health grading unit assesses the health status of the lining based on the severity score. The assessment criteria divide the lining health status into five levels: Excellent (0-20), indicating good lining condition with no obvious defects; Good (20-40), indicating a few minor defects that do not affect structural safety; Medium (40-60), indicating a certain number of defects requiring attention and monitoring; Poor (60-80), indicating relatively serious defects requiring timely repair; and Critical (80-100), indicating serious defects threatening structural safety and requiring immediate action.

[0052] The quantitative defect assessment platform 3 generates a quantitative defect assessment report, which includes information such as the geometric feature parameters of each defect, severity score, lining health level, and defect distribution heatmap. The platform transmits the assessment report to the maintenance plan generator 4, providing a quantitative basis for maintenance decisions. Simultaneously, the platform feeds back defect feature statistics to the automatic lining defect identification engine 2. These statistics include the geometric feature distribution range and typical feature values ​​of each type of defect. The engine optimizes the discrimination threshold based on this statistical information; for example, when the statistics show a change in the feature value range of a certain type of defect, the engine adjusts the discrimination threshold accordingly to improve identification accuracy.

[0053] like Figure 5 As shown, the maintenance plan generator 4 is used to automatically generate maintenance plans based on the defect assessment results, providing a scientific basis for tunnel maintenance decisions.

[0054] The maintenance plan generator 4 includes a maintenance method matching unit, a maintenance area planning unit, and a work quantity estimation unit.

[0055] The repair method matching unit matches the corresponding repair method based on the defect type and severity. The repair method library pre-stores various repair processes for different defect types, including crack sealing grouting, surface repair, structural reinforcement, and drainage treatment. For crack defects, if the crack width is less than 0.2mm, the surface sealing method is used, employing polymer materials to seal the crack; if the crack width is between 0.2mm and 2.0mm, the pressure grouting method is used, injecting resin grout into the crack for sealing and reinforcement; if the crack width is greater than 2.0mm or the crack depth is significant, the slotting and caulking method is used, slotting along the crack and filling it with elastic sealing material. For honeycomb defects, if the honeycomb area is less than 0.1 square meters and the depth is shallow, the mortar repair method is used, removing the loose parts and filling with high-strength mortar; if the honeycomb area is large or the depth is deep, the concrete repair method is used, chiseling away the defective part and pouring high-performance concrete for repair. For void defects, the injection filling method is used, filling the void by drilling and injecting expanding mortar or polyurethane material. For water seepage defects, leak sealing and waterproofing treatment is adopted, which includes a combination of applying waterproof coating to the surface and grouting to seal the leak.

[0056] The maintenance area planning unit plans maintenance areas based on the spatial location and distribution characteristics of defects. The planning considers the density and spatial continuity of defects; if multiple defects are spatially adjacent and close together, they are merged into one maintenance area for overall treatment to improve maintenance efficiency and effectiveness. The planning also considers the structural characteristics of the tunnel, prioritizing defects located in critical load-bearing areas and giving higher priority to defects in important areas such as the lining arch and waist. The maintenance area planning results are presented graphically, with the location, extent, and maintenance priority of each maintenance area marked on the tunnel lining development diagram.

[0057] The quantity estimation unit estimates the repair work volume and cost based on the repair method and area. Quantity estimation includes material consumption estimation and labor hour estimation. Material consumption is calculated based on the geometric dimensions of the defect and the material consumption quota for the repair method. For example, the grout consumption for crack grouting is calculated based on the crack length, depth, and grouting filling rate; the mortar consumption for honeycomb repair is calculated based on the defect volume and material loss rate. Labor hours are calculated based on the procedures and unit quantity quota of the repair method; different repair methods correspond to different labor hour quotas. Repair costs are calculated by summarizing the material unit price and labor unit price to generate a repair cost budget table.

[0058] The Maintenance Plan Generator 4 outputs a complete maintenance plan, including maintenance methods, maintenance areas, workload, and costs. The plan is provided to the user in a structured report format, including text descriptions, tabular data, and graphical annotations, facilitating user understanding and implementation. The maintenance plan provides a scientific basis for tunnel management departments' maintenance decisions, helping to rationally plan maintenance schedules and budgets.

[0059] The system of this invention achieves a nonlinear improvement in overall performance by forming a closed-loop collaborative mechanism through deep coupling between modules.

[0060] The output of the multispectral image acquisition module 1 is directly used as the input of the lining defect automatic identification engine 2, and the improvement in image quality directly determines the accuracy of defect identification. The output of the lining defect automatic identification engine 2 is used as the input of the defect quantitative assessment platform 3, and the improved identification accuracy makes the assessment results more reliable. The output of the defect quantitative assessment platform 3 is used as the input of the maintenance plan generator 4, and the quantitative indicators of the assessment provide a scientific basis for maintenance decisions.

[0061] More importantly, the defect quantitative assessment platform 3 feeds back statistical information on defect features to the lining defect automatic identification engine 2. This feedback information is used to optimize the discrimination threshold of the identification engine. Specifically, after the system has been running for a period of time, the defect quantitative assessment platform 3 accumulates a large amount of geometric feature data of defect samples. Statistical analysis of this data reveals the characteristic distribution patterns of various defects, such as the distribution range of crack width and typical values ​​of honeycomb area. The identification engine adjusts the discrimination threshold based on these statistical patterns, making the threshold more consistent with the actual distribution of defect features, thereby improving the accuracy and robustness of identification. This closed-loop feedback mechanism enables the system to have self-learning and adaptive capabilities, continuously optimizing performance as usage time increases.

[0062] The deep coupling between modules is also reflected in the mutual influence of parameters. The enhancement parameters of the multispectral image acquisition module 1 affect the image contrast, which in turn affects the feature extraction effect of the automatic lining defect recognition engine 2; the confidence threshold of the recognition engine affects the input quality of the defect quantitative assessment platform 3, which in turn affects the reliability of the assessment results; the severity scoring criteria of the assessment platform affects the selection of maintenance methods by the maintenance plan generator 4, which in turn affects maintenance costs and effectiveness. Through the synergistic optimization of the parameters of each module, the overall performance of the system is significantly improved, achieving a synergistic effect where 1+1 is greater than 2.

[0063] Experimental verification shows that, compared with the independent operation of each module, the closed-loop collaborative system of the present invention has significantly improved the accuracy of defect identification, the precision of quantitative assessment, and the scientific nature of maintenance decisions. The overall performance of the system has been improved by more than 30%, which fully demonstrates the advantages of the deep coupling collaborative mechanism.

[0064] Application Examples

[0065] In its application to a highway tunnel in a karst region, this invention's system conducted a comprehensive inspection of a 3.2-kilometer-long tunnel. The multispectral image acquisition module 1 acquired over 50,000 high-resolution images, and the automatic lining defect identification engine 2 identified a total of 1,253 defects, including 728 cracks, 312 honeycomb defects, 98 cavities, and 115 water seepage defects. The defect quantitative assessment platform 3 measured the geometric features and severity of all defects, classifying the lining health level as medium, with 42% of sections rated as excellent, 35% as good, 18% as medium, 5% as poor, and no sections classified as hazardous. The maintenance plan generator 4 generated detailed maintenance plans, planned 63 maintenance areas, estimated the amount of maintenance work and costs, providing a scientific basis for tunnel management department maintenance decisions.

[0066] The detection results were compared and verified with those obtained manually. The detection accuracy of the system of this invention reached 0.15mm, the recognition accuracy rate reached 96.8%, and the detection time was 1 / 12 of that of manual detection, significantly improving detection efficiency and accuracy. Some early minor defects identified by the system were missed by manual detection, demonstrating the superiority of the intelligent detection system.

[0067] In a preferred embodiment of the present invention, the visible light camera of the multispectral image acquisition module 1 is an industrial-grade area array camera with a 1-inch sensor size, a pixel size of 3.45μm, a fixed-focus lens with an F1.4 aperture, a field of view of 45 degrees, and an exposure time that can be automatically adjusted within the range of 10μs to 100ms. The near-infrared camera uses a cooled InGaAs sensor with a sensitivity wavelength range of 900nm to 1700nm and is equipped with a bandpass filter to filter out visible light interference. The inspection vehicle is equipped with an LED supplementary lighting array with a color temperature of 5000K, and the brightness can be automatically adjusted according to the lighting conditions inside the tunnel.

[0068] In a preferred embodiment of the present invention, the Transformer encoder of the automatic lining defect recognition engine 2 adopts a 6-layer stacked structure, with each layer containing 8 attention heads and a hidden layer dimension of 512. The convolutional neural network uses ResNet50 as the backbone network and extracts feature maps at 4 different scales to construct a feature pyramid. The feature fusion network adopts a feature pyramid network (FPN) structure to achieve top-down feature fusion. The defect classification and localization head uses Focal Loss as the classification loss function to effectively alleviate the class imbalance problem.

[0069] In a preferred embodiment of the present invention, the edge detection of the defect quantitative assessment platform 3 adopts the Canny algorithm, with a high threshold of 100, a low threshold of 50, and a Gaussian filter standard deviation of 1.4. The skeleton extraction employs the Zhang-Suen thinning algorithm, which preserves the topological structure of the skeleton. Texture roughness calculation uses the gray-level co-occurrence matrix method to extract texture features such as contrast, correlation, energy, and entropy.

[0070] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An intelligent detection and image analysis system for the lining quality of karst tunnels, characterized in that, include: The multispectral image acquisition module is used to acquire visible light and near-infrared images of the tunnel lining, register and fuse the acquired images, divide the registered and fused images into several local regions and calculate the contrast feature value of each local region, determine the adaptive enhancement parameters based on the contrast feature value, perform adaptive enhancement processing on the fused image, generate preprocessed image data, and transmit the preprocessed image data to the lining defect automatic recognition engine. The automatic lining defect identification engine is connected to the multispectral image acquisition module. It is used to receive the preprocessed image data, construct a dual-branch attention feature extraction network to extract multi-scale features from the preprocessed image, adaptively weight and fuse the extracted global semantic features and local detail features, identify the defect type and locate the defect area through the defect classification head, generate a defect identification result containing defect type, location coordinates and confidence level, and transmit the defect identification result to the defect quantitative assessment platform. The defect quantitative assessment platform is connected to the automatic defect identification engine of the lining. It is used to receive the defect identification results, perform edge detection and contour extraction on the identified defect areas, calculate the geometric feature parameters of the defects based on the defect contours, determine the severity score of the defects according to the defect type, geometric feature parameters and distribution density, classify and assess the health status of the lining according to the score results, and generate a defect quantitative assessment report.

2. The intelligent detection and image analysis system for karst tunnel lining quality according to claim 1, characterized in that, The multispectral image acquisition module includes: The spectral acquisition unit is equipped with a visible light camera and a near-infrared camera to simultaneously acquire visible light and near-infrared images of the tunnel lining surface, wherein the wavelength range of the visible light images is 400 nm to 700 nm, and the wavelength range of the near-infrared images is 700 nm to 1400 nm. The image preprocessing unit is used to perform geometric registration of the visible light image and the near-infrared image, dynamically determine the fusion weight based on the response intensity of each pixel in the visible light and near-infrared bands using a weighted fusion strategy, generate a fused image, divide the fused image into several local regions, calculate the contrast feature value for each local region, determine the adaptive enhancement parameter for that region based on the contrast feature value, perform adaptive histogram equalization on the fused image, and perform bilateral filtering denoising and normalization on the enhanced image to generate the preprocessed image data.

3. The intelligent detection and image analysis system for karst tunnel lining quality according to claim 2, characterized in that, When determining the adaptive enhancement parameters, the image preprocessing unit determines the adaptive enhancement factor based on the contrast feature value of the local region through a nonlinear function. When the contrast feature value is close to the target contrast threshold, the adaptive enhancement factor reaches its maximum value. When the contrast feature value is far from the target contrast threshold, the adaptive enhancement factor gradually decays to the minimum enhancement factor to avoid over-enhancement of high contrast regions.

4. The intelligent detection and image analysis system for karst tunnel lining quality according to claim 1, characterized in that, The dual-branch attention feature extraction network of the automatic lining defect identification engine includes: The global semantic branch adopts a Transformer encoder structure and captures the long-distance dependencies and global semantic information of the preprocessed image through a self-attention mechanism; The local detail branch employs a convolutional neural network structure, using multi-layer convolution and pooling operations to extract local texture features and edge details from the preprocessed image; The feature fusion layer is used to receive the global semantic feature map extracted by the global semantic branch and the local detail feature map extracted by the local detail branch, determine the adaptive fusion weight based on the information content evaluation of the feature maps, and generate the fused feature map.

5. The intelligent detection and image analysis system for karst tunnel lining quality according to claim 4, characterized in that, When determining the adaptive fusion weights, the feature fusion layer calculates the information entropy of the global semantic feature map and the local detail feature map. Based on the information entropy, the fusion weights of each feature map are determined through an exponential function. The fusion weights satisfy the normalization condition. For defects such as cellular and void types, the fusion weight of the global semantic feature map is increased. For defects such as cracks and spalling types, the fusion weight of the local detail feature map is increased.

6. The intelligent detection and image analysis system for karst tunnel lining quality according to claim 1, characterized in that, The automatic lining defect identification engine also includes: A multi-scale feature pyramid is used to extract features at different levels of the dual-branch attention feature extraction network to construct a feature pyramid containing shallow and deep features, wherein the shallow features are used to detect small-sized defects and the deep features are used to detect large-sized defects. The defect classification and localization head includes a classification subnetwork and a localization subnetwork. The classification subnetwork uses a fully connected layer and a Softmax activation function to output the probability distribution of defect types, and the localization subnetwork uses a fully connected layer to output the bounding box coordinates of the defect region.

7. The intelligent detection and image analysis system for karst tunnel lining quality according to claim 1, characterized in that, The defect quantitative assessment platform includes: The geometric feature extraction unit is used to extract sub-images of the defect region from the preprocessed image based on the defect recognition result, perform Canny edge detection and morphological processing on the sub-images to obtain a complete and continuous defect contour, and calculate the geometric feature parameters based on the defect contour. Specifically, for crack-type defects, the unit extracts crack length, average width, maximum width and orientation angle; for honeycomb-type defects, the unit extracts defect area, perimeter, roundness and texture roughness; and for cavity-type defects, the unit extracts cavity area, depth estimate and shape irregularity. The severity scoring unit is used to determine the severity score of the defect based on the defect type, the geometric feature parameters, and the distribution density using a weighted scoring model, wherein different defect types and feature parameters correspond to different weight coefficients, and a severity score is added for densely distributed defects; The health grading unit is used to classify the lining health status into five levels: excellent, good, medium, poor, and critical, based on the severity of the defects.

8. The intelligent detection and image analysis system for karst tunnel lining quality according to claim 7, characterized in that, When calculating the crack length, the geometric feature extraction unit uses a skeleton extraction algorithm to refine the crack outline into a skeleton line with a single pixel width, counts the number of pixels on the skeleton line and multiplies it by the pixel resolution to obtain the actual length; when calculating the crack width, the width of the outline is measured in the direction perpendicular to the skeleton line, the average value of all measurement points is calculated to obtain the average width, and the maximum measurement value is taken to obtain the maximum width.

9. The intelligent detection and image analysis system for karst tunnel lining quality according to claim 1, characterized in that, The quantitative assessment report of defects is transmitted to the maintenance plan generator and the automatic identification engine of lining defects, wherein the defect feature statistics transmitted to the automatic identification engine of lining defects are used to optimize the discrimination threshold of the defect identification model; A repair plan generator, connected to the defect quantitative assessment platform, is used to receive the defect quantitative assessment report, match the corresponding repair method according to the defect type and severity, plan the repair area according to the spatial location and distribution characteristics of the defect, estimate the repair workload and repair cost, generate a repair plan including repair method, repair area, workload and cost, and output the repair plan for user reference. The repair scheme generator includes: The repair method matching unit is used to score the defect type and the severity of the defect and match the corresponding repair method from the repair method library. For crack defects, the surface sealing method, pressure grouting method or groove caulking method is selected according to the crack width. For honeycomb defects, the mortar repair method or concrete repair method is selected according to the honeycomb area and depth. For void defects, the injection filling method is used. For water seepage defects, the leak sealing and waterproofing treatment is used. The maintenance area planning unit is used to plan maintenance areas based on the spatial location and distribution characteristics of defects, merging multiple spatially adjacent defects with small spacing into one maintenance area, and prioritizing the treatment of defects located in critical load-bearing parts; The work quantity estimation unit is used to estimate the work quantity and cost of maintenance based on the maintenance method and the maintenance area, including material usage estimation and labor hour estimation.

10. The intelligent detection and image analysis system for karst tunnel lining quality according to claim 1, characterized in that, When the defect quantitative assessment platform feeds back defect feature statistics to the lining defect automatic identification engine, the statistics include the geometric feature distribution range and typical feature values ​​of each type of defect. The lining defect automatic identification engine adjusts the discrimination threshold for defect classification based on the statistics, so that the discrimination threshold is more consistent with the actual distribution of defect features, forming a closed-loop feedback mechanism to achieve adaptive learning and continuous optimization.

Images