Intelligent Detection Method for Tunnel Lining Cracks Based on Image Analysis
By establishing the correspondence between scanning units and surface positions and observing both time phases, and combining brightness changes to determine the source of pixels, the problem of time interference in the identification of tunnel lining cracks was solved, and the accuracy and reliability of detection were improved.
Patent Information
- Application Number
- CN202611602361.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-16
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies struggle to effectively distinguish between cracks on tunnel lining surfaces and linear responses formed by changes in lighting or moving water droplets, resulting in insufficient accuracy in crack identification.
By establishing the correspondence between the scanning unit and the scanning time and the position of the lining surface, and combining the results of repeated observations in two phases with changes in background brightness, the source state of the pixels is determined, time-interference pixels are separated, and the probability of linear responses formed by changes in illumination and moving water droplets being identified as cracks is reduced.
It improves the accuracy of tunnel lining crack detection, ensures that crack detection results correspond to the lining surface location, and reduces the probability of false identification.
Smart Images

Figure CN123118618A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of image processing and tunnel detection technology, and more specifically, to an intelligent detection method for tunnel lining cracks based on image analysis. Background Technology
[0002] As highway and railway tunnels age, cracks may appear on the lining surface. The location, length, width, and direction of these cracks can provide a basis for assessing the condition of the lining and arranging maintenance. To improve the inspection efficiency of long-distance tunnels, existing inspection methods are gradually adopting mobile inspection vehicles equipped with image acquisition and lighting devices. These vehicles collect images of the lining surface while traveling and use image processing or deep learning methods to identify cracks.
[0003] Existing technologies have enabled the use of vehicle-mounted cameras to acquire images of tunnel linings and automatically identify cracks. For example, Chinese patent application CN114113118A discloses a rapid detection device and method for water leakage defects in subway tunnel linings. This patent application discloses an image acquisition system consisting of an area array camera, a line array camera, and a detection light source. It adjusts the travel speed of the detection device, the camera frame rate, and the image resolution according to the detection area, and then uses a deep learning model to identify cracks and water leakage defects. From the disclosed content, the relevant technologies mainly address the coordination between image acquisition efficiency, image resolution, and defect identification speed.
[0004] However, the two-dimensional unfolded image of the lining formed by a line scan camera or a rolling scan camera is composed of scanning units obtained at different acquisition times arranged sequentially. The spatial extension structure in the image does not necessarily correspond to the fixed structure of the lining surface. When the inspection vehicle passes through the seepage area, moving water droplets may occupy different positions at adjacent scanning times, and changes in illumination output may also cause adjacent scanning units to produce strip-shaped grayscale boundaries. The above-mentioned time changes will be converted into slender and interconnected spatial targets in the unfolded image. For example, when the water droplets move down along the lining surface and the inspection vehicle moves along the tunnel axis, the water droplet pixels obtained at multiple times may be connected into a diagonal dark line in the unfolded image. The dark line is similar to a crack in terms of width, grayscale, and extension shape. Existing methods mainly identify cracks based on grayscale, edge, texture, or depth features in a single unfolded image. It is difficult to determine whether the linear target belongs to a crack fixed at the same position on the lining surface or a spatial trajectory formed by time changes during acquisition. Therefore, changes in illumination or moving water droplets may be identified as cracks.
[0005] In view of this, this application proposes an intelligent detection method for tunnel lining cracks based on image analysis to solve the above problems. Summary of the Invention
[0006] To address the shortcomings of existing technologies, this application aims to provide an intelligent detection method for tunnel lining cracks based on image analysis. By establishing a correspondence between scanning units, scanning time, and lining surface position, and combining dual-temporal repeated observation results, background brightness changes, and candidate linear target position changes to determine the pixel source state, and separating time-interference pixels before crack identification, the method aims to reduce the probability of linear responses formed by lighting changes and moving water droplets being identified as lining cracks and to obtain crack detection results corresponding to the lining surface position.
[0007] This application provides the following technical solution: an intelligent detection method for tunnel lining cracks based on image analysis, including:
[0008] Acquire detection task data and perform tunnel scanning according to the detection task data to obtain scanning records; acquire calibration acquisition data and perform spatiotemporal conversion on the scanning records based on the calibration acquisition data to obtain scanning spatiotemporal correlation data;
[0009] The scan records are rearranged according to the scan spatiotemporal correlation data to obtain dual-temporal observation images; the position correspondence and image quality of the dual-temporal observation images and scan records are determined to obtain the qualification for repeated observation; candidate linear targets are detected in the dual-temporal observation images to obtain candidate linear targets; the scan formation time corresponding to the candidate linear targets is queried according to the scan spatiotemporal correlation data to obtain the formation time of the candidate targets.
[0010] Compare dual-temporal candidate linear targets at locations eligible for repeated observation to obtain surface repeat support data; calculate the correlation between candidate target formation time and background brightness changes, and the correlation between candidate target formation time and target position changes to obtain illumination time correlation data and movement time correlation data.
[0011] The state of repeated observation eligibility, surface repeated support data, illumination time-related data, and movement time-related data is determined to obtain the pixel source state; the observed pixels are selected and compensated according to the pixel source state to obtain the interference separation image; crack identification and geometric calculation are performed on the interference separation image to obtain the crack detection result.
[0012] The technical effects and advantages of the intelligent detection method for tunnel lining cracks based on image analysis in this application are as follows:
[0013] This application establishes a correspondence between the scanning units in the scanning record and the scanning formation time and the lining surface position, compares dual-temporal candidate linear targets at the same lining surface position, and calculates the correlation between the candidate target formation time and changes in background brightness and target position. Based on this, it distinguishes between fixed surface pixels, illumination time interference pixels, and movement time interference pixels. Before crack identification, it selects observation pixels according to the pixel source state and performs brightness compensation, which can reduce the linear responses formed by illumination changes and moving water droplets in the unfolded image from entering the crack identification process, reduce the probability of time-varying targets being identified as lining cracks, and make the crack detection results correspond to the lining surface position. Attached Figure Description
[0014] Figure 1 This is a schematic diagram of the intelligent detection method for tunnel lining cracks based on image analysis according to this application;
[0015] Figure 2 This is a schematic diagram of the tunnel inspection vehicle inspection scenario in this application;
[0016] Figure 3 This is a schematic diagram of dual-temporal observation at the same lining surface location in this application;
[0017] Figure 4 This is a schematic diagram of dual-temporal observation of different targets at the same lining surface location in this application. Detailed Implementation
[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0019] Example 1:
[0020] Please see Figure 1 and Figure 2As shown, this embodiment provides an intelligent detection method for tunnel lining cracks based on image analysis, applied to a tunnel inspection vehicle. The tunnel inspection vehicle is equipped with a first linear array imaging unit, a second linear array imaging unit, an illumination device, an illumination drive recording unit, a travel displacement detection device, an attitude detection device, an imaging distance detection device, an acquisition control device, and an image processing device. The first and second linear array imaging units are installed at intervals along the travel direction of the inspection vehicle, so that the same lining surface position passes through the first and second scan lines sequentially. The imaging distance detection device uses a line laser profile measuring device installed coplanar with the imaging plane of the linear array imaging unit. The line laser profile measuring device outputs the imaging distance in units of length for the position of the linear array pixels. The acquisition control device uses a common clock to record the sampling times of line exposure, illumination drive signal, travel displacement, inspection vehicle attitude, and imaging distance. One row of pixels obtained from a single line exposure is taken as a scanning unit. The start time of the exposure of the scanning unit is recorded as... The end time of exposure is recorded as ,according to Calculate the scan formation time.
[0021] The methods include:
[0022] Step 1: Perform tunnel scanning according to the detection task data to obtain scanning records; perform spatiotemporal conversion on the scanning records based on the calibration acquisition data to obtain the spatiotemporal correlation data of the scanning.
[0023] In specific implementations, methods for obtaining spatiotemporal correlation data include:
[0024] The minimum crack width required to be identified can be obtained from the testing specifications, testing contract, or acceptance documents. Minimum crack length to be identified and allowable error of crack geometry When multiple documents describe similar requirements, select the smallest values for minimum crack width, minimum crack length, and crack geometry tolerance; retrieve the tunnel cross-section profile, track centerline, design mileage benchmark, and mileage increment direction from the tunnel design data; use the distance along the track centerline as the longitudinal surface coordinate. The arc length from the starting point of the lining section along the section profile to the target position is used as the circumferential surface coordinate. Obtain detection task data; read the currently enabled crack segmentation model version identifier from the model deployment record, and write the currently enabled crack segmentation model version identifier into the detection task data.
[0025] The method for obtaining mileage benchmarks includes: selecting tunnel mileage markers with design mileage records as mileage benchmark points; measuring the spatial coordinates of adjacent mileage benchmark points using a metrologically calibrated total station; using the projected distance between adjacent mileage benchmark points on the centerline of the tunnel as the mileage marker spacing; calculating the absolute difference between the measured mileage marker spacing and the design mileage difference; using the design mileage when the absolute difference does not exceed the allowable error for mileage measurement recorded in the testing specifications; and re-measuring and using the re-measurement confirmation result when the absolute difference exceeds the allowable error for mileage measurement; the known spacing represents the mileage marker spacing obtained through the aforementioned measurement process, and the preset spacing that has not been measured is not used.
[0026] The method for obtaining calibration parameters of the linear array imaging unit includes: after the linear array imaging unit is installed, establishing an imaging unit coordinate system with the optical center of the linear array imaging unit, taking the optical axis direction as the Z-axis, the imaging direction corresponding to the increasing direction of the linear array pixel number as the X-axis, and the direction perpendicular to the X-axis and Z-axis as the Y-axis; reading the effective pixel count and effective pixel number range from the device parameters of the linear array imaging unit; and performing imaging calibration using a planar calibration plate with a known mark spacing, wherein each mark point on the planar calibration plate has calibration plate coordinates that can be determined according to the mark spacing.
[0027] Set the planar calibration board within the imaging distance working range of the linear array imaging unit, and acquire calibration images at the lower limit, upper limit, and midpoint of the imaging distance working range. Move the calibration board laterally at each distance position so that the markers at different lateral positions on the calibration board sequentially pass through the effective scan lines of the linear array imaging unit. Record the pixel number and common clock time when each marker passes through the scan line. Use a metrologically calibrated total station to measure the spatial position of the markers, and verify the spatial distance between the markers according to the known marker spacing on the calibration board. Only records that can determine the marker identification, actual pixel number, and corresponding spatial position are considered valid calibration records.
[0028] Based on the installation design record of the linear array imaging unit, the initial installation position and initial installation direction of the optical center relative to the fixed coordinate system of the inspection vehicle are obtained; the spatial position of the calibration plate marker point in the fixed coordinate system of the inspection vehicle is measured using a metrologically calibrated total station; according to the initial installation position and initial installation direction, the spatial position of the marker point is transformed to the coordinate system of the linear array imaging unit to obtain the initial lateral coordinate and initial axial coordinate; when the initial axial coordinate is greater than zero, the ratio of the initial lateral coordinate to the initial axial coordinate is calculated to obtain the initial normalized lateral coordinate of the marker point; the initial normalized lateral coordinate and the actual acquired pixel coordinates are used to establish the initial fitting relationship of the imaging parameters.
[0029] A one-dimensional radial distortion imaging model is used: ;in, For the first Calculate the pixel coordinates of each marker point. These are the principal pixel coordinates of the linear array imaging unit. The equivalent focal length is expressed in pixels. and These are the first-order and second-order radial distortion coefficients, respectively.
[0030] The objective is to minimize the sum of the squared differences between the calculated and actual pixel coordinates of each marker point. , , , The mounting pose parameters of the linear array imaging unit relative to the detection vehicle are fitted using least-squares to obtain the corresponding calibration parameters; the obtained focal length parameters and principal point parameters are then combined to form the imaging intrinsic parameter matrix. , radial distortion coefficient and Composition of imaging distortion parameters .
[0031] The imaging intrinsic parameter matrix is established according to the following formula: The second row of the matrix is used to maintain the fixed scanning plane coordinates of the linear array imaging unit and is not used as an independent vertical pixel focal length parameter.
[0032] The valid calibration records are divided into parameter fitting records and calibration verification records according to the marker point identification. At each imaging distance position, the marker points are numbered in ascending order of their lateral coordinates. Odd-numbered marker points are used for parameter fitting, and even-numbered marker points are used for calibration verification. If any distance position lacks a parameter fitting record or a calibration verification record, the calibration data acquisition for that distance position is supplemented.
[0033] After obtaining the imaging intrinsic parameter matrix, imaging distortion parameters, and installation pose parameters by using parameter fitting records, the corresponding imaging direction is calculated using the actual pixel positions in the calibration verification records, and the lateral position of the marker points is calculated by combining the known axial distance of the marker points. The absolute difference between the calculated lateral position and the measured lateral position of the corresponding marker point is taken as the calibration verification position error. When all calibration verification position errors do not exceed the allowable error of crack geometry in the detection task data, the obtained calibration parameters are adopted. When any verification position error exceeds the allowable error of crack geometry, the acquisition position of the calibration board is readjusted and the calibration is re-executed.
[0034] The installation pose parameters of the linear array imaging unit obtained by the aforementioned least-squares joint fitting are used as the final installation pose parameters of the linear array imaging unit relative to the detection vehicle coordinate system. The rotation matrix from the linear array imaging unit coordinate system to the detection vehicle coordinate system is determined based on the installation direction in the final installation pose parameters, and the displacement vector from the linear array imaging unit coordinate system to the detection vehicle coordinate system is determined based on the installation position in the final installation pose parameters. The rotation matrix and displacement vector are used as the installation parameters of the linear array imaging unit after the aforementioned calibration and verification, and are used for the coordinate transformation of the subsequent imaging rays.
[0035] The rotation matrix and displacement vector of the linear imaging unit to the vehicle coordinate system, and the imaging intrinsic parameter matrix. Imaging distortion parameters The corresponding linear array imaging unit identifier and calibration verification results are combined to obtain the calibration parameters of the linear array imaging unit.
[0036] The method for obtaining attitude detection data includes: mounting the attitude detection device on an angle calibration platform capable of changing the pitch, roll, and yaw angles respectively; determining the attitude working range based on the minimum and maximum values of each attitude angle obtained by the test vehicle during unloaded pre-running and load pre-running in the section to be tested; sequentially setting the angle calibration platform at the lower limit, upper limit, and midpoint of the attitude working range; using the angle readings from the angle calibration platform as known attitude angles; calculating the difference between the output angle of the attitude detection device and the known attitude angles; and converting any angle difference into the lateral distance using the maximum imaging distance. When the allowable error in crack geometry is exceeded, the corresponding attitude detection device is not used for formal testing, and recalibration is performed according to the calibration process provided by the attitude detection device manufacturer; when all angular positions meet the requirements, the angle output of the attitude detection device, the corresponding common clock sampling time, and the installation direction of the attitude detection device relative to the testing vehicle are combined to obtain attitude detection data; where The maximum imaging distance output by the imaging distance detection device during the pre-operation period. This corresponds to the attitude angle difference.
[0037] The method for obtaining imaging distance detection data includes: setting up a planar calibration plate perpendicular to the distance measurement direction in front of the imaging distance detection device; using a metrologically calibrated laser rangefinder to measure the distance from the optical center of the linear array imaging unit to the planar calibration plate at the lower limit, upper limit, and midpoint of the imaging distance working range to obtain known distances; the lower and upper limits of the imaging distance working range are obtained by using the minimum and maximum imaging distances obtained by the detection vehicle during pre-running in the section to be detected, respectively; reading the physical distance output by the imaging distance detection device at the same position; using the absolute difference between the output physical distance and the known distance as the distance verification error; when the surface position error formed by the distance verification error after projection onto the corresponding pixel imaging direction does not exceed the allowable error of crack geometry, the physical distance output by the imaging distance detection device is used; if the aforementioned conditions are not met, formal testing is not performed and recalibration is performed according to the calibration process provided by the imaging distance detection device manufacturer.
[0038] The method for obtaining spatial calibration parameters between the imaging distance detection device and the linear array imaging unit includes: setting a calibration plate with known geometric dimensions within the common measurement range of both devices; sequentially setting the calibration plate at at least three different spatial positions to obtain measurement points that can be identified by marking on the calibration plate and are not located on the same straight line; reading the spatial coordinates of the corresponding measurement points in the coordinate system of the imaging distance detection device, and the spatial coordinates in the coordinate system of the imaging unit calculated based on the spatial position of the calibration plate and the installation parameters of the linear array imaging unit; solving for the rotation matrix and displacement vector with the goal of minimizing the sum of the squared distances between the two sets of corresponding coordinates, under the condition that the rotation matrix satisfies the orthogonal constraint and the determinant is one; verifying the spatial coordinate transformation results using the calibration plate position that did not participate in parameter fitting, and calculating the Euclidean distance between the transformed coordinates and the corresponding measured coordinates; if all verified distances do not exceed the allowable error of crack geometry, using the obtained rotation matrix and displacement vector as the spatial calibration parameters between the devices; otherwise, re-performing the calibration; and using the verified spatial calibration parameters between the devices to transform the measurement points to the coordinate system of the linear array imaging unit during formal testing.
[0039] The method for obtaining the lining surface coordinate calculation parameters includes: reading the imaging intrinsic parameter matrix, imaging distortion parameter, rotation matrix and displacement vector from the coordinate system of the linear array imaging unit to the coordinate system of the detection vehicle after calibration and verification of the first linear array imaging unit and the second linear array imaging unit; reading the installation direction of the attitude detection device relative to the coordinate system of the detection vehicle, and the spatial calibration parameters between the imaging distance detection device and the corresponding linear array imaging unit; reading the tunnel line centerline, tunnel cross-section profile and mileage reference from the detection task data; establishing a correspondence between the aforementioned data according to the corresponding linear array imaging unit identifier to obtain the lining surface coordinate calculation parameters; wherein, the imaging intrinsic parameter matrix and imaging distortion parameter are used to calculate the imaging direction of the pixel, the rotation matrix, displacement vector and attitude detection device installation direction are used for coordinate transformation, the spatial calibration parameters between the imaging distance detection device and the linear array imaging unit are used to determine the position of the measurement point relative to the imaging ray, and the tunnel line centerline, tunnel cross-section profile and mileage reference are used to convert the pixel spatial position into lining surface coordinates.
[0040] Methods for calculating pixel surface coordinates based on lining surface coordinate calculation parameters include: for linear array pixel coordinates Read the imaging intrinsic parameter matrix of the corresponding linear array imaging unit. and imaging distortion parameters The equivalent focal length is obtained from the imaging intrinsic parameter matrix. and principal pixel coordinates The radial distortion coefficient is obtained from the imaging distortion parameters. and Calculate the normalized distortion coordinates of the pixel according to the following formula. : ; Greater than zero, when When the value is not greater than zero, the corresponding imaging parameters are marked as invalid, and pixel imaging direction calculation is not performed.
[0041] The correspondence between normalized distorted coordinates and dedistorted coordinates is established according to the following formula: ;in, Let be the distortion-free normalized coordinates to be obtained.
[0042] All distortion-free normalized lateral coordinates were obtained from the effective calibration records of the linear array imaging units, and the minimum and maximum values were used as the lower limits for the distortion-free coordinate search. and search limit During the calibration phase, it is confirmed that the pixel coordinates of the aforementioned search range, after being converted by the distortion model, cover all valid pixel numbers of the linear array imaging unit. If they cannot be covered, the calibration records at the edge of the field of view are supplemented.
[0043] according to Establish the distortion transformation function, and follow the... Calculate its derivative; obtain the derivatives within the search range respectively. The minimum and maximum values, and take them as... The boundary values of the quadratic function; calculate the quadratic function. The minimum value among the function values at the two boundaries and the stationary point within the boundaries is taken as the minimum derivative of the distortion map; when the minimum derivative is not greater than zero, the distortion map is marked as not satisfying the monotonicity requirement, and the corresponding pixel inverse calculation is not performed.
[0044] When the distortion mapping satisfies the monotonicity requirement, and As the initial upper and lower bounds for the binary search, calculate the distortion transform values corresponding to the upper and lower bounds; when If the distortion value is less than the lower search limit or greater than the upper search limit, the corresponding pixel is marked as unattainable in the imaging direction.
[0045] For those within the search range Calculate the arithmetic mean of the current search upper and lower limits, use this mean as the search coordinates, and calculate the corresponding... If the calculated value is less than The current search coordinates will be used as the new lower limit of the search; if the calculated value is greater than... The current search coordinates will be used as the new upper limit of the search; if the calculated value equals The coordinates of this search are directly used as the coordinates for distortion correction.
[0046] Read the allowable error of crack geometry from the inspection task data. The maximum imaging distance is read from the pre-operation record of the imaging distance detection device. When the product of the difference between the upper and lower limits of the binary search and the maximum imaging distance does not exceed half of the allowable error of the crack geometry, the search is stopped, and the arithmetic mean of the final search upper and lower limits is used as the distortion-corrected coordinates. ;in, If it is a positive number, the difference between the upper and lower limits of the binary search is a dimensionless value, and its product with the maximum imaging distance is a length dimension, which is used to limit the imaging direction position error generated by the inverse calculation.
[0047] The unit imaging direction in the coordinate system of the linear array imaging element is calculated using the following formula. : ; where superscript This represents the transpose of the vector; a second component of the vector being zero indicates that the effective imaging ray lies within the fixed scanning plane of the linear array imaging unit.
[0048] A rotation matrix from the detection vehicle coordinate system to the tunnel coordinate system is generated based on the pitch, roll, and yaw angles at the moment the scan is formed. Based on the travel displacement and mileage reference at the moment of scan formation, the position vector of the origin of the detection vehicle coordinate system in the tunnel coordinate system is obtained. Read the rotation matrix from the linear imaging unit coordinate system to the detection vehicle coordinate system. and displacement vector .
[0049] according to Computational imaging ray origin ;according to Calculate the direction of the imaging ray Reading pixel coordinates Corresponding imaging distance ;according to Calculate the spatial position of the lining surface corresponding to each pixel; orthogonally project the spatial position of the lining surface onto the centerline of the tunnel line, and use the length of the line from the mileage reference to the projected position as the longitudinal surface coordinate. Within the cross-section passing through the projection position, the arc length from the starting point of the cross-section along the cross-section outline to the spatial position of the lining surface is taken as the circumferential surface coordinate. , to obtain the pixel surface coordinates.
[0050] When the imaging direction is unavailable, the imaging distance is unavailable, or the imaging distance is not greater than zero, the corresponding pixel is marked as having unavailable surface coordinates and is not involved in dual-temporal position correspondence and automatic crack identification.
[0051] The distance data output by the imaging distance detection device is calibrated and converted into the distance along the corresponding pixel unit imaging direction. When the device directly outputs the distance from the optical center of the linear array imaging unit to the measurement point along this direction, it is directly used as the imaging distance. When the device outputs the axial depth obtained along the Z-axis of the imaging unit with the same optical center as the origin, the axial depth is divided by the Z-axis component of the unit imaging direction to obtain the imaging distance. When the imaging distance detection device has an independent measurement origin, the measurement point is first transformed to the coordinate system of the linear array imaging unit according to the calibrated inter-device rotation matrix and displacement vector, and then the projected distance from the optical center to the measurement point along the corresponding unit imaging direction is calculated. When the perpendicular distance from the measurement point to the imaging ray exceeds the allowable error of the crack geometry, the corresponding distance record is marked as invalid.
[0052] The method for obtaining the surface coordinate calculation error includes: setting verification marker points at the arch crown, arch waist, and sidewalls of the tunnel lining, with surface coordinates measurable by a total station; for each verification marker point, obtaining the corresponding calculated surface coordinates according to the aforementioned pixel surface coordinate calculation method, and measuring the spatial position of the verification marker point using a metrologically calibrated total station, converting the measured position into measured surface coordinates according to the tunnel centerline and cross-sectional profile; and calculating the surface distance between the calculated coordinates and measured coordinates of each verification marker point according to the following formula. ; ;in, and These are the longitudinal and circumferential surface coordinates of the verification marker points obtained using the aforementioned surface coordinate calculation process, respectively; and These are the longitudinal and circumferential surface coordinates of the same verification marker point, calculated based on the total station measurement results and the tunnel centerline and tunnel cross-sectional profile. The surface distances of all verification marker points are then... The maximum value is used as the error in surface coordinate calculation. When any verification marker cannot obtain valid calculated or measured coordinates, the verification marker is not used to generate numerical error, and the corresponding verification position is marked as incomplete coordinate verification. After all verification markers have completed coordinate verification, the obtained surface coordinate calculation error is used to participate in the subsequent position search radius calculation.
[0053] The method for obtaining time synchronization error includes: the acquisition control device sending the same test trigger to two linear imaging units, a travel displacement detection device, an attitude detection device, an imaging distance detection device, and an illumination drive recording unit; recording the common clock time at which each device generates the corresponding sampling record; calculating the difference between the latest and earliest sampling times in the same test; repeatedly sending the test trigger according to adjacent times that can be distinguished by the common clock; and taking the maximum value among all time differences as the time synchronization error. The common clock resolution is read from the timer configuration register of the acquisition control device and verified using the difference between adjacent distinguishable timestamps.
[0054] The method for obtaining the lighting drive cycle and lighting drive phase includes: the lighting drive recording unit acquiring the pulse control signal output by the lighting drive device; the position where the pulse control signal changes from low to high level as the rising drive edge, or the position where it changes from high to low level as the falling drive edge; the same type of drive edge is consistently selected in a single detection; and the median value of the common clock time difference between adjacent drives of the same type is taken as the lighting drive cycle. The difference between the scan formation time and the most recent same driving edge time before the scan formation time is used as the lighting drive phase; the lighting drive delay adjustment resolution is read from the delay setting register of the lighting controller and verified by the difference between two consecutive executable delay setting values; the larger value between the common clock resolution and the lighting drive delay adjustment resolution is used as the lighting phase sampling interval.
[0055] The method for obtaining calibration acquisition data includes: obtaining the imaging intrinsic parameter matrix, imaging distortion parameters, rotation matrix and displacement vector from the coordinate system of the linear array imaging unit to the coordinate system of the detection vehicle after calibration and verification of the first and second linear array imaging units; obtaining the installation direction and attitude calibration results of the attitude detection device relative to the detection vehicle, the distance calibration results of the imaging distance detection device, and the displacement conversion parameters and displacement verification results of the traveling displacement detection device; obtaining the tunnel centerline, tunnel cross-sectional profile and mileage benchmark from the detection task data; obtaining the surface coordinate calculation error calculated by the verification marker points, the time synchronization error calculated by the common clock trigger test, the common clock resolution, the lighting drive cycle and the lighting phase sampling interval; and associating the above data according to the device identifier and the corresponding calibration version to obtain calibration acquisition data.
[0056] Among them, the imaging intrinsic parameter matrix, imaging distortion parameters, and linear array imaging unit mounting pose parameters are used to determine the pixel imaging direction and imaging ray position; the attitude calibration results, distance calibration results, displacement conversion parameters, and tunnel geometric data are used to convert the measurement position corresponding to the pixel to the lining surface coordinates; the surface coordinate calculation error and time synchronization error are used to calculate the corresponding range of the dual-phase position; the common clock resolution, illumination drive cycle, and illumination phase sampling interval are used to determine the correspondence between the scan formation time and the illumination drive phase.
[0057] The calibration acquisition data does not include the actual pixel grayscale, exposure start and end times, running posture, and imaging distance obtained for each scanning unit during the actual scanning; the actual running data forms a scanning record according to the scanning unit identifier. The image processing device reads the calibration acquisition data and the scanning record, and performs coordinate conversion and time correlation calculation based on the corresponding device identifier and the scanning formation time to obtain the scanning spatiotemporal correlation data.
[0058] The method for obtaining parameter sample data includes: selecting a fixed crack location confirmed by close observation, and acquiring corresponding scan images under different detection speeds, exposure times, and illumination output conditions; when the crack's two sides can be continuously acquired, and the difference between the crack width calculated based on the two sides and the close-range measurement result does not exceed the allowable error of crack geometry, the corresponding image is marked as a identifiable crack sample; when the crack's two sides cannot be acquired or the crack width difference exceeds the allowable error of crack geometry, the corresponding image is marked as an indistinguishable crack sample; the sample identifier, image data, and corresponding acquisition parameters are associated to obtain image quality sample data; for confirmed crack-free background areas, the illumination output is changed according to a predetermined illumination driving cycle, and linear array images are acquired synchronously; the background brightness change and corresponding linear response formed during the illumination output change are associated with surface coordinates and the scanning formation time to obtain illumination change sample data; and fixed crack sample data are obtained. The data includes samples of lighting changes, moving water droplets, image quality, correct movement, and incorrect movement. For fixed crack samples, the linear location to be identified is cleaned and observed closely with illumination incident from both sides of the linear location. If the linear location remains fixed on the same surface after the illumination direction is changed, and the edges of the linear opening can be obtained using a crack width measuring gauge, then the corresponding location is identified as a crack location. For crack-free background areas, the surface is cleaned and observed with illumination from different incident directions. If the edges of the opening cannot be obtained within the minimum crack width required for identification, there are no construction joints or edges of ancillary facilities in the area, and there are no water droplets or watermarks with changes in position in the synchronous area array observation image, then the corresponding area is identified as a crack-free background area. The identification results, surface coordinates, and corresponding scanned images are combined to obtain fixed crack sample data and lighting change sample data, respectively.
[0059] The method for obtaining moving water droplet sample data includes: arranging marker points on the lining surface to be observed, whose surface coordinates have been measured by a total station; acquiring images of the marker points using a surface array camera; and setting the homogeneous pixel coordinates of the surface array image as follows: The corresponding homogeneous coordinates of the surface are By minimizing all marked points Obtain the transformation matrix from pixel coordinates of the area array to coordinates of the locally unfolded surface. Transformation matrix Used as surface coordinate calibration parameters for the area array observation camera; forming a closed contour along the visible water droplet boundary in the area array observation image; transforming all pixels within the closed contour using a transformation matrix. The arithmetic mean of the calculated surface coordinates is used as the surface coordinates of the water droplet center. The surface coordinates of the water droplet center are arranged according to the exposure midpoint of the area array observation image. When the distance between the surface coordinates of the water droplet center at adjacent observation times is greater than the sum of the errors of the two surface coordinate measurements, it is confirmed that the surface position of the water droplet has changed. The water droplet outline, the surface coordinates of the center, the observation time, and the synchronous linear array image are combined to obtain the moving water droplet sample data.
[0060] The method for obtaining unified judgment boundary generation rules includes: firstly, defining the target sample, control sample, and comparison direction for each judgment index; for the illumination time-related evaluation index, using illumination change samples as target samples and fixed crack samples as control samples, and taking the direction of numerical increase as the target direction; for the moving centerline coincidence ratio, using the correct moving corresponding sample as the target sample and the incorrect moving corresponding sample as the control sample, and taking the direction of numerical increase as the target direction; for the moving direction difference and moving width difference, using the correct moving corresponding sample as the target sample and the incorrect moving corresponding sample as the control sample, and taking the direction of numerical decrease as the target direction; for the surface repeated observation direction difference and width difference, using fixed crack samples as target samples and illumination change samples and moving water droplet samples as control samples, and taking the direction of numerical decrease as the target direction; for the image quality index, using discernible crack samples as target samples and indistinguishable crack samples as control samples, where gradient energy takes the direction of numerical increase, and the highest output pixel ratio, lowest output pixel ratio, and local grayscale noise take the direction of numerical decrease.
[0061] For the determination index of the target sample corresponding to the direction of numerical increase, obtain the minimum value among all target parameter determination sample index values. Determine the maximum value among all control parameters of the sample indicator. ;when At that time, according to Obtain the corresponding decision boundary, and take the index value reaching the decision boundary as meeting the target condition; for the decision index of the target sample corresponding to the direction of numerical decrease, obtain the maximum value among all target parameters and determine the sample index value. Determine the minimum value among all control parameters for the sample indicator. ;when At that time, according to Obtain the corresponding judgment boundary and take the index value not exceeding the judgment boundary as meeting the target condition; when the parameter determination sample cannot form the aforementioned size relationship, mark the corresponding judgment boundary as not being generated; substitute the verification sample into the generated judgment boundary one by one; when any target verification sample does not meet the target condition or any control verification sample meets the target condition, mark the corresponding judgment boundary as invalid; judgment boundaries that cannot be generated or are invalid shall not be used to automatically eliminate cracks.
[0062] The method for obtaining displacement conversion parameters includes: using a metrologically calibrated total station to measure the spatial position of adjacent mileage markers in the section to be inspected, determining the longitudinal surface distance between adjacent mileage markers according to the tunnel centerline; controlling the inspection vehicle to pass through the corresponding mileage markers sequentially along the tunnel centerline, collecting the effective displacement count generated by the traveling displacement detection device between two mileage markers, and using the acquisition control device to accumulate the effective displacement count to obtain the cumulative count increment corresponding to each measurement interval.
[0063] The first The measured longitudinal surface distance of each measurement interval is denoted as . The cumulative count increment corresponding to this interval is denoted as ;when Greater than zero and When greater than zero, according to Calculate the longitudinal displacement corresponding to the unit count. Longitudinal displacement counted per unit across all valid measurement intervals The values are compared, and the maximum value is used as the displacement conversion parameter. The purpose of using the maximum value is to conservatively convert the scan trigger interval, and to avoid generating the trigger interval by counting displacement in smaller units, which would cause the actual scan spacing to exceed the detection requirements.
[0064] The effective displacement count is accumulated from the displacement pulses output by the traveling displacement detection device. When the traveling displacement detection device directly outputs a count value with counting loop characteristics, the counting range is determined according to the number of bits in the counting register recorded by the device interface, and the count increment is calculated based on the count changes at adjacent sampling times and the register counting range. If more than one count loop may occur within adjacent sampling intervals and the count increment cannot be uniquely determined, the corresponding displacement record is marked as invalid and is not used for displacement conversion parameter calculation. When the traveling displacement detection device does not obtain an effective count increment, the displacement count is not inferred based on the vehicle's travel time.
[0065] After completing the displacement conversion, the calculated displacement of each measurement interval is calculated by multiplying the displacement conversion parameters by the corresponding cumulative count increment. The absolute difference between the calculated displacement and the measured longitudinal surface distance is taken as the displacement conversion verification error. The allowable error of mileage measurement is read from the detection specification, and the displacement conversion verification error is compared with the allowable error of mileage measurement. When the displacement conversion verification error exceeds the allowable error of mileage measurement, the travel displacement detection device is recalibrated or the cumulative displacement is recalibrated using mileage markers. After the calibration is completed, the acquisition quality is still checked according to the surface spacing of adjacent actual scanning units and the exposure movement distance, and the cumulative mileage verification result is not used to replace the local sampling quality judgment.
[0066] The method for obtaining scanning operation configuration data includes: reading the lower limit of the vehicle's executable speed, the upper limit of the vehicle's executable speed, and the difference between adjacent executable speed settings from the vehicle controller, and using the difference between adjacent executable speed settings as the speed adjustment resolution; generating candidate detection speeds by gradually increasing the speed adjustment resolution starting from the lower limit of the vehicle's executable speed; generating candidate exposure times by reading the upper and lower limits of exposure time and the difference between adjacent executable exposure times from the linear array imaging unit; generating candidate illumination outputs by reading the upper and lower limits of illumination output and the difference between adjacent executable illumination outputs from the illumination driving device; using the candidate detection speed, candidate exposure time, and candidate illumination output for image quality sample acquisition; and retaining only candidate combinations that ensure acceptable image quality, that the surface distance between the exposure start position and the exposure end position does not exceed the allowable error of crack geometry, and that the measured surface distance between adjacent scanning units does not exceed the minimum crack width.
[0067] The longitudinal surface coordinate differences between the center of the first scan line and the center of the second scan line were measured at multiple circumferential locations using a total station, and the median value of all coordinate differences was used as the interval for repeated observations. Regarding candidate detection speed ,according to Calculate the observation time difference When the lighting driving cycle is available, according to Calculate observed phase separation data ,in To minimize the absolute difference, select the candidate combination with the best image quality and the largest observed phase separation data. When multiple candidate combinations correspond to the same observed phase separation data, select the candidate combination with the lowest detection speed and the shortest exposure time in sequence. Combine the selected detection speed, exposure time, illumination output, longitudinal displacement trigger interval, and observation time difference to obtain the scanning operation configuration data.
[0068] Methods for obtaining the longitudinal displacement trigger interval include: reading the minimum crack width to be identified from the detection task data. Read the unit count longitudinal displacement from the displacement conversion parameters. ;according to Calculate the trigger count interval, where, Indicates rounding down. This is the cumulative displacement count that needs to be accumulated between two adjacent line exposure triggers.
[0069] when When the count resolution is less than one, it is determined that the counting resolution of the travel displacement detection device does not meet the requirements of the current detection task; when Not less than one hour, according to Calculate the longitudinal displacement trigger interval .
[0070] The data acquisition and control device re-accumulates the effective displacement count after the detection vehicle passes the current line exposure trigger position. When the accumulated count reaches... The next line exposure trigger is generated at the specified time; the actual exposure start time and exposure end time are recorded, and the theoretical trigger time is not used to replace the actual exposure time.
[0071] Perform a pre-scan according to the obtained longitudinal displacement trigger interval, and calculate the longitudinal surface coordinate difference between adjacent actual scanning units at the same circumferential position using the calibrated imaging parameters, attitude data and imaging distance data; when the longitudinal surface coordinate difference between all adjacent scanning units does not exceed the minimum crack width, and the surface displacement corresponding to the exposure start time and exposure end time of each scanning unit does not exceed the allowable error of crack geometry, the current longitudinal displacement trigger interval is taken as the effective trigger interval.
[0072] When the longitudinal surface coordinate difference between any two adjacent scanning units exceeds the minimum crack width, the trigger counting interval will be reduced by one counting unit, and a pre-scan will be performed again. When the surface displacement generated during exposure exceeds the allowable error of crack geometry, the candidate exposure time will be shortened or the candidate detection speed will be reduced, and a pre-scan will be performed again. If the longitudinal sampling interval requirement is still not met when the trigger counting interval is reduced to one counting unit, the current scanning operation configuration will not be used for formal detection.
[0073] Optionally, when multiple candidate combinations satisfy the requirements for image quality, exposure movement distance, and surface distance between adjacent scanning units exist, the selection criterion for scanning operation configuration data is not based on a single image sharpness, but rather on the repeated observation interval between the first and second scan lines and the candidate detection speed. The corresponding observation time difference is calculated, and the scanning operation configuration data is selected based on the phase relationship between the observation time difference and the illumination drive cycle. This ensures that the first and second observations maintain spatial correspondence on the same lining surface position while having an observation time interval for characterizing time changes. For structures fixed to the lining surface, the surface position correspondence between the two observations does not change due to the observation time interval. For targets that undergo positional changes during detection or are affected by periodic illumination changes, the two observation results can produce positional differences or brightness response differences. The aforementioned observation time difference is only used to establish observable conditions for time changes and is not used as a separate criterion for whether a crack exists or not.
[0074] The scanning is performed according to the scanning operation configuration data; for each line exposure, the pixel grayscale, imaging unit identifier, exposure start time, exposure end time, travel displacement, pitch angle, roll angle, yaw angle, imaging distance and illumination drive phase are recorded to obtain the scanning record; for each scanning unit, the longitudinal surface coordinates and circumferential surface coordinates of each pixel are obtained by using the aforementioned surface coordinate calculation process; the scanning unit identifier, pixel grayscale, exposure time interval, scan formation time, lining surface coordinates and illumination drive phase are combined to obtain the scanning spatiotemporal correlation data.
[0075] Optionally, a surface position identifier is set for each grid position rearranged to a unified lining surface sampling grid. The surface position identifier consists of the longitudinal grid number to which the longitudinal surface coordinates belong and the circumferential grid number to which the circumferential surface coordinates belong. The actual scanning units falling into the same surface position identifier in the first and second observation images are recorded as the first and second observation sources for that surface position, respectively. Subsequent surface repetition support determination, illumination time correlation determination, movement time correlation determination, and observation pixel selection all use the surface position identifier as the position index. Pixels in the first and second observation images with the same image array coordinates but different surface position identifiers are not compared as the same lining surface position.
[0076] The scan records are rearranged according to the scan spatiotemporal correlation data to obtain dual-temporal observation images; the position correspondence and image quality are determined for the dual-temporal observation images and scan records to obtain the qualification for repeated observation; candidate linear targets are detected in the dual-temporal observation images to obtain candidate linear targets, and the scan formation time corresponding to the candidate linear targets is queried according to the scan spatiotemporal correlation data to obtain the formation time of the candidate targets; the dual-temporal observation images include a first observation image and a second observation image.
[0077] In a specific implementation, refer to Figure 3 and Figure 4 Methods for obtaining candidate linear target data and formation time include:
[0078] A unified lining surface sampling grid is established according to the longitudinal displacement trigger interval in the scanning operation configuration data; for the effective scanning units obtained from the pre-scan, the circumferential surface coordinates corresponding to adjacent effective pixels are read in the order of the linear array pixel numbers; the absolute value of the difference between the circumferential surface coordinates of each adjacent effective pixel is calculated, and records with a difference equal to zero or adjacent pixels having invalid surface coordinates are excluded; the median value of the remaining differences is used as the circumferential grid interval; when there is no effective adjacent pixel coordinate difference, or the obtained circumferential grid interval exceeds the minimum crack width in the detection task data, the current imaging configuration is not used; the longitudinal displacement trigger interval is used as the longitudinal grid interval, and the circumferential grid interval is used as the circumferential sampling interval to establish a unified lining surface sampling grid; target grid position. Located at surface coordinates and Between two actual scanning units, according to Calculate the grayscale value at the target grid location; where and The ratio of the surface distance from the target grid position to the two actual scanning units to the distance between the two actual scanning units is defined as follows: when the distance between the two actual scanning units exceeds the allowable missing scanning distance, resampling is not performed, and the corresponding grid position is marked as the missing acquisition position; the allowable missing scanning distance is obtained by successively increasing the number of consecutively deleted scanning units in the identifiable crack sample; after each deletion, the edges on both sides of the crack are reacquired; the product of the maximum number of consecutive deletions where the edges on both sides of all identifiable crack samples can be acquired and the error of the edges on both sides does not exceed the allowable error of the crack geometry and the longitudinal displacement trigger interval is used as the allowable missing scanning distance.
[0079] According to the data of the first linear array imaging unit and the second linear array imaging unit arranged according to the unified lining surface sampling grid, the first observation image and the second observation image are obtained respectively; the complete scan record means that the corresponding grid position has the actual pixel gray level, exposure start and end time, travel displacement, attitude, imaging distance and imaging unit identifier, and the distance between adjacent actual scan units does not exceed the allowable scan missing distance; the exposure movement exceeds the limit means that the distance between the exposure start surface position and the exposure end surface position obtained by linear interpolation exceeds the allowable error of crack geometry.
[0080] Methods for obtaining image quality data include: reading the minimum digital output value from the metadata of the linear array imaging unit. and maximum digital output value ;according to Obtain normalized grayscale Calculate the horizontal gray-level gradient based on the normalized gray-level difference between adjacent pixels. and vertical grayscale gradient ; to convert all pixels within an image block The arithmetic mean of the values is used as the gradient energy; the number of pixels with grayscale equal to the maximum digital output value is divided by the total number of pixels in the image block to obtain the highest output pixel ratio; the number of pixels with grayscale equal to the minimum digital output value is divided by the total number of pixels in the image block to obtain the lowest output pixel ratio; the smallest odd number not less than the minimum crack pixel width is selected as the side length of the median filtering window; the median value of the normalized grayscale within the window is used as the median filtered grayscale at the center of the window; the median value of the absolute value of the difference between the original normalized grayscale and the median filtered grayscale is used as the local grayscale noise; based on the discernible crack samples and the indistinguishable crack samples, the lower boundary of the gradient energy, the upper boundary of the highest output pixel ratio, the upper boundary of the lowest output pixel ratio, and the upper boundary of the local grayscale noise are obtained according to the unified judgment boundary generation rules; when the gradient energy reaches the corresponding lower boundary and the other three items do not exceed the corresponding upper boundaries, the image quality is judged to be qualified.
[0081] If the scan record of the grid position corresponding to the second observation is complete, does not belong to the missing acquisition position or the exposure movement exceeding the limit position, the image quality judgment boundary is valid, and the image quality of the second observation is qualified, the qualification for repeated observation of the corresponding position will be set to valid; if any condition is not met, the qualification for repeated observation will be set to invalid; if the qualification for repeated observation is invalid, the candidate linear target of the first observation shall not be excluded based on the result that no linear response appears in the second observation.
[0082] Optionally, repeat observation eligibility serves as a prerequisite for the second observation result to negate the first observation result, but not as a prerequisite for the first observation candidate linear target to enter subsequent processing. When the corresponding position of the candidate linear target in the first observation image has valid repeat observation eligibility, it is permissible to determine whether surface repeat support is valid or invalid based on the second observation result; when the corresponding position does not have valid repeat observation eligibility, the first observation candidate linear target is retained, and surface repeat support is set to undecidable, and the first observation candidate linear target is not deleted simply because no candidate linear target was detected in the second observation image. This distinguishes between targets that were not observed again and targets that met the conditions for re-observation but did not reappear.
[0083] Candidate directions are generated based on the minimum crack length and the allowable error of crack geometry; those that satisfy... The smallest positive integer ;according to Generate candidate directions, where Incrementing from zero to less than The integer; the pixel width corresponding to the minimum crack width is obtained based on the imaging calibration results. Never less than Starting from the smallest integer pixel distance, increase by one pixel distance until it is not less than The minimum integer pixel distance; delete the pixel distances that correspond to the same surface distance after surface resolution conversion, and obtain the normal sampling distance.
[0084] For candidate directions and normal sampling distance Obtain the normalized grayscale of the current pixel. and normalized grayscale on both sides of the normal direction and When the normal sampling position is located within a grid enclosed by four pixels, let the horizontal ratio of the sampling position relative to the top-left pixel be . The vertical ratio is ,according to Obtain sampled grayscale ;in, and These are the ratios obtained by dividing the horizontal and vertical distances between the sampling position and the top left corner of the grid by the grid interval in the corresponding direction, respectively, with values ranging from 0 to 1; , , and These are the normalized grayscale values of the top-left, top-right, bottom-left, and bottom-right pixels among the four adjacent pixels surrounding the normal sampling position; according to... Calculate the linear response When the gray level of the center pixel is lower than the average of the gray levels of the two side samples, a signed linear response occurs. A negative value indicates a signed linear response when the gray level of the center pixel is higher than the average of the gray levels sampled on both sides. It is a positive value.
[0085] Iterate through all the aforementioned candidate directions and normal sampling distances, and take the maximum value in the corresponding linear response as the linear response value of the current pixel; take the candidate direction that obtains the maximum linear response value as the pixel candidate direction, and take the corresponding... The sign is used as the signed normal response of the pixel; when multiple candidate directions and normal sampling distances correspond to the same maximum linear response value, the one with the smaller normal sampling distance is selected first, and then the one with the smaller candidate direction angle is selected as the corresponding parameter of the current pixel.
[0086] When any normal sampling position falls outside the sampling missing position, the exposure movement exceeding the limit position, or the image boundary, the linear response of the corresponding candidate direction and normal sampling distance is not calculated; when no effective linear response can be obtained from all candidate combinations, the corresponding pixel is marked as having no linear response.
[0087] Based on the aforementioned artificially labeled samples of fixed cracks, lighting variation structures, and moving water droplet structures, the maximum linear response in each artificially labeled region is recalculated, and the minimum value among the maximum linear responses in all artificially labeled regions is used as the linear response determination boundary. When a valid sample response cannot be obtained or the obtained determination boundary is not greater than zero, the linear response determination boundary is marked as invalid, and candidate linear targets are not automatically excluded based on this boundary.
[0088] Pixels whose linear responses reach the linear response determination boundary are connected by eight neighbors to obtain candidate linear regions; boundary pixels that do not change the number of connected eight neighbors, do not delete endpoints, and do not break the region are iteratively deleted from the candidate linear regions; when there are no boundary pixels that meet the deletion conditions, the deletion is stopped, and the candidate linear target centerline is obtained.
[0089] For each centerline location, obtain the centerline point extending along both sides of the centerline with a cumulative surface length not exceeding the minimum crack length; calculate the arithmetic mean of the surface coordinates of the centerline point. Calculate the covariance matrix The unit eigenvector corresponding to the larger eigenvalue of the covariance matrix is taken as the direction of the local centerline; when the centerline position is a branch point or the larger eigenvalue and the smaller eigenvalue cannot be distinguished, the direction of the local centerline is marked as unattainable.
[0090] A measurement line perpendicular to the local centerline is established at the centerline position; linear responses are sequentially obtained along both sides of the measurement line according to the surface sampling grid interval; two adjacent sampling positions that change from reaching the linear response judgment boundary to falling below the linear response judgment boundary are taken as the state transition interval; let the two adjacent positions be... , The corresponding linear response is , ,according to Calculate the state transition position, where The boundary is determined for the linear response; when both sides obtain the state transition position, the surface distance between the two state transition positions is taken as the local width; when the state transition position is not obtained before reaching the image boundary, the missing acquisition position, or the exposure movement exceeds the limit position on either side, the local width is marked as unobtainable; the candidate linear region, center line, local center line direction, local width, and signed normal response are combined to obtain the candidate linear target data.
[0091] Query each scanning unit through which the candidate linear target passes; combine the exposure start time, exposure end time, scan formation time of the scanning unit, and the surface coordinates of the center line within the scanning unit to obtain the formation time of the candidate target.
[0092] Compare dual-temporal candidate linear targets at locations eligible for repeated observation to obtain surface repeat support data; calculate the correlation between the candidate target formation time and background brightness changes, and the correlation between the candidate target formation time and target position changes to obtain illumination time correlation data and movement time correlation data.
[0093] In a specific implementation, the method for obtaining surface repetition support data, illumination time-related data, and movement time-related data includes:
[0094] Methods for obtaining surface repeatability support data include: reading the surface coordinate calculation error from the surface coordinate calculation error data. Read the time synchronization error from the time synchronization error data The maximum permissible detection speed for the actual scanning is read from the scanning operation configuration data and the detection vehicle controller. The maximum permissible detection speed is set at the upper limit of the vehicle's executable speed, which covers all candidate operating speeds for the current detection task. Calculate the search radius of the location The surface coordinate calculation error is added to the maximum travel position deviation caused by time synchronization to obtain the conservative surface position deviation for a single observation; the conservative surface position deviation for a single observation is multiplied by 2 to obtain the position search radius used to cover the position deviations of two observations; for each surface position of the first observation candidate centerline, only the centerline position whose surface distance does not exceed the position search radius is selected from the second observation candidate linear target data to participate in the subsequent morphological comparison; when there are multiple centerline positions in the second observation candidate linear target data, the centerline position with the smallest surface distance is selected; the smallest surface distance is used as the surface position difference; when both local centerline directions are obtainable, according to... Calculate the direction difference When both local widths are available, according to Calculate the width difference .
[0095] Using a fixed crack sample as the target sample and a changing illumination sample and a moving water droplet sample as control samples, the upper boundaries of the direction difference and width difference are obtained according to a unified decision boundary generation rule. Surface repeat support is set to valid when the repeat observation qualification is valid, the surface position difference does not exceed the position search radius, at least one of the direction difference and width difference can be obtained from the two observation results, all of which can be obtained and the difference between the corresponding decision boundaries does not exceed the corresponding upper boundary, and the signed normal response signs of the two observation results are the same. Surface repeat support is set to invalid when the repeat observation qualification is valid but there is no linear response reaching the linear response decision boundary within the second observation search range. Surface repeat support is set to undecidable when the repeat observation qualification is invalid or neither the direction difference nor the width difference can be obtained.
[0096] Optionally, for cases where the local centerline direction and local width cannot be obtained simultaneously due to the candidate linear target being located at the image boundary, bifurcation position, or locally missing position, a layered consistency determination is adopted: First, it is determined whether the surface position difference between the second observed candidate linear target and the first observed candidate linear target is within the position search radius; if so, the morphological parameters that can be obtained from both observations are then determined; when the local centerline direction is obtained jointly, the direction difference is compared; when the local width is obtained jointly, the width difference is compared; when at least one morphological parameter can be obtained and all obtainable morphological parameters satisfy the corresponding determination boundary, the signed normal response signs of the two observations are then compared; surface repetition support is only set when the above position condition, obtainable morphological condition, and grayscale polarity condition are simultaneously satisfied. Unobtainable morphological parameters are not replaced with preset values.
[0097] The method for obtaining background brightness includes: acquiring data from a uniform calibration board under constant DC illumination; repeating the acquisition process multiple times, equal to the number of phase positions generated according to the illumination phase sampling interval within one illumination driving cycle; calculating the normalized grayscale median value of the central effective area of the uniform calibration board in each acquisition; and using the maximum absolute difference between each median value and the median value of all medians as the grayscale repetition error. For each scanning unit traversed by the candidate linear target, the surface area corresponding to the larger of the local width and the minimum crack width is excluded, centered on the candidate linear target's centerline. Starting from both sides of the excluded area, a surface sampling grid interval is added to form a background statistical region. Pixels reaching the linear response judgment boundary, digital output boundary pixels, missing acquisition pixels, and pixels exceeding the exposure movement limit are excluded. The normalized grayscale median value of the remaining pixels is calculated. Another surface sampling grid interval is added to both sides, and the median value is recalculated. The process stops when the absolute value of the difference between two adjacent median values does not exceed the grayscale repetition error, and the last median value is used as the background brightness of the current scanning unit. If the stopping condition cannot be met even after reaching the image boundary, or if there are no remaining pixels on either side, the background brightness is marked as unattainable.
[0098] When the background brightness of adjacent scanning units can be obtained, according to Calculate background brightness changes The median value of the signed normal response of the candidate linear target within the corresponding scan cell is used as the candidate response. ;in, These are the scan unit numbers arranged according to the scan formation time; The background brightness of the previous scan unit is used as the reference. From the scan units covered at the time of candidate target formation, scan units whose candidate responses and background brightness changes are both obtainable are selected. The serial numbers of the selected scan units are used to form a set of correlation calculation serial numbers. The arithmetic mean of all candidate responses in the correlation calculation serial number set is calculated to obtain the mean of the candidate responses. Calculate the arithmetic mean of all background brightness changes within the same set to obtain the mean background brightness change. The scan formation time range involved in the calculation is not less than one illumination driving cycle, the set of associated calculation numbers contains at least two scan units, and the sum of the squared deviations of both the candidate response sequence and the background brightness change sequence is greater than zero, according to... Calculate the normalized correlation coefficient , The value range is -1 to 1; Scanning units whose candidate responses and background brightness changes are both non-zero are selected from the associated calculation sequence set. The selected scanning unit indices are grouped into a symbolic calculation sequence set, and the number of scanning units in the set whose candidate responses and background brightness changes have the same sign is counted. And the number of scan units with opposite signs. ; Increase the number of scanning units Divide by To obtain the same sign ratio ; Increase the number of scanning units Divide by To obtain the inverse sign ratio ;according to Calculate the time-related evaluation value of lighting When the lighting drive cycle is unavailable, the correlation calculation time range is shorter than the lighting drive cycle, the correlation calculation sequence number set is less than two scan units, the sum of squared deviations of candidate responses or the sum of squared deviations of background brightness changes is zero, or the symbol calculation sequence number set is empty, the lighting time correlation evaluation value is marked as unavailable.
[0099] The above lighting time-related evaluation value is compared with the lighting interference judgment boundary obtained according to the unified judgment boundary generation rule; when the lighting time-related evaluation value reaches the effective lighting interference judgment boundary, the corresponding lighting time-related state is set to established; when the lighting time-related evaluation value is obtainable but does not reach the effective judgment boundary, the corresponding lighting time-related state is set to invalid; when the lighting time-related evaluation value is unobtainable or the lighting interference judgment boundary is invalid, the corresponding lighting time-related state is set to undeterminable.
[0100] Using the aforementioned methods for calculating signed normal response and normalized correlation coefficient, the lighting time correlation evaluation values for lighting variation samples and fixed crack samples are calculated respectively, and the lighting interference judgment boundary is obtained according to the unified judgment boundary generation rule; when the judgment boundary fails independent verification, the corresponding lighting time correlation state is set to undecidable.
[0101] The methods for obtaining candidate movement trajectories include: obtaining the arithmetic mean of the centerline surface coordinates of the candidate linear target in different scanning units, and arranging them according to the scanning formation time; and solving... and Obtain longitudinal surface velocity and circumferential surface velocity When the number of different scan formation times cannot uniquely determine the two slopes, the candidate movement trajectory is marked as unattainable; when the two slopes can be uniquely determined, As a candidate surface velocity vector.
[0102] The surface coordinates of the same water droplet in the sample at two observation times were determined based on the parameters of the moving water droplet. , and scan formation time , ,according to Calculate the surface displacement distance; according to Calculate the direction of surface displacement; according to Calculate the surface displacement per unit time; use the moving water droplet parameters to determine the minimum and maximum values of the surface displacement per unit time in the sample as the lower boundary of the moving velocity. and upper boundary Arrange the surface displacement directions in a circular pattern, and select the direction range with the smallest angle span that includes all parameters determining the sample direction as the movement direction range; if the velocity or direction of any moving water droplet verification sample is outside the corresponding range, mark the movement velocity range or movement direction range as invalid.
[0103] Movement association is performed only on the first observation candidate linear targets whose repeat observation eligibility is valid and whose surface repeat support is invalid; the center position of the first observation candidate linear target is taken as the first observation center position, and the corresponding first observation time and second observation time are read. Let the first observation center position be... The difference between the second observation time and the first observation time is ,when When the value is not greater than zero, the corresponding movement time association status is set to undeterminable; when... When the value is greater than zero, the lower boundary of the movement speed is read from the movement speed range. and the upper boundary of movement speed The location search radius is read from the surface repeating support data. ,according to and Calculate the radial range of the moving search. and These are the lower and upper limits of the radial distance for the moving search, respectively; when Greater than At that time, according to Calculate the expansion angle of the movement search direction, and expand the range of the movement direction to both sides. To obtain the range of movement search directions; when Not greater than When the second observation search area is not limited by the direction of movement, the entire circumference direction is used as the range of the movement search direction; with the first observation center position as the center, the second observation surface area whose surface distance is within the radial range of the movement search and whose direction is within the range of the movement search direction is used as the second observation movement search area; when the range of the movement search direction is the entire circumference direction, the corresponding annular area is used as the second observation movement search area.
[0104] Read the centerline surface coordinates of each candidate linear target one by one from the second observation candidate linear target data; take the candidate linear targets whose centerline at least one position falls into the second observation moving search area as candidate moving targets; when there are no candidate linear targets that meet the above conditions, record the candidate moving target set as an empty set, and determine the moving time association state according to the moving speed range, moving direction range and the validity of the corresponding judgment boundary.
[0105] The method for calculating the total length of the first observation centerline includes: establishing adjacent centerline pixels as undirected graph edges; recording each undirected graph edge only once; calculating the surface distance of the graph edge based on the surface coordinates at both ends of the graph edge; and taking the sum of the surface distances of all graph edges as the total length of the first observation centerline. The method for calculating the centerline coincidence ratio includes: calculating the displacement vector between the average coordinate of the first observation centerline and the average coordinate of the candidate moving target centerline; translating the first observation centerline using the displacement vector; marking graph edges on the translated first observation centerline whose minimum surface distance to the candidate moving target centerline does not exceed the position search radius; and dividing the sum of the surface distances of the marked graph edges by the total length of the first observation centerline to obtain the centerline coincidence ratio.
[0106] The methods for obtaining the difference in movement direction and width include: for the target to be judged and the candidate moving targets, selecting the centerline position on their respective centerlines that is closest to the arithmetic mean of the corresponding centerline coordinates and can obtain the local centerline direction; and reading the unit local centerline direction vector of the selected position. and ,according to Calculate the difference in movement direction; where, The angle between the two centerline directions is undirected and expressed in radians.
[0107] Read the local width corresponding to the two selected positions respectively and ,according to Calculate the movement width difference; when the local centerline direction of any target cannot be obtained, mark the movement direction difference as unobtainable; when the local width of any target cannot be obtained, mark the movement width difference as unobtainable; when both the movement direction difference and the movement width difference are unobtainable, do not confirm the movement correspondence based on shape consistency.
[0108] Two observed targets confirmed to belong to the same water droplet through continuous manual observation are designated as correct movement corresponding samples; targets not confirmed to belong to the same water droplet within the second observation movement search area are grouped into incorrect movement corresponding samples; using the correct movement corresponding samples as target samples and the incorrect movement corresponding samples as control samples, the lower boundary of the centerline coincidence ratio, the upper boundary of the movement direction difference, and the upper boundary of the movement width difference are obtained according to the unified judgment boundary generation rules; movement association is allowed when the centerline coincidence ratio boundary is valid, and at least one of the direction difference boundary and the width difference boundary is valid; the movement direction difference and movement width difference of the candidate moving targets are defined as follows: the centerline coincidence ratio of the candidate moving targets reaches the lower boundary of the centerline coincidence ratio, all are obtainable, and the corresponding judgment boundaries are valid. If neither of the two observations exceeds the corresponding upper boundary and the signs of the signed normal responses are the same, the movement time association status of the target to be judged is set to true; if movement association is allowed and there is no candidate moving target that meets the aforementioned conditions, the movement time association status is set to false; if the candidate movement trajectory, movement speed range, movement direction range, or the judgment boundary required to perform movement association is invalid, the movement time association status is set to undecidable; the target to be judged identifier, the lining surface position corresponding to the target to be judged, the candidate moving target identifier, the candidate surface velocity vector, the centerline coincidence ratio, the movement direction difference, the movement width difference, and the movement time association status are combined to obtain the movement time association data.
[0109] Optionally, the illumination temporal correlation data is only used to indicate whether there is a temporal correspondence between the grayscale response of a candidate linear target and the background brightness change within the same scanning period, and the illumination temporal correlation is not determined based on the absolute grayscale value of the candidate linear target itself. When the background brightness is unavailable, the candidate response sequence or background brightness change sequence cannot form an effective deviation, or the illumination driving cycle cannot be obtained, the illumination temporal correlation is set to undeterminable, and the candidate linear target is retained for surface repetition support and movement time correlation processing. The undeterminability of the illumination temporal correlation is not equated with the absence of illumination interference.
[0110] Optionally, the movement time association is only performed on candidate linear targets that are eligible for repeated observation but whose surface repeatability support is not valid; candidate linear targets whose surface repeatability support is valid are not directly set to a time interference state because there are nearby moving targets. When there are multiple candidate moving targets, the centerline overlap ratio, movement direction difference, and movement width difference of each candidate moving target are calculated separately, and only candidate targets that meet the movement speed range, movement direction range, and corresponding movement matching determination conditions are determined as the corresponding moving targets; when a unique and valid movement correspondence cannot be formed, the movement time association is set to undeterminable.
[0111] The state of repeated observation eligibility, surface repeated support data, illumination time-related data, and movement time-related data is determined to obtain the pixel source state; the observed pixels are selected and compensated according to the pixel source state to obtain the interference separation image; crack identification and geometric calculation are performed on the interference separation image to obtain the crack detection result.
[0112] In a specific implementation, the method for obtaining crack detection results includes:
[0113] The surface repeat support status is read from the surface repeat support data, the illumination time association status is read from the illumination time association data, and the movement time association status is read from the movement time association data. When the surface repeat support status is valid and the illumination time association status is invalid, the pixel source status is set to the surface fixed status. When the surface repeat support status is valid and the illumination time association status is valid, the pixel source status is set to the illumination coincidence status. When the surface repeat support status is invalid, the repeat observation qualification is valid, and the illumination time association status or the movement time association status is valid, the pixel source status is set to the time interference status. When the surface repeat support status is undeterminable, the repeat observation qualification is invalid, the illumination time association status is undeterminable, or the above status conditions are not met, the pixel source status is set to the pending verification status to obtain the pixel source status.
[0114] For pixels in a fixed surface state, the actual observed pixel with the larger gradient energy is selected when both observed images are of acceptable quality; when the gradient energies are the same, the arithmetic mean of the normalized gray levels of the two actual observed pixels is used; and when only one observed image is of acceptable quality, the corresponding observed pixel is used. For pixels in a state of overlapping illumination, the background brightness at both observation times is obtained. , ,according to Compensate the first observed pixel to the brightness of the second observed background, or according to The second observation pixel is compensated to the brightness of the first observation background, and the observation pixel with a large gradient energy that is within the normalized grayscale range after compensation is selected; when the compensation value exceeds the normalized grayscale range, the compensation value is limited to the corresponding boundary and marked as a brightness compensation limited position; for pixels in the temporal interference state, when the repeated observation qualification is valid, the quality of the second observation image is qualified, and there is no response at the same surface position of the second observation that reaches the linear response judgment boundary, the actual pixel of the second observation is used to replace the first observation pixel; when the replacement condition cannot be met, it is marked as a temporal interference position where a replacement image cannot be obtained; background interpolation and pixel speculation are not performed on pixels in the state to be verified; the selected or compensated actual observation pixels are arranged according to the lining surface coordinates to obtain the interference separation image.
[0115] Optionally, when generating the interference separation image, only the actual observed pixels obtained in the first or second observation image and the brightness compensation value formed based on the brightness difference of the corresponding observation background are used, and spatial interpolation is not performed on the missing observed pixels based on adjacent crack pixels, adjacent background pixels, or crack model output; for the fixed surface state, actual observed pixels with repeated observation support are retained first; for the illumination coincidence state, only the background brightness reference compensation between two actual observation times is performed; for the temporal interference state, actual observed pixels of another time phase are used only when the corresponding surface position of another time phase has valid observation qualifications and there is no corresponding temporal interference response; positions that do not meet the above conditions are marked as pending verification; the pending verification marks participate in the subsequent crack detection result output, so that positions that have not formed reliable observation evidence are not converted into a definite non-crack conclusion.
[0116] An effective observation mask is established based on the sampling grid of the lining surface corresponding to the interference separation image; the mask value corresponding to the pixel with detection shielding mark or unable to obtain alternative image mark is set to zero, and the mask value corresponding to the pixel without detection shielding mark or unable to obtain alternative image mark is set to one, thus obtaining the effective observation mask; the effective observation mask is used to determine the crack segmentation processing area and construct the crack segmentation model input, and the effective observation mask is not modified according to the model pending verification state generated by the crack segmentation model output.
[0117] During crack segmentation model training or crack adaptation, record the crack segmentation model type, model input construction method, input items included in the training input, and the order of each input item; for machine learning models based on explicit image features, record pixel feature items and their order; for deep learning image segmentation models, record the model input channels, their order, the number of input rows, and the number of input columns; for visual segmentation models that complete crack adaptation based on pre-trained parameters, record image patches, cue items, the number of input rows, and the number of input columns; when the training input includes spatiotemporal auxiliary feature maps, set the first auxiliary input identifier, and train... When the input does not contain a spatiotemporal auxiliary feature map, a second auxiliary input identifier is set; after the crack segmentation model training or crack adaptation is completed, the sequential number of the existing crack segmentation model version identifier in the model library is read; if the sequential number does not exist in the model library, the first sequential number is created; if the sequential number exists in the model library, the maximum sequential number is increased by one number unit; the crack segmentation model type identifier and the obtained sequential number are combined to obtain the crack segmentation model version identifier; the crack segmentation model version identifier, crack segmentation model type, model input construction method, auxiliary input identifier, and model input structure are combined accordingly to obtain the crack segmentation model configuration data.
[0118] When performing crack segmentation, the crack segmentation model version identifier in the detection task data is read. This identifier is then compared item by item with the crack segmentation model version identifier in the crack segmentation model configuration data. Crack segmentation model configuration data with identical identifiers is used as the current model configuration data. If no crack segmentation model configuration data with identical identifiers is found, or if more than two crack segmentation model configuration data with identical identifiers are found, crack segmentation model inference is not performed, and pixels within the effective observation area are set as the model's verification location. The crack segmentation processing area is determined based on the model input construction method in the current model configuration data. The model input construction method includes candidate target guidance and effective observation area block division. For each candidate linear target, the centerline and the corresponding local widths at each position of the centerline are read from the candidate linear target data, and the minimum crack width is read from the detection task data. For each scanning unit traversed by the candidate linear target, the centerline position corresponding to that scanning unit is selected. The larger of the obtainable local width and the minimum crack width is used as the linear target exclusion width. If the local width is unobtainable, the minimum crack width is used as the linear target exclusion width.
[0119] Using the current centerline position as the center, extend the linear target exclusion width to both sides along the local centerline normal by half. The range between the two extended positions is taken as the linear target exclusion range. The local centerline normal is obtained by rotating the local centerline direction by 90 degrees. If the local centerline direction is unavailable, read the two centerline points in the candidate linear target centerline that are adjacent to but do not coincide with the current centerline position, and take the direction of the line connecting the two centerline points as the current centerline direction. If the aforementioned direction cannot be obtained, mark the background statistical band width at the current centerline position as unavailable.
[0120] On both sides of the exclusion range, a surface sampling grid interval is added outward along the center line normal to obtain the initial background statistical area; pixels whose linear response reaches the linear response judgment boundary, digital output boundary pixels, missing acquisition pixels, and exposure movement exceeding the limit are excluded from the background statistical area; when there are remaining pixels on both sides of the exclusion range, the median value of the normalized grayscale of all remaining pixels is calculated to obtain the initial background brightness median value.
[0121] Add one surface sampling grid interval to each side of the background statistical area again, and calculate the second background brightness median value according to the same pixel exclusion rule; take the absolute value of the difference between the second background brightness median value and the first background brightness median value as the background brightness change; read the grayscale repetition error from the aforementioned uniform calibration plate repeated acquisition results, and compare the background brightness change with the grayscale repetition error.
[0122] When the change in background brightness does not exceed the grayscale repetition error, stop expanding the background statistical area; when the change in background brightness exceeds the grayscale repetition error, continue to add one surface sampling grid interval to each side of the background statistical area, and recalculate the absolute difference between the current and previous median values of background brightness; each time it is expanded, only the actual observed pixels within the current effective image range are used, and fill values or inferred pixels outside the image boundary are not used in the median value calculation.
[0123] When the absolute value of the difference between two adjacent median values of background brightness does not exceed the grayscale repetition error, the normal surface distances from the outer boundary of the current background statistical area to the current center line position are recorded as the left background distance and the right background distance, and the larger of the two values is taken as the background statistical distance of the current center line position; when the expansion to the image boundary still does not reach the stopping condition, or there are no remaining effective pixels on either side, the background statistical distance of the current center line position is marked as unattainable.
[0124] For a candidate linear target, obtain the centerline positions of all its possible effective background statistical distances, and take the maximum value among the corresponding background statistical distances as the background statistical band width of the candidate linear target. When the background statistical distances at all centerline positions are unavailable, mark the background statistical band width as unavailable.
[0125] When using the candidate target guidance method, calculate the bounding rectangle that can enclose all pixels of the candidate linear target; read the longitudinal grid interval from the sampling grid of the lining surface. and circumferential grid spacing According to , Calculate the number of pixels expanded vertically and circumferentially, where, Indicates rounding up. and It is a dimensionless integer. , and All use the same unit of length.
[0126] Expand the rectangular circumscribed region of the candidate linear target on both sides longitudinally. Each pixel expands on both sides of the ring. 1 pixel is used to obtain an extended rectangular region; the extended rectangular region is restricted to the actual image boundary, and the corresponding effective observation mask within the region is retained to obtain a candidate segmentation region; the candidate segmentation region is used as the crack segmentation processing region.
[0127] When the background statistical band width of a candidate linear target is unavailable, the corresponding candidate linear target is marked as a target to be reviewed, and no candidate segmentation region is generated based on the candidate linear target; the corresponding candidate linear target identifier, lining surface position, and reason why the background statistical band width is unavailable are written into the review record.
[0128] When using the effective observation region segmentation method, pixels with a mask value of one are selected from the effective observation mask. Pixels connected within their eight neighborhoods are combined to form the effective observation region. The effective observation region is divided into overlapping image blocks according to the number of input rows and columns recorded in the current model configuration data. Each image block is then used as the crack segmentation processing region. Machine learning models based on explicit image features employ a candidate target guidance method. Visual segmentation models that complete crack adaptation based on pre-trained parameters also employ a candidate target guidance method. Deep learning image segmentation models employ either the candidate target guidance method or the effective observation region segmentation method recorded in the current model configuration data. When using a machine learning model based on explicit image features, pixel feature vectors calculated from the interference separation images and various associated data corresponding to the crack segmentation processing region are used as model input. When using a deep learning image segmentation model… When using a visual segmentation model that adapts to cracks, the image patch corresponding to the crack segmentation region and the effective observation mask are used as model inputs, or the image patch, the effective observation mask, and the spatiotemporal auxiliary feature map are used as model inputs. When using a visual segmentation model that adapts to cracks based on pre-trained parameters, the image patch corresponding to the candidate segmentation region, the crack positive cue point, the non-crack negative cue point, and the region cue are used as model inputs. The crack segmentation model output is the crack probability corresponding to each effective observation pixel within the crack segmentation region. For pixels with detection masking markers or for which alternative image markers cannot be obtained, the corresponding pixels are set as the model verification positions. When the model input structure corresponding to the crack segmentation region is inconsistent with the model input structure recorded in the current model configuration data, the pixels within the crack segmentation region are set as the model verification positions, and crack pixel states or non-crack pixel states are not formed based on the model output.
[0129] The method for obtaining effective observation region block input includes: selecting pixels with a mask value of one from the effective observation mask, and combining pixels connected within eight neighborhoods into an effective observation region; reading the number of input rows and columns specified by the input layer of the deep learning image segmentation model, and setting the image block size according to the number of input rows and columns; moving the image block along the row and column directions within the effective observation region according to a block step size smaller than the size of the image block in the corresponding direction, so that adjacent image blocks have overlapping areas; when the remaining range at the edge of the effective observation region is smaller than the size of the image block, moving the last image block into the effective observation region so that the outer boundary of the last image block coincides with the outer boundary of the effective observation region; when the size of the effective observation region in the corresponding direction is smaller than the size of the image block, mirroring and filling the missing pixel positions of the image block, and setting the effective observation mask value corresponding to the filled pixels to zero; combining the starting row, starting column, original row number, original column number, filling range, and image block identifier of the image block in the original lining surface sampling grid to obtain the image block position record.
[0130] The block step size is calculated based on the image block size and the image block overlap width. The method for setting the image block overlap width includes: sequentially selecting candidate overlap widths from positive integers smaller than the size of the image block in the corresponding direction; dividing the verification sample into blocks according to each candidate overlap width and inputting them into the deep learning image segmentation model; calculating the arithmetic mean of the crack probability of the same pixel in the overlapping area of adjacent image blocks to obtain the verification segmentation result corresponding to each candidate overlap width; calculating the number of pixels that are inconsistent with the pixel label of the verification segmentation result and the verification sample, and selecting the candidate overlap width with the fewest inconsistent pixels; when there are multiple candidate overlap widths with the same number of inconsistent pixels, selecting the candidate overlap width with the smaller value; subtracting the image block overlap width from the size of the image block in the corresponding direction to obtain the block step size in the corresponding direction.
[0131] The method for obtaining crack segmentation sample data includes: converting pixel positions in the interference separation image into lining surface positions based on the spatiotemporal correlation data of the scan; performing close-range inspection on the lining surface positions, marking pixels confirmed by close-range inspection to be located within the crack range as crack pixels, and marking pixels confirmed by close-range inspection to be located within the construction joint, edge of ancillary facilities, ordinary lining texture, water stains, and crack-free background range as non-crack pixels; dividing the interference separation images with completed pixel marking into training samples and validation samples according to the tunnel section, so that the first and second observation image pixels corresponding to the same lining surface position are only assigned to training and validation samples. One set of training samples or validation samples is selected, and both training and validation samples include crack pixels and non-crack pixels. If any set of tunnel segments lacks crack pixels or non-crack pixels after the initial segmentation, the tunnel segment with the fewest pixels containing the missing pixel category is selected from the other set and transferred to the corresponding set. The two sets are then checked again to see if they both contain crack pixels and non-crack pixels. The interference separation image, pixel labels, lining surface location, pixel source status, sample segment identifier, and sample purpose identifier are combined to obtain crack segmentation sample data. The sample purpose identifier is used to indicate whether the corresponding sample belongs to the training sample or the validation sample.
[0132] In one optional implementation, the crack segmentation model employs a machine learning model based on explicit image features. For each valid observation pixel within a candidate segmentation region, normalized darkness, linear response, and normalized gradient magnitude are calculated from the interference separation image. For valid observation pixels belonging to candidate linear targets, the local width and local centerline direction change corresponding to the candidate linear target are read. The local width is divided by the minimum crack width to obtain the local width ratio. The surface repeat support value, illumination time-related evaluation value, and movement time-related value of the corresponding pixel are read from the surface repeat support data, illumination time-related data, and movement time-related data. For valid observation pixels not belonging to candidate linear targets, the local width ratio, local centerline direction change, surface repeat support value, illumination time-related evaluation value, and movement time-related value are set to zero. The normalized darkness, linear response, normalized gradient magnitude, and local centerline direction change are calculated according to the following parameters: The pixel feature vector is obtained by sequentially arranging the values of local width ratio, local centerline direction change, surface repeatability support, illumination time-related evaluation value, and movement time-related value. If any required feature value cannot be obtained, the corresponding pixel is set as the model's verification position and is not input into the machine learning model. The feature items and feature order in the pixel feature vector are compared with the pixel feature items and pixel feature order recorded in the current model configuration data. If the comparison results are the same, the pixel feature vector that can obtain all feature values is input into the machine learning model based on explicit image features to obtain the crack probability of the corresponding pixel. If any comparison result is different, the corresponding pixel is set as the model's verification position and crack probability calculation is not performed. The machine learning model based on explicit image features adopts any one of the following: logistic regression model, random forest model, or gradient boosting tree model.
[0133] When using a logistic regression model, the pixel crack probability is calculated using the following formula: ;
[0134] in, Indicates the first The probability of a crack in a pixel; Indicates the first The th pixel feature vector Item features; Indicates the first Model coefficients corresponding to the features; Represents the model intercept; This represents the number of feature terms contained in the pixel feature vector; the crack pixel labels in the training samples are set to one, and the non-crack pixel labels are set to zero. The model coefficients and model intercept are obtained by reducing the cross-entropy loss between the predicted probability of the training samples and the pixel labels; when there are parameter combinations with the same cross-entropy loss, the parameter combination with the smaller sum of squared model coefficients is selected.
[0135] When using the random forest model, sampling with replacement is performed from the training samples to form multiple subsamples. A decision tree is built for each subsample. At each branch node of the decision tree, some features are selected from the pixel feature vector. Crack pixels and non-crack pixels in the current node are divided using candidate feature values. The feature and feature value that maximizes the decrease in Gini impurity after the division are selected as the node splitting condition. The number of crack pixels falling into the same leaf node is divided by the total number of pixels in the corresponding leaf node to obtain the crack probability output by a single decision tree. The arithmetic mean of the crack probabilities output by multiple decision trees is calculated, and the arithmetic mean is used as the pixel crack probability output by the random forest model.
[0136] When using the gradient boosting tree model, the initial classification value is set according to the proportion of crack pixels in the training samples; the negative gradient of the cross-entropy loss between the current model output and the pixel label is used as the fitting target for the current round according to the pixel feature vector, a regression decision tree is built, and the output of the regression decision tree is accumulated to the current classification value; the regression decision trees are built in the same way, and the cross-entropy loss is calculated using the validation samples after each round of training; the regression decision tree combination corresponding to the lowest cross-entropy loss of the validation samples is selected, the classification value output by the regression decision tree combination is converted into a value between zero and one through a logistic function, and the converted value is used as the pixel crack probability output by the gradient boosting tree model.
[0137] The method for obtaining a deep learning image segmentation model includes: reading training and validation samples from crack segmentation sample data; when the model input construction method is candidate target-guided, determining the candidate segmentation region according to the outer range of the candidate linear targets in the training samples, and extracting the corresponding training image blocks, training mask blocks, and pixel label blocks from the interference separation image, effective observation mask, and pixel label blocks; when the model input construction method is effective observation region segmentation, dividing the effective observation region in the training samples into overlapping training image blocks according to the number of input rows and columns specified by the deep learning image segmentation model input layer, and extracting the corresponding training mask blocks and pixel label blocks from the effective observation mask and pixel label blocks; when the number of rows or columns of the training image blocks is different from the number of input rows or columns specified by the model input layer, adjusting the size of the training image blocks using bilinear interpolation, and adjusting the size of the training mask blocks and pixel label blocks using nearest neighbor interpolation.
[0138] When the training input does not contain spatiotemporal auxiliary feature maps, the training image blocks and training mask blocks are arranged according to the model input channel order set during training to obtain the first training input. When the training input contains spatiotemporal auxiliary feature maps, the surface repetition support data, illumination time-related data, and movement time-related data in the training samples are converted into spatiotemporal auxiliary feature maps corresponding to the pixel positions of the training image blocks according to the lining surface coordinates. The training image blocks, training mask blocks, and spatiotemporal auxiliary feature maps are arranged according to the model input channel order set during training to obtain the second training input. The first training input and the second training input correspond to the second auxiliary input identifier and the first auxiliary input identifier, respectively. The training process and inference process of the same deep learning image segmentation model use the same auxiliary input identifier, input items, and channel arrangement order.
[0139] The first or second training input is provided to the deep learning image segmentation model to be trained to obtain the training crack probability corresponding to each pixel in the training image block. Only pixels with a mask value of one are selected from the training mask block, and the cross-entropy of the training crack probability is calculated according to the crack pixel label and non-crack pixel label in the pixel label block. The arithmetic mean of the cross-entropy of the selected pixels is used as the training loss. Backpropagation is used to calculate the gradient of the training loss with respect to the model parameters, and the model parameters are updated in the direction that reduces the training loss. After each round of model parameter update, the validation samples are provided to the updated deep learning image segmentation model using the same input construction method, and the validation loss is obtained according to the same calculation process. When the validation loss in this round is not less than the minimum validation loss obtained, the model parameter update is stopped, and the model parameters corresponding to the minimum validation loss are used as the deep learning image segmentation model parameters to obtain the deep learning image segmentation model.
[0140] If the first-round validation loss cannot be obtained, or if there are no pixels with a valid observation mask value of one in the validation samples, the deep learning image segmentation model version identifier is not generated, and the training sample inspection information is output; if the validation loss can be obtained and the model parameter selection is completed, the model input construction method, auxiliary input identifier, model input channel order, number of input rows and number of input columns are recorded, and combined with the deep learning image segmentation model version identifier, and written into the crack segmentation model configuration data.
[0141] In another optional implementation, the crack segmentation model employs a deep learning image segmentation model. When the model input construction method in the current model configuration data is candidate target-guided, image blocks are extracted from the interference separation image according to the pixel boundaries of the candidate segmentation region. The starting row, starting column, original row number, and original column number of the candidate segmentation region in the original lining surface sampling grid are combined with the corresponding image block identifiers. The filling range is recorded as an empty set to obtain the image block position record. When the model input construction method in the current model configuration data is effective observation region segmentation, the image block position record obtained during the effective observation region segmentation process is read and processed according to the image... The block position record extracts image blocks from the interference separation image and performs mirror filling on the filling range in the image block position record; according to the image block position record, the corresponding mask range is extracted from the effective observation mask, and the mask value corresponding to the filling range is set to zero to obtain the effective observation mask of the image block; when the number of rows or columns of the image block is different from the number of input rows or columns of the current model configuration data record, the image block is resized using bilinear interpolation, and the effective observation mask of the image block is resized using nearest neighbor interpolation; the resized image block is used as the image channel, and the resized effective observation mask of the image block is used as the effective observation mask channel.
[0142] When the auxiliary input identifier in the current model configuration data is the second auxiliary input identifier, the image channel and the effective observation mask channel are arranged according to the model input channel order recorded in the current model configuration data to obtain the deep learning model input; when the auxiliary input identifier in the current model configuration data is the first auxiliary input identifier, the lining surface coordinates corresponding to the surface repeating support data, illumination time-related data, and movement time-related data are converted into surface sampling grid row and column numbers according to the scan spatiotemporal correlation data, and the surface repeating support value, illumination time-related evaluation value, and movement time-related value are written into the corresponding row and column positions to obtain the surface repeating support feature map, illumination time-related feature map, and movement time-related feature map; when the corresponding correlation data can be obtained, the corresponding effective correlation value is set to one, and the corresponding key... When the associated data is unavailable, the associated evaluation value and the associated valid value at the corresponding position are both set to zero. Feature ranges corresponding to the image blocks are extracted from each feature map according to the image block position records. The illumination time-related feature map and the movement time-related feature map are resized using the same bilinear interpolation position relationship as the image channel. The surface repeating support feature map and the associated valid value are resized using the same nearest neighbor interpolation position relationship as the effective observation mask channel. The resized surface repeating support feature map, illumination time-related feature map, movement time-related feature map, and associated valid value are combined to form a spatiotemporal auxiliary feature map. The image channel, the effective observation mask channel, and the spatiotemporal auxiliary feature map are arranged according to the model input channel order recorded in the current model configuration data to obtain the deep learning model input.
[0143] The deep learning model input channels, number of channels, channel order, number of input rows, and number of input columns are compared item by item with the model input structure in the current model configuration data. If all comparison results are the same, the deep learning model input is provided to the deep learning image segmentation model corresponding to the current model configuration data to obtain the crack probability map. If any comparison result is different, crack probability calculation is not performed, and the pixels in the corresponding crack segmentation processing area are set as the model to be reviewed positions.
[0144] Deep learning image segmentation models can employ encoder-decoder convolutional segmentation models with skip connections, dilated convolutional image segmentation models, or self-attention structures. When an interference-separated image is divided into multiple overlapping image blocks, each image block is input into the deep learning image segmentation model. The arithmetic mean of the crack probabilities obtained at the same pixel location in different image blocks is calculated, and the arithmetic mean is used as the crack probability of the corresponding pixel.
[0145] Before using the pre-trained visual segmentation model for crack segmentation, the parameters of the image encoder, cue encoder, and mask decoder of the pre-trained visual segmentation model are read. For the training samples in the crack segmentation sample data, for the crack pixel regions in the training samples that have been confirmed by close inspection, the center line of the candidate linear target corresponding to the crack pixel region is extracted, and pixels located within the confirmed crack region are selected from the center line as crack positive cue points. Non-crack negative cue points are selected from the background positions manually marked as non-crack. The positive and negative cue points and the corresponding manually marked crack regions are used together for crack adaptation of the pre-trained visual segmentation model. Non-crack negative cue points are generated according to the background positions on both sides of the candidate linear target, and region cue is generated according to the outer range of the candidate linear target. The training sample image, crack positive cue points, non-crack negative cue points, and region cue are input into the pre-trained visual segmentation model to obtain the candidate segmentation mask corresponding to the training sample. This is only for cases where there is no detection masking mark and no alternative image can be obtained. For pixels marked as identifiers and those to be verified by the model, calculate the pixel classification loss and region overlap loss between the candidate segmentation mask and the pixel marker, and update the mask decoder parameters based on the sum of the two losses. After each round of parameter updates, calculate the validation loss using validation samples, select the model parameters corresponding to the lowest validation loss, and obtain a visual segmentation model that completes crack adaptation based on pre-trained parameters. Read the output type of the visual segmentation model. When the output type of the visual segmentation model indicates that the output is an unnormalized mask value, convert the unnormalized mask value to a value between zero and one using a logical function. When the output type of the visual segmentation model indicates that the output is a probability between zero and one, directly read the model output value. Use the converted or read value as the crack probability of the corresponding pixel. The output type of the visual segmentation model is set according to whether the output layer of the visual segmentation model contains probability normalization operations. When the output layer contains probability normalization operations, it is set to probability output type; when the output layer does not contain probability normalization operations, it is set to unnormalized output type.
[0146] In another optional implementation, the crack segmentation model employs a visual segmentation model that adapts to cracks based on pre-trained parameters; image patches containing the candidate linear target and the background regions on both sides of the candidate linear target are extracted from the interference separation image; the two endpoints of the centerline of the candidate linear target and the midpoint determined according to the cumulative length of the centerline are used as positive crack cue points; the pixel positions on both sides of the candidate linear target that do not belong to the candidate linear target for the first time are searched along the normal direction of the centerline of the candidate linear target, and the searched pixel positions are used as negative non-crack cue points; a rectangular region that can enclose the candidate linear target and the negative cue points on both sides is calculated. The model is configured to input images, crack positive points, non-crack negative points, and region prompts, as well as the number of input rows and columns of the images, into the model input structure in the current model configuration data. If the comparison results are the same, the images, crack positive points, non-crack negative points, and region prompts are input into a visual segmentation model that adapts to cracks based on pre-trained parameters to obtain candidate segmentation masks and corresponding pixel crack probabilities. If any comparison result is different, the pixels within the candidate segmentation region are set as the model's verification positions, and candidate segmentation mask calculation is not performed.
[0147] When a visual segmentation model that performs crack adaptation based on pre-trained parameters outputs multiple candidate segmentation masks, the number of candidate linear target centerline pixels and the number of non-crack negative cue points covered by each candidate segmentation mask are counted. Candidate segmentation masks that cover non-crack negative cue points are deleted. The number of candidate linear target centerline pixels covered by the remaining candidate segmentation masks is divided by the total number of candidate linear target centerline pixels to obtain the centerline coverage ratio. The candidate segmentation mask with the largest centerline coverage ratio is selected, and the crack probability of the selected candidate segmentation mask is used as the crack probability of the corresponding image block. If there is no candidate segmentation mask that does not cover non-crack negative cue points, the corresponding candidate linear target is set as the position to be verified by the model.
[0148] The crack probability output by the crack segmentation model is mapped to the original lining surface sampling grid. For machine learning models based on explicit image features, the row and column numbers of the surface sampling grid corresponding to the pixel feature vector are used as the corresponding crack probability positions. For image blocks without size adjustment, the row and column positions of the model output pixels are added to the original row and column positions in the image block position record to obtain the surface sampling grid positions corresponding to the model output pixels. For image blocks with size adjustment, the crack probability map output by the model is reverse-adjusted to the original row and column numbers of the image block according to the ratio of the number of rows and columns before and after size adjustment, and the crack probability corresponding to the original row and column positions is calculated using bilinear interpolation. The model output corresponding to the filling range in the image block position record is deleted. When the same surface sampling grid position corresponds to multiple image block outputs, the arithmetic mean of the multiple crack probabilities is calculated, and the arithmetic mean is used as the pixel crack probability of the corresponding surface sampling grid position. The surface sampling grid row and column numbers, lining surface coordinates, pixel crack probabilities, crack segmentation model version identifier, and image block identifier are combined to obtain the original grid crack probability data.
[0149] Input the validation samples into the selected crack segmentation model to obtain the crack probability of each pixel in the validation samples; obtain the minimum crack probability from pixels manually marked as cracks and the maximum crack probability from pixels manually marked as non-cracks; when the maximum crack probability of a non-crack pixel is less than the minimum crack probability of a crack pixel, the midpoint between the two values is used as the single crack probability judgment boundary; when the maximum crack probability of a non-crack pixel is not less than the minimum crack probability of a crack pixel, the maximum crack probability of a non-crack pixel is used as the lower boundary for crack confirmation, the minimum crack probability of a crack pixel is used as the upper boundary for non-crack confirmation, and the range between the two boundaries is used as the model to be reviewed range; combine the crack segmentation model type, crack segmentation model version identifier, judgment mode, corresponding boundary, training sample identifier, and validation sample identifier to obtain crack probability judgment data.
[0150] The method for obtaining model deployment records includes: when the detection task sets only one crack segmentation model, setting the corresponding crack segmentation model version identifier to the currently enabled crack segmentation model version identifier; when the detection task sets two or more crack segmentation models, determining the segmentation state of the verification sample using the same verification sample and the crack probability judgment data corresponding to each crack segmentation model; setting pixels in the verification sample with a valid observation mask value of one but without obtaining pixel crack probability as models awaiting verification; counting the number of pixels with crack pixel state or non-crack pixel state whose segmentation state is inconsistent with the manual pixel label, and separately counting the number of pixels in the models awaiting verification state, adding the two numbers to obtain the model selection evaluation quantity; selecting the crack segmentation model version identifier with the smallest model selection evaluation quantity; when the model selection evaluation quantity is the same, recording the time elapsed from inputting the same verification sample to outputting all pixel crack probabilities on the same image processing device, selecting the crack segmentation model version identifier with the shortest elapsed time; combining the selected crack segmentation model version identifier with the currently enabled identifier to obtain the model deployment record.
[0151] The method for obtaining crack segmentation data includes: querying the original mesh crack probability data one by one according to the row and column numbers of the surface sampling mesh corresponding to the crack segmentation processing area, and checking whether the corresponding pixel has a valid pixel crack probability; when the corresponding pixel does not obtain a pixel crack probability, or the obtained pixel crack probability is not a valid real number in the range of zero to one, the segmentation status of the corresponding pixel is set to the model pending verification state, and the comparison between the crack probability and the crack probability judgment boundary is not performed; for pixels that obtain a valid pixel crack probability, it is checked whether they have a missing acquisition location marker, an exposure movement exceeding the limit location marker, a brightness compensation limited location marker, a detection shielding marker, and an inability to obtain a replacement image marker; When any of the aforementioned location markers are present, the corresponding pixel is set to the model pending verification state. When no such location marker is present, pixels with crack probabilities reaching the single crack probability determination boundary and whose source state is either a fixed surface state or a lighting overlap state are set to the crack pixel state; pixels with crack probabilities below the single crack probability determination boundary are set to the non-crack pixel state; pixels with crack probabilities reaching the single crack probability determination boundary but whose source state is neither a fixed surface state nor a lighting overlap state are set to the model pending verification state. When using the crack confirmation lower boundary and the non-crack confirmation upper boundary, the pixel crack probability validity and location marker of each pixel are checked in the same way, and pixels that do not obtain valid results are removed. Pixels with crack probability or any of the aforementioned location identifiers are set to the model's pending verification state. For pixels with valid crack probabilities but no aforementioned location identifiers, pixels with crack probabilities strictly greater than the crack confirmation lower boundary and whose pixel source state is a fixed surface state or a lighting coincidence state are set to the crack pixel state; pixels with crack probabilities strictly less than the non-crack confirmation upper boundary are set to the non-crack pixel state; pixels with crack probabilities greater than or equal to the non-crack confirmation upper boundary and less than or equal to the crack confirmation lower boundary, and pixels with crack probabilities strictly greater than the crack confirmation lower boundary but whose pixel source state is not a fixed surface state or a lighting coincidence state are set to the model's pending verification state. After determining the segmentation state, the lining surface coordinates of each pixel are obtained one by one according to the row and column numbers of the surface sampling grid corresponding to the crack segmentation processing area, and the original grid crack probability data is queried to obtain the corresponding pixel crack probability. For pixels for which no valid pixel crack probability is obtained, the corresponding pixel crack probability is recorded as unobtainable, and the model is kept in a state of pending verification. The surface sampling grid row and column numbers, lining surface coordinates, pixel crack probability, segmentation state, pixel source state, observation source, and crack segmentation model version identifier of each pixel in the crack segmentation processing area are combined according to the pixel position to obtain crack segmentation data. The segmentation state includes crack pixel state, non-crack pixel state, and model pending verification state.
[0152] Crack regions are formed by connecting crack pixels according to their eight neighborhoods; the crack centerline is obtained using a centerline thinning method; adjacent centerline pixels are established as undirected graph edges; the edge length is calculated based on the surface coordinates at both ends of the edge; the length of each undirected graph edge is accumulated to obtain the crack length; a normal measurement line is established at the crack centerline location along the local centerline direction; adjacent positions where the crack pixel state changes to a non-crack pixel state are searched along both sides of the normal; when using a single crack probability determination boundary, the adjacent positions are set as... , The corresponding crack probability is , ,according to Calculate the crack edge location; when using two confirmation boundaries, the midpoint of the surface coordinates of two adjacent locations is taken as the crack edge location only when there is no model verification location between adjacent crack pixels and non-crack pixels, and the surface distance between adjacent locations does not exceed the allowable error of crack geometry; when the crack edge cannot be obtained on either side, the corresponding local width is marked as unobtainable; the surface distance between the crack edges on both sides is taken as the local crack width; compare all obtainable local crack widths, and take the local crack width with the largest value as the maximum crack width.
[0153] Calculate the covariance matrix of all surface coordinates along the crack centerline; take the angle between the eigenvector corresponding to the larger eigenvalue and the longitudinal surface coordinate axis as the crack direction; take the minimum, maximum, and median values of the circumferential surface coordinates along the crack centerline as the circumferential position of the crack; when the detection direction is the same as the design mileage increasing direction, add the mileage reference to the longitudinal surface coordinates of the crack centerline to obtain the crack mileage; when the detection direction is opposite to the design mileage increasing direction, subtract the longitudinal surface coordinates of the crack centerline from the mileage reference to obtain the crack mileage; combine the crack identifier, crack surface coordinate range, crack length, local crack width, maximum crack width, crack direction, circumferential position, crack mileage, pixel source status, and observation source to obtain the crack detection result.
[0154] The system collects locations with incomplete scan records, missing acquisition locations, locations where exposure movement exceeds limits, locations with invalid repeated observation qualifications, locations with temporal interference where no alternative image can be obtained, and locations where the model needs to be verified. Locations with the same cause and adjacent at edges or corners in the surface sampling grid are connected to form supplementary detection areas. When the percentage of the highest output pixels exceeds the corresponding upper boundary, the exposure time is adjusted to an adjacent candidate value of the next lower exposure time adjustment resolution; when the exposure time is already at the lower limit, the illumination output is adjusted to an adjacent candidate value of the next lower illumination output adjustment resolution. When the percentage of the lowest output pixels exceeds the corresponding upper boundary, the exposure time is adjusted to an adjacent candidate value of the next higher exposure time adjustment resolution and re-checked. When the exposure movement distance exceeds the limit, the illumination output is adjusted to an adjacent candidate value with a higher illumination output adjustment resolution. When local grayscale noise exceeds the limit or there are missing acquisition locations, the detection speed is adjusted to an adjacent candidate speed with a lower speed adjustment resolution that meets the scanning operation configuration conditions. When it is necessary to change the repeated observation phase, the candidate speed with the smallest speed difference and observation phase separation data greater than the current value is selected from the qualified candidate speeds. When no candidate parameters meet the conditions, on-site close-range verification requirements are generated. Supplementary detection task data is obtained by combining the supplementary detection area, the reason for not forming an automatic conclusion, the adjusted acquisition parameters, or the on-site close-range verification requirements.
[0155] In one optional implementation, after completing a tunnel inspection, the location to be reviewed is obtained from the pixel source status and crack probability determination results, and adjacent locations to be reviewed are merged into a supplementary observation area according to the lining surface coordinates. For supplementary observation areas formed due to insufficient image quality, the detection speed is reduced or the longitudinal displacement trigger interval is shortened. For supplementary observation areas formed due to undeterminable illumination time correlation, the detection speed or illumination drive phase is adjusted so that the observation time difference corresponding to the re-observation is different from the previous detection. For supplementary observation areas formed due to undeterminable movement time correlation, the detection speed is changed while maintaining the surface sampling interval to change the observation time difference between the two observations at the same lining surface location. The supplementary observation area is re-scanned according to the adjusted scan operation configuration data, and the re-acquired scan record is matched with the original detection result according to the lining surface coordinates. If the supplementary observation still cannot form a valid determination, the status to be reviewed is retained, and a non-crack conclusion is not automatically generated.
[0156] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. An intelligent detection method for tunnel lining cracks based on image analysis, characterized in that, include: Acquire detection task data and perform tunnel scanning according to the detection task data to obtain scan records; Acquire calibration data and perform spatiotemporal conversion on the scan records based on the calibration data to obtain scan spatiotemporal correlation data; The scan records are rearranged according to the scan spatiotemporal correlation data to obtain dual-temporal observation images; the position correspondence and image quality of the dual-temporal observation images and scan records are determined to obtain the qualification for repeated observation; candidate linear targets are detected in the dual-temporal observation images to obtain candidate linear targets; the scan formation time corresponding to the candidate linear targets is queried according to the scan spatiotemporal correlation data to obtain the formation time of the candidate targets. By comparing bi-temporal candidate linear targets at locations eligible for repeated observation, surface repeatability support data is obtained. The correlation between the candidate target formation time and the background brightness change, and the correlation between the candidate target formation time and the target position change are calculated respectively to obtain illumination time correlation data and movement time correlation data; The state of pixel source is determined by analyzing repeated observation eligibility, surface repeated support data, illumination time-related data, and movement time-related data. Select and compensate the observed pixels according to the pixel source state to obtain an interference separation image; Crack detection results are obtained by performing crack identification and geometric calculations on the interference separation image.
2. The intelligent detection method for tunnel lining cracks based on image analysis according to claim 1, characterized in that, The method for obtaining scan records includes: controlling the movement of the detection vehicle, driving the lighting device, and exposing the image acquisition device according to the detection task data; synchronously recording the pixel value, exposure start and end time, vehicle movement displacement, vehicle posture, and imaging distance of each scanning unit; and combining the pixel value, exposure start and end time, vehicle movement displacement, vehicle posture, and imaging distance of each scanning unit with the scanning unit identifier to obtain the scan record.
3. The intelligent detection method for tunnel lining cracks based on image analysis according to claim 1, characterized in that, The method for obtaining the spatiotemporal correlation data of scanning includes: calculating the midpoint between the start and end times of the exposure of the scanning unit, reading the travel displacement, vehicle attitude and imaging distance corresponding to the midpoint time; performing coordinate transformation on the pixel imaging rays of the scanning unit using calibration acquisition data, determining the spatial position of the lining surface on the pixel imaging rays according to the imaging distance corresponding to the midpoint time, and converting the spatial position of the lining surface into lining surface coordinates; and combining the lining surface coordinates and the midpoint time with the scanning unit identifier to obtain the spatiotemporal correlation data of scanning.
4. The intelligent detection method for tunnel lining cracks based on image analysis according to claim 1, characterized in that, The method for obtaining dual-temporal observation images includes: rearranging the scanning units in the scanning record to the surface sampling grid according to the lining surface coordinates in the scanning spatiotemporal correlation data; combining the scanning units obtained at the first observation position into a first observation image, and combining the scanning units obtained at the second observation position into a second observation image, thereby obtaining a dual-temporal observation image including the first observation image and the second observation image.
5. The intelligent detection method for tunnel lining cracks based on image analysis according to claim 1, characterized in that, The method for obtaining the qualification for repeated observation includes: establishing a correspondence between the positions of the same lining surface in the first observation image and the second observation image according to the coordinates of the lining surface; when the scanning record of the corresponding position of the second observation is complete, does not belong to the position of missing acquisition or position of exposure movement exceeding the limit, the image quality judgment boundary is valid and the quality of the second observation image is qualified, the corresponding lining surface position is marked as having the qualification for repeated observation, and the qualification for repeated observation is obtained.
6. The intelligent detection method for tunnel lining cracks based on image analysis according to claim 1, characterized in that, The method for obtaining candidate linear targets and candidate target formation time includes: for each pixel in the dual-phase observation image, obtaining the grayscale of the current pixel and the two sides of the normal direction according to the candidate direction and normal sampling distance; calculating the linear response based on the grayscale of the current pixel and the two sides of the normal direction; taking the maximum value of the linear response corresponding to different candidate directions and normal sampling distances as the pixel linear response value, and obtaining the corresponding candidate direction; connecting and combining pixels whose pixel linear response values reach the linear response judgment boundary and are interconnected in the eight-neighborhood to obtain candidate linear targets; querying the scan formation time in the scan spatiotemporal correlation data according to the pixel position of the candidate linear target to obtain the candidate target formation time.
7. The intelligent detection method for tunnel lining cracks based on image analysis according to claim 6, characterized in that, The method for obtaining surface repetition support data includes: searching for candidate linear targets at the centerline position of the candidate linear target in the first observation image and at the corresponding lining surface position in the second observation image, and calculating the surface position difference; calculating the direction difference when both observation results can obtain the local centerline direction, and calculating the width difference when both can obtain the local width; setting surface repetition support as valid when the repetition observation qualification is valid, the surface position difference does not exceed the position search radius, at least one of the direction difference and width difference can be obtained and all the differences that can be obtained do not exceed the corresponding decision boundary, and the signed normal response signs of the two observation results are the same; setting surface repetition support as invalid when the repetition observation qualification is valid but there is no linear response reaching the linear response decision boundary within the corresponding search range; setting surface repetition support as undecidable when the repetition observation qualification is invalid or neither the direction difference nor the width difference can be obtained, thereby obtaining surface repetition support data.
8. The intelligent detection method for tunnel lining cracks based on image analysis according to claim 1, characterized in that, The method for obtaining lighting time-related data includes: selecting background areas on both sides of the candidate linear target, obtaining background brightness change sequences and candidate response sequences according to the formation time of the candidate target; calculating the normalized correlation coefficient and sign correspondence of the background brightness change sequence and candidate response sequence, obtaining the lighting time-related evaluation value based on the normalized correlation coefficient and sign correspondence, and comparing the lighting time-related evaluation value with the lighting interference judgment boundary to obtain lighting time-related data.
9. The intelligent detection method for tunnel lining cracks based on image analysis according to claim 1, characterized in that, The method for obtaining time-of-motion correlation data includes: identifying a target to be judged from candidate linear targets whose repeated observation eligibility is valid and whose surface repeated support is not valid; obtaining candidate surface velocity vectors based on the centerline surface coordinates of the target to be judged in different scanning units and the scanning formation time; determining a second observation movement search area based on the candidate surface velocity vectors; identifying candidate moving targets from the second observation movement search area; calculating the centerline overlap ratio, movement direction difference, and movement width difference between the candidate moving targets and the target to be judged; and comparing these with the corresponding judgment boundary to obtain time-of-motion correlation data.
10. The intelligent detection method for tunnel lining cracks based on image analysis according to claim 1, characterized in that, The method for obtaining the pixel source state includes: marking pixels with valid repeated observation eligibility, valid surface repeated support, and invalid illumination temporal association as surface fixed state; marking pixels with valid surface repeated support and valid illumination temporal association as illumination coincident state; marking pixels with valid repeated observation eligibility, invalid surface repeated support, and valid illumination temporal association or movement temporal association as temporal interference state; and marking pixels with invalid repeated observation eligibility, undeterminable surface repeated support, undeterminable illumination temporal association, or that do not meet the surface fixed state, illumination coincident state, and temporal interference state as pending verification state, thereby obtaining the pixel source state including surface fixed state, illumination coincident state, temporal interference state, and pending verification state.
11. The intelligent detection method for tunnel lining cracks based on image analysis according to claim 10, characterized in that, The method for obtaining the interference separation image includes: using two consecutive observation images from a dual-temporal observation image as the first observation image and the second observation image, respectively; selecting observation pixels with acceptable image quality for pixels in the fixed surface state; compensating for pixels in the illumination overlap state according to the background brightness difference between the two observations and selecting the compensated observation pixels; for pixels in the temporal interference state, if the repeated observation qualification is valid, the quality of the second observation image is acceptable, and the linear response at the corresponding surface position is lower than the linear response judgment boundary used for candidate linear target detection, replacing the first observation pixel with the second observation pixel; otherwise, setting a sign that a replacement image cannot be obtained; setting a detection masking sign for pixels in the state to be reviewed, thereby obtaining the interference separation image.
12. The intelligent detection method for tunnel lining cracks based on image analysis according to claim 1, characterized in that, The method for obtaining crack detection results includes: constructing a crack segmentation model input from an interference separation image; providing the crack segmentation model input to the crack segmentation model to obtain pixel crack probabilities, wherein the crack segmentation model is a machine learning model based on explicit image features, a deep learning image segmentation model, or a visual segmentation model that completes crack adaptation based on pre-trained parameters; setting boundaries for pixel crack probabilities using validation samples marked by close-range inspection to obtain crack probability determination data; determining the state of pixel crack probabilities according to the crack probability determination data to obtain crack segmentation data; performing connectivity processing, centerline processing, and surface coordinate conversion on the crack segmentation data, calculating crack length, width, and extension direction, and combining them with crack location to obtain crack detection results.
13. The intelligent detection method for tunnel lining cracks based on image analysis according to claim 12, characterized in that, The method for obtaining the input to the crack segmentation model includes: checking pixel-by-pixel whether the interference separation image contains selected or compensated observation pixels; setting the mask value corresponding to the position containing the observation pixel to one and the mask value corresponding to the position not containing the observation pixel to zero to obtain the effective observation mask; reading the model configuration data corresponding to the selected crack segmentation model, which records the model input construction method and the auxiliary input identifier set according to whether the model training input contains the spatiotemporal auxiliary feature map; according to the model input construction method, the candidate segmentation region or the overlapping image block divided by the effective observation region is used as the crack segmentation processing region; the auxiliary input identifier indicates that when using the spatiotemporal auxiliary feature map, the surface repetition support data, illumination time-related data, and movement time-related data are converted into the spatiotemporal auxiliary feature map; and arranging the image channels, effective observation mask channels, and the spatiotemporal auxiliary feature map to be used according to the channel order recorded in the model configuration data to obtain the crack segmentation model input.
14. The intelligent detection method for tunnel lining cracks based on image analysis according to claim 12, characterized in that, The method for obtaining crack segmentation data includes: adjusting the crack probability map output by the crack segmentation model back to the original lining surface sampling grid according to the image block position record and size adjustment ratio; deleting the model output corresponding to the filling range and calculating the arithmetic mean of the crack probabilities corresponding to the overlapping image blocks; determining the segmentation state according to the crack probability judgment data, and combining the lining surface coordinates, pixel crack probability, segmentation state, pixel source state, observation source and crack segmentation model version identifier according to the pixel position to obtain crack segmentation data.
15. The intelligent detection method for tunnel lining cracks based on image analysis according to claim 10, characterized in that, The method further includes: reading the state pixels to be reviewed from the pixel source state, and reading the incomplete scan record positions, missing acquisition positions, exposure movement exceeding limits positions, and invalid repeated observation positions from the scan record and scan spatiotemporal correlation data; converting the state pixels to be reviewed and the scan unit positions into lining surface positions according to the scan spatiotemporal correlation data, connecting and merging adjacent lining surface positions, and configuring supplementary observation parameters according to the reasons for not forming detection conclusions to obtain supplementary detection task data.
Citation Information
Patent Citations
Subway tunnel lining crack water leakage disease rapid detection device and detection method thereof
CN114113118A
