Intelligent detection method and system for apparent defect images of shield tunnel segments

Abstract

The application discloses a shield tunnel segment apparent defect image intelligent detection method and system, and belongs to the technical field of tunnel engineering detection. The method comprises the following steps: acquiring a tunnel inner wall ring scanning image in an image acquisition step; performing illumination correction and contrast enhancement in an image preprocessing step; synchronously identifying four types of defects, i.e. cracks, damages, misalignments and water seepage, through a deep learning network in a multi-class defect synchronous detection step; extracting a crack skeleton line and measuring the length, width and trend of the crack in a crack geometric parameter measurement step; calculating the misalignment amount based on gray gradient analysis in a misalignment quantitative analysis step; and performing threat degree scoring and generating a detection report in a defect evaluation and report generation step, while feeding back optimized detection parameters. The application realizes synchronous detection of multiple types of defects, the detection accuracy is above 95%, and the crack width detection precision reaches 0.2 mm.

Description

Technical Field

[0001] This invention belongs to the field of tunnel engineering inspection technology, specifically relating to an intelligent detection method and system for surface defects in shield tunnel segments. Background Technology

[0002] With the rapid development of urban rail transit and underground space development, shield tunneling has been widely used in the construction of subway tunnels, municipal utility tunnels, and underwater passages due to its advantages such as high efficiency, safety, and minimal impact on the ground. Shield tunnels are typically assembled from precast concrete segments. As the permanent support structure of the tunnel, the quality of these segments directly affects the structural safety and service life of the tunnel. However, due to the long-term coupled effects of various factors such as ground pressure, groundwater erosion, train vibration, and temperature stress, surface defects such as cracks, damage, misalignment, and water leakage inevitably appear on the segments. If these defects are not detected and addressed in a timely manner, they will gradually worsen, seriously threatening the operational safety of the tunnel.

[0003] Existing methods for inspecting tunnel segments mainly fall into two categories: manual visual inspection and sensor-based automated inspection. Manual visual inspection is currently the most common method, where inspectors walk or ride in inspection vehicles inside the tunnel, visually observing the surface of the segments and recording any defects found. This method has significant drawbacks, including low inspection efficiency, high labor intensity, strong subjectivity, and the tendency to miss minute defects. In particular, it is difficult for the human eye to accurately identify minute cracks less than 0.3 mm wide in the dim environment of the tunnel; for misalignment defects at segment joints, manual measurement also struggles to obtain accurate quantitative data. Furthermore, manual inspection must be conducted during tunnel shutdowns, severely impacting normal line operation and incurring high economic and time costs.

[0004] Chinese invention CN117934473A discloses a deep learning-based method for detecting apparent cracks in highway tunnels. This method constructs an S-LKA network containing LKA and SCA layers and replaces the C3 module in the YOLOv5 backbone network with this network to improve crack detection accuracy in complex environments. While this approach improves crack detection accuracy to some extent, it still has the following technical limitations: First, this method only detects a single crack type and cannot simultaneously identify multiple apparent defects common in shield tunnels, such as damage, misalignment, and water leakage, making it difficult to meet the practical needs of comprehensive tunnel inspection. Second, this method can only determine the presence or absence of cracks and lacks the ability to automatically measure geometric parameters such as crack length, width, and direction, thus failing to provide quantitative data for defect assessment. Third, while this method optimizes the inspection environment for highway tunnels, shield tunnels have unique characteristics such as segment splicing, obvious joints, and curved surface geometry, making it difficult to achieve ideal results when directly applying this method. Fourth, this method does not consider the quantitative detection of misalignment, a unique defect in shield tunnels, while misalignment is a key indicator for evaluating segment assembly quality.

[0005] In summary, existing technologies are insufficient for the efficient and simultaneous detection and accurate quantitative measurement of various surface defects in shield tunnel segments. There is an urgent need to develop an intelligent detection method and system that can automatically identify multiple defect types, accurately measure crack geometric parameters, accurately quantify misalignment, and automatically generate quality assessment reports. Summary of the Invention

[0006] To address the technical problems of existing shield tunnel segment appearance defect detection technologies, such as low detection efficiency, incomplete coverage of defect types, difficulty in automatically measuring geometric parameters, and inability to quantify misalignment defects, this invention provides an intelligent image detection method and system for shield tunnel segment appearance defects.

[0007] The technical solution of the present invention is as follows:

[0008] A method for intelligent detection of surface defects in shield tunnel segments includes: an image acquisition step, which acquires high-definition ring-scan image data of the tunnel inner wall. This high-definition ring-scan image data is continuously acquired along the tunnel axis by a ring-scan camera mounted on a tunnel inspection vehicle, with mileage and ring number information recorded synchronously during acquisition; an image preprocessing step, which performs illumination non-uniformity correction and contrast enhancement processing on the high-definition ring-scan image data of the tunnel inner wall to generate a preprocessed image, wherein the illumination non-uniformity correction adaptively adjusts the brightness values ​​of each region of the image based on a preset illumination compensation coefficient; a multi-type defect synchronous detection step, which inputs the preprocessed image into a trained multi-type defect synchronous detection network. This multi-type defect synchronous detection network uses a feature pyramid structure to fuse multi-scale features, synchronously identifying and locating four types of surface defects: cracks, breaks, misalignments, and water leakage, and outputting the bounding box positions, category labels, and confidence scores for each type of defect; and a crack geometric parameter measurement step, which measures the detected crack defect areas. The process involves several steps: First, a crack skeleton line is extracted, and the crack width is calculated along its normal direction. The actual crack length is calculated based on the pixel length of the skeleton line and the image scale. The crack direction is determined according to the principal direction angle of the skeleton line. Second, a misalignment quantification analysis step is performed. For detected misalignment defect areas, grayscale profile curves are extracted along the joint direction of adjacent segments. The misalignment boundary is determined based on the gradient peak position of the grayscale profile curve. The misalignment amount is calculated based on the difference in grayscale mean on both sides of the misalignment boundary and the image scale. Third, a defect assessment and report generation step integrates crack geometric parameters and misalignment quantification data. The crack geometric parameters include crack width, actual crack length, and crack direction. The misalignment quantification data includes the misalignment amount. Combining segment design parameter data and quality acceptance standard data, threat level scores and quality grade determinations are performed on each defect. A segment defect distribution unfolding map and quality grade assessment report are generated, and the assessment results are fed back to the multi-type defect synchronous detection step to optimize detection parameters.

[0009] This invention also provides an intelligent image detection system for surface defects of shield tunnel segments, comprising: an image acquisition module for acquiring high-definition ring-scan image data of the tunnel inner wall; an image preprocessing module for performing illumination non-uniformity correction and contrast enhancement processing on the high-definition ring-scan image data of the tunnel inner wall to generate a preprocessed image; a multi-type defect synchronous detection module for inputting the preprocessed image into a trained multi-type defect synchronous detection network to simultaneously identify and locate four types of surface defects: segment cracks, damage, misalignment, and water leakage; a crack geometric parameter measurement module for extracting crack skeleton lines and calculating crack length, width, and direction for detected crack defect areas; a misalignment quantification analysis module for calculating the misalignment value based on gradient analysis of grayscale profile curves; and a defect assessment and report generation module for performing threat level scoring and quality level determination and generating an inspection report, while simultaneously feeding back the assessment results to optimize the inspection parameters.

[0010] The beneficial effects of this invention include: First, it achieves simultaneous detection of four types of surface defects in tunnel segments: cracks, damage, misalignment, and water leakage, with an accuracy rate exceeding 95%, significantly improving detection efficiency and comprehensiveness. Second, the crack width detection accuracy reaches 0.2 mm, enabling the identification of minute cracks that are difficult to detect with the human eye, and automatically measuring geometric parameters such as crack length, width, and direction. Third, through grayscale gradient analysis, it achieves precise quantitative measurement of misalignment, with a measurement accuracy of up to 1 mm, providing reliable quantitative data for tunnel segment assembly quality assessment. Fourth, the system automatically generates a tunnel segment defect distribution map and a quality grade assessment report, and continuously optimizes detection parameters through a closed-loop feedback mechanism, improving the automation and intelligence level of the detection. Attached Figure Description

[0011] Figure 1 This is a flowchart of the intelligent detection method for surface defects of shield tunnel segments according to the present invention;

[0012] Figure 2 This is an architecture diagram of the intelligent image detection system for surface defects in shield tunnel segments according to the present invention. Detailed Implementation

[0013] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0014] See Figure 1 The intelligent detection method for surface defects of shield tunnel segments provided by this invention includes six core steps: image acquisition, image preprocessing, simultaneous detection of multiple types of defects, measurement of crack geometric parameters, quantitative analysis of misalignment, and defect assessment and report generation. These steps form a deeply coupled processing flow, and the assessment results of subsequent steps can be fed back to the preceding steps to optimize the processing parameters, thus forming a complete closed-loop collaborative detection system.

[0015] Step S1: Image acquisition step.

[0016] The image acquisition step is the data input stage of the entire inspection process, and its task is to acquire high-resolution circumferential scan images of the tunnel interior. In one embodiment of the present invention, image acquisition is accomplished using a circumferential scan camera mounted on a tunnel inspection vehicle.

[0017] In practice, the tunnel inspection vehicle travels at a preset speed along the tunnel axis. Preferably, the speed is set to 5 to 15 kilometers per hour, which ensures image acquisition quality while improving inspection efficiency. The circumferential scanning camera is installed at the center of the top of the inspection vehicle, with its optical axis aligned with the tunnel axis. It achieves 360-degree circumferential coverage of the tunnel wall through rotational scanning or a multi-camera array.

[0018] Regarding the technical parameters for image acquisition, the resolution of the panoramic camera should be no less than 4096×2160 pixels to ensure the capture of minute defect features on the surface of the tube segment. The sampling frame rate is set according to the speed of the inspection vehicle and the image overlap requirements. Preferably, at a speed of 10 kilometers per hour, the frame rate is set to 10 frames per second, and the overlap rate of adjacent frames is no less than 20%. The camera lens uses a low-distortion wide-angle lens, with the distortion coefficient controlled within one percent to reduce the impact of image geometric deformation on the accuracy of subsequent measurements.

[0019] During image acquisition, mileage and ring number information are recorded simultaneously. Mileage information is acquired using encoders mounted on the wheels of the inspection vehicle, with a measurement accuracy of up to 0.1 meters; ring number information is obtained by reading ring number markings on the tunnel sidewalls or by conversion based on mileage. This location information is stored in association with the image data, providing a spatial reference basis for subsequent defect localization and distribution mapping.

[0020] In addition, to ensure image quality under the complex lighting conditions inside the tunnel, the inspection vehicle is equipped with a high-brightness LED ring lighting system with an illumination power of no less than 5000 lumens and a neutral white light color temperature of 5000 to 6000 Kelvin to achieve image color reproduction effects close to natural light conditions. The lighting system is triggered synchronously with the camera to ensure that each frame of the image has consistent lighting conditions.

[0021] Step S2: Image preprocessing step.

[0022] The image preprocessing step receives high-definition circumferential scan image data of the tunnel inner wall from the image acquisition step, performs illumination non-uniformity correction and contrast enhancement processing on it, and generates a preprocessed image, providing high-quality input data for subsequent defect detection.

[0023] The interior of tunnels is dimly lit and unevenly lit. Even with artificial lighting, significant differences in brightness exist between different areas of the image, with areas near the light source being too bright and areas farther away being too dark. This uneven lighting severely affects the accuracy of defect detection, thus requiring targeted correction. This invention proposes an adaptive lighting compensation algorithm, the core idea of ​​which is to adaptively adjust the brightness values ​​of each region based on the relationship between the local and global brightness of image blocks.

[0024] In a preferred embodiment of the present invention, the specific implementation process of the adaptive illumination compensation algorithm is as follows. First, the input image is divided into several non-overlapping rectangular blocks, preferably with a block size of 64×64 pixels. Then, the local grayscale mean of each block and the global grayscale mean of the entire image are calculated. Next, the local illumination compensation factor of each block is calculated based on the following formula:

[0025] ,

[0026] in, For the first Local illumination compensation factor for each block, This represents the global grayscale mean of the entire image. For the first The local grayscale average of each block.

[0027] To avoid image distortion caused by excessively large or small compensation factors in extreme cases, the compensation factor is limited to a preset value range, preferably between 0.8 and 1.5. Then, bilinear interpolation is used to expand the compensation factors of each block into a compensation factor map of the same size as the original image. Finally, each pixel value of the original image is multiplied by the corresponding compensation factor to obtain the illumination-corrected image.

[0028] ,

[0029] in, These are the pixel values ​​after illumination correction. These are the original pixel values. This is the compensation factor for this position obtained through bilinear interpolation.

[0030] After correcting for uneven illumination, further contrast enhancement processing is performed to highlight the defect features on the tunnel segment surface. This invention employs a contrast-limited adaptive histogram equalization method, which divides the image into multiple small regions, performs histogram equalization on each region separately, and eliminates discontinuities at region boundaries through bilinear interpolation. Preferably, the block size is set to 8×8 pixels, and the contrast limiting factor is set to 2.0 to 4.0. After contrast enhancement processing, the contrast between defects such as cracks and breaks on the tunnel segment surface and normal areas is significantly improved, which is beneficial for the accurate identification of subsequent detection algorithms.

[0031] Step S3: Simultaneous detection of multiple types of defects.

[0032] The multi-type defect synchronous detection step is the core of the entire detection method. Its task is to input the preprocessed image into a trained multi-type defect synchronous detection network to simultaneously identify and locate four types of surface defects in the tunnel segments: cracks, damage, misalignment, and water leakage. This invention designs a dedicated multi-type defect synchronous detection network specifically for the characteristics of defects in shield tunnel segments. This network can simultaneously detect multiple types of defects in a single forward inference, significantly improving detection efficiency.

[0033] In a preferred embodiment of the present invention, the multi-defect synchronous detection network adopts an encoder-decoder structure, mainly composed of a backbone network, a neck network, and a detection head. The backbone network is responsible for extracting multi-scale features from the input image and uses a residual connection structure to solve the gradient vanishing problem in deep network training. Preferably, the backbone network contains 5 stages, each stage consisting of a convolutional module and a residual block. From stage 1 to stage 5, the feature map size is successively reduced to 1 / 2, 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the input size, and the corresponding number of feature channels increases successively to 64, 128, 256, 512, and 1024.

[0034] The neck network employs a bidirectional feature pyramid structure to achieve bidirectional fusion of shallow detailed features and deep semantic features. Specifically, the neck network includes two feature propagation paths: top-down and bottom-up. The top-down path propagates deep, highly semantic features to the shallow layers through upsampling and lateral connections, enhancing the semantic expressiveness of shallow features. The bottom-up path propagates shallow, high-resolution features to the deep layers through downsampling and lateral connections, enhancing the localization accuracy of deep features. This bidirectional feature fusion strategy is particularly important for detecting defects of different sizes on pipe segments: fine cracks require precise localization using shallow, high-resolution features, while large-area damage requires accurate identification using deep, highly semantic features.

[0035] The detection head is responsible for outputting defect detection results based on the fused multi-scale features. This invention employs a decoupled detection head design, processing the classification and localization tasks separately through two parallel convolutional branches. The classification branch outputs the confidence score for each defect category, while the localization branch outputs the position parameters of the defect bounding boxes. In the post-processing stage, detection results with confidence scores not lower than a preset confidence threshold are first selected; preferably, the preset confidence threshold is set to 0.5 to 0.8. Then, a non-maximum suppression algorithm is used to remove duplicate detected bounding boxes, outputting the final defect detection results.

[0036] The training of the multi-class defect synchronous detection network utilizes supervised learning based on historical defect sample data. The training dataset contains a large number of segment defect images annotated by professionals, covering four types of defects: cracks, breaks, misalignments, and leaks. In one embodiment of the invention, the training dataset contains no fewer than 10,000 annotated images, with the number of samples for each type of defect remaining relatively balanced. The training process employs a stochastic gradient descent optimizer, with an initial learning rate set to 0.01, a cosine annealing strategy for learning rate decay, a batch size of 16, and 300 training epochs. The training loss function consists of a weighted sum of classification loss and localization loss.

[0037] The trained multi-class defect synchronous detection network achieved the following performance metrics on the test set: For crack defects, the accuracy reached 96.2% and the recall reached 94.5%; for breakage defects, the accuracy reached 97.1% and the recall reached 95.8%; for misalignment defects, the accuracy reached 95.3% and the recall reached 93.2%; and for water leakage defects, the accuracy reached 94.6% and the recall reached 92.1%. The average detection accuracy for all four types of defects reached 95.8%, meeting the accuracy requirements for practical engineering applications.

[0038] Step S4: Crack geometry parameter measurement procedure.

[0039] The crack geometry parameter measurement step automatically measures the length, width, and orientation of cracks in the defect area detected by the multi-type defect synchronous detection step. These geometric parameters are key indicators for evaluating the severity of cracks and are of significant reference value for developing maintenance strategies.

[0040] In a preferred embodiment of the present invention, the specific implementation process of the crack geometric parameter measurement step is as follows: First, a binary mask image of the crack is extracted from the crack defect region. Then, morphological thinning processing is performed on the binary image to extract the crack skeleton line with a single pixel width. The morphological thinning adopts an iterative erosion algorithm, which gradually peels away edge pixels while maintaining the crack connectivity until all pixels become skeleton points. The determination criterion for skeleton points is: the removal of the point will not cause the crack connectivity to break or the loop structure to be destroyed.

[0041] After obtaining the crack skeleton line, the local tangent direction is calculated for each skeleton point along the skeleton line. Specifically, taking the current skeleton point as the center, n adjacent skeleton points before and after it (preferably, n is set to 5 to 10) are selected. A linear fit is then performed on these 2n+1 skeleton points, and the direction of the fitted line is the local tangent direction at that skeleton point. Then, the normal direction perpendicular to the tangent is determined.

[0042] The crack width is measured using the orthogonal projection-based crack width measurement algorithm proposed in this invention. For each skeleton point, the crack's edge points are found by searching both sides of the binary mask image along its normal direction. Edge points are determined by the position where the pixel value changes from 1 (crack region) to 0 (background region). The Euclidean distance between the two edge points is the crack width pixel value at that skeleton point. This can be expressed mathematically as:

[0043] ,

[0044] in, skeleton points The width of the crack in pixels at that location. and These are the coordinates of the two edge points searched along the normal direction.

[0045] To obtain the actual physical width of the crack, pixel values ​​need to be converted into actual length units. This invention uses an image scale for this conversion, calculated based on the distance from the camera to the tunnel segment surface recorded during image acquisition and the camera's intrinsic parameters. Let the camera focal length be... (Unit: pixels), the distance from the camera to the surface of the tube is... (Unit: millimeters), then the image scale. (Unit: mm / pixel) The calculation formula is:

[0046] ,

[0047] Actual width of the crack The calculation formula is:

[0048] ,

[0049] in, skeleton points The actual width of the crack at that location, in millimeters.

[0050] For a complete crack, the width values ​​of all its skeleton points are calculated, and the maximum width, average width, and other indicators are statistically analyzed as the width parameters of the crack. In a preferred embodiment of the present invention, the crack width detection accuracy is not less than 0.2 mm, which can identify minute cracks that are difficult for the human eye to detect.

[0051] The crack length is measured based on the pixel length of the crack skeleton line. The pixel length of the crack skeleton line is equal to the sum of the distances between all adjacent skeleton points:

[0052] ,

[0053] in, The pixel length of the crack skeleton line. The total number of skeleton points. For the first The coordinates of the skeleton points.

[0054] Actual length of crack The calculation formula is:

[0055] ,

[0056] The measurement error of crack length shall not exceed five percent of the actual length.

[0057] The crack orientation is determined based on the principal direction angle of the crack skeleton line. Principal component analysis (PCA) is used to analyze the coordinates of all skeleton points, and the direction of the first principal component is extracted as the principal direction of the crack. The principal direction angle is defined as the angle between this direction and the horizontal axis of the image, with a value ranging from 0 to 180 degrees.

[0058] Step S5: Quantitative analysis of misalignment.

[0059] The misalignment quantitative analysis step precisely measures the misalignment between adjacent segments in the misalignment defect areas detected by the multi-type defect synchronous detection step. Misalignment is a common defect that occurs during the assembly of shield tunnel segments, manifested as radial positional deviation between adjacent segments. Severe misalignment can affect the structural stability and waterproofing performance of the tunnel.

[0060] In a preferred embodiment of the present invention, the misalignment quantification analysis step employs a grayscale gradient analysis-based method to measure the misalignment amount. This method utilizes the difference in light reflection caused by the height difference of the pipe segment surface at the misalignment location, and locates the misalignment boundary and calculates the misalignment amount by analyzing the gradient changes in the grayscale image.

[0061] The specific implementation process is as follows. First, the analysis area is determined based on the bounding box position of the misalignment defect, and the direction of the adjacent segment joints is identified. Then, a grayscale profile curve is extracted along the normal direction of the adjacent segment joints. The grayscale profile curve is a sequence of grayscale values ​​of all pixels along the normal line, and the curve length covers 50 to 100 pixels on both sides of the joint, preferably set to 75 pixels.

[0062] Due to image noise, the original grayscale profile curve will have some fluctuations. To accurately detect misalignment boundaries, the grayscale profile curve is first subjected to Gaussian smoothing filtering to remove noise interference. The standard deviation of the Gaussian filter kernel is set to 2 to 5 pixels, preferably 3 pixels.

[0063] Next, the first-order gradient of the smoothed grayscale profile curve is calculated. The gradient curve reflects the rate of change of grayscale values ​​along the profile direction, and obvious gradient peaks appear at the misalignment boundaries. The gradient is calculated using the central difference method:

[0064] ,

[0065] in, For the first Gradient values ​​at each position, This is the grayscale value at that location.

[0066] The location of the peak with the largest absolute value in the detected gradient curve is taken as the misalignment boundary location. To improve the robustness of the detection, a threshold condition for the gradient peak is set: the absolute value of the peak must be greater than a preset multiple of the standard deviation of the gradient curve, preferably set to 2.5 times.

[0067] After determining the location of the misalignment boundary, calculate the average grayscale value within a range of 20 to 50 pixels on both sides of the boundary, preferably set to 30 pixels. Let the average grayscale value on the left side of the boundary be... The mean gray level on the right is The difference in grayscale mean The calculation is as follows:

[0068] ,

[0069] A mapping relationship exists between the difference in grayscale mean and the misalignment height, which is established through calibration experiments. In the calibration experiments, images are acquired using misalignment samples of known heights, and a curve showing the correspondence between the difference in grayscale mean and the misalignment height is established. Preferably, a linear mapping model is used:

[0070] ,

[0071] in, This refers to the misalignment height (i.e., the misalignment amount). and These are the mapping coefficients obtained through calibration. In one embodiment of the present invention... The typical value range is 0.02 to 0.05 mm / gray level. The typical value is close to 0.

[0072] The final misalignment value is calculated using the aforementioned mapping relationship combined with the image scale. When the calculated misalignment value exceeds a preset misalignment threshold, the misalignment is marked as an excessive misalignment and highlighted in the quality grade assessment report. Preferably, the preset misalignment threshold ranges from 3 mm to 10 mm, with the specific value determined according to the engineering acceptance standards.

[0073] To improve measurement accuracy, multiple profile lines are selected along a direction parallel to the joint within the misalignment defect area for measurement. The average value of all profile line measurements is calculated as the final misalignment value. Preferably, the number of profile lines is set to 5 to 10, and the spacing between adjacent profile lines is 10 to 20 pixels.

[0074] Step S6: Defect assessment and report generation steps.

[0075] The defect assessment and report generation step is the output stage of the entire inspection process. Its task is to integrate crack geometric parameters and misalignment quantitative data, combined with segment design parameters and quality acceptance standard data, to score the threat level and determine the quality grade of each defect, generating a segment defect distribution map and a quality grade assessment report. Simultaneously, this step feeds the assessment results back to the multi-type defect synchronous inspection steps to optimize inspection parameters and form a closed-loop collaboration.

[0076] In a preferred embodiment of the present invention, the defect threat level score employs the multi-factor weighted evaluation algorithm proposed in this invention. This algorithm comprehensively considers three factors: defect type, defect geometry, and defect location, to calculate the overall threat level score for each defect. The threat level score calculation formula is as follows:

[0077] ,

[0078] in, Score the threat level of the defect. This is the defect type weighting coefficient. This is the normalized value of the defect geometry. This is the importance coefficient of the defect location. , , The weight coefficients of the three factors satisfy the following conditions: Preferably, , , Set them to 0.3, 0.4, and 0.3 respectively.

[0079] Defect type weighting coefficient The design is based on the degree of impact of different types of defects on structural safety. In one embodiment of the invention, crack-type defects... Set to 0.9, for breakage defects. Set to 0.8, for misalignment defects. Set to 0.7 for water leakage defects. Set it to 0.6.

[0080] Defect geometry normalized value Calculated based on the geometric measurement parameters of the defect. For crack-type defects, the maximum width and length of the crack are considered together:

[0081] ,

[0082] in, The maximum width of the crack. The reference width threshold (preferably set to 1 mm) is used. The length of the crack. This is a reference length threshold (preferably set to 500 mm). The normalization value is set to 1 if it exceeds 1.

[0083] For misalignment defects, the calculation is based on the misalignment amount:

[0084] ,

[0085] in, This represents the misalignment value. The reference misalignment threshold is set to 10 mm.

[0086] Defect location importance coefficient The importance of the defect is determined based on its location on the segment. Different parts of the segment have different levels of importance to structural safety; the vault area and the area near the joints are of higher importance, while the sidewall areas are of relatively lower importance. In one embodiment of the invention, the vault area... Set to 1.0, for the area near the seam. Set to 0.9, sidewall area Set to 0.7, for the inverted arch area. Set it to 0.8.

[0087] The quality level is determined by comparing the threat score with the level threshold in the quality acceptance standard data. In one embodiment of the present invention, the quality level is divided into four levels: A, B, C, and D: a threat score less than 0.3 is level A, indicating a minor defect that does not require immediate action; a threat score between 0.3 and 0.5 is level B, indicating a minor defect that requires observation and monitoring; a threat score between 0.5 and 0.7 is level C, indicating a more serious defect that requires planned maintenance; and a threat score greater than 0.7 is level D, indicating a severe defect that requires immediate action.

[0088] The segment defect distribution map is generated using a cylindrical surface unfolding method, unfolding the tunnel's cylindrical surface along the generatrix into a plane. The horizontal axis of the unfolded map represents the tunnel's axial position (mileage or ring number), and the vertical axis represents its circumferential position (angle, 0 to 360 degrees). On the unfolded plane, the location, type, and size information of various defects are marked according to their actual positions. Different types of defects are distinguished by different colors: cracks are represented in red, damage in orange, misalignment in yellow, and leakage in blue. Defects of different quality levels are distinguished by different patterns or line types.

[0089] The quality rating report is generated in a structured format and mainly includes: basic information about the testing project (project name, testing date, testing scope, etc.), testing equipment parameter information, defect statistics summary (quantity, location distribution, severity distribution, etc. of various defects), detailed information on key defects (location, parameters, image screenshots, etc. of defects exceeding standards), comprehensive quality rating results, and maintenance recommendations. The report can be exported as PDF or Word format for easy archiving and transmission.

[0090] The evaluation results from the defect assessment and report generation steps are fed back to the multi-type defect synchronous detection steps to optimize detection parameters. Specific feedback optimization mechanisms include: adjusting the confidence threshold based on the false negative rate statistics; if the false negative rate for a certain type of defect is high, the confidence threshold for that type of defect is appropriately lowered; adjusting the non-maximum suppression parameters based on the false positive rate statistics; if the false positive rate is high, the suppression threshold is appropriately increased; and determining whether incremental training or fine-tuning of the detection network is needed based on the trend of detection accuracy changes.

[0091] See Figure 2 This invention also provides an intelligent image detection system for surface defects in shield tunnel segments. This system includes an image acquisition module 1, an image preprocessing module 2, a multi-type defect synchronous detection module 3, a crack geometric parameter measurement module 4, a misalignment quantitative analysis module 5, and a defect assessment and report generation module 6. The functions of each module correspond one-to-one with the functions of the corresponding steps in the method embodiments. The modules are interconnected through data interfaces to form a complete detection system.

[0092] Image acquisition module 1 is used to acquire high-definition circumferential scan image data of the tunnel inner wall. In one embodiment of the present invention, image acquisition module 1 includes a circumferential scan camera unit, an illumination unit, an odometer encoder unit, and a data storage unit. The circumferential scan camera unit adopts an industrial-grade high-resolution camera with a resolution of not less than 4096×2160 pixels; the illumination unit adopts a high-brightness LED ring illumination system with an illumination power of not less than 5000 lumens; the odometer encoder unit is used to record the mileage of the inspection vehicle, with a measurement accuracy of up to 0.1 meters; the data storage unit is used to store the acquired image data and location information in real time, adopting a high-speed solid-state drive with a storage bandwidth of not less than 500MB / s.

[0093] Image preprocessing module 2 is used to perform illumination non-uniformity correction and contrast enhancement processing on the high-definition circular scan image data of the tunnel inner wall. This module implements the adaptive illumination compensation algorithm and contrast-limited adaptive histogram equalization algorithm described in step S2 of the method embodiment.

[0094] The multi-type defect synchronous detection module 3 is used to input the preprocessed image into a trained multi-type defect synchronous detection network to simultaneously identify and locate four types of surface defects: segment cracks, damage, misalignment, and water leakage. This module implements the deep learning-based multi-type defect synchronous detection function described in step S3 of the method embodiment. In one embodiment of the present invention, the multi-type defect synchronous detection module 3 is deployed on an industrial control computer equipped with a high-performance GPU, with a GPU memory of not less than 8GB, enabling real-time detection and processing.

[0095] The crack geometry parameter measurement module 4 is used to extract crack skeleton lines for detected crack defect areas, calculate the crack width along the normal of the crack skeleton lines, calculate the actual crack length based on the pixel length of the crack skeleton lines and the image scale, and determine the crack direction according to the principal direction angle of the crack skeleton lines. This module implements the crack geometry parameter measurement function described in step S4 of the method embodiment.

[0096] The misalignment quantification analysis module 5 is used to extract grayscale profile curves along the joint direction of adjacent segments for the detected misalignment defect area. Based on the gradient peak position of the grayscale profile curve, the misalignment boundary is determined. The misalignment amount is calculated based on the difference in grayscale mean on both sides of the misalignment boundary and the image scale. This module implements the misalignment quantification measurement function based on grayscale gradient analysis described in step S5 of the method embodiment.

[0097] The defect assessment and report generation module 6 integrates crack geometric parameters and misalignment quantification data, combines segment design parameter data and quality acceptance standard data to perform threat level scoring and quality grade determination, and generates a segment defect distribution map and quality grade assessment report. This module implements the defect assessment and report generation functions described in step S6 of the method embodiment. Furthermore, the defect assessment and report generation module 6 also feeds back the assessment results to the multi-type defect synchronous detection module 3 to optimize detection parameters, forming a closed-loop feedback optimization mechanism.

[0098] In a preferred embodiment of the present invention, the above-mentioned modules are integrated and deployed on the onboard computing platform of the tunnel inspection vehicle. The onboard computing platform adopts an industrial-grade embedded computer or industrial control computer, equipped with a high-performance CPU and GPU, which can realize real-time acquisition, processing and analysis of inspection data. The onboard computing platform is connected to the back-end management system through a wired or wireless network, and can upload the inspection results to the cloud database in real time for storage and management.

[0099] The intelligent image detection method and system for surface defects of shield tunnel segments of the present invention realizes the simultaneous detection of four types of surface defects of segments: cracks, damage, misalignment, and water leakage, with a detection accuracy of over 95%. The crack width detection accuracy reaches 0.2 mm, which can identify minute cracks that are difficult to detect by the human eye. Through gray-scale gradient analysis, the system realizes the precise quantitative measurement of misalignment, with a measurement accuracy of up to 1 mm. The system automatically generates a distribution map of segment defects and a quality grade assessment report, and continuously optimizes the detection parameters through a closed-loop feedback mechanism, which significantly improves the efficiency and intelligence level of shield tunnel segment detection.

[0100] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. A method for intelligent image detection of apparent defects in shield tunnel segments, characterized in that, include: The image acquisition step involves obtaining high-definition ring-scan image data of the tunnel inner wall. The high-definition ring-scan image data of the tunnel inner wall is obtained by continuously acquiring ring-scanning camera installed on the tunnel inspection vehicle along the tunnel axis. During the acquisition process, mileage information and ring number information are recorded simultaneously. The image preprocessing step involves performing illumination non-uniformity correction and contrast enhancement processing on the high-definition ring scan image data of the tunnel inner wall to generate a preprocessed image. The illumination non-uniformity correction is based on a preset illumination compensation coefficient to adaptively adjust the brightness value of each region of the image. The multi-type defect synchronous detection step involves inputting the preprocessed image into a trained multi-type defect synchronous detection network. The multi-type defect synchronous detection network uses a feature pyramid structure to fuse multi-scale features, synchronously identify and locate four types of surface defects: segment cracks, damage, misalignment, and water leakage, and outputs the bounding box position, category label, and confidence score of each type of defect. The crack geometry parameter measurement steps are as follows: for the detected crack defect area, the crack skeleton line is extracted and the crack width is calculated along the normal of the crack skeleton line. The actual length of the crack is calculated based on the pixel length of the crack skeleton line and the image scale. The crack direction is determined according to the main direction angle of the crack skeleton line. The misalignment quantitative analysis step involves extracting a grayscale profile curve along the joint direction of adjacent segments for the detected misalignment defect area, determining the misalignment boundary based on the gradient peak position of the grayscale profile curve, and calculating the misalignment amount based on the difference in grayscale mean on both sides of the misalignment boundary and the image scale. The defect assessment and report generation steps integrate crack geometric parameters and misalignment quantitative data, combine segment design parameter data and quality acceptance standard data, score the threat level of each defect and determine its quality level, generate a segment defect distribution map and quality level assessment report, and feed the assessment results back to the multi-type defect synchronous detection steps to optimize the detection parameters.

2. The method according to claim 1, characterized in that, In the image preprocessing step, the preset illumination compensation coefficient is determined based on the ratio of the local mean to the global mean of the image block region. The pixel value after illumination compensation is equal to the original pixel value multiplied by the local illumination compensation factor, and the value range of the local illumination compensation factor is 0.8 to 1.

5.

3. The method according to claim 1, characterized in that, In the multi-type defect synchronous detection step, the backbone network of the multi-type defect synchronous detection network adopts a residual connection structure, and the neck network adopts a bidirectional feature pyramid structure to realize the fusion of shallow detail features and deep semantic features. The detection head outputs defect detection results with a confidence score not lower than a preset confidence threshold, and the preset confidence threshold ranges from 0.5 to 0.

8.

4. The method according to claim 1, characterized in that, In the crack geometry parameter measurement step, a subpixel edge detection algorithm is used to extract crack edge points. The subpixel edge detection algorithm determines the subpixel-level edge position by performing a quadratic curve fitting on the gray values ​​in the neighborhood of the edge pixel. The crack width is calculated based on the distance between the subpixel edge positions on both sides, and the crack length is calculated based on the sum of the Euclidean distances between each skeleton point of the crack skeleton line and combined with the image scale.

5. The method according to claim 1, characterized in that, In the step of quantitative analysis of misalignment, when the calculated misalignment value is greater than the preset misalignment threshold, the misalignment is marked as an excessive misalignment and highlighted in the quality grade assessment report. The preset misalignment threshold ranges from 3 mm to 10 mm.

6. The method according to claim 1, characterized in that, The specific steps for measuring the geometric parameters of the crack include: performing morphological thinning processing on the binarized image of the crack defect area to extract the crack skeleton line with a single pixel width; calculating the local tangent direction along each skeleton point of the crack skeleton line to determine the normal direction perpendicular to the tangent; searching for crack edge points in the binarized image along the normal direction and calculating the pixel distance between the two edge points as the crack width pixel value at the skeleton point; determining the image scale based on the distance from the camera to the surface of the segment recorded during image acquisition and the camera's intra-camera parameters, and converting the crack width pixel value into the actual width value.

7. The method according to claim 1, characterized in that, The specific steps of the misalignment quantification analysis include: extracting a grayscale profile curve along the normal direction of the joint between adjacent segments, wherein the length of the grayscale profile curve covers 50 to 100 pixels on each side of the joint; performing Gaussian smoothing filtering on the grayscale profile curve to remove noise interference, and calculating the first-order gradient of the smoothed curve; detecting the peak position with the largest absolute value in the gradient curve as the misalignment boundary position, and calculating the grayscale mean value within a range of 20 to 50 pixels on each side of the boundary; and calculating the misalignment amount based on the mapping relationship between the difference in grayscale mean values ​​on both sides of the boundary and the misalignment height, combined with the image scale.

8. The method according to claim 1, characterized in that, In the defect assessment and report generation steps, the threat score is calculated based on the weighted sum of the defect type weight coefficient, the defect geometric size normalized value, and the defect location importance coefficient. The quality level is determined based on the comparison between the threat score and the level threshold in the quality acceptance standard data.

9. The method according to claim 1, characterized in that, The segment defect distribution unfolded diagram unfolds the tunnel cylindrical surface into a plane along the generatrix. On the unfolded plane, the location, type and size information of various defects are marked according to their actual positions, and different colors and patterns are used to distinguish different types and different levels of defects.

10. An intelligent image detection system for surface defects of shield tunnel segments, used to implement the method described in any one of claims 1-9, characterized in that, include: The image acquisition module is used to acquire high-definition circular scan image data of the tunnel inner wall. The high-definition circular scan image data of the tunnel inner wall is continuously acquired along the tunnel axis by a circular scan camera installed on the tunnel inspection vehicle. During the acquisition process, mileage information and ring number information are recorded simultaneously. The image preprocessing module is used to perform illumination non-uniformity correction and contrast enhancement processing on the high-definition ring scan image data of the tunnel inner wall to generate a preprocessed image. The multi-type defect synchronous detection module is used to input the preprocessed image into the trained multi-type defect synchronous detection network, synchronously identify and locate four types of surface defects: segment cracks, damage, misalignment and water leakage, and output the bounding box position, category label and confidence score of each type of defect. The crack geometry parameter measurement module is used to extract crack skeleton lines for the detected crack defect area, calculate the crack width along the normal of the crack skeleton lines, calculate the actual length of the crack based on the pixel length of the crack skeleton lines and the image scale, and determine the crack direction according to the main direction angle of the crack skeleton lines. The misalignment quantitative analysis module is used to extract grayscale profile curves along the joint direction of adjacent segments for the detected misalignment defect area, determine the misalignment boundary based on the gradient peak position of the grayscale profile curve, and calculate the misalignment value according to the difference in grayscale mean on both sides of the misalignment boundary and the image scale. The defect assessment and report generation module integrates crack geometric parameters and misalignment quantification data, combines segment design parameter data and quality acceptance standard data to perform threat level scoring and quality grade determination, generates segment defect distribution unfolding diagram and quality grade assessment report, and feeds the assessment results back to the multi-type defect synchronous detection module to optimize detection parameters.

Images