Tunnel lining multi-defect intelligent identification and detection method based on target detection

Abstract

The present application relates to the technical field of image recognition, and particularly relates to a tunnel lining multi-defect intelligent recognition detection method based on target detection, which comprises the following steps: preprocessing a tunnel lining image and extracting a lining region; identifying a lining inherent structure feature region in the lining region based on the symmetry of left and right images of the same tunnel section and the periodicity of tunnel longitudinal construction; for each lining region, segmenting a suspected defect region in the lining region, determining the defect complexity of the lining region according to the distance correlation and structure similarity between the suspected defect region obtained by segmentation and the lining inherent structure feature region; determining a target detection outlet corresponding to the defect complexity in a preset multi-outlet detection network, and inputting the lining region into the preset multi-outlet detection network to obtain a defect detection result of the tunnel lining corresponding to the target detection outlet. The present application improves the detection accuracy and efficiency when detecting defects of the tunnel lining.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and specifically to an intelligent identification and detection method for multiple defects in tunnel lining based on target detection. Background Technology

[0002] As the core load-bearing and waterproofing structure of tunnel engineering, the health of tunnel lining is directly related to the operational safety of the tunnel. During long-term service, the lining is subjected to the coupling effects of multiple physical fields such as train dynamic loads, groundwater erosion, and stress redistribution in the surrounding rock, which can easily lead to typical defects such as cracks, water leakage, spalling, and exposed rebar, seriously threatening the structural safety and operational durability of the tunnel.

[0003] In recent years, deep learning-based computer vision methods have been widely used in the field of tunnel defect detection. Among them, object detection-based methods can simultaneously locate and classify multiple defects in images, and have become the mainstream technical route for tunnel lining defect identification.

[0004] Existing methods for detecting defects in tunnel lining typically employ a unified depth target detection network to perform full-network inference on all acquired images, enabling the identification and localization of various defects. However, the surface environment of tunnel lining is complex. In some areas, defects have simple morphologies and significant contrast with the background, while in others, defects are highly similar to the inherent structural features of the lining surface, making identification difficult. The unified network inference approach struggles to differentiate between areas of varying complexity, resulting in a trade-off between detection accuracy and efficiency, thus hindering the practical engineering application of intelligent tunnel lining defect detection technology. Summary of the Invention

[0005] To address the technical challenge of balancing detection accuracy and efficiency in tunnel lining defect detection, this invention aims to provide an intelligent identification and detection method for multiple defects in tunnel lining based on target detection. The specific technical solution adopted is as follows: Firstly, a method for intelligent identification and detection of multiple defects in tunnel lining based on target detection is provided. This method includes: simultaneously acquiring tunnel lining images using image acquisition devices symmetrically installed on both sides of an inspection vehicle; preprocessing the tunnel lining images to extract lining regions; identifying inherent structural feature regions of the lining within the lining region based on the symmetry of left and right images of the same tunnel cross-section and the periodicity of longitudinal tunnel construction; segmenting the lining region into suspected defect areas for each lining region; determining the defect complexity of the lining region based on the distance correlation and structural similarity between the segmented suspected defect areas and the inherent structural feature regions of the lining; determining the target detection exit in a preset multi-exit detection network corresponding to the defect complexity, and inputting the lining region into the preset multi-exit detection network to obtain the defect detection result of the tunnel lining corresponding to the target detection exit.

[0006] In one possible design, images of the tunnel lining are simultaneously acquired by image acquisition devices symmetrically mounted on both sides of the inspection vehicle. The tunnel lining images are then preprocessed to extract the lining region. This process includes: the image acquisition devices are area array cameras, installed at the same height as the tunnel arch, with the shooting angle perpendicular to the lining surface. The overlapping area of ​​the fields of view of the two area array cameras covers the symmetrical central axis of the tunnel lining. A synchronous trigger is used to control the two image acquisition devices to simultaneously acquire tunnel lining images. The acquired tunnel lining images are then subjected to distortion correction and dehazing processing, and a preset segmentation algorithm is used to remove interfering pixel areas corresponding to the inspection vehicle body, track, and lighting fixtures, thereby extracting the lining region.

[0007] In one possible design, based on the symmetry of left and right images of the same tunnel cross-section and the periodicity of longitudinal tunnel construction, the inherent structural feature regions of the lining within the lining area are identified. This includes: registering feature points of the left and right images of the lining area of ​​the same tunnel cross-section; performing symmetry matching based on the registered images to obtain initial symmetrical feature regions; stitching the continuously acquired lining areas into a longitudinal unfolded image according to the time sequence; performing periodic analysis on the longitudinal unfolded image to obtain periodic feature regions; and determining the intersection of the initial symmetrical feature regions and the periodic feature regions as the inherent structural feature regions of the lining.

[0008] In one possible design, feature points of the lining area of ​​the same tunnel cross section are registered between left and right images. Based on the registered images, symmetry matching is performed to obtain an initial symmetrical feature region. This includes: extracting feature points from the left and right images of the same cross section using a preset feature matching algorithm, removing mismatched points based on a preset mismatch removal algorithm, registering the left and right images to the same coordinate system to obtain a registered and stitched image, where the pixel grayscale value of the overlapping area in the registered and stitched image is the average of the pixel grayscale values ​​at corresponding positions in the left and right images; using the vertical midline of the registered and stitched image as the axis of symmetry, determining the symmetry similarity of each pixel based on the difference between the grayscale value of each pixel in the registered and stitched image and the grayscale value of the pixel at the symmetrical position; marking pixels with a symmetry similarity greater than or equal to a preset symmetry threshold as symmetrical feature points, and extracting connected components from all symmetrical feature points to obtain the initial symmetrical feature region.

[0009] In one possible design, continuously acquired lining areas are stitched together in a time sequence to form a longitudinal unfolded map. Periodic analysis is then performed on the longitudinal unfolded map to obtain periodic feature regions. This includes: calculating grayscale projection curves along the forward direction of the inspection vehicle on the longitudinal unfolded map; the grayscale projection curves characterize the grayscale distribution of the lining area along the longitudinal direction; performing a pre-defined correlation analysis on the grayscale projection curves to obtain an autocorrelation coefficient sequence; the autocorrelation coefficient sequence characterizes the similarity of the grayscale projection curves under different longitudinal displacements; determining the longitudinal period length based on the maximum value of the autocorrelation coefficient sequence; and locating periodically recurring areas in the longitudinal unfolded map based on the longitudinal period length to obtain the periodic feature regions.

[0010] In one possible design, the lining area is segmented into suspected defect areas, including: for each pixel in the lining area, determining the average gray value of the pixel's neighborhood based on the gray values ​​of each pixel within a preset neighborhood range; if the absolute difference between the pixel's gray value and the average gray value of its neighborhood exceeds a preset gray value difference threshold, the pixel is marked as a suspected defect pixel; and connected component extraction is performed on all suspected defect pixels to obtain the suspected defect area.

[0011] In one possible design, determining the distance correlation between the segmented suspected defect areas and the inherent structural feature areas of the lining includes: for each suspected defect area, determining the Euclidean distance from the geometric center of the suspected defect area to the edge of each inherent structural feature area of ​​the lining, and taking the minimum value among the Euclidean distances as the minimum Euclidean distance of the suspected defect area; normalizing the minimum Euclidean distance of the suspected defect area using the diagonal length of the lining area as the normalization benchmark to obtain the normalized distance of the suspected defect area; and determining the distance correlation between the suspected defect areas based on the normalized distance, wherein the distance correlation is negatively correlated with the normalized distance.

[0012] In one possible design, determining the structural similarity between the segmented suspected defect regions and the inherent structural feature regions of the lining includes: extracting the aspect ratio, area ratio, mean edge gradient, and mean grayscale value of all inherent structural feature regions of the lining, and taking the average of each to form a first multidimensional feature vector; for each suspected defect region, extracting the aspect ratio, area ratio, mean edge gradient, and mean grayscale value of the suspected defect region to form a second multidimensional feature vector corresponding to the suspected defect region; and determining the cosine similarity between the second multidimensional feature vector corresponding to the suspected defect region and the first multidimensional feature vector as the structural similarity of the suspected defect region.

[0013] In one possible design, determining the defect complexity of the lining region includes: determining the maximum distance correlation among all suspected defective regions within the lining region as the maximum distance correlation; determining the maximum structural similarity among all suspected defective regions within the lining region as the maximum structural similarity; determining a first weight and a second weight based on the sum of the distance correlations and the sum of the structural similarities of all suspected defective regions within the lining region; and determining the defect complexity of the lining region based on the maximum distance correlation, the maximum structural similarity, the first weight, and the second weight.

[0014] In one possible design, a pre-defined multi-exit detection network includes a shallow detection head, a middle detection head, a mid-deep detection head, and a deep main detection head, corresponding to different network depths from shallow to deep. Determining the target detection exit in the pre-defined multi-exit detection network corresponding to the defect complexity includes: based on the defect complexity distribution of labeled test samples, using a pre-defined threshold determination algorithm to determine a first exit threshold, a second exit threshold, and a third exit threshold, where the first exit threshold is less than the second exit threshold, and the second exit threshold is less than the third exit threshold; if the defect complexity in the lining area is less than or equal to the first exit threshold, the exit corresponding to the shallow detection head is determined as the target detection exit; if the defect complexity in the lining area is greater than the first exit threshold but less than or equal to the second exit threshold, the exit corresponding to the middle detection head is determined as the target detection exit; if the defect complexity in the lining area is greater than the second exit threshold but less than or equal to the third exit threshold, the exit corresponding to the mid-deep detection head is determined as the target detection exit; if the defect complexity in the lining area is greater than the third exit threshold, the exit corresponding to the deep main detection head is determined as the target detection exit.

[0015] The present invention has the following beneficial effects: In the intelligent identification and detection method for multiple defects in tunnel lining based on target detection provided by this invention, images of the tunnel lining are simultaneously acquired by image acquisition devices symmetrically installed on both sides of the inspection vehicle, and the lining area is extracted after preprocessing, providing a high-quality image data foundation for subsequent analysis. Based on this, the inherent structural features within the lining area are identified and extracted using the inherent left-right symmetry and longitudinal construction periodicity of the tunnel lining, effectively distinguishing between inherent structural features and actual defects on the lining surface. This solves the problem of general detection networks misclassifying inherent structures such as construction joints, embedded channels, and drainage holes as defects, significantly reducing the false detection rate and providing a reliable reference benchmark for subsequent defect complexity assessment. Based on the distinction between inherent structures and actual defects, each lining area is further segmented into suspected defect regions. The distance correlation and structural similarity between suspected defect regions and inherent structural feature regions are comprehensively considered to adaptively determine the defect complexity of each lining region. This defect complexity accurately reflects the degree of confusion between defects and inherent structures within the lining area, providing a quantitative basis for the dynamic reasoning of the detection network. Finally, based on the complexity of the defects, different lining areas are matched to detection exits at different depths in a pre-set multi-exit detection network for inference. Lining areas with lower defect complexity can be detected at the shallow layer of the network and exit the inference process early, while lining areas with higher defect complexity are detected using deep features. This effectively reduces the waste of computing resources for simple defects while ensuring that all types of defects can be accurately identified. It achieves a balance between detection accuracy and detection efficiency, meeting the engineering requirements for detection efficiency and recognition accuracy in the periodic inspection of tunnel linings. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a method for intelligent identification and detection of multiple defects in tunnel lining based on target detection, provided in one embodiment of the present invention. Detailed Implementation

[0017] The following description, in conjunction with the accompanying drawings, details a specific scheme for an intelligent identification and detection method for multiple defects in tunnel lining based on target detection, provided by the present invention.

[0018] Please see Figure 1 The diagram illustrates a flowchart of an intelligent identification and detection method for multiple defects in tunnel lining based on target detection, according to an embodiment of the present invention, including the following steps S101-S104.

[0019] S101. The tunnel lining images are simultaneously acquired by image acquisition devices symmetrically installed on both sides of the inspection vehicle. The tunnel lining images are preprocessed to extract the lining area.

[0020] In some embodiments, before conducting tunnel lining inspection, the image acquisition equipment is first installed and debugged. This equipment uses a regional array camera, which can be a global shutter regional array camera. One regional array camera is installed on each of the left and right sides of the inspection vehicle, symmetrically arranged about the vehicle's longitudinal central axis. The installation height of the regional array cameras is flush with the tunnel arch, and the shooting angle is perpendicular to the lining surface to ensure minimal geometric distortion of the lining area in the acquired images. The overlapping area of ​​the two regional array cameras covers the symmetrical central axis of the tunnel lining to ensure no blind spots appear in the central area of ​​the tunnel cross-section. In this embodiment, the resolution and frame rate settings of the regional array cameras should meet the requirements of non-overlapping and non-missing image acquisition at a certain speed. For example, at a vehicle speed of 30 km / h, the resolution is set to 1920×1080, and the frame rate is set to 30fps.

[0021] After the area array camera is installed, a synchronous trigger is used to control the two image acquisition devices to synchronously acquire images of the tunnel lining. The synchronous trigger sends acquisition signals to the two image acquisition devices simultaneously, controlling the acquisition time difference between the two image acquisition devices to be less than 1ms, so as to ensure that the left and right images acquired at the same time correspond to the lining areas on both sides of the same tunnel section.

[0022] After image acquisition is completed, the acquired tunnel lining images are preprocessed, which includes three sub-steps: distortion correction, defogging, and lining area extraction.

[0023] In the distortion correction sub-step, based on the pre-calibrated camera calibration parameters, geometric correction is performed on each original image to eliminate image geometric deformation caused by lens distortion and restore the true geometry of the lining surface. The camera calibration parameters, including the camera's intrinsic parameter matrix and distortion coefficients, can be obtained through the offline calibration process of a standard calibration board.

[0024] In the dehazing sub-step, the dark channel prior algorithm can be used to dehaze the distortion-corrected image. The tunnel environment contains interference factors such as dust, water mist, and light glare, leading to decreased contrast and blurred details in the acquired images. The dark channel prior algorithm, based on the statistical regularity of dark pixels in local areas of fog-free images, estimates atmospheric light values ​​and transmittance to dehaze and restore the image, improving image clarity and contrast, and making the texture details and defect features of the lining surface more prominent.

[0025] In the sub-step of lining area extraction, a preset segmentation algorithm is used to segment the image after dehazing, extract the lining area, and remove the interfering pixel areas corresponding to the detection vehicle body, track and lighting fixtures.

[0026] Optionally, in one exemplary implementation, since the spatial positions of the detection vehicle body, track, and lighting fixtures in the acquired image are relatively fixed when the installation position and shooting angle of the image acquisition device are fixed, the position range of the aforementioned interference pixel area in the image can be pre-marked. When preprocessing each image, the pixels within this position range are directly removed, and the remaining pixel area is the lining area.

[0027] In another exemplary implementation, an adaptive threshold segmentation algorithm is adopted based on the grayscale distribution characteristics of different local regions of the image, and different threshold strategies are applied for dark area interference and bright area interference to effectively remove various types of interference pixels and obtain the lining region.

[0028] In another exemplary implementation, the two methods described above can be combined. First, an initial interfering pixel region is determined based on pre-defined spatial location constraints. Then, pixels within the initial interfering pixel region are further discriminated based on their grayscale distribution characteristics. Only pixels that simultaneously satisfy both the spatial location constraints and grayscale distribution characteristics are identified as interfering pixel regions and removed, thus obtaining the lining region. This combined approach reduces the risk of false rejection under complex lighting or occlusion conditions by a single discrimination method, improving the accuracy of lining region extraction.

[0029] S102. Based on the symmetry of the left and right images of the same tunnel cross section and the periodicity of the tunnel's longitudinal construction, identify the inherent structural feature areas of the lining within the lining area.

[0030] It should be noted that the construction process of tunnel lining leaves numerous inherent structural features on the lining surface, including construction joints formed during the pouring process to release stress, splicing gaps between precast components during the prefabrication process, as well as embedded channels, drainage holes, and expansion joints. The grayscale and morphology of these inherent structural features are highly similar to actual defects, interfering with defect identification. Therefore, before defect identification, it is necessary to first identify the inherent structural feature areas within the lining area as a reference benchmark for defect complexity assessment. The inherent structural features of tunnel lining exhibit two typical distribution patterns: symmetrical distribution about the tunnel's central axis on the same cross-section, and periodic repetition along the longitudinal direction according to construction sections. The direction along the tunnel axis is considered longitudinal, while the direction perpendicular to the tunnel axis is considered transverse. This embodiment identifies the inherent structural feature areas of the lining based on these two distribution patterns through symmetrical matching of left and right images and periodic analysis of consecutive longitudinal frames.

[0031] As one possible approach, feature points of the left and right images of the lining area of ​​the same tunnel section are registered, and symmetrical matching is performed based on the registered images to obtain the initial symmetrical feature region.

[0032] In some embodiments, a preset feature matching algorithm is first used to extract feature points from the left and right images of the same cross section, and a preset mismatch elimination algorithm is used to remove mismatch points, thereby registering the left and right images to the same coordinate system.

[0033] Optionally, the preset feature matching algorithm adopts the ORB (Oriented FAST and Rotated BRIEF) feature matching algorithm. This algorithm extracts key points in the image through a FAST corner detector and generates a rotation-invariant binary descriptor for each key point. The descriptors of the left and right images are matched using Hamming distance to obtain initial feature point pairs. Since mismatches are inevitable in the initial matching results, the preset mismatch elimination algorithm in this embodiment adopts the random sampling consensus algorithm. This algorithm estimates the transformation model by randomly sampling a minimum point set and filters interior points according to the consensus criterion, eliminating mismatched points that do not meet the geometric constraints, and controlling the registration error to within 2 pixels.

[0034] After feature point registration, the left and right images are mapped to the same coordinate system. Since the fields of view of the two area-scan cameras overlap, some spatial locations after registration have corresponding pixels in both the left and right images. In this embodiment, a registered stitched image is generated from the registered left and right images. The pixel grayscale values ​​in the overlapping fields of view of this registered stitched image are taken as the average of the pixel grayscale values ​​at corresponding positions in the left and right images, while the pixel grayscale values ​​in the non-overlapping areas retain the original grayscale values ​​of the corresponding single-sided image.

[0035] After obtaining the registered and stitched image, the gray value of each pixel in the registered and stitched image is compared with the gray value of the pixel at the symmetrical position about the vertical center line, using the vertical center line of the registered and stitched image as the axis of symmetry, to determine the symmetry similarity of each pixel.

[0036] It should be noted that the symmetry similarity calculation is used to quantify the degree of left-right symmetry at each position in the registered and stitched images. Because the inherent structural features of tunnel linings (construction joints, embedded channels, drainage holes, etc.) have highly consistent grayscale distributions on both sides of the same cross-section, their symmetry similarity values ​​are high; while real defects (cracks, spalling, etc.) usually do not possess strict left-right symmetry, resulting in lower symmetry similarity values. Therefore, symmetry similarity can effectively distinguish between inherent structural features and real defects.

[0037] For example, let the width of the registered and stitched image be... 1 pixel, for coordinates in the registered and stitched image The pixel at that location, whose grayscale value is... Its symmetrical position coordinates about the vertical median are: The grayscale value of the pixel at the symmetrical position is Symmetric similarity of the pixel The calculation formula is as follows: In the formula, To register coordinates in the stitched image The symmetric similarity of pixels, with values ​​ranging from 0 to 1. The closer to 1, the smaller the difference in grayscale between the left and right sides at that position, and the higher the degree of symmetry; To register coordinates in the stitched image The grayscale value of the pixel at that location; To register coordinates in the stitched image The grayscale value of the pixel at that location, i.e. The grayscale value of the pixel at the position symmetrical to the vertical midline; To register the width of the stitched images, the unit is pixels; 1 is the absolute value of the difference between the gray values ​​of two symmetrically positioned pixels; 255 is the maximum possible difference in gray values, used for normalization processing to map the gray value difference between two symmetrically positioned pixels to 0 to 1.

[0038] After calculating the symmetry similarity of each pixel, pixels with a symmetry similarity greater than or equal to a preset symmetry threshold are marked as symmetric feature points. Then, connected component extraction is performed on all symmetric feature points, grouping spatially adjacent symmetric feature points into the same region to obtain an initial symmetric feature region. This initial symmetric feature region preliminarily identifies the locations of structural features on the lining surface that satisfy the left-right symmetry condition.

[0039] Optionally, a preset symmetry threshold can be set to 0.85, meaning that pixels with a symmetry similarity greater than or equal to 0.85 are considered structural feature locations that satisfy the symmetry condition. The preset symmetry threshold can be adjusted according to the symmetry accuracy of the tunnel design; the higher the symmetry accuracy, the more appropriate the preset symmetry threshold can be.

[0040] Furthermore, the continuously collected lining areas are stitched together in a time series to form a longitudinal unfolded map, and periodic analysis is performed on the longitudinal unfolded map to obtain periodic characteristic regions.

[0041] It should be noted that when the tunnel lining is constructed using the formwork method, the length of each construction segment is fixed. Therefore, the longitudinal construction joints, expansion joints, and embedded components are all distributed according to a fixed cycle. For example, if the length of a construction segment is 12m, and the inspection vehicle travels at a speed of 30km / h and the area array camera has a frame rate of 30fps, the number of image frames corresponding to each construction segment is approximately 43 frames.

[0042] In some embodiments, during the process of stitching continuously collected lining areas into a longitudinal unfolded image according to the time sequence of the inspection vehicle's travel, the wheel speed encoder data of the inspection vehicle can be used to correct the longitudinal displacement deviation between adjacent frames, so that the longitudinal stitching error is controlled within 5 pixels.

[0043] Because the inherent structural features of the lining repeat periodically along the longitudinal direction, the grayscale distribution in the longitudinal unfolded diagram along the direction of the inspection vehicle's movement will exhibit a periodic fluctuation pattern. Therefore, after obtaining the longitudinal unfolded diagram, a grayscale projection curve is calculated for the longitudinal unfolded diagram along the direction of the inspection vehicle's movement. The grayscale projection curve is used to characterize the grayscale distribution of the lining area along the longitudinal direction.

[0044] Optionally, for each vertical position in the vertical unfolded diagram The grayscale projection value at that vertical position is obtained by summing and averaging the grayscale values ​​of all pixels in the horizontal direction. .

[0045] After obtaining the grayscale projection curve, a pre-defined correlation analysis is performed on the grayscale projection curve to obtain the autocorrelation coefficient sequence, so as to determine the repetition period of the lining structure along the longitudinal direction.

[0046] It should be noted that the periodic fluctuations in the grayscale projection curve reflect the longitudinal repetition intervals of the inherent structural characteristics of the lining. When the displacement is exactly equal to the length of one period, the two segments of the grayscale projection curve are highly similar, and the autocorrelation coefficient reaches its maximum value. Therefore, by performing autocorrelation analysis on the grayscale projection curve, the length of the period can be accurately located.

[0047] For example, let the total number of pixels in the vertical direction of the unfolded image be... The average grayscale projection value at all locations in the grayscale projection curve is [number] pixels. For a given displacement Autocorrelation coefficient The calculation formula is as follows: In the formula, The displacement is The autocorrelation coefficient at time is used to characterize the grayscale projection curve's shift relative to itself. The degree of linear correlation between versions after each pixel; The vertical position in the vertical unfolded diagram The grayscale projection value at that location; The vertical position in the vertical unfolded diagram The grayscale projection value at that location; This is the average grayscale projection value at all locations in the grayscale projection curve; This represents the total number of pixels along the vertical direction of the unfolded image.

[0048] in, and It can be considered as each constituting a length of The sequence is given by the following formula: The numerator is the inner product of the original sequence vector and the translation vector; the denominator is the arithmetic square root of the product of the squared deviations of the original sequence and the translated sequence from their respective means. According to the Cauchy-Schwarz inequality, the autocorrelation coefficient... The value range of is strictly constrained to between -1 and 1. A value of 1 indicates perfect positive correlation, a value of 0 indicates no correlation, and a value of -1 indicates perfect negative correlation. When the denominator is zero, it means that the gray-scale projection values ​​of all sampling points in the gray-scale projection curve are equal, there is no gray-scale fluctuation, and the autocorrelation coefficient is directly set to zero. When the displacement When the shifted grayscale projection curve is exactly equal to or an integer multiple of the longitudinal period length, it highly coincides with the original grayscale projection curve at the peaks and troughs, and their trends are consistent. At this point, the autocorrelation coefficient approaches 1. ... When the length of the longitudinal period is deviated, the peaks and troughs of the two curves are misaligned, the correlation weakens, and the autocorrelation coefficient decreases accordingly.

[0049] Based on the autocorrelation coefficients corresponding to different displacements calculated above, an autocorrelation coefficient sequence is constructed, and items are excluded from this autocorrelation coefficient sequence. After the value is zero, the displacement corresponding to the first maximum value is... The longitudinal period length is determined.

[0050] Search for all peak locations in the grayscale projection curve; these peaks correspond to locations where the grayscale value changes significantly along the longitudinal direction. Among the obtained peak locations, some peaks correspond to inherent structural features of the lining (such as construction joints), while others may correspond to random fluctuations caused by defects or noise. Use the longitudinal period length determined by autocorrelation analysis to filter the peaks, retaining those that repeat according to that period length and excluding non-periodic peaks. Using each filtered peak location as the center, expand longitudinally to both sides until the grayscale projection value falls back to the mean level, determining the pixel region range of each inherent structural feature in the longitudinal unfolded diagram. Merge all located pixel regions to obtain the periodic feature region.

[0051] Finally, the intersection of the initial symmetrical feature region and the periodic feature region is determined as the inherent structural feature region of the tunnel lining. Since the initial symmetrical feature region originates from the symmetry constraint of the left and right images of the same cross section, and the periodic feature region originates from the repetitive constraint of the longitudinal construction cycle, the intersection of the two can effectively eliminate false detections caused by using symmetry or periodicity alone, and obtain the stable inherent structural feature region in the tunnel lining, including non-defect structures such as construction joints, embedded channels, drainage holes, and expansion joints.

[0052] S103. For each lining area, the lining area is segmented into suspected defect areas. Based on the distance correlation and structural similarity between the segmented suspected defect areas and the inherent structural feature areas of the lining, the defect complexity of the lining area is determined.

[0053] As one possible approach, for each lining area, the lining area is segmented into suspected defect areas.

[0054] It should be noted that there is a significant grayscale difference between the lining defect area and the normal lining area. The grayscale value of cracks and spalled areas is usually lower than that of normal lining, while the grayscale value of water stains caused by water leakage is usually higher than that of normal lining. Based on this grayscale difference characteristic, the grayscale mean of its local neighborhood is calculated for each pixel in the lining area. When the grayscale value of a pixel deviates from the neighborhood mean by more than a certain range, the pixel is determined to be a suspected defect pixel.

[0055] In some embodiments, for each pixel in the lining region, the average gray value of the pixel's neighborhood is determined based on the gray values ​​of all pixels within a preset neighborhood range. For example, the preset neighborhood range is a 5×5 pixel area centered on the pixel, i.e., a rectangular neighborhood formed by extending 2 pixels horizontally and vertically from the pixel.

[0056] After determining the average grayscale value of each pixel's neighborhood, it is determined whether the absolute difference between the pixel's grayscale value and the average grayscale value of its neighborhood exceeds a preset grayscale difference threshold. If the absolute difference between the pixel's grayscale value and the average grayscale value of its neighborhood exceeds the preset grayscale difference threshold, the pixel is marked as a suspected defect pixel. Among them, suspected defect pixels with grayscale values ​​lower than the average grayscale value of its neighborhood correspond to dark area defects, including cracks and chipping; suspected defect pixels with grayscale values ​​higher than the average grayscale value of its neighborhood correspond to bright area defects, including water leakage.

[0057] Optionally, the preset grayscale difference threshold is set to 20. This threshold can be adjusted according to the actual lighting conditions in the tunnel. It can be appropriately increased when the degree of uneven lighting is high.

[0058] Then, connected component extraction is performed on all suspected defective pixels, and spatially adjacent suspected defective pixels are grouped into the same region, resulting in multiple suspected defective regions, denoted as the i-th region. One suspected defective area, among which This is the number for the suspected defective area.

[0059] Furthermore, the distance correlation between the suspected defect areas obtained from the segmentation and the inherent structural feature areas of the lining is determined.

[0060] It should be noted that before calculating the distance correlation and structural similarity of each suspected defect area, it is first determined whether the extracted inherent structural feature areas of the lining are empty sets. If it is determined to be an empty set, it means that there are no inherent structural features such as construction joints, embedded channels, drainage holes, and expansion joints in the current lining area, and no inherent structure interferes with defect identification. At this time, there is no need to perform distance calculation and structural similarity extraction operations. The distance correlation of all suspected defect areas in the lining area is directly set to zero, and the structural similarity of all suspected defect areas is also directly set to zero. Subsequently, the defect complexity of the lining area is set to zero accordingly.

[0061] If the inherent structural feature regions of the lining are determined to be non-empty sets, then for each suspected defect region, the Euclidean distance from the geometric center of the suspected defect region to the edge of each inherent structural feature region of the lining is determined, and the minimum value among the Euclidean distances is taken as the minimum Euclidean distance of the suspected defect region, and so on. Taking a suspected defect region as an example, the minimum Euclidean distance of the suspected defect region is denoted as: .

[0062] After obtaining the minimum Euclidean distance, the minimum Euclidean distance is normalized to a uniform numerical range so that the distance correlation between lining areas of different sizes is comparable.

[0063] Optionally, since the diagonal length of the lining region represents the maximum possible distance between two points in that region, the minimum Euclidean distance of the suspected defect region is divided by the diagonal length of the lining region to obtain the normalized distance of the suspected defect region. The normalized distance ranges from 0 to 1. A normalized distance of 0 indicates that the geometric center of the suspected defect region is exactly located on the edge of a certain inherent structural feature area of ​​the lining. A normalized distance of 1 indicates that the distance from the geometric center of the suspected defect region to the nearest edge of the inherent structural feature area of ​​the lining reaches the diagonal length of the lining region, which is the maximum possible distance in that lining region.

[0064] Furthermore, based on the normalized distance of the suspected defect area, the distance correlation degree of the suspected defect area is determined. The distance correlation degree is negatively correlated with the normalized distance, and its calculation formula can be expressed as follows: , For the first Normalized distance of a suspected defective region For the first The distance correlation degree of a suspected defect area. The distance correlation degree ranges from 0 to 1. The closer the distance correlation degree is to 1, the higher the spatial proximity of the suspected defect area to the inherent structural feature area of ​​the lining.

[0065] Furthermore, the structural similarity between the segmented suspected defect areas and the inherent structural feature areas of the lining is determined.

[0066] In some embodiments, provided that the inherent structural feature regions of the lining are not empty sets, morphological statistical features of all inherent structural feature regions of the lining are extracted. These morphological statistical features include aspect ratio, area ratio, mean edge gradient, and mean grayscale. The aspect ratio is the ratio of the long side to the short side of the minimum bounding rectangle of each structural feature region, used to characterize the degree of shape extension of the region. The area ratio is the proportion of the pixel area of ​​each structural feature region to the pixel area of ​​the entire lining region, used to characterize the relative size of the region. The mean edge gradient is the average value of the grayscale gradient amplitude at the edge pixels of each structural feature region, used to characterize the clarity of the region's edges. The mean grayscale is the average value of the grayscale values ​​of pixels within each structural feature region, used to characterize the overall brightness level of the region.

[0067] Then, the average value of the above-mentioned morphological statistical features of all the inherent structural feature areas of the lining is taken to form a first multidimensional feature vector, which is used to characterize the overall morphological statistical features of the inherent structural feature areas of the lining.

[0068] Secondly, for each suspected defect region, the aspect ratio, area ratio, mean edge gradient, and mean gray value of the suspected defect region are extracted to form the second multidimensional feature vector corresponding to the suspected defect region. The second multidimensional feature vector is used to characterize the morphological statistical features of the suspected defect region.

[0069] Subsequently, the cosine similarity between the second multidimensional feature vector and the first multidimensional feature vector corresponding to the suspected defect region is determined as the structural similarity of the suspected defect region.

[0070] It should be noted that cosine similarity measures the directional consistency of two feature vectors by calculating the cosine of the angle between them, with a value ranging from -1 to 1. The smaller the angle between the two vectors, the closer the cosine similarity is to 1. Since all four morphological statistical features mentioned above are non-negative, and all components of the first and second multidimensional feature vectors are non-negative, the actual range of the structural similarity of the suspected defect region is 0 to 1. The closer the structural similarity is to 1, the more similar the morphological features of the suspected defect region are to the inherent structural features of the lining, and the more difficult it is to distinguish them.

[0071] Finally, the complexity of defects in the lining area is determined.

[0072] It should be noted that in order to assess the complexity of defects in each lining area, it is necessary to consider the influence of all suspected defect areas within the lining area. Since the high complexity of any suspected defect area may affect the overall detection difficulty of the lining area, the maximum distance correlation and the maximum structural similarity among all suspected defect areas within the lining area are taken as representative indicators of the complexity of the lining area.

[0073] In some embodiments, the maximum distance correlation among all suspected defective areas within the lining area is determined as the maximum distance correlation; the maximum structural similarity among all suspected defective areas within the lining area is determined as the maximum structural similarity.

[0074] Then, the sum of the distance correlation scores and the sum of the structural similarities of all suspected defect areas within the lining area are determined. The sum of the distance correlation scores reflects the overall spatial proximity between the suspected defects and the inherent structure within the lining area, while the sum of the structural similarities reflects the overall morphological similarity between the suspected defects and the inherent structure within the lining area. Furthermore, the ratio of the sum of the distance correlation scores to the total sum of the distance correlation scores and structural similarities of all suspected defect areas within the lining area is used as the first weight, and the ratio of the sum of the structural similarities to the total sum of the distance correlation scores and structural similarities of all suspected defect areas within the lining area is used as the second weight. That is, the larger the sum of the distance correlation scores, the larger the first weight; the larger the sum of the structural similarities, the larger the second weight, and the sum of the first and second weights is 1. In addition, if there are no suspected defect areas within the lining area, it is not necessary to determine the first and second weights corresponding to that lining area, and the defect complexity of that lining area is directly set to zero.

[0075] Finally, the complexity of the defects in the lining area is determined based on the maximum distance correlation, the maximum structural similarity, the first weight, and the second weight.

[0076] In some embodiments, the formula for calculating the complexity of defects in the lining area is as follows: In the formula, The maximum value among all suspected defect areas within the lining area is the distance correlation coefficient. The first weight is the ratio of the sum of the distance correlations of all suspected defective areas within the lining area to the sum of the distance correlations and structural similarities of all suspected defective areas within the lining area. This represents the maximum structural similarity among all suspected defective areas within the lining region. The second weight is the ratio of the sum of the structural similarities of all suspected defective areas within the lining area to the sum of the distance correlation and structural similarities of all suspected defective areas within the lining area. The value represents the complexity of the defects in the lining area, ranging from 0 to 1. The closer the value is to 1, the closer the suspected defects in the lining area are to the inherent structural features of the lining in space and the more similar they are in shape. The greater the difficulty of identification, the more advanced the network reasoning is required. The closer the value is to 0, the lower the risk of confusion and the easier it is to identify.

[0077] S104. Determine the target detection exit in the preset multi-exit detection network that corresponds to the complexity of the defect, and input the lining area into the preset multi-exit detection network to obtain the defect detection result of the tunnel lining corresponding to the target detection exit.

[0078] The pre-defined multi-exit detection network uses a pre-defined basic target detection network as its framework, with multiple detection heads set at different network depths to form multiple detection exits. For example, the pre-defined basic target detection network can adopt the YOLOv8 (You Only Look Once version 8) network structure. Auxiliary detection heads are set at layers 3, 6, and 9 of this network framework, along with the main detection head at the end of the network, resulting in a total of four detection exits: a shallow detection head at layer 3, a medium-depth detection head at layer 6, a medium-deep detection head at layer 9, and a deep main detection head at the end of the network. Each detection head corresponds to a different network depth from shallow to deep. The shallow detection head is suitable for simple defect identification with significant morphological differences from the background; the medium-depth detection head is suitable for defect identification with a certain degree of morphological complexity; the medium-deep detection head is suitable for defect identification with morphological features similar to inherent structural features; and the deep main detection head is suitable for defect identification that is highly confused with inherent structural features.

[0079] In some embodiments, to automatically select the corresponding target detection exit based on the complexity of the defect, it is necessary to predetermine the exit threshold for each detection exit. The exit threshold is used to divide the range of defect complexity, assigning lining region images of different complexities to detection heads of different depths for inference. Optionally, by acquiring a batch of test samples with labeled defects, the defect complexity of each test sample is calculated according to the aforementioned steps. Values, to obtain all test samples Value distribution, and a preset threshold determination algorithm is used to determine the value distribution. The value distribution is analyzed to determine the first exit threshold. Second exit threshold and the third exit threshold And satisfy .

[0080] For example, the preset threshold determination algorithm can adopt Otsu's method. Its basic principle is to iterate through all possible threshold division methods, calculate the inter-class variance between categories under each division, and determine the threshold that maximizes the inter-class variance as the exit threshold. Specifically, the test samples... After sorting the values ​​by size, try to sort them one by one. The values ​​are divided into four intervals. The inter-class variance corresponding to each division method is calculated. The three boundary values ​​corresponding to the maximum inter-class variance are taken as the first exit thresholds. Second exit threshold and the third exit threshold .in, This corresponds to the optimal threshold for separating the lowest complexity sample from the rest. This corresponds to the optimal threshold for separating low-to-medium complexity samples from high-to-medium complexity samples. This corresponds to the optimal threshold for separating samples with medium to high complexity from those with the highest complexity.

[0081] After obtaining three exit thresholds, the complexity of the defects in the current lining area is determined. The corresponding target detection exit is determined by comparing the results with each exit threshold.

[0082] For example, when When the condition is low, it indicates that the defects in the lining area are less complex and the risk of confusion between the suspected defects and the inherent structural features of the lining is small. Only shallow features are needed to complete the identification. At this time, the outlet corresponding to the shallow detection head is determined as the target detection outlet.

[0083] when When the condition is met, it indicates that the defect complexity of the lining area is moderate and requires the assistance of the middle layer features for identification. In this case, the outlet corresponding to the middle layer detection head is determined as the target detection outlet.

[0084] when When the condition is met, it indicates that the defects in the lining area are highly complex and require identification using medium-deep features. In this case, the outlet corresponding to the medium-deep detection head is determined as the target detection outlet.

[0085] when When the condition is met, it indicates that the lining area has the highest degree of defect complexity. The suspected defects are highly confused with the inherent structural features of the lining, requiring deep semantic features for refined identification. In this case, the exit corresponding to the deep main detection head is determined as the target detection exit.

[0086] After determining the target detection exit, the lining area is input into a preset multi-exit detection network. The network performs forward reasoning to the target detection exit, and the detection head corresponding to that exit outputs the defect detection result.

[0087] Optionally, for images of lining regions where the target detection exit is the exit corresponding to a shallow detection head, the preset multi-exit detection network performs forward inference only up to layer 3, exiting after the shallow detection head outputs the detection result, with subsequent network layers not participating in the calculation; for images of lining regions where the target detection exit is the exit corresponding to a mid-layer detection head, the preset multi-exit detection network performs forward inference up to layer 6, exiting after the mid-layer detection head outputs the detection result; for images of lining regions where the target detection exit is the exit corresponding to a mid-deep detection head, the preset multi-exit detection network performs forward inference up to layer 9, exiting after the mid-deep detection head outputs the detection result; for images of lining regions where the target detection exit is the exit corresponding to a deep main detection head, the preset multi-exit detection network performs complete forward inference up to the last layer, with the deep main detection head outputting the detection result. Thus, images of simple defects complete inference and exit at the shallow layer of the preset multi-exit detection network, avoiding the waste of computational resources caused by full-network deep inference; images of complex defects continue to infer to deeper layers, utilizing deep semantic features to achieve accurate recognition.

[0088] It should be noted that since there are multiple detection heads in the preset multi-exit detection network, and some detection heads may output defect detection results for the same defect when exiting early, it is necessary to deduplicate the output results of all executed detection exits.

[0089] Optionally, deduplication can be performed using a non-maximum suppression algorithm. This algorithm sorts detection boxes of the same category from highest to lowest confidence, sequentially selects the detection box with the highest confidence, and calculates the intersection-union ratio (IU / R) of this detection box with the remaining detection boxes of the same category. Detection boxes with IU / R exceeding a preset threshold are removed. This process is repeated until all detection boxes have been processed, thus eliminating duplicate detection boxes. After deduplication, detection results with confidence levels below a preset confidence threshold are considered false alarms and removed. The remaining detection results are used as the defect detection results for the tunnel lining, and the final defect category, bounding box location, and confidence level are output. Defect categories include cracks, spalling, water leakage, and exposed rebar, etc. The preset IU / R threshold is set to 0.5, and the preset confidence threshold is set to 0.7.

[0090] Understandably, in the intelligent identification and detection method for multiple defects in tunnel lining based on target detection provided in this embodiment of the invention, image acquisition devices symmetrically installed on both sides of the inspection vehicle synchronously acquire tunnel lining images and extract the lining area after preprocessing, providing a high-quality image data foundation for subsequent analysis. Based on this, utilizing the inherent left-right symmetry and longitudinal construction periodicity of the tunnel lining, the inherent structural features within the lining area are identified and extracted, effectively distinguishing inherent structural features from actual defects on the lining surface. This solves the problem of general detection networks misjudging inherent structures such as construction joints, embedded channels, and drainage holes as defects, significantly reducing the false detection rate and providing a reliable reference benchmark for subsequent defect complexity assessment. Based on distinguishing inherent structures from actual defects, each lining area is further segmented into suspected defect regions. Considering the distance correlation and structural similarity between suspected defect regions and inherent structural feature regions, the defect complexity of each lining region is adaptively determined. This defect complexity accurately reflects the degree of confusion between defects and inherent structures in the lining region, providing a quantitative basis for the dynamic reasoning of the detection network. Finally, based on the complexity of the defects, different lining areas are matched to detection exits at different depths in a pre-set multi-exit detection network for inference. Lining areas with lower defect complexity can be detected at the shallow layer of the network and exit the inference process early, while lining areas with higher defect complexity are detected using deep features. This effectively reduces the waste of computing resources for simple defects while ensuring that all types of defects can be accurately identified. It achieves a balance between detection accuracy and detection efficiency, meeting the engineering requirements for detection efficiency and recognition accuracy in the periodic inspection of tunnel linings.

Claims

1. A method for intelligent identification and detection of multiple defects in tunnel lining based on target detection, characterized in that, The method includes: Images of the tunnel lining are simultaneously acquired by image acquisition devices symmetrically installed on both sides of the inspection vehicle. The tunnel lining images are then preprocessed to extract the lining area. Based on the symmetry of left and right images of the same tunnel cross section and the periodicity of tunnel longitudinal construction, the inherent structural feature areas of the lining within the lining area are identified. For each lining area, the lining area is segmented into suspected defect areas. The complexity of the defects in the lining area is determined based on the distance correlation and structural similarity between the segmented suspected defect areas and the inherent structural feature areas of the lining. A target detection exit corresponding to the complexity of the defect is determined in a preset multi-exit detection network, and the lining area is input into the preset multi-exit detection network to obtain the defect detection result of the tunnel lining corresponding to the target detection exit.

2. The intelligent identification and detection method for multiple defects in tunnel lining based on target detection according to claim 1, characterized in that, Images of the tunnel lining are simultaneously acquired by image acquisition devices symmetrically installed on both sides of the inspection vehicle. These images are then preprocessed to extract the lining area, including: The image acquisition device is an area array camera. The installation height of the area array camera is flush with the position of the tunnel arch, and the shooting angle is perpendicular to the lining surface. The overlapping area of ​​the fields of view of the two area array cameras covers the symmetrical central axis of the tunnel lining. A synchronous trigger is used to control the two image acquisition devices to synchronously acquire tunnel lining images; The acquired tunnel lining images are subjected to distortion correction and defogging processing, and a preset segmentation algorithm is used to remove the interfering pixel areas corresponding to the detection vehicle body, track and lighting fixtures, and the lining area is extracted.

3. The intelligent identification and detection method for multiple defects in tunnel lining based on target detection according to claim 1, characterized in that, Based on the symmetry of left and right images of the same tunnel cross-section and the periodicity of tunnel longitudinal construction, the inherent structural feature regions of the lining within the lining area are identified, including: Feature points of the left and right images of the lining area of ​​the same tunnel section are registered, and symmetric matching is performed based on the registered images to obtain the initial symmetrical feature region. The continuously collected data of the lining area are stitched together in a time series to form a longitudinal unfolded map. Periodic analysis is then performed on the longitudinal unfolded map to obtain periodic feature regions. The intersection of the initial symmetrical feature region and the periodic feature region is determined as the inherent structural feature region of the lining.

4. The intelligent identification and detection method for multiple defects in tunnel lining based on target detection according to claim 3, characterized in that, Feature points of the lining area of ​​the same tunnel cross section are registered between left and right images. Based on the registered images, symmetrical matching is performed to obtain an initial symmetrical feature region, including: Feature points of the left and right images of the same cross section are extracted using a preset feature matching algorithm, and mismatched points are removed based on a preset mismatch elimination algorithm. The left and right images are registered to the same coordinate system to obtain a registered stitched image. The pixel gray value of the overlapping area of ​​the field of view in the registered stitched image is the average of the pixel gray values ​​of the corresponding positions of the left and right images. Using the vertical midline of the registered and stitched image as the axis of symmetry, the symmetry similarity of each pixel is determined based on the difference between the gray values ​​of each pixel in the registered and stitched image and the gray values ​​of pixels at symmetrical positions. Pixels with a symmetry similarity greater than or equal to a preset symmetry threshold are marked as symmetric feature points. Connected component extraction is performed on all symmetric feature points to obtain the initial symmetric feature region.

5. The intelligent identification and detection method for multiple defects in tunnel lining based on target detection according to claim 3, characterized in that, The continuously collected data on the lining area are stitched together in a time series to form a longitudinal unfolded map. Periodic analysis is then performed on the longitudinal unfolded map to obtain periodic characteristic regions, including: A grayscale projection curve is calculated for the longitudinal unfolded diagram along the direction of travel of the inspection vehicle. The grayscale projection curve is used to characterize the grayscale distribution of the lining area along the longitudinal direction. A preset correlation analysis is performed on the grayscale projection curve to obtain an autocorrelation coefficient sequence, which is used to characterize the similarity of the grayscale projection curve under different longitudinal displacements. The longitudinal period length is determined based on the maximum value of the autocorrelation coefficient sequence, and the periodic feature region is obtained by locating the area where the periodicity repeats in the longitudinal unfolded diagram based on the longitudinal period length.

6. The intelligent identification and detection method for multiple defects in tunnel lining based on target detection according to claim 1, characterized in that, The lining area is segmented into suspected defect areas, including: For each pixel in the lining area, the average gray value of the pixel's neighborhood is determined based on the gray values ​​of each pixel within a preset neighborhood range. If the absolute difference between the gray value of a pixel and the average gray value of its neighborhood exceeds a preset gray value difference threshold, the pixel is marked as a suspected defective pixel. Connected component extraction is performed on all suspected defective pixels to obtain the suspected defective region.

7. The intelligent identification and detection method for multiple defects in tunnel lining based on target detection according to claim 1, characterized in that, Determining the distance correlation between the segmented suspected defect areas and the inherent structural feature areas of the lining includes: For each suspected defect area, determine the Euclidean distance from the geometric center of the suspected defect area to the edge of each inherent structural feature area of ​​the lining, and take the minimum value among the Euclidean distances as the minimum Euclidean distance of the suspected defect area. Using the diagonal length of the lining area as a normalization benchmark, the minimum Euclidean distance of the suspected defect area is normalized to obtain the normalized distance of the suspected defect area. The distance correlation degree of the suspected defective region is determined based on the normalized distance of the suspected defective region, and the distance correlation degree is negatively correlated with the normalized distance.

8. The intelligent identification and detection method for multiple defects in tunnel lining based on target detection according to claim 1, characterized in that, Determining the structural similarity between the segmented suspected defect regions and the inherent structural feature regions of the lining includes: Extract the aspect ratio, area ratio, mean edge gradient, and mean gray level of all inherent structural feature regions of the lining, and then take the mean values ​​to form the first multidimensional feature vector. For each suspected defect region, the aspect ratio, area ratio, mean edge gradient, and mean gray level of the suspected defect region are extracted to form the second multidimensional feature vector corresponding to the suspected defect region. The cosine similarity between the second multidimensional feature vector corresponding to the suspected defect region and the first multidimensional feature vector is determined as the structural similarity of the suspected defect region.

9. The intelligent identification and detection method for multiple defects in tunnel lining based on target detection according to claim 1, characterized in that, Determining the complexity of defects in the lining area includes: The maximum distance correlation among all suspected defect areas within the lining area is determined as the maximum distance correlation. The maximum value among the structural similarities of all suspected defect areas within the lining area is determined as the maximum structural similarity. The first weight and the second weight are determined based on the sum of the distance correlation and the sum of the structural similarity of all suspected defect areas within the lining area. The defect complexity of the lining area is determined based on the maximum distance correlation, the maximum structural similarity, the first weight, and the second weight.

10. The intelligent identification and detection method for multiple defects in tunnel lining based on target detection according to claim 1, characterized in that, The preset multi-exit detection network includes a shallow detection head, a middle detection head, a medium-deep detection head, and a deep main detection head, which correspond to different network depths from shallow to deep. Determining the target detection exit in the preset multi-exit detection network corresponding to the complexity of the defect includes: Based on the defect complexity distribution of the labeled test samples, a preset threshold determination algorithm is used to determine the first exit threshold, the second exit threshold, and the third exit threshold. The first exit threshold is less than the second exit threshold, and the second exit threshold is less than the third exit threshold. If the complexity of the defects in the lining area is less than or equal to the first exit threshold, the exit corresponding to the shallow detection head is determined as the target detection exit. If the complexity of the defects in the lining area is greater than the first exit threshold and less than or equal to the second exit threshold, the exit corresponding to the middle layer detection head is determined as the target detection exit. If the complexity of the defects in the lining area is greater than the second exit threshold and less than or equal to the third exit threshold, the exit corresponding to the medium-deep detection head is determined as the target detection exit. If the complexity of the defects in the lining area is greater than the third exit threshold, the exit corresponding to the deep main detection head is determined as the target detection exit.

Images