A Machine Vision-Based Method and System for Automatic Crack Identification in Subway Tunnels

By correcting surface distortion and adaptive illumination compensation in subway tunnel images, combined with multi-scale segmentation networks and a false defect filtering mechanism, high-precision identification and parameter quantification of subway tunnel cracks are achieved. This solves the problems of low identification efficiency and high false detection rate in existing technologies, and supports the refined maintenance of subway tunnels.

CN123118481APending Publication Date: 2026-11-17SHANDONG HUIHONG INTELLIGENT TECHNOLOGY CO LTD
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202611231635.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-14
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing machine vision-based crack recognition technologies suffer from problems such as image distortion, uneven lighting, crack-like interference, complex topological morphology, and insufficient pixel-level measurement accuracy in subway tunnel scenarios. This results in low recognition efficiency, high false detection rate, and an inability to meet the needs of refined maintenance of subway tunnels.

Method used

By employing surface distortion correction, adaptive illumination compensation, and edge-preserving denoising, combined with a multi-scale segmentation network and a pseudo-defect filtering mechanism, high-precision segmentation and topology reconstruction of cracks are achieved. Subpixel edge detection is used for parameter quantization, and defect assessment is performed by combining subway tunnel mileage and segment ring number information.

Benefits of technology

It significantly improves the recall rate and segmentation integrity of sub-millimeter cracks, reduces the false detection rate, provides high-precision crack parameter quantification and spatial positioning, and supports precise maintenance management of subway tunnels.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118481A_ABST
    Figure CN123118481A_ABST
Patent Text Reader

Abstract

This invention relates to the technical field of safety monitoring and discloses a method and system for automatic identification of cracks in subway tunnels based on machine vision. A standardized preprocessing system is constructed through cylindrical inverse projection distortion correction, adaptive illumination compensation, and edge-preserving denoising to correct morphological distortion caused by curved surfaces, while effectively improving the identifiability of low-contrast fine cracks. A multi-scale feature fusion segmentation network with an attention mechanism is combined, taking into account both global semantics and subtle texture features of the cracks, effectively solving the problems of easy missed detection and breakage of sub-millimeter-level cracks. A pseudo-defect classifier is constructed through multi-dimensional features of shape, texture, and spatial distribution to accurately filter out non-defect interference. Simultaneously, relying on breakpoint connection and topology parsing mechanisms, continuous and complete reconstruction of discontinuous, branching, and mesh-like cracks is achieved. Combined with cubic B-spline curve completion technology for occluded areas, the covered crack segments are restored, realistically restoring the complete topological morphology of the cracks.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the technical field of safety monitoring, specifically to a method and system for automatic identification of cracks in subway tunnels based on machine vision. Background Technology

[0002] Subway tunnels are core infrastructure of urban rail transit. During long-term service, they are subjected to multiple factors such as surrounding rock pressure, uneven soil settlement, train cyclic vibration, and material aging, making their lining structures prone to cracking. Continued crack propagation can further lead to secondary defects such as water leakage, lining spalling, and steel corrosion, directly threatening the tunnel's structural safety and operational stability. Traditional subway tunnel crack detection mainly relies on manual foot patrols combined with handheld measuring instruments, which suffers from low efficiency, long inspection cycles, large subjective judgment errors, and high risks associated with working at heights. Furthermore, the limited operating windows for subways make it difficult for manual inspections to achieve full-line coverage and high-frequency defect detection, failing to meet the current management needs for refined maintenance of subway tunnels. Existing machine vision-based crack recognition technology... There are still significant limitations in the engineering application of subway tunnels: First, the curved surface of the tunnel lining and the acquisition environment with low light and uneven lighting can easily cause image distortion and decreased contrast. Submillimeter-level fine cracks are easily submerged by background noise, and pixel-level segmentation is prone to breakage and missed detection. Second, the subway lining surface has a large number of crack-like interferences such as segment joints, bolt holes, construction cold joints, water stains, and scratches, which are highly similar to the appearance of real cracks, resulting in a high false detection rate. Third, cracks with complex topological forms such as branches, discontinuities, and networks are difficult to extract continuously and completely, making it impossible to restore the true structure of the cracks. Fourth, conventional pixel-level measurement accuracy is insufficient, and the obstruction of pipelines and attachments on the lining surface can easily cause the loss of crack information, making it difficult for the quantitative results to meet the engineering requirements for refined assessment of defects in subway tunnels.

[0003] Chinese invention patent application CN119477886A, published on February 18, 2025, discloses a machine learning-based method and system for monitoring building cracks. This method employs multi-band image data enhancement and preprocessing; adaptive crack region segmentation; vectorized crack feature representation; self-supervised crack feature learning; crack generation based on Generative Adversarial Networks (GANs); crack morphological dynamics modeling; and graph-based crack topology analysis. Through adaptive segmentation, self-supervised learning, and crack morphological dynamics modeling, the accuracy and robustness of crack monitoring are effectively improved, solving the false detection and missed detection problems of traditional methods. The system can adaptively learn the crack behavior under different environments and materials, accurately extracting crack features and predicting their propagation paths. However, the above technical solutions cannot effectively filter environmental interference factors, reducing the quality of building crack monitoring. Summary of the Invention

[0004] (a) Technical problems to be solved To address the significant limitations of existing machine vision-based crack recognition technologies in subway tunnel engineering applications: First, the curved surface of the tunnel lining and the low-light, uneven lighting conditions easily cause image distortion and decreased contrast; sub-millimeter-level fine cracks are easily submerged by background noise, and pixel-level segmentation is prone to breakage and missed detection. Second, the subway lining surface contains numerous crack-like interferences such as segment joints, bolt holes, construction cold joints, water stains, and scratches, which have high similarity to the appearance of real cracks, resulting in a high false detection rate. Third, cracks with complex topological forms such as branches, discontinuities, and networks are difficult to extract continuously and completely, making it impossible to reconstruct the true crack structure. Fourth, conventional... The problem of insufficient pixel-level measurement accuracy and the potential for missing crack information due to obstruction by pipelines and attachments on the lining surface makes it difficult to meet the engineering requirements for refined defect assessment in subway tunnels. To address this, the above measures are implemented by standardizing images through surface distortion correction, adaptive illumination compensation, and edge-preserving denoising. This is combined with a multi-scale segmentation network with an attention mechanism to effectively improve the recall rate of minute cracks. Furthermore, a multi-dimensional feature pseudo-defect classifier and topology reconstruction mechanism are used to reduce the false detection rate and restore the complete crack morphology. Subpixel edge detection is employed to achieve high-precision parameter quantification. Combined with spatial positioning, defect classification, and historical extension tracking, the goal is to output accurate detection results that are compatible with the subway maintenance system.

[0005] (II) Technical Solution This invention is achieved through the following technical solution: an automatic identification method for crack defects in subway tunnel lining based on machine vision, the method comprising the following steps: Step S1: Acquire the original visual image of the subway tunnel lining surface, perform surface distortion correction and preprocessing on the original image to obtain a standardized lining image; Step S2: Perform multi-scale crack feature enhancement and pixel-level semantic segmentation on the standardized lining image to obtain the initial crack segmentation mask; Step S3: Perform crack-like pseudo-defect filtering and crack topology reconstruction on the initial crack segmentation mask to obtain a continuous and complete crack target region. Step S4: Perform sub-pixel level edge extraction and geometric parameter quantization on the crack target area to obtain the crack length, width, direction, and area feature parameters; Step S5: Combine crack characteristic parameters with tunnel mileage and segment ring number information to assess the defect level and output crack defect identification results and spatial positioning information.

[0006] Preferably, step S1 includes the following steps: S11. Adaptive gamma correction and homomorphic filtering are used to compensate for non-uniform illumination in the original image to eliminate local overexposure and shadows caused by tunnel lighting. S12. An edge-preserving noise reduction algorithm is used to suppress noise in the image, preserving the edge details of sub-millimeter cracks; S13. Construct a cylindrical inverse projection mapping matrix based on the parameters of the curved surface of the subway tunnel lining, perform perspective distortion correction on the image, and map the curved surface image into a scaled plane standardized image.

[0007] Preferably, step S2 includes the following steps: S21. Construct a multi-scale feature fusion segmentation network with an attention mechanism. Extract global semantic features of cracks through a deep network and extract fine texture features of cracks through a shallow network. S22. Upsample and merge feature maps of different scales with channels, and output a pixel-level crack probability map with the same size as the original image after weighting by the attention module. S23. Set a segmentation threshold and perform binarization on the crack probability map to obtain the initial crack segmentation mask.

[0008] Preferably, the crack-like pseudo-defect filtering in step S3 includes the following steps: S31. Extract the shape features, grayscale texture features and spatial distribution features of each connected component in the initial segmentation mask; S32. Input the features into the pre-trained pseudo-defect classifier to distinguish between real cracks and pseudo-defects such as subway segment joints, bolt holes, water stains, scratches, and construction cold joints, and filter out non-crack interference areas.

[0009] Preferably, the crack topology reconstruction in step S3 includes the following steps: S33. Extract the skeleton and detect breakpoints in the connected domain of the crack after filtering out pseudo-defects. S34. Based on the constraints of crack orientation consistency and grayscale continuity, the fracture cracks within the adaptive distance threshold are connected and repaired. S35. Perform topological analysis on branching, intersecting, and network cracks to distinguish between main and branch cracks, restore the complete crack morphology, and obtain continuous crack target areas.

[0010] Preferably, after obtaining the target crack area in step S3, the method further includes a crack completion step for the occluded area: S36. Identify areas obscured by pipelines, attachments, and concrete spalling on the lining surface; S37. Based on the crack orientation and width variation trend on both sides of the shielded area, the crack morphology of the shielded section is inferred by fitting a cubic B-spline curve, and the target crack area is completed.

[0011] Preferably, step S4 includes the following steps: S41. The Zernike moment subpixel edge detection algorithm is used to accurately locate the crack target area and obtain the upper and lower edge coordinates of the crack with subpixel accuracy. S42. Calculate the normal width point by point along the crack skeleton, and statistically analyze the maximum width, average width and spatial distribution of the crack. S43. Calculate the actual length and area of ​​the crack by integrating the skeleton chain code and converting it to pixel equivalent. S44. The main direction of the crack is obtained by fitting the edge direction.

[0012] Preferably, step S5 includes the following steps: S51. Based on the maximum width, length and expansion trend of the cracks, the cracks are classified into different defect levels in accordance with the urban rail transit structure inspection specifications. S52. Combining the mileage code of the inspection equipment with the visual recognition results of the segment ring number, mark the spatial coordinates of the crack and the corresponding lining ring number and circumferential position; S53. Generate a graded list of crack defects and their location distribution information, and output it as the identification result.

[0013] Preferably, after outputting the crack defect identification result in step S5, the method further includes: S54. Perform spatiotemporal registration between the current identification results and historical inspection data, match historical cracks at the same location, track crack expansion changes, and generate crack defect development trend analysis results.

[0014] An automatic identification system for crack defects in subway tunnel lining based on machine vision is provided to implement the aforementioned automatic identification method for crack defects in subway tunnel lining based on machine vision. The system includes: The image preprocessing unit is used to acquire the original visual image of the surface of the subway tunnel lining, perform surface distortion correction and preprocessing on the original image, and obtain a standardized lining image. The crack segmentation unit is used to perform multi-scale crack feature enhancement and pixel-level semantic segmentation on the standardized lining image to obtain the initial crack segmentation mask. The topology reconstruction unit is used to perform crack-like pseudo-defect filtering and crack topology reconstruction on the initial crack segmentation mask to obtain a continuous and complete crack target region. The parameter quantization unit is used to perform sub-pixel level edge extraction and geometric parameter quantization on the crack target area to obtain the crack's length, width, orientation, and area feature parameters. The defect assessment unit is used to assess the defect level by combining crack characteristic parameters with tunnel mileage and segment ring number information, and outputs crack defect identification results and spatial positioning information.

[0015] (III) Beneficial Effects This invention provides a machine vision-based method and system for automatic identification of cracks in subway tunnels. It offers the following advantages: I. For the acquisition scenario of curved lining and uneven lighting in subway tunnels, a standardized preprocessing system is constructed by cylindrical inverse projection distortion correction, adaptive illumination compensation and edge-preserving denoising to correct the morphological distortion caused by the curved surface and effectively improve the recognizability of low-contrast fine cracks. Combined with a multi-scale feature fusion segmentation network with attention mechanism, which takes into account both the global semantics and fine texture features of cracks, the problem of easy missed detection and easy breakage of sub-millimeter cracks is effectively solved, and the recall rate and completeness of crack segmentation under complex imaging conditions are greatly improved.

[0016] Second, considering the dense interference from cracks such as joints, bolt holes, and cold joints in the subway lining surface, a pseudo-defect classifier is constructed using multi-dimensional features of shape, texture, and spatial distribution to accurately filter out non-defect interference. At the same time, relying on the breakpoint connection and topology analysis mechanism, continuous and complete reconstruction of discontinuous, branching, and network cracks is achieved. Combined with cubic B-spline curve completion technology for occluded areas, the covered crack segments are restored, significantly reducing the false detection rate and truly restoring the complete topological morphology of cracks.

[0017] Third, Zernike matrix subpixel edge detection is used to achieve high-precision quantification of crack geometric parameters, improving the precision of dimensional measurement; combined with the spatial positioning mechanism of subway tunnel mileage coding and segment ring number visual recognition, automatic defect classification is completed in accordance with industry inspection standards, while supporting spatiotemporal registration and extended tracking of multi-period inspection data. The output results are directly adapted to the subway tunnel maintenance management system, providing accurate and reliable data support for structural safety assessment and maintenance decisions. Attached Figure Description

[0018] Figure 1 A flowchart of a machine vision-based automatic crack identification method for subway tunnels provided by the present invention; Figure 2 This is a schematic diagram of a machine vision-based automatic crack identification system for subway tunnels provided by the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] An embodiment of a machine vision-based automatic crack identification method and system for subway tunnels is as follows: Example 1: Please see Figure 1 An automatic identification method for crack defects in subway tunnel lining based on machine vision, the method comprising the following steps: Step S1: Acquire the original visual image of the subway tunnel lining surface, perform surface distortion correction and preprocessing on the original image to obtain a standardized lining image; Step S2: Perform multi-scale crack feature enhancement and pixel-level semantic segmentation on the standardized lining image to obtain the initial crack segmentation mask; Step S3: Perform crack-like pseudo-defect filtering and crack topology reconstruction on the initial crack segmentation mask to obtain a continuous and complete crack target region. Step S4: Perform sub-pixel level edge extraction and geometric parameter quantization on the crack target area to obtain the crack length, width, direction, and area feature parameters; Step S5: Combine crack characteristic parameters with tunnel mileage and segment ring number information to assess the defect level and output crack defect identification results and spatial positioning information.

[0021] Step S1 includes the following steps: S11. Adaptive gamma correction and homomorphic filtering are used to compensate for non-uniform illumination in the original image to eliminate local overexposure and shadows caused by tunnel lighting. Specifically, in embodiment S11, the inspection robot is equipped with five area array cameras and corresponding diffuse reflection strip lighting sources arranged in a circumferential direction. The field of view of a single camera covers a 72° circumferential range. The frame rate of acquisition is matched with the robot's travel speed to ensure that the axial overlap rate of adjacent frames is not less than 30% to avoid motion blur. After acquisition, a 360° panoramic lining unfolded image is generated through an image stitching algorithm, covering the entire surface of the arch, left and right arch waists, and sidewall segments, as the original visual image input. To address the problem of uneven lighting in the tunnel environment and the easy generation of light spots and shadows by supplementary lighting, the original image is converted to grayscale space. The local grayscale mean is calculated using a 32×32 pixel sliding window. The gamma correction coefficient is dynamically adjusted according to the local grayscale to enhance the brightness of dark areas and suppress the brightness of bright areas, thus completing adaptive gamma correction. Then, Gaussian-homomorphic filtering is used to separate the illuminance component and reflection component of the image, compressing the dynamic range of the illuminance component and enhancing the detail contrast of the reflection component. This effectively eliminates the light spots on the arch and the dark areas on the sidewalls caused by supplementary lighting, improving the global contrast of the image and the recognition of crack edges.

[0022] S12. An edge-preserving noise reduction algorithm is used to suppress noise in the image, preserving the edge details of sub-millimeter cracks; Specifically, in embodiment S12, a guided filtering algorithm is used to suppress image noise caused by dust and water vapor in the tunnel. The original grayscale image is used as the guide image, and the filter window radius is set to 5 pixels and the regularization coefficient is 0.02. While filtering out random noise, the edge gradient information of minute cracks with a width of 0.1mm is completely preserved. Compared with bilateral filtering, guided filtering has higher computational efficiency and is more suitable for the real-time processing requirements of the edge. At the same time, it avoids the problem of crack edge blurring and detail loss caused by traditional Gaussian filtering, and ensures the effectiveness of subsequent segmentation in identifying minute cracks.

[0023] S13. Construct a cylindrical inverse projection mapping matrix based on the parameters of the curved surface of the subway tunnel lining, perform perspective distortion correction on the image, and map the curved surface image into a scaled plane standardized image.

[0024] Specifically, in embodiment S13, based on the design inner diameter parameters of the subway tunnel segment and the camera imaging geometric model, an inverse projection mapping matrix from the cylindrical surface to the plane is constructed. The image is resampled using bilinear interpolation, and the lateral stretching and edge distortion caused by the curved surface imaging are corrected into a proportional planar image. After correction, the straightness error of the segment joint is controlled within 1 pixel, eliminating the distortion effect of the arc wall on the crack shape and size, ensuring the accuracy of subsequent geometric parameter quantification, and finally obtaining a standardized lining image.

[0025] For acquisition scenarios involving curved linings and uneven lighting in subway tunnels, a standardized preprocessing system is constructed using cylindrical inverse projection distortion correction, adaptive illumination compensation, and edge-preserving denoising. This system corrects the morphological distortion caused by the curved surface and effectively improves the discernibility of low-contrast fine cracks. Combined with a multi-scale feature fusion segmentation network with an attention mechanism, which takes into account both the global semantics and fine texture features of cracks, the system effectively solves the problems of easy missed detection and easy breakage of sub-millimeter cracks, and significantly improves the recall and completeness of crack segmentation under complex imaging conditions.

[0026] Step S2 includes the following steps: S21. Construct a multi-scale feature fusion segmentation network with an attention mechanism. Extract global semantic features of cracks through a deep network and extract fine texture features of cracks through a shallow network. Specifically, in embodiment S21, an encoder-decoder segmentation network based on an improved U-Net is constructed. The encoder uses ResNet-18 as its backbone. After inputting a standardized lining image, it undergoes downsampling through four residual stages, outputting feature maps at four scales: 1 / 2, 1 / 4, 1 / 8, and 1 / 16. The 1 / 16 deep feature map extracts global semantic features of the cracks, used to distinguish cracks from large background areas. The 1 / 2 shallow feature map retains high-resolution detail information, used to capture the subtle textures and edge features of the cracks, compensating for detail loss caused by the downsampling process. A channel attention module is introduced in the skip connection between the encoder and decoder to adaptively strengthen the weights of the crack feature channels and suppress background noise channels, achieving a balance between global semantic enhancement and local detail enhancement.

[0027] S22. Upsample and merge feature maps of different scales with channels, and output a pixel-level crack probability map with the same size as the original image after weighting by the attention module. Specifically, in embodiment S22, the decoding end adopts a fusion structure of progressive upsampling and splicing. After bilinear upsampling of the deep feature map, it is spliced ​​with the shallow feature map of the corresponding scale, and then the feature is fused through a 3×3 convolutional layer. This process is repeated to progressively restore the resolution of the feature map. Finally, a single-channel crack probability map with the same size as the input image is output through a 1×1 convolutional layer. The value of each pixel is in the range of [0,1]. The larger the value, the higher the probability that the pixel belongs to the crack, thus realizing pixel-level crack probability prediction.

[0028] S23. Set a segmentation threshold and perform binarization on the crack probability map to obtain the initial crack segmentation mask.

[0029] Specifically, in embodiment S23, a segmentation threshold of 0.5 is set. Regions with pixel values ​​greater than or equal to the threshold in the probability map are marked as crack pixels, and regions with pixel values ​​less than the threshold are marked as background pixels, generating a binarized initial crack segmentation mask. The network training uses labeled subway tunnel crack image samples, and the loss function is a weighted combination of cross-entropy loss and Dice loss to alleviate the imbalance problem of extremely low crack sample ratio. After training, the network's segmentation recall rate for cracks with a width of 0.1 mm or more meets the requirements of engineering applications and can adapt to tunnel scenarios with different lining materials and different levels of pollution.

[0030] The crack-like pseudo-defect filtering in step S3 includes the following steps: S31. Extract the shape features, grayscale texture features and spatial distribution features of each connected component in the initial segmentation mask; Specifically, in embodiment S31, connected component analysis is performed on the initial crack segmentation mask to extract a 12-dimensional feature vector for each independent connected component: shape features include aspect ratio, rectangularity, circularity, and contour complexity (perimeter² / 4π area); grayscale texture features include mean grayscale value, grayscale variance, mean gradient magnitude, and mean local binary pattern within the connected component; spatial distribution features include circumferential position proportion, axial length proportion, and positional overlap with the segment joint, providing multi-dimensional basis for subsequent classification and discrimination.

[0031] S32. Input the features into the pre-trained pseudo-defect classifier to distinguish between real cracks and pseudo-defects such as subway segment joints, bolt holes, water stains, scratches, and construction cold joints, and filter out non-crack interference areas.

[0032] Specifically, in embodiment S32, the feature vector is input into a pre-trained support vector machine classifier. The classifier uses a radial basis kernel function and is trained with 5000 labeled samples. The samples cover five typical types of subway tunnel cracks and pseudo-defects: real cracks and segment joints, bolt holes, construction cold joints, water seepage stains, and surface scratches. The classifier outputs the confidence score of each connected component as a real crack. A confidence score ≥ 0.7 is judged as a real crack and retained, while a confidence score < 0.7 is judged as a pseudo-defect and removed from the mask, which greatly reduces the false detection rate of crack identification.

[0033] The crack topology reconstruction in step S3 includes the following steps: S33. Extract the skeleton and detect breakpoints in the connected domain of the crack after filtering out pseudo-defects. Specifically, in embodiment S33, the Zhang-Suen thinning algorithm is used to extract the skeleton of the preserved crack connected domain to obtain a crack skeleton with a width of one pixel; all skeleton pixels are traversed, and the endpoints containing only one neighboring skeleton point and the intersections containing three or more neighboring skeleton points are marked to provide basic structural data for subsequent fracture repair and topology analysis.

[0034] S34. Based on the constraints of crack orientation consistency and grayscale continuity, the fracture cracks within the adaptive distance threshold are connected and repaired. Specifically, in embodiment S34, the Euclidean distance and direction difference between all pairs of endpoints are calculated. The distance threshold is adaptively set to 20 pixels corresponding to the physical size based on the image pixel equivalent, and the direction threshold is set to 30°. When the distance between the two endpoints is within the threshold and the direction deviation is less than the threshold, it is determined to be the break point of the same crack. A straight line segment is used to connect the two endpoints to repair the discontinuous cracks caused by low contrast and noise interference, thereby improving the continuity and integrity of the cracks.

[0035] S35. Perform topological analysis on branching, intersecting, and network cracks to distinguish between main and branch cracks, restore the complete crack morphology, and obtain continuous crack target areas.

[0036] Specifically, in embodiment S35, the cross points on the skeleton are used as nodes to divide the crack skeleton into multiple independent crack segments. The main cracks and branch cracks are distinguished according to the length and direction of each segment. The branch positions and branch angles are recorded to restore the complete topological structure of the branch and network cracks, so as to avoid complex cracks being mistakenly divided into multiple independent cracks and ensure the authenticity of crack morphology restoration.

[0037] After obtaining the target crack area in step S3, the method further includes a crack completion step for the occluded area: S36. Identify areas obscured by pipelines, attachments, and concrete spalling on the lining surface; Specifically, in embodiment S36, a lightweight target detection model is used to identify three typical obstructions on the lining surface: fire pipelines, cable boxes, and concrete blocks, and outputs a rectangular frame outside the obstruction area; when the crack skeleton passes through the obstruction area, it is determined to be an obstruction crack, and the position of the break point to be filled is marked.

[0038] S37. Based on the crack orientation and width variation trend on both sides of the shielded area, the crack morphology of the shielded section is inferred by fitting a cubic B-spline curve, and the target crack area is completed.

[0039] Specifically, in embodiment S37, the coordinates of the crack endpoints, tangent direction and width values ​​on both sides of the obscured area are extracted, and the center line of the crack in the obscured section is fitted with a cubic B-spline curve to ensure a smooth transition. At the same time, based on the linear change trend of the width on both sides, the crack width at each position of the obscured section is interpolated to complete the crack edge and skeleton of the obscured area, restore the covered crack segment, and reduce the impact of the obscured object on the crack integrity detection.

[0040] Step S4 includes the following steps: S41. The Zernike moment subpixel edge detection algorithm is used to accurately locate the crack target area and obtain the upper and lower edge coordinates of the crack with subpixel accuracy. Specifically, in embodiment S41, a Zernike moment subpixel edge detection algorithm with a 7×7 template is used to perform subpixel-level fine positioning of the initial pixel-level edge of the crack target area; by calculating the Zernike moments of the image area to reconstruct the subpixel position of the edge, the edge coordinate accuracy is improved to 0.1 pixel level, and a high-precision coordinate point set of the upper and lower edges of the crack is obtained; in this embodiment, the image pixel equivalent is 0.08 mm / pixel, and after subpixel processing, the edge positioning accuracy can reach 0.008 mm, providing an accurate edge reference for subsequent size parameter quantization and solving the problem of insufficient traditional pixel-level detection accuracy.

[0041] S42. Calculate the normal width point by point along the crack skeleton, and statistically analyze the maximum width, average width and spatial distribution of the crack. Specifically, in embodiment S42, the normal direction of the skeleton tangent is calculated point by point along the crack skeleton, and the distance between the intersection points of the normal and the upper and lower edges of the crack is solved, which is the normal width of the crack at that location. The normal width calculation method can avoid the measurement error of the width of inclined cracks and is more in line with the engineering definition than the horizontal / vertical width measurement. By traversing all skeleton points, the width sequence of the entire crack is obtained, and the maximum width, average width, and width distribution data along the length direction of the crack are statistically output, which comprehensively reflects the width characteristics of the crack.

[0042] S43. Calculate the actual length and area of ​​the crack by integrating the skeleton chain code and converting it to pixel equivalent. Specifically, in embodiment S43, an eight-neighbor chain code is used to encode the crack skeleton. The pixel length of the skeleton is calculated by chain code integration, and the actual length of the crack is obtained by combining the pixel equivalent of the image. This method has higher accuracy than simple pixel accumulation. The total number of pixels in the crack target area is counted and multiplied by the square of the pixel equivalent to obtain the actual area of ​​the crack, thereby achieving accurate quantification of the crack geometry.

[0043] S44. The main direction of the crack is obtained by fitting the edge direction.

[0044] Specifically, in embodiment S44, least squares straight line fitting is performed on all sub-pixel points of the crack skeleton, and the angle between the fitted straight line and the horizontal direction is calculated, which is the main direction angle of the crack. The angle range is defined as 0°-180°. This parameter can characterize the extension direction of the crack and help to determine the stress cause and development trend of the crack.

[0045] To address the issue of dense interference from cracks such as joints, bolt holes, and cold joints in subway lining surfaces, a pseudo-defect classifier is constructed using multi-dimensional features of shape, texture, and spatial distribution to accurately filter out non-defect interference. Simultaneously, relying on breakpoint connection and topology analysis mechanisms, continuous and complete reconstruction of discontinuous, branching, and network cracks is achieved. Combined with cubic B-spline curve completion technology for obscured areas, the covered crack segments are restored, significantly reducing the false detection rate and accurately restoring the complete topological morphology of cracks.

[0046] Step S5 includes the following steps: S51. Based on the maximum width, length and expansion trend of the cracks, the cracks are classified into different defect levels in accordance with the urban rail transit structure inspection specifications. Specifically, in Example S51, in accordance with the "Technical Standard for Testing Urban Rail Transit Structures", the maximum width of the crack is used as the core classification index to classify crack defects into four levels: Level I (minor), Level II (general), Level III (relatively serious), and Level IV (serious), corresponding to different treatment requirements. At the same time, the level is corrected by combining the crack length and historical expansion rate to ensure that the defect classification results meet the requirements of engineering specifications and provide a classification basis for subsequent maintenance decisions.

[0047] S52. Combining the mileage code of the inspection equipment with the visual recognition results of the segment ring number, mark the spatial coordinates of the crack and the corresponding lining ring number and circumferential position; Specifically, in embodiment S52, the track mileage encoder data (positioning accuracy ±1cm) of the inspection robot is combined with the visual recognition results of the segment ring number to convert the pixel position of the crack into the actual mileage and circumferential angle of the tunnel, and the segment ring number to which the crack belongs and the specific circumferential position (such as 0° at the arch top, 90° at the left arch waist, and 180° at the sidewall) are marked. The segment ring number is identified by matching the joint features with the preset ring number identifier. The overall spatial positioning error meets the maintenance management requirements and makes it easy for maintenance personnel to quickly locate the defect position.

[0048] S53. Generate a graded list of crack defects and their location distribution information, and output it as the identification result.

[0049] Specifically, in embodiment S53, a standardized list of crack defects is generated, including fields such as defect number, mileage location, segment ring number, circumferential location, length, maximum width, average width, direction, and defect level. At the same time, a statistical map of the circumferential distribution of cracks is output. The data format is compatible with mainstream subway tunnel maintenance management systems and can be directly imported synchronously to realize the digital management of inspection results.

[0050] After outputting the crack defect identification result in step S5, the method further includes: S54. Perform spatiotemporal registration between the current identification results and historical inspection data, match historical cracks at the same location, track crack expansion changes, and generate crack defect development trend analysis results.

[0051] Specifically, in embodiment S54, the current identification results are spatiotemporally registered with historical inspection data. First, coarse registration is performed based on the segment ring number, and then fine registration is completed by matching crack feature points. Historical crack data at the same location are matched, and the changes in crack length and width and monthly expansion rate are calculated. Cracks that are not matched are marked as new defects. Finally, a crack defect development trend analysis report is generated, and an early warning is issued for cracks with abnormal expansion rates to support the long-term structural health assessment of the tunnel.

[0052] The Zernike matrix subpixel edge detection method is used to achieve high-precision quantification of crack geometric parameters, improving the accuracy of dimensional measurement. Combined with the spatial positioning mechanism of subway tunnel mileage coding and segment ring number visual recognition, the method completes automatic defect classification according to industry inspection standards. At the same time, it supports spatiotemporal registration and extended tracking of multi-period inspection data. The output results are directly adapted to the subway tunnel maintenance management system, providing accurate and reliable data support for structural safety assessment and maintenance decision-making.

[0053] Example 2: Please see Figure 2 An automatic identification system for crack defects in subway tunnel lining based on machine vision is provided to implement the aforementioned automatic identification method for crack defects in subway tunnel lining. The system includes: The image preprocessing unit is used to acquire the original visual image of the surface of the subway tunnel lining, perform surface distortion correction and preprocessing on the original image, and obtain a standardized lining image. The crack segmentation unit is used to perform multi-scale crack feature enhancement and pixel-level semantic segmentation on the standardized lining image to obtain the initial crack segmentation mask. The topology reconstruction unit is used to perform crack-like pseudo-defect filtering and crack topology reconstruction on the initial crack segmentation mask to obtain a continuous and complete crack target region. The parameter quantization unit is used to perform sub-pixel level edge extraction and geometric parameter quantization on the crack target area to obtain the crack's length, width, orientation, and area feature parameters. The defect assessment unit is used to assess the defect level by combining crack characteristic parameters with tunnel mileage and segment ring number information, and outputs crack defect identification results and spatial positioning information.

[0054] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for automatic identification of crack defects in subway tunnel lining based on machine vision, characterized in that, The method includes the following steps: Step S1: Acquire the original visual image of the subway tunnel lining surface, perform surface distortion correction and preprocessing on the original image to obtain a standardized lining image; Step S2: Perform multi-scale crack feature enhancement and pixel-level semantic segmentation on the standardized lining image to obtain the initial crack segmentation mask; Step S3: Perform crack-like pseudo-defect filtering and crack topology reconstruction on the initial crack segmentation mask to obtain a continuous and complete crack target region. Step S4: Perform sub-pixel level edge extraction and geometric parameter quantization on the crack target area to obtain the crack length, width, direction, and area feature parameters; Step S5: Combine crack characteristic parameters with tunnel mileage and segment ring number information to assess the defect level and output crack defect identification results and spatial positioning information.

2. The automatic identification method for crack defects in subway tunnel lining based on machine vision according to claim 1, characterized in that, Step S1 includes the following steps: S11. Adaptive gamma correction and homomorphic filtering are used to compensate for non-uniform illumination in the original image to eliminate local overexposure and shadows caused by tunnel lighting. S12. An edge-preserving noise reduction algorithm is used to suppress noise in the image, preserving the edge details of sub-millimeter cracks; S13. Construct a cylindrical inverse projection mapping matrix based on the parameters of the curved surface of the subway tunnel lining, perform perspective distortion correction on the image, and map the curved surface image into a scaled plane standardized image.

3. The automatic identification method for crack defects in subway tunnel lining based on machine vision according to claim 1, characterized in that, Step S2 includes the following steps: S21. Construct a multi-scale feature fusion segmentation network with an attention mechanism. Extract global semantic features of cracks through a deep network and extract fine texture features of cracks through a shallow network. S22. Upsample and merge feature maps of different scales with channels, and output a pixel-level crack probability map with the same size as the original image after weighting by the attention module. S23. Set a segmentation threshold and perform binarization on the crack probability map to obtain the initial crack segmentation mask.

4. The automatic identification method for crack defects in subway tunnel lining based on machine vision according to claim 1, characterized in that, The crack-like pseudo-defect filtering in step S3 includes the following steps: S31. Extract the shape features, grayscale texture features and spatial distribution features of each connected component in the initial segmentation mask; S32. Input the features into the pre-trained pseudo-defect classifier to distinguish between real cracks and pseudo-defects such as subway segment joints, bolt holes, water stains, scratches, and construction cold joints, and filter out non-crack interference areas.

5. The automatic identification method for crack defects in subway tunnel lining based on machine vision according to claim 4, characterized in that, The crack topology reconstruction in step S3 includes the following steps: S33. Extract the skeleton and detect breakpoints in the connected domain of the crack after filtering out pseudo-defects. S34. Based on the constraints of crack orientation consistency and grayscale continuity, the fracture cracks within the adaptive distance threshold are connected and repaired. S35. Perform topological analysis on branching, intersecting, and network cracks to distinguish between main and branch cracks, restore the complete crack morphology, and obtain continuous crack target areas.

6. The automatic identification method for crack defects in subway tunnel lining based on machine vision according to claim 1, characterized in that, After obtaining the target crack area in step S3, the method further includes a crack completion step for the occluded area: S36. Identify areas obscured by pipelines, attachments, and concrete spalling on the lining surface; S37. Based on the crack orientation and width variation trend on both sides of the shielded area, the crack morphology of the shielded section is inferred by fitting a cubic B-spline curve, and the target crack area is completed.

7. The automatic identification method for crack defects in subway tunnel lining based on machine vision according to claim 1, characterized in that, Step S4 includes the following steps: S41. The Zernike moment subpixel edge detection algorithm is used to accurately locate the crack target area and obtain the upper and lower edge coordinates of the crack with subpixel accuracy. S42. Calculate the normal width point by point along the crack skeleton, and statistically analyze the maximum width, average width and spatial distribution of the crack. S43. Calculate the actual length and area of ​​the crack by integrating the skeleton chain code and converting it to pixel equivalent. S44. The main direction of the crack is obtained by fitting the edge direction.

8. The automatic identification method for crack defects in subway tunnel lining based on machine vision according to claim 1, characterized in that, Step S5 includes the following steps: S51. Based on the maximum width, length and expansion trend of the cracks, the cracks are classified into different defect levels in accordance with the urban rail transit structure inspection specifications. S52. Combining the mileage code of the inspection equipment with the visual recognition results of the segment ring number, mark the spatial coordinates of the crack and the corresponding lining ring number and circumferential position; S53. Generate a graded list of crack defects and their location distribution information, and output it as the identification result.

9. The automatic identification method for crack defects in subway tunnel lining based on machine vision according to claim 8, characterized in that, After outputting the crack defect identification result in step S5, the method further includes: S54. Perform spatiotemporal registration between the current identification results and historical inspection data, match historical cracks at the same location, track crack expansion changes, and generate crack defect development trend analysis results.

10. A machine vision-based automatic identification system for crack defects in subway tunnel lining, used to implement the machine vision-based automatic identification method for crack defects in subway tunnel lining as described in any one of claims 1-9, characterized in that, The system includes: The image preprocessing unit is used to acquire the original visual image of the surface of the subway tunnel lining, perform surface distortion correction and preprocessing on the original image, and obtain a standardized lining image. The crack segmentation unit is used to perform multi-scale crack feature enhancement and pixel-level semantic segmentation on the standardized lining image to obtain the initial crack segmentation mask. The topology reconstruction unit is used to perform crack-like pseudo-defect filtering and crack topology reconstruction on the initial crack segmentation mask to obtain a continuous and complete crack target region. The parameter quantization unit is used to perform sub-pixel level edge extraction and geometric parameter quantization on the crack target area to obtain the crack's length, width, orientation, and area feature parameters. The defect assessment unit is used to assess the defect level by combining crack characteristic parameters with tunnel mileage and segment ring number information, and outputs crack defect identification results and spatial positioning information.

Citation Information

Patent Citations

  • Building crack monitoring method and system based on machine learning

    CN119477886A