Image recognition-based deep foundation pit crack feature coding detection method

Abstract

The application discloses a kind of deep foundation pit crack feature coding detection methods based on image recognition, to solve the problem that two-dimensional image is difficult to accurately characterize three-dimensional crack, opening width and the estimation of height of fault block is not accurate, the application is collected by short time sequence and multi-polar imaging calibration, crack is extracted center line and is divided into left and right side domain, local reconstruction is carried out using neural radiation field or three-dimensional Gaussian sputtering in mask neighborhood, left and right implicit surface is established and fused polar solution algorithm line, gap area is applied to occupy priori and cross-slit repulsive potential, combined with reference object, the absolute scale is obtained, millimeter level opening and fault block are measured along the center line section, and feature code is generated, threshold and time sequence comparison determination is carried out, the technical effects of high-precision three-dimensional measurement, reliable risk assessment and traceable management to deep foundation pit crack are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of geotechnical engineering safety monitoring, and particularly relates to a deep foundation pit crack feature coding detection method based on image recognition. BACKGROUND

[0002] In deep foundation pit engineering, the supporting structure (such as underground continuous wall, row pile and sprayed concrete surface layer) is prone to cracks due to excavation disturbance, environmental load and material shrinkage.

[0003] Engineering practice and related specifications usually use quantitative indicators such as crack opening width and bench height as important basis for risk assessment, so there is an urgent need for crack detection and measurement means with millimeter-level precision, repeatability and traceability.

[0004] In recent years, image visual technology has developed rapidly: crack segmentation and skeleton extraction methods based on deep learning have been used for automatic identification; multi-view geometry and dense reconstruction can recover three-dimensional topography from short-time sequence images; new local reconstruction methods such as neural radiance field or three-dimensional Gaussian splatting show good adaptability on low-cost devices; research on polarization imaging for estimating surface normal and material properties has also gradually increased.

[0005] However, the prior art still has deficiencies in reliable acquisition and engineering expression of key three-dimensional features of cracks:

[0006] 1. Two-dimensional representation and scale uncertainty make it difficult to accurately estimate key quantities: crack width calculation based on image plane is significantly affected by viewing angle, distance and imaging distortion, making it difficult to recover the true physical width, bench height is a three-dimensional quantity that is difficult to obtain by simply two-dimensional methods, and multi-view geometry has absolute scale uncertainty when there is no reference, affecting millimeter-level measurement;

[0007] 2. Lack of structural prior in three-dimensional reconstruction of crack neighborhood: general photometric consistency driven reconstruction is prone to "bridging" across cracks and over-smoothing on slender, texture-poor or reflective surfaces, making it difficult to maintain the gap between cracks empty, existing methods rarely explicitly decouple the two sides of the crack, and lack of occupancy prior and cross-crack repulsion constraints in the gap area, resulting in systematic underestimation of opening width and bench height or instability;

[0008] 3. Observation fusion and measurement coding chain is not perfect: polarization normal is usually used independently, less combined with implicit geometry to improve the consistency of normal and shape, cross-sectional measurement and statistics along the center line of the crack lack a unified process, lack of reconstruction confidence based on reprojection residual and effectiveness of acquisition conditions, and lack of uniformity in historical time series comparison and feature coding, affecting the comparability and traceability of risk determination.

[0009] Therefore, a deep foundation pit crack feature coding detection method capable of solving the above-mentioned problems of the prior art is needed to be solved by those skilled in the art. SUMMARY

[0010] An object of the present application is to provide a deep foundation pit crack feature coding detection method based on image recognition. In view of the problems that the existing two-dimensional image is difficult to accurately estimate the crack opening width and the step height, the absolute scale is uncertain, and the reconstruction is easy to bridge across the crack, a technical scheme is proposed, which comprises short-time sequence and multi-polarization imaging acquisition and calibration, crack segmentation to obtain a center line and divide left and right side domains, local reconstruction of left and right implicit surfaces by using neural radiation field or three-dimensional Gaussian sputtering in the mask neighborhood and fusion of polarization normal, joint optimization by applying occupancy prior and cross-crack repulsive potential in the gap area, identification of reference objects to obtain absolute scale and correction, generation of cross-sectional sequences along the center line to calculate millimeter-level opening and step height and form feature coding and threshold and time sequence determination. The present application has the technical effects of realizing high-precision three-dimensional measurement, effectively inhibiting cross-crack misreconstruction, improving scale consistency and confidence evaluation, and supporting traceable risk determination.

[0011] According to the deep foundation pit crack feature coding detection method based on image recognition, the method comprises the following steps:

[0012] S1, a short-time sequence image sequence and a polarization image group with different polarization angles of a deep foundation pit supporting structure surface are acquired, and camera internal and camera external parameter calibration is completed;

[0013] S2, the image sequence is taken as input to generate a crack mask by crack semantic segmentation, and a crack center line and left and right side domains thereof are determined and output;

[0014] S3, the crack mask, the crack center line, the left side domain, the right side domain, the image sequence, the camera internal and external parameters, and the polarization image group are inputted, local three-dimensional reconstruction is performed in the crack mask neighborhood, left and right implicit surfaces are established, and a normal field and a re-projection residual statistical value are obtained by joint optimization combined with polarization information;

[0015] S4, the left and right implicit surfaces, the image sequence, and the camera internal and external parameters are taken as input, a reference object with a known size is identified, and an absolute scale parameter is obtained therefrom, scale consistency correction is performed on the two implicit surfaces, and the absolute scale parameter and the left and right implicit surfaces after scale correction are outputted;

[0016] S5, the crack center line, the left and right implicit surfaces after scale correction, the normal field, and the absolute scale parameter are taken as input, cross-sectional sequences are generated along the center line, and opening width and step height are calculated based on the two implicit surfaces and the normal field at each cross section;

[0017] S6, generating crack feature codes according to engineering rules with the cross-section sequence, the opening width and the crack height value of each cross-section, the crack center line and the re-projection residual statistical value as inputs;

[0018] S7, inputting the crack feature codes and calling historical codes, judging and generating detection conclusions according to threshold and time sequence comparison.

[0019] Optionally, step S1 is specifically:

[0020] At least two frames of images with short time sequence attribute are continuously collected in the same field of view on the surface of the deep foundation pit support structure to form an image sequence, and corresponding images are collected to form a polarized image group under the setting of polarization transmission axes of at least two different polarization angles, and the polarization angle parameters of each polarized image are recorded;

[0021] Camera calibration is completed to obtain camera intrinsic parameters including focal length, principal point coordinates, radial distortion coefficient and tangential distortion coefficient, and camera extrinsic parameters including rotation and translation are estimated for each frame in the image sequence;

[0022] The image sequence and the polarized image group are subjected to distortion correction and resolution unification, so as to match the imaging model consistent with the camera intrinsic parameters and the camera extrinsic parameters;

[0023] The image sequence, the polarized image group, the camera intrinsic parameters and the camera extrinsic parameters are output.

[0024] Terminology interpretation:

[0025] The surface of the deep foundation pit support structure is the exposed surface of the deep foundation pit engineering support member, which is used for crack observation and measurement;

[0026] The short time sequence image sequence is a collection of at least two frames of images continuously collected at a short time interval in the same field of view, which is used to provide multi-view geometric parallax information;

[0027] The polarized image group is a collection of images collected under the setting of polarization transmission axes of at least two different polarization angles and recorded with polarization angles;

[0028] The polarization angle is the orientation angle of the polarization transmission axis of the polarizer relative to the reference direction, recorded as an angle value;

[0029] The polarization transmission axis is the axis corresponding to the transmission direction of the polarizer, which is used to define the polarization orientation of the incident light;

[0030] The camera intrinsic parameters are the internal parameters of the camera imaging model, including focal length, principal point coordinates, radial distortion coefficient and tangential distortion coefficient;

[0031] The camera extrinsic parameters are the pose parameters of the camera relative to the scene coordinate system, including rotation and translation;

[0032] The camera calibration is a process of obtaining camera intrinsic parameters and extrinsic parameters and establishing an imaging model consistent with the collected data.

[0033] The same field of view is a scene area where the imaging ranges of each collected frame overlap, to ensure consistency of geometric and polarization observation.

[0034] The focal length is an equivalent focal length parameter in the imaging model that represents the ratio of object distance to image distance.

[0035] The principal point coordinates are the intersection positions of the optical axis and the image plane on the imaging plane, expressed in pixel coordinates.

[0036] The radial distortion coefficient is a parameter that describes the degree of radial distortion of the lens, used to correct barrel or pincushion distortion.

[0037] The tangential distortion coefficient is a parameter that describes the degree of tangential distortion of the lens due to misalignment.

[0038] The distortion correction is a processing step that corrects distortion of the collected image according to the distortion coefficient.

[0039] The resolution unification is a process of unifying the pixel size and sampling rate of the image sequence and the polarization image group to match the imaging model.

[0040] The imaging model is a projection and distortion description model consistent with the camera intrinsic and extrinsic parameters, used to support subsequent reconstruction and measurement.

[0041] Optionally, step S2 is specifically:

[0042] The image sequence is input into an image recognition network to perform crack semantic segmentation and generate a pixel-level crack probability map.

[0043] The pixel-level crack probability map is converted into a crack mask according to a preset rule, and the crack mask is subjected to connectivity screening and morphological refinement to form a continuous crack region.

[0044] The crack mask is skeletonized to obtain a single-pixel-wide connected path, and when there are branches in the path, the main trunk is selected as the criterion for path continuity and length to generate a crack centerline.

[0045] At each discrete point of the crack centerline, the normal direction is determined on the image plane according to the local tangential direction of the centerline, and the two side regions are divided in the crack mask with the normal direction as the boundary and merged to obtain the left and right side regions.

[0046] Terminology:

[0047] The image recognition network is a model or algorithm used to perform crack-related visual tasks, including but not limited to deep learning segmentation networks, used to produce segmentation results required for subsequent processing.

[0048] The crack semantic segmentation is a process of determining pixels in an image according to a "crack / non-crack" category, used to form an initial semantic result of a crack region;

[0049] The pixel-level crack probability map is a map numerically representing the probability of each pixel belonging to the "crack" category, used for subsequent thresholding to generate a mask;

[0050] The preset rule is a set of rules for converting a probability map to a mask and post-processing, including a probability threshold, a minimum connected domain area, a morphological structure element size, and an iteration number, etc.;

[0051] The crack mask is a binary image obtained by thresholding and regularizing the probability map, used to indicate a pixel set of a crack region;

[0052] The connectivity screening is a process of screening and retaining mask regions based on pixel connectivity, used to remove isolated or small noise regions and retain main connected regions;

[0053] The morphological thinning is a morphological processing procedure performed on the mask, including erosion, dilation, opening and closing operations, and thinning operations, to improve the region coherence and boundary regularity;

[0054] The continuous crack region is a main region of the mask that remains stably connected after connectivity screening and morphological thinning;

[0055] The skeletonization is an image processing step of contracting the crack mask to a centerline form to obtain a single-pixel-wide connected skeleton path;

[0056] The single-pixel-wide connected path is a pixel-level thin line structure formed inside the mask after skeletonization, used to represent the center direction of the crack;

[0057] The path branch is a situation where the skeleton path has a bifurcation, including a bifurcation point and its branch paths;

[0058] The path continuity is an index for measuring the smoothness and coherence of the skeleton path, used to assist in distinguishing the main path from the noise path when there is a branch;

[0059] The length is a criterion for measuring the total length of the skeleton path, used to determine the main stem together with the path continuity when selecting a branch;

[0060] The main stem is the main connected path selected according to the path continuity and length criterion when the skeleton has a branch;

[0061] The crack centerline is the main stem path determined after skeletonization and branch selection, representing the direction of the crack;

[0062] The discrete points are a sequence of points obtained by sampling the center line of the crack in pixels or equal intervals, used for calculating the local direction and dividing the region;

[0063] The local tangent direction is a tangent direction calculated at the discrete points of the center line according to the sequence of adjacent points, used for defining the normal direction and the cross-sectional direction;

[0064] The image plane is a two-dimensional pixel coordinate plane corresponding to the camera imaging model, serving as a reference plane defined by the tangent direction and the normal direction;

[0065] The normal direction is a direction perpendicular to the local tangent direction in the image plane, used as a boundary direction of the two side regions;

[0066] The boundary is a process of dividing the pixels in the crack mask into left and right sides according to the normal direction;

[0067] The two side regions are left and right side pixel sets divided in the crack mask according to the normal direction;

[0068] The left and right domains are two side region sets respectively labeled according to the normal direction, used for decoupled modeling of the two sides of the crack in subsequent reconstruction.

[0069] Optionally, step S3 is specifically:

[0070] In the neighborhood defined by the crack mask, a local three-dimensional reconstruction is performed using neural radiance field or three-dimensional Gaussian scattering with the image sequence, the camera internal parameters and the camera external parameters as inputs, and a photometric consistency weighted by a semantic mask is used as a basic optimization objective to calculate a reprojection residual;

[0071] A left implicit surface and a right implicit surface are respectively established in the left domain and the right domain to form a bilateral representation;

[0072] A linear observation is solved by a polarized image group, and a consistency constraint is imposed on the normal output of the left implicit surface and the right implicit surface with the linear observation;

[0073] A gap region is determined according to the normal direction of the center line of the crack, a occupancy prior is imposed on the body density or occupancy value in the gap region to make it tend to a low value, and a cross-slit repulsive potential is defined to limit the minimum distance between the left implicit surface and the right implicit surface in the gap region;

[0074] The photometric consistency, the polarization consistency, the semantic mask weighting, the occupancy prior and the cross-slit repulsive potential are jointly optimized in a differentiable manner, and the left implicit surface, the right implicit surface and the normal field are output, and the reprojection residual statistics are updated.

[0075] Terminology:

[0076] The crack mask defined neighborhood is a local range set around the crack mask as spatial constraint, used to define the scope of 3D reconstruction and optimization;

[0077] The neural radiance field is a 3D field model represented by continuous density and color, optimized by differentiable body rendering, used to recover local geometry and appearance from multi-view data;

[0078] The 3D Gaussian Splatting is a 3D representation method that represents the scene by a set of Gaussian splats and realizes fast reconstruction by projection rendering;

[0079] The local 3D reconstruction is a process that recovers the 3D geometry and appearance of the scene within the crack neighborhood, rather than the full scene reconstruction;

[0080] The semantic mask weighted photometric consistency is a measure that weights the photometric consistency error by the semantics of the crack mask, used to strengthen the target area constraint in optimization;

[0081] The re-projection residual is a difference measure between the reconstructed result projected in each view and the original observation, used to evaluate the fitting degree of reconstruction;

[0082] The left implicit surface is a surface represented by the zero level set of an implicit function in the left domain of the crack, used to describe the geometry on the left side of the crack;

[0083] The right implicit surface is a surface represented by the zero level set of an implicit function in the right domain of the crack, used to describe the geometry on the right side of the crack;

[0084] The bilateral representation is a pair of geometric expressions that simultaneously establish the left implicit surface and the right implicit surface, used to explicitly decouple the two sides of the crack;

[0085] The polarized normal observation is the surface normal direction observation data obtained from the polarized image group, used as a constraint for geometric optimization;

[0086] The normal output is the normal direction result calculated by the implicit surface according to the gradient, used for comparison with the polarized normal observation;

[0087] The consistency constraint is an optimization constraint that makes the normal output and the polarized normal observation as consistent as possible in direction;

[0088] The polarization consistency is an error term that constrains the consistency of the normal direction of the implicit surface with the polarized normal observation, used to improve the consistency of geometry and optics;

[0089] The normal band of the crack center line is a band-shaped range set in the normal direction of the center line, used to define the spatial position of the crack gap;

[0090] The gap region is a void space in the normal band corresponding to the crack opening, used to impose an occupancy and repulsion prior;

[0091] The volume density is a numerical measure of the occupancy degree of a spatial point in the reconstructed volume representation;

[0092] The occupancy value is a numerical representation of whether a spatial point or voxel is occupied by geometry;

[0093] The occupancy prior is a prior constraint that requires the volume density or occupancy value in the gap region to approach a low value, used to keep the crack empty;

[0094] The cross-slit repulsion potential is a potential function constraint that penalizes the excessive proximity or overlap of left and right implicit surfaces in the gap region, used to avoid cross-slit bridging;

[0095] The minimum distance is a minimum distance threshold that must be maintained between left and right implicit surfaces in the gap region;

[0096] The differentiable way joint optimization is a method of jointly solving the photometric consistency, polarization consistency, semantic mask weighting, occupancy prior, and cross-slit repulsion potential loss terms in a differentiable optimization process;

[0097] The normal field is a spatial normal distribution synthesized by the left and right implicit surfaces in the visible domain, used for subsequent measurement and encoding;

[0098] The reprojection residual statistical value is a statistical indicator of the reprojection residual in the optimization process, including mean and variance or its equivalent measure, used for subsequent confidence evaluation.

[0099] Optionally, step S4 is specifically:

[0100] Taking the left and right implicit surfaces and the image sequence and camera intrinsic and extrinsic parameters as inputs, at least one reference object with a known size is identified in the image sequence;

[0101] According to the camera intrinsic and extrinsic parameters, the pixel boundaries of the reference object are projected onto the left and right implicit surfaces to obtain the corresponding geometric positions and current sizes of the reference object on the left and right implicit surfaces;

[0102] The absolute scale parameter is obtained from the proportional relationship between the known size and the corresponding current size of the reference object, and when multiple reference objects exist, a unified absolute scale parameter is obtained by a consistency criterion;

[0103] The left and right implicit surfaces are corrected for coordinate scale consistency with the absolute scale parameter, so that the two implicit surfaces are expressed in the same measurement scale;

[0104] The absolute scale parameter and the scale-corrected left and right implicit surfaces are output.

[0105] Terminology:

[0106] The reference object is a target recognizable in the image sequence and with known physical dimension, used to establish the proportional relationship between the reconstruction result and the actual measurement;

[0107] The known dimension is the nominal or measured physical length, diameter or spacing of the reference object in the real world, used as the basis for scale calibration;

[0108] The pixel boundary is a pixel-level description of the contour edge or feature edge of the reference object in the image, used to correlate with the imaging geometry to recover the three-dimensional correspondence;

[0109] The projection to the implicit surface is a process of intersecting the line of sight corresponding to the pixel boundary with the implicit surface using the camera internal and external parameters to determine the three-dimensional correspondence point;

[0110] The corresponding geometric position is the point or point set position of the reference object pixel boundary after being projected to the implicit surface in three-dimensional space, used to calculate the current dimension;

[0111] The current dimension is the three-dimensional measurement result calculated by the corresponding geometric position under the existing reconstruction scale, which has not been converted to absolute physical scale;

[0112] The absolute scale parameter is the proportional coefficient that maps the current reconstruction scale to the physical measurement scale, used to eliminate the scale uncertainty of multi-view reconstruction;

[0113] The consistency criterion is a criterion for obtaining a single absolute scale parameter by minimizing error, robust consistency or outlier rejection when there are multiple reference objects or multi-view observations;

[0114] The coordinate scale consistency correction is a linear scale transformation of the implicit surface coordinates according to the absolute scale parameter, making the reconstruction result consistent with the physical unit;

[0115] The same measurement scale is the measurement state of the left and right implicit surfaces expressed in the same physical unit and scale reference after correction;

[0116] The scale-corrected left implicit surface is the geometric representation obtained by applying the absolute scale parameter to the left implicit surface, used for subsequent millimeter-level measurement and comparison;

[0117] The scale-corrected right implicit surface is the geometric representation obtained by applying the absolute scale parameter to the right implicit surface, used for subsequent millimeter-level measurement and comparison;

[0118] The unified absolute scale parameter is a single scale coefficient obtained under the condition of multiple reference objects according to the consistency criterion and applicable to all data.

[0119] Optionally, step S5 is specifically:

[0120] The fracture center line, the scale-corrected left implicit surface, the scale-corrected right implicit surface, the normal field, and the absolute scale parameter are taken as inputs;

[0121] A local coordinate system is established along the fracture center line in a local three-dimensional space, with the center line tangent as the axis and the center line normal as the cross-sectional normal, and a cross-sectional sequence is generated in the local coordinate system and orthogonal to the center line normal;

[0122] The opening width is calculated in each cross section based on the nearest point pairing of the scale-corrected left implicit surface and the scale-corrected right implicit surface, and the faulted height is calculated by the difference in the coordinate components of the two implicit surfaces in the height direction defined by the normal field in the cross section;

[0123] The opening width and the faulted height are converted into millimeter-level numerical values according to the absolute scale parameter;

[0124] The cross-sectional sequence, the opening width numerical value, and the faulted height numerical value of each cross section are output.

[0125] Terminology interpretation:

[0126] The local three-dimensional space is a three-dimensional measurement region with the fracture center line and its mask neighborhood as the range, used for constructing cross sections and implementing measurement;

[0127] The local coordinate system is a coordinate reference system established along the fracture center line, with the center line tangent as the axis and the center line normal as the cross-sectional normal, used for normalizing cross section generation and measurement direction;

[0128] The center line tangent is the tangent direction along the fracture center line, used to define the axial arrangement of the cross-sectional sequence;

[0129] The cross-sectional normal is the normal direction perpendicular to the center line tangent, used as the normal direction of each measurement cross section;

[0130] The cross-sectional sequence is a set of multiple measurement planes generated in order along the center line in the local coordinate system, orthogonal to the center line normal;

[0131] The cross section is a single measurement plane in the cross-sectional sequence, used to perform fracture opening and faulted calculation in the plane;

[0132] The nearest point pairing is a process of finding and matching the nearest points between the scale-corrected left implicit surface and the right implicit surface in each cross section;

[0133] The opening width is the distance between the two sides calculated by the nearest point pairing in each cross section, used to represent the actual width of the fracture opening;

[0134] The height direction is the measurement direction determined by the normal field within the cross section, used to calculate the elevation difference between the two sides in this direction;

[0135] The coordinate component difference is the difference between the coordinate components of the two implicit surfaces in the height direction, which is used to quantify the misalignment height.

[0136] The misalignment height is the difference in elevation between the two sides calculated along the height direction within each cross section, used to characterize the degree of misalignment on both sides of the crack.

[0137] The millimeter-level values ​​are measurement results that convert the opening width and misalignment height to millimeters based on absolute dimensional parameters, and are used for engineering judgment and comparison.

[0138] Optionally, step S6 specifically includes:

[0139] The inputs include the cross-section sequence, the opening width and misalignment height of each cross-section, the crack centerline, the reprojection residual statistics, and the external polarization angle parameters of the camera.

[0140] The opening width and misalignment height of each section are statistically calculated to obtain the opening width distribution statistics and misalignment height distribution statistics;

[0141] The orientation of the crack is calculated based on the tangential direction of the crack centerline, and curvature statistics are performed using the curvature sequence of the crack centerline.

[0142] The centerline of the crack is mapped to the scene coordinate system using camera extrinsic parameters to generate spatial positioning information;

[0143] The reconstruction confidence level is obtained by mapping the reprojection residual statistics with a monotonically decreasing method.

[0144] Crack numbers are assigned according to preset unique identifier generation rules;

[0145] The crack feature code is composed of crack number, spatial location information, opening width distribution statistics, misalignment height distribution statistics, direction and curvature, reconstruction confidence and acquisition conditions, and then outputs the crack feature code.

[0146] Terminology definition:

[0147] The distribution statistics are the results of statistical feature calculations on the opening width and misalignment height values ​​of each cross section, and are used for threshold determination and comparison.

[0148] The orientation is an overall orientation parameter calculated based on the tangential direction of the crack centerline, used to describe the dominant orientation of the crack.

[0149] The curvature sequence is a discrete curvature value sequence calculated along the crack centerline, used to characterize the degree of bending and the trend of change.

[0150] The curvature statistics are statistical features obtained based on the curvature sequence, used for encoding the crack morphology and changes;

[0151] The scene coordinate system is a three-dimensional reference coordinate system consistent with the camera external parameters, used for carrying the spatial positioning and measurement results;

[0152] The spatial positioning information is the position and range information obtained after mapping the crack center line to the scene coordinate system, used for positioning and tracing;

[0153] The monotonic decreasing mapping is a mapping method that converts the re-projection residual statistics value according to the relationship that the greater the residual, the lower the confidence;

[0154] The reconstruction confidence is a reliability measure obtained by monotonic decreasing mapping of the re-projection residual statistics value, used for effectiveness review and weighted decision;

[0155] The preset unique identification generation rule is a rule set for generating non-repeating crack numbers, used to ensure the uniqueness and traceability of the numbers;

[0156] The crack number is a unique identifier assigned according to the preset unique identification generation rule, used for historical association and management;

[0157] The crack feature code is a data set that structurally encapsulates the key quantitative information of the crack, including crack number, spatial positioning information, opening width distribution statistics, fault height distribution statistics, strike and curvature, reconstruction confidence and acquisition conditions;

[0158] The acquisition conditions are a parameter set representing the state of this data acquisition, including camera external parameters and polarization angle parameters of the participating views.

[0159] Optionally, step S7 specifically comprises:

[0160] The crack feature code is taken as input, and when there is a historical crack feature code, the crack number is used for association and retrieval;

[0161] The opening width distribution statistics and fault height distribution statistics in the crack feature code are compared according to the preset threshold or adaptive threshold to obtain the determination of whether it is out of limit;

[0162] The determination of whether it is out of limit is corrected for effectiveness according to the reconstruction confidence and acquisition conditions, and a recheck marker is generated when the reconstruction confidence is lower than the threshold or the acquisition conditions do not meet the requirements;

[0163] When there is a historical crack feature code, the opening width distribution statistics and fault height distribution statistics are compared in time sequence, the change amount or change rate is calculated and compared with the growth threshold to obtain the determination of whether it is a developing crack;

[0164] corresponding to whether the crack is a developing crack, and output a detection conclusion, the detection conclusion including whether the crack is over-limit determination, whether the crack is a developing crack determination, risk level and re-inspection mark.

[0165] Terminology interpretation:

[0166] The historical crack feature code is a set of coding data of previous detection records associated with crack numbers, used for time series comparison;

[0167] The preset threshold is a fixed determination limit value set according to engineering specifications or design requirements, used for over-limit determination of opening width distribution statistics and bench height distribution statistics;

[0168] The adaptive threshold is a determination limit value formed by modifying the preset threshold based on factors such as acquisition conditions and reconstruction confidence, used to improve determination applicability;

[0169] The over-limit determination is a judgment result of whether the current opening width distribution statistics and bench height distribution statistics are over-limit compared with the corresponding threshold;

[0170] The validity correction is a process of checking and adjusting the over-limit determination according to the reconstruction confidence and acquisition conditions, used to ensure the reliability of the determination;

[0171] The re-inspection mark is a review prompt identifier triggered when the reconstruction confidence is lower than the confidence threshold or the acquisition conditions do not meet the requirements, used to arrange for re-acquisition and review;

[0172] The time series comparison is a process of comparing the opening and bench statistics in the current and historical crack feature codes in the time dimension, used to identify the development trend;

[0173] The change amount is the increase or decrease between the current statistical value and the historical statistical value, used as a quantitative basis for time series comparison;

[0174] The change rate is the change proportion or speed of the statistical value per unit time, used to measure the speed of crack development;

[0175] The growth threshold is a limit value of the change amount or change rate for determining whether the crack is developing;

[0176] The developing crack is a crack type whose opening width or bench height change amount or change rate exceeds the growth threshold in time series comparison;

[0177] The level mapping rule is a rule set for mapping whether the crack is over-limit and whether the crack is a developing crack to risk level;

[0178] The risk level is the result of grading the crack risk according to the level mapping rule, and is used to guide treatment and management.

[0179] The detection conclusion is the structured result finally output by the method, including whether to exceed limit determination, whether to be a developing crack determination, risk level and reexamination marking;

[0180] The confidence threshold is the minimum value for judging whether the reconstruction confidence meets the effectiveness requirement;

[0181] The acquisition working condition requirement is a criterion set for determining whether the acquisition condition meets the effectiveness standard, including items such as view angle coverage, stable illumination and polarization angle setting.

[0182] The beneficial effects of the present application are:

[0183] 1. Realize millimeter-level high-precision three-dimensional measurement: obtain absolute scale by recognizing reference objects and perform scale consistency correction on left and right implicit surfaces, generate a cross-section sequence along the crack center line, calculate the opening width and the fault height and convert them into millimeter-level values, and significantly reduce the errors caused by view angle, distance and imaging distortion;

[0184] 2. Effectively suppress cross-crack misreconstruction and improve geometric consistency: left and right implicit surfaces are established in the crack neighborhood, combined with semantic mask weighted photometric consistency, polarization solution normal consistency, gap area occupation priori and cross-crack repulsive potential for joint optimization, to reduce bridging and false filling, and improve reconstruction stability and normal geometric consistency under weak texture and reflective material conditions;

[0185] 3. Provide traceable risk determination and management capability: build crack feature coding including spatial positioning, trend and curvature, opening and fault distribution statistics, reconstruction confidence and acquisition working condition, combined with threshold and historical time series comparison to output over-limit and developing determination and risk level, realize on-site rapid decision and reexamination closed-loop management. BRIEF DESCRIPTION OF DRAWINGS

[0186] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, illustrate the present application and are used to explain the present application together with the embodiments of the present application, and do not constitute a limitation on the present application. In the drawings:

[0187] Figure 1 A flowchart of a deep foundation crack feature coding detection method based on image recognition is provided. DETAILED DESCRIPTION

[0188] The present application will now be further described in detail in conjunction with the drawings. These drawings are all simplified schematic diagrams, and only illustrate the basic structure of the present application in a schematic manner, and therefore only show the components related to the present application.

[0189] Reference Figure 1 A deep foundation crack feature coding detection method based on image recognition, comprising:

[0190] S1, collect short-time sequence image sequence and polarization image group of different polarization angles on the surface of deep foundation pit supporting structure, and complete camera internal and camera external parameter calibration;

[0191] S2, input image sequence, perform crack semantic segmentation to generate crack mask, determine crack center line and left and right side domains and output according to the crack mask;

[0192] S3, input crack mask, crack center line, left and right side domains, image sequence, camera internal and external parameters, and polarization image group, perform local three-dimensional reconstruction in the neighborhood of the crack mask, establish left and right implicit surfaces, combine polarization information for joint optimization, obtain normal field and reprojection residual statistical value;

[0193] S4, input left and right implicit surfaces, image sequence, camera internal and external parameters, identify reference objects with known size and obtain absolute scale parameters according to the reference objects, perform scale consistency correction on the two implicit surfaces, and output absolute scale parameters and scale-corrected left and right implicit surfaces;

[0194] S5, input crack center line, scale-corrected left and right implicit surfaces, normal field and absolute scale parameters, generate cross-section sequence along the center line, and calculate opening width and bench height based on the two implicit surfaces and the normal field at each cross-section;

[0195] S6, input cross-section sequence, opening width and bench height values of each cross-section, crack center line and reprojection residual statistical value, generate crack feature coding according to engineering rules;

[0196] S7, input crack feature coding, call historical coding, determine and generate detection conclusion according to threshold and time sequence comparison;

[0197] In the specific embodiment, S1 is specifically:

[0198] In the same field of view on the surface of the deep foundation pit supporting structure, short-time interval continuous acquisition frame images to form short-time sequence image sequence (hereinafter referred to as image sequence):

[0199] ;

[0200] Wherein represents the number of image frames, represents the two-dimensional image obtained at the acquisition time ;

[0201] a set of polarization images are acquired under at least different polarization angle parameter settings, under the same field of view and the same camera mounting posture conditions:

[0202] ;

[0203] wherein is the number of polarization angles, is a two-dimensional polarization image acquired under the th polarization transmission axis orientation condition, is a polarization angle parameter corresponding to the polarization image , used to represent the orientation angle of the polarization transmission axis relative to a reference direction;

[0204] A calibration process is performed on the camera used to acquire the above-mentioned images to obtain a camera intrinsic matrix , imaging distortion parameters, and camera extrinsics of each frame in the image sequence, wherein the camera intrinsic matrix can be written as:

[0205] ;

[0206] wherein and are the equivalent focal lengths along the horizontal and vertical directions of the image, respectively, in units of pixels, and are the pixel coordinates of the principal point in the image coordinate system, and the matrix is used to map the dimensionless coordinates on the normalized imaging plane to pixel coordinates;

[0207] The distortion parameters include a set of radial distortion coefficients and a set of tangential distortion coefficients , wherein is used to describe the degree of radial distortion of the lens pixels relative to the optical axis, is used to describe the tangential distortion caused by lens assembly eccentricity, and through these distortion coefficients, a functional relationship between the normalized ideal coordinates and the distorted coordinates can be established, for example, for a certain normalized ideal coordinate , the radial distance , and the distorted normalized coordinate can be as follows:

[0208] ;

[0209] wherein is the normalized coordinate after lens imaging distortion, is the dimensionless radial distance of the point to the imaging optical axis;

[0210] The camera extrinsics are for each frame in the image sequence a set of wherein is a three-dimensional rotation matrix representing a rigid body rotation from the scene coordinate system to the camera coordinate system of the frame, is a corresponding translation vector, the projections of a three-dimensional scene point in different frames satisfy the imaging relationship in homogeneous coordinates:

[0211] ;

[0212] wherein denotes the coordinates of a three-dimensional point in the scene coordinate system, are the coordinate components of the point in the direction of the three coordinate axes, respectively, is the homogeneous pixel coordinate of the point in the image of the frame, and are the horizontal and vertical pixel coordinates, respectively, the superscript T represents vector transposition, and the symbol represents that the vectors on both sides are equal in the sense of homogeneous coordinates;

[0213] After obtaining the above intrinsic parameters, extrinsic parameters and distortion parameters, for each image in the image sequence and the polarized image group , distortion correction is performed according to the distortion model parameters , the distorted normalized coordinates are inversely calculated into ideal normalized coordinates , and the intrinsic matrix is used to map into corrected pixel coordinates, and meanwhile, all the corrected images are resampled according to a preset target resolution, so that the images have uniform pixel size and sampling rate in the horizontal and vertical directions, thereby ensuring that the short-time sequence image sequence and the polarized image group together with the camera intrinsic parameters and the extrinsic parameters of each frame form a consistent imaging model;

[0214] Finally, the short-time sequence image sequence , the polarized image group , the camera intrinsic parameters and the corresponding camera extrinsic parameter set after the distortion correction and resolution uniformity processing are taken as the output of step S1.

[0215] In the specific embodiment, S2 is specifically:

[0216] Let the short-time sequence image sequence be denoted as , wherein denotes the number of image frames, denotes a single two-dimensional image obtained at the acquisition time , and each frame of image in the image sequence is respectively, to obtain a pixel-level crack probability map consistent with the size of the frame image , i.e.

[0217] ;

[0218] wherein denotes an image recognition network for crack semantic segmentation, denotes a parameter set of the network, denotes a pixel with pixel coordinates in the frame image , and denotes a probability that the pixel belongs to the "crack" category, and are the horizontal and vertical pixel coordinates of the pixel on the image plane, respectively;

[0219] Then, according to a pre-set probability threshold , the crack probability map is converted into a binary crack mask , which can be thresholded by an indicator function to obtain:

[0220] ;

[0221] wherein is an indicator function, which takes a value of 1 when the condition in the parentheses is true, and otherwise takes a value of denotes whether the pixel in the frame image is determined to be a crack pixel; Next, based on the crack mask

[0222] , the connected regions therein are analyzed according to the connectivity rule, and a minimum connected domain area threshold is introduced. When the number of pixels in a connected region is less than , the connected region is regarded as noise and removed, and only the main connected region with an area not less than is retained to form an updated crack mask. Subsequently, a morphological thinning operation is applied to the mask, which can be formally expressed as:

[0223] ;

[0224] wherein is a continuous crack region mask obtained after connectivity screening and morphological processing, denotes an operator that integrates morphological operations such as erosion, dilation, opening-closing operation, and thinning;

[0225] ​​Obtaining a mask for a continuous crack region Then, skeletonization is performed on it to obtain a set of crack skeletons with a width of one pixel. It can be written as:

[0226] ;

[0227] in For the first The set of pixels representing the crack skeleton in a frame image. For skeletonization operators;

[0228] skeleton pixel set Consider the image as a discrete set of points on the image plane, and construct an undirected graph based on pixel adjacency relationships. Several connected paths can be extracted from this graph. To select the main path representing the overall direction of the crack from multiple paths, each candidate path in the skeleton path is denoted as:

[0229] ;

[0230] superscript For candidate path indexing, This represents the number of discrete points on the candidate path. Indicates the first [item] on this path pixel coordinate vectors of skeleton points and These are the horizontal and vertical pixel coordinates of the skeleton point, respectively, and the superscript T indicates vector transpose;

[0231] Define candidate paths Total path length The sum of the Euclidean distances between adjacent skeleton points:

[0232] ;

[0233] in The Euclidean norm of a two-dimensional vector is used to determine the main path of the crack, with the candidate path being longer and having a smoother change in direction. This ultimately determines the crack centerline. ,in The number of sampling points on the central line, For the first The pixel coordinates of the center line point;

[0234] After obtaining the crack centerline, the local tangential and normal directions of the centerline are defined in the image plane for each internal point (excluding endpoints) on the centerline. You can use adjacent points before and after. and Perform difference estimation on the local tangential vector:

[0235] ;

[0236] where is the unit tangent vector at the centerline point , and are the components of the tangent vector in the transverse and longitudinal directions of the image plane, respectively;

[0237] Based on the unit tangent vector , a unit normal vector orthogonal to it is constructed to define the dividing direction on the two sides of the crack in the image plane, such that:

[0238] ;

[0239] where is the unit normal vector at the centerline point , and are the components of the normal vector in the transverse and longitudinal directions of the image plane, respectively;

[0240] After obtaining the crack centerline and its local normal direction , for each crack pixel in the continuous crack region mask , the nearest centerline point in the image plane can be found, and the signed distance of the pixel relative to the centerline point in the normal direction is constructed , for example, let the pixel coordinate vector be , then:

[0241] ;

[0242] where is the signed distance of the pixel point relative to the centerline point in the normal direction, when , it is considered that the pixel is located on the side where the normal vector points, when , it is considered that the pixel is located on the side where the normal vector is reversed, according to which all the pixels in the crack mask can be divided into left and right domains, denoted as and , respectively, where is the set of crack pixels satisfying , is the set of crack pixels satisfying , thereby completing the division of the regions on the two sides of the crack centerline in the image plane, and finally outputting the crack mask , the crack centerline and the corresponding left-hand domain and right-side domain This provides input for the subsequent step S3, which involves local 3D reconstruction within the crack mask neighborhood and the creation of the left and right implicit surfaces.

[0243] In this specific embodiment, S3 specifically refers to:

[0244] A short time-series image sequence after distortion correction and resolution unification is denoted as... ,in For the number of image frames, For the first A two-dimensional image frame, denoted by a group of polarization images. ,in This refers to the number of polarization angles. For polarization angle parameters Two-dimensional polarization images acquired below, These are the polarization angle parameters corresponding to the polarization image;

[0245] Then use the camera intrinsic parameter matrix With camera extrinsic set For imaging geometry input, where For the camera intrinsic parameter matrix, For the first The 3D rotation matrix of the frame camera, For the first The 3D translation vector of the frame camera;

[0246] Simultaneously, the crack mask, crack centerline, and left and right lateral regions are used as geometric prior inputs, where the first... The mask for the continuous crack region of the frame is denoted as Indicates pixel coordinates Whether a pixel belongs to the crack region, the crack centerline is denoted as ,in The number of discrete points on the center line, For the first pixel coordinate vector of center line point and These are the horizontal and vertical pixel coordinates of the center line point, respectively. The superscript T indicates vector transpose. The left and right domains are denoted as... and , This refers to the set of crack pixels located on one side after being divided according to the centerline normal direction. This is the set of crack pixels located on the other side;

[0247] Based on this, a local 3D reconstruction model is constructed within the neighborhood defined by the crack mask, for each frame of image. Let the set of pixels in the neighborhood of the crack be denoted as:

[0248] ;

[0249] wherein denotes the region of interest in the frame image defined by the crack mask;

[0250] Using the camera intrinsic matrix and the camera extrinsic , a camera view ray through three-dimensional space for each pixel can be constructed and denoted as:

[0251] ;

[0252] wherein denotes the three-dimensional position vector of the camera in the frame in a unified scene coordinate system, is the unit view direction vector corresponding to the pixel , and

[0253] In the present embodiment, a neural radiance field is preferably adopted as the local three-dimensional representation, which realizes the mapping from a three-dimensional space point and its observation direction to the volume density and color through a multi-layer perceptron with parameters , which can be formalized as:

[0254] ;

[0255] wherein denotes a three-dimensional point in the three-dimensional scene coordinate system, are the coordinate components of the point in the three coordinate axis directions, respectively, is the unit observation direction vector from the point to the camera, is the volume density scalar of the three-dimensional point , and is the RGB color vector of the point, are the color components of the red, green, and blue channels, respectively;

[0256] For each view ray , the neural radiance field can be volumetrically integrated based on the volume rendering principle to synthesize the pixel color corresponding to the view ray. The present embodiment abstracts the volume rendering process as an operator , that is, the predicted color of the pixel in the frame image is denoted as:

[0257] ; ​

[0258] wherein is the rendering color of the pixel at the current neural radiance field parameter , is the volume rendering operator that integrates the volume density and color along the line of sight to obtain the pixel color;

[0259] To constrain the photometric consistency between the reconstruction result and the multi-view observation, and to increase the constraint strength within the crack neighborhood, a semantic mask weighted photometric consistency loss is constructed based on the crack mask . First, the re-projection residual of the pixel in the first frame is recorded as:

[0260] ;

[0261] wherein is the pixel color vector of the measured image, is the predicted pixel color vector rendered by the neural radiance field, is the difference between the two;

[0262] A semantic weight function is then defined, for example:

[0263] ;

[0264] wherein is a coefficient to enhance the weight of the crack area, and when , the weight of the crack area pixel is increased, and when , the weight of the non-crack area pixel is the reference value 1;

[0265] Thus, a semantic mask weighted photometric consistency loss function can be constructed:

[0266] ;

[0267] wherein denotes the Euclidean norm of the vector, is the weighted squared error that measures the difference between the reconstruction result and the measured image within the crack neighborhood;

[0268] On the basis of the local volume field reconstruction, to explicitly express the structure surfaces on both sides of the crack, left and right implicit surfaces are respectively established within the crack mask neighborhood, which are respectively given by the zero level set of the scalar implicit functions and , that is:

[0269] ; ​

[0270] where is a scalar function of the left implicit surface, is a scalar function of the right implicit surface, and are the parameter sets of the left and right implicit surfaces respectively, the sign of the function value can be used to distinguish whether a three-dimensional point is inside or outside the surface;

[0271] According to the pixel division of the left field and the right field , the pixel view line samples from are mainly used to constrain the left implicit function , and the pixel view line samples from are mainly used to constrain the right implicit function , so as to form a bilateral geometric representation with the left and right implicit surfaces as the core in the crack neighborhood;

[0272] The corresponding unit normal vector field can be defined by the implicit function. For a three-dimensional point falling on the left implicit surface, the left surface normal can be expressed as:

[0273] ;

[0274] For a three-dimensional point falling on the right implicit surface, the right surface normal can be expressed as:

[0275] ;

[0276] where denotes the gradient operator of the three-dimensional coordinates , and and are the unit normal vectors of the left and right implicit surfaces at the point respectively;

[0277] By performing visibility judgment and side selection on the left and right surfaces in space, the normal field of the entire crack neighborhood can be obtained by combining the two , where denotes the unit normal vector obtained by integrating the left and right implicit surfaces at the three-dimensional point ;

[0278] In order to further utilize polarization information to improve the consistency of the normal and geometry, according to the polarization image group , a pre-calibrated polarization solving function is used to perform surface normal estimation on the pixels in each frame, for example, the surface normal of pixel in the i-th frame can be expressed as:

[0279] ​ ;

[0280] where is the unit normal vector of the pixel at the th frame calculated from the polarization image group, is the function of surface normal inversion based on the polarization intensity variation model;

[0281] With the camera intrinsic matrix and camera extrinsic parameters , the corresponding pixel position of the three-dimensional point in the th frame can be determined by its projection, so as to map the polarization normal observation to the three-dimensional space, and the polarization normal corresponding to the three-dimensional point is recorded as ;

[0282] For the set of three-dimensional points visible under the current multi-view geometry, denoted as , the polarization normal consistency loss is constructed:

[0283] ;

[0284] where represents the sum of squared errors between the normal field derived from the implicit surface and the polarization normal observation;

[0285] To maintain a low occupancy in the gap area where the crack opening is located and avoid false geometric filling, according to the crack center line and its unit normal direction in the image plane, combined with the camera extrinsic parameters and the current implicit geometry estimation, the center line is expanded into a strip-shaped area along the normal direction in the three-dimensional space, which surrounds the crack direction, and the set of three-dimensional points covered by the area is recorded as the gap area , where contains the void space corresponding to the expected crack opening;

[0286] In this gap area, a low occupancy prior constraint is imposed on the volume density or occupancy value of the neural radiance field, and for this purpose, the occupancy value is defined, for example, by monotonically mapping the volume density :

[0287] ;

[0288] where represents the degree of geometric occupancy of the three-dimensional point , the smaller the value, the closer the point is to the empty space;

[0289] Accordingly, the gap region occupancy loss is constructed as:

[0290]

[0291] wherein is used to encourage the occupancy value in the gap region to approach zero, thus keeping the crack gap empty at the model level;

[0292] On the other hand, to avoid the left implicit surface and the right implicit surface from excessively approaching or even overlapping in the gap region, a cross-slit repulsive potential is imposed on the minimum distance between the left and right surface points near the crack centerline. Specifically, at each discrete index of the crack centerline , the nearest points on the left and right implicit surfaces corresponding to the position are searched along the local normal direction in three-dimensional space, respectively denoted as and , and the Euclidean distance between the two is calculated:

[0293]

[0294] wherein and are the three-dimensional point coordinate vectors on the left and right implicit surfaces at index , respectively, is the distance estimate between the surfaces on both sides of the corresponding position;

[0295] According to a pre-set minimum distance threshold , the cross-slit repulsive potential loss is constructed as:

[0296]

[0297] wherein is the minimum allowed distance that the left and right implicit surfaces must maintain in the gap region, is a function that takes the larger value of the two, and when , a positive penalty term is generated at the corresponding position, thus driving the optimization process to push the left and right implicit surfaces away from each other at that position;

[0298] In summary, the semantic mask weighted photometric consistency loss , the polarization normal consistency loss , the gap region occupancy prior loss , and the cross-slit repulsive potential loss are combined in a differentiable manner to form the total optimization objective function:

[0299]

[0300] wherein ​​​​These are the weighting coefficients for each loss term, used to balance the relative importance of photometric consistency and normal consistency, occupied prior, and trans-slit repulsion potential;

[0301] Through the Perform gradient-based iterative optimization while updating the neural radiation field parameters. And the parameters of the left implicit surface and the right implicit surface. After optimization and convergence, the left implicit surface in the neighborhood of the crack mask can be obtained. With right implicit face And the normal field derived from both and corrected by polarization normal observation. Simultaneously, based on the pixel-level reprojection residual in the photometric uniformity loss... The reprojection residual statistics are obtained, and the mean value of the reprojection residual magnitude is calculated for all pixels involved in the optimization. With variance As a statistical indicator describing the degree of fit of the reconstruction, it is used in subsequent steps for reconstruction confidence assessment and result validity determination.

[0302] In this specific embodiment, S4 specifically refers to:

[0303] The left and right implicit surfaces obtained from step S3, along with the short temporal image sequence and camera parameters output from step S1, are used as input. The left implicit surface is derived from a scalar implicit function. The zero-level set representation, the right implicit surface is represented by scalar implicit functions. The zero-level set representation, that is, satisfying or 3D points Belongs to the corresponding implicit surface, where To unify the 3D point coordinate vectors in the scene coordinate system, These represent the coordinate components of the point along the three coordinate axes. and These are the parameter sets corresponding to the left implicit surface and the right implicit surface, respectively;

[0304] Short time series image sequences are denoted as ,in For the number of image frames, For the first A two-dimensional image frame, the camera intrinsic parameter matrix is ​​denoted as... The set of camera extrinsic parameters is denoted as ,in For the first The 3D rotation matrix of the camera coordinate system relative to the scene coordinate system. This is the corresponding three-dimensional translation vector;

[0305] Then, in the image sequence At least one reference object with known physical dimension is identified by target detection or template matching, etc. The number of reference objects is denoted as where The known dimension of the i-th reference object in the real world is pre-defined as , e.g. the actual length of a certain feature edge, in meter or millimeter;

[0306] To relate the reference object from pixel space to the three-dimensional space where the implicit surface lies, for the i-th reference object, a pair of end-point pixels corresponding to the known dimension is selected in a certain frame (or multiple frames) of images where the reference object is visible, denoted as and where and , respectively, are the horizontal and vertical pixel coordinates of the two end-point pixels in the image plane; Using the camera intrinsic matrix and the i-th frame camera extrinsic

[0307] , a camera view ray through the three-dimensional space can be constructed for each end-point pixel, denoted as where is the three-dimensional position vector of the i-th frame camera in the scene coordinate system, is the unit view ray direction vector obtained by back-projection of the end-point pixel

[0308] ;

[0309] where is the three-dimensional position vector of the i-th frame camera in the scene coordinate system, is the unit view ray direction vector obtained by back-projection of the end-point pixel , and is the scalar parameter along the view ray; The view ray

[0310] is intersected with the left and right implicit surfaces respectively, i.e. finding the parameter value satisfying or , and the coordinates of the intersection points in the three-dimensional space are denoted as where is the intersection parameter value corresponding to the end-point , and is the three-dimensional geometric position of the end-point corresponding to the reference object in the current reconstruction scale; For the i-th reference object, the three-dimensional distance between the two end-points in the reconstruction result is denoted as

[0311] , which can be calculated as the Euclidean norm: ​​​​

[0312] ;

[0313] wherein is the Euclidean norm of a three-dimensional vector, denotes the reconstructed length of the corresponding edge of the th reference object with known size under the current implicit geometric scale;

[0314] Based on the known physical size of the th reference object and the reconstructed length , the scale factor corresponding to the reference object is defined as:

[0315] ;

[0316] wherein is the scale coefficient that needs to be applied to the current reconstructed scale, for converting the reconstructed length of the th reference object to its real physical length;

[0317] When there are multiple reference objects in the scene, i.e. , in order to obtain a unified absolute scale parameter applicable in a global sense, the scale factors of all reference objects are consistently estimated, for example, the unified absolute scale parameter may be determined by using the least square criterion, which can be written as:

[0318] ;

[0319] wherein is the absolute scale parameter, denotes the value on the scalar that makes the objective function take the minimum value, and the least square solution can be simplified as the arithmetic average of all . The present embodiment can also use robust statistical methods such as median or weighted average to suppress abnormal values as needed;

[0320] After obtaining the unified absolute scale parameter , the left implicit surface and the right implicit surface are corrected for coordinate scale consistency, i.e. the coordinates of any three-dimensional point on the implicit surface are linearly scaled, and the corrected coordinates are denoted as , and have:

[0321] ;

[0322] wherein is the physical coordinate corrected by the absolute scale parameter, ​​The introduction of the absolute scale parameter guarantees the left and right implicit surfaces to be expressed in a unified physical metric scale, thus eliminating the scale ambiguity inherent in multi-view reconstruction.

[0323] In practical implementation, the above scaling of three-dimensional coordinate input can be performed while keeping the implicit function form unchanged, or the scale can be absorbed into the implicit function parameters through equivalent transformation With In this embodiment, the geometric representation corresponding to the left implicit surface after scale correction is collectively referred to as "scale-corrected left implicit surface", and the geometric representation corresponding to the right implicit surface after scale correction is collectively referred to as "scale-corrected right implicit surface", and finally the unified absolute scale parameter and the scale-corrected left implicit surface and the scale-corrected right implicit surface are taken as outputs.

[0324] In this specific embodiment, S5 is specifically:

[0325] The foregoing crack center line and the scale-corrected left implicit surface and the scale-corrected right implicit surface, the normal field and the absolute scale parameter are taken as inputs, wherein the discrete representation of the crack center line on the image plane is denoted as is the number of center line discrete points, is the pixel coordinate vector of the th center line point, and respectively represent the pixel coordinates of the point in the horizontal direction and the vertical direction of the image, and the superscript T represents the vector transpose;

[0326] The scale-corrected left implicit surface and the right implicit surface are still represented by the zero level set of the implicit functions and , i.e. the three-dimensional point or satisfying belongs to the corresponding implicit surface, wherein represents the three-dimensional point coordinates in the unified scene coordinate system, are the coordinate components of the point in the three coordinate axis directions, respectively;

[0327] The normal field is denoted as , which represents the unit surface normal vector obtained by the foregoing joint optimization at the three-dimensional point ;

[0328] The absolute scale parameter is denoted as This parameter represents the scaling factor between the reconstruction scale and the physical scale obtained from step S4. After this parameter is applied, the three-dimensional coordinates on the implicit surface are under a unified physical metric scale. To establish a cross-sectional sequence along the crack centerline in the local three-dimensional space, the crack centerline is first mapped from the image plane to the scene coordinate system to obtain the corresponding three-dimensional centerline point sequence. ,in Represents the point relative to the pixel center line. The corresponding three-dimensional point coordinates These are the coordinate components of the centerline point along the three coordinate axes;

[0329] At the three-dimensional centerline Above, the local tangential direction is discretized based on adjacent centerline points, and each internal point is indexed. Local tangent vectors can be defined (excluding endpoints). The normalized result of the difference between adjacent points:

[0330] ;

[0331] in For the first The unit tangential vector at each centerline point These are the components of the vector along the three coordinate axes. The Euclidean norm of a three-dimensional vector;

[0332] Using the unit tangential vector Establish in three-dimensional space Centered on, with The cross-section plane is normal to the plane. Its mathematical expression can be written as:

[0333] ;

[0334] in Indicates the first A cross-sectional plane, composed of all planes satisfying vector difference With tangential vector Vertical three-dimensional points constitute;

[0335] For each cross-sectional plane Find the intersections of the scale-corrected left implicit surface and the scale-corrected right implicit surface with the plane, denoted as:

[0336] ;

[0337] in For the first The set of intersection points of the left implicit surface on the cross section and the cross section plane. For the first The set of intersection points of the right implicit surface on the cross section and the cross section plane;

[0338] In a given cross-sectional plane To calculate the crack opening width at this section, the left-side point set is used based on the principle of nearest-point pairing. With the right-hand point set Find the pair of points that are closest to each other, and denote the pair of points as follows: and ,in And thus define the first Opening width at each cross section The Euclidean distance between the pair of closest points is expressed as:

[0339] ;

[0340] in To the first The scalar value of the crack opening width of a cross section;

[0341] To calculate the misalignment height, it is necessary to consider each cross-sectional plane. An internal "height direction" is defined, which is determined by the distribution of the normal field near the centerline. In this embodiment, it can be defined at the three-dimensional centerline point. Take the unit normal vector of the normal field. And project it onto the cross-sectional plane. Within this section, the direction vector per unit height is obtained. Take this vector so that it is both perpendicular to the cross section normal. Orthogonal, and as close as possible to the original normal direction. , denoted as:

[0342] These are the components of the height direction on the three coordinate axes, and they satisfy... as well as ;

[0343] Based on this, connect the vectors of the left and right nearest points. Defined as:

[0344] ;

[0345] in Indicates the first A three-dimensional vector pointing from the left surface point to the right surface point at each cross section. These are the components of the vector along the three coordinate axes, respectively;

[0346] The vector of the line in the height direction The difference of the coordinate components in the height direction of the projection on the upper surface is the height difference of the misaligned surface , which can be written as

[0347] ;

[0348] wherein is the height difference scalar of the misaligned surface at the i-th cross section, and when is positive, it represents the relative uplift of the right surface relative to the left surface in the height direction, and when it is negative, it represents the relative subsidence of the right surface relative to the left surface;

[0349] Since the absolute scale parameter is calibrated through step S4, the three-dimensional coordinates on the implicit surface are in a unified physical unit (for example, in meters), and the above opening width and the height difference of the misaligned surface are also in the physical unit. In order to meet the requirement of millimeter-level precision in engineering interpretation, a unit conversion coefficient is introduced to convert the current physical unit into millimeter unit, for example, when the current physical length unit is meter, let , then the millimeter-level values of the opening width and the height difference of the misaligned surface at the i-th cross section can be represented as:

[0350] ;

[0351] wherein is the opening width of the crack in millimeters at the i-th cross section, is the height difference of the misaligned surface in millimeters at the i-th cross section, , and the value of can be set according to the physical unit corresponding to the absolute scale parameter , so as to ensure the consistency of the numerical dimension; By repeating the above process of constructing the cross section plane, finding the nearest point pair of the left and right implicit surfaces in the cross section, calculating the opening width

[0352] and the height difference of the misaligned surface along the three-dimensional center line for all discrete indexes , and performing millimeter-level conversion, the cross section sequence , the opening width value sequence corresponding to each cross section, and the height difference value sequence of the misaligned surface are finally obtained. .

[0353] In the specific embodiment, the S6 is specifically:

[0354] Let the cross section sequence be​​ wherein is the number of sections generated along the fracture centerline, denotes the i-th section plane;

[0355] a sequence of millimeter-level values of the fracture opening width at each corresponding section is denoted as wherein is the opening width at the i-th section in millimeter;

[0356] a sequence of millimeter-level values of the heave height at each section is denoted as wherein is the heave height at the i-th section in millimeter;

[0357] a three-dimensional representation of the fracture centerline in the unified scene coordinate system is denoted as wherein is the three-dimensional coordinate vector of the centerline point corresponding to the i-th section, are the coordinate components of the point in the directions of the three coordinate axes, respectively;

[0358] the re-projection residual statistics value obtained in step S3 based on the pixel-level re-projection residual statistics is denoted as and wherein is the mean value of the re-projection residual amplitude, is the variance of the re-projection residual amplitude;

[0359] the camera extrinsic parameter set and the polarization angle parameter set are simultaneously called as the acquisition geometry and polarization condition inputs, wherein is the three-dimensional rotation matrix of the i-th frame camera, is the three-dimensional translation vector of the i-th frame camera, is the polarization angle parameter corresponding to the i-th polarization image; On the basis of the above inputs, firstly, the opening width values and the heave height values on the section sequence are statistically calculated to form the opening width distribution statistics and the heave height distribution statistics, specifically, the arithmetic mean of the opening width sequence is denoted as

[0360] , the standard deviation is denoted as , and the maximum value is denoted as , which can be calculated as:

[0361] ;​​​​​​​

[0362] where is the average value of the opening width, is the variance of the opening width, is the standard deviation, is the maximum value of the opening width among all cross sections;

[0363] Similarly, for the misalignment height sequence its average value, variance and maximum value are calculated and denoted as and where is the average value of the misalignment height, is the variance of the misalignment height, is the standard deviation, is the maximum value of the misalignment height among all cross sections, and the set of opening width related statistics is denoted as and the set of misalignment height related statistics is denoted as ;

[0364] Secondly, to describe the overall orientation and curvature of the crack, based on the three-dimensional centerline point set the crack orientation and curvature statistics are calculated, assuming that in the scene coordinate system the axis is the vertical direction, the plane is the horizontal plane, and the overall orientation vector is defined as the vector of the line connecting the first and last points of the centerline where are the components of the vector in the three coordinate axis directions, to obtain the orientation in the horizontal projection, the vector is projected in the plane as and the crack orientation angle is defined as:

[0365] ; where

[0366] is the azimuth angle of the crack in the horizontal plane relative to the axis, is the two-parameter arctangent function; To characterize the bending degree of the crack, the tangent vector is constructed using the discrete point sequence on the centerline and the discrete curvature is calculated accordingly. For each internal index

[0367] the local tangent vector is defined as where , , are the components of the unit tangent vector at the th centerline point in the three coordinate axis directions, Euclidean norm of three-dimensional vector, and the discrete curvature of the first point is denoted as:

[0368] ;

[0369] wherein is the discrete curvature scalar at the first centerline point, and the larger the value is, the more severe the bending is;

[0370] The average value and the maximum value of the curvature sequence are calculated, denoted as and , wherein is the average curvature of the crack along the line, is the maximum curvature of the crack along the line, and the set of curvature-related statistics is denoted as ;

[0371] In terms of spatial positioning information, since the three-dimensional centerline point set has been in the scene coordinate system consistent with the camera extrinsic parameters, the spatial position and range features of the crack can be extracted based on , such as calculating the minimum and maximum values of the centerline points in each coordinate direction, denoted as:

[0372] ;

[0373] wherein are the spatial range boundaries of the crack in three directions, respectively, and these quantities and the coordinates of the first and last points of the centerline are combined to form a set of spatial positioning information , used to indicate the position and scale of the crack in the scene;

[0374] In terms of reconstruction confidence, starting from the statistical values of the re-projection residuals and , a monotonically decreasing mapping function about the residuals is constructed, and the larger the amplitude of the re-projection residual is, the lower the confidence is. In the present embodiment, the root mean square value of the re-projection residual is defined as:

[0375] ;

[0376] wherein is a re-projection residual scalar index considering the mean and variance comprehensively, and the larger the value is, the larger the overall fitting error is; on this basis, the reconstruction confidence is defined as:

[0377] ;

[0378] wherein represents the confidence of the reconstruction result, is a scale parameter, used to adjust the decay speed of the residual to the confidence, when increases, monotonically decreases, so as to meet the requirement that "the greater the residual, the lower the confidence";

[0379] For the acquisition condition, the camera extrinsic parameter set participating in the current detection , the polarization angle parameter set and the optional other acquisition state information (such as the number of participating views, the acquisition time period, the light condition mark, etc.) are uniformly encapsulated into the acquisition condition set , wherein is used to describe the geometric layout and polarization setting of the current crack detection data acquisition;

[0380] After having the above-mentioned geometric and statistical information, the crack number is assigned to the current crack according to the preset unique identification generation rule, and the number is recorded as , which can be generated by combining the project number, the measurement point area number, the acquisition date and time stamp, and the serial number in the same detection, to ensure that it is not repeated in the whole engineering period; finally, the crack number , the spatial positioning information , the opening width distribution statistics , the joint height distribution statistics , the strike and curvature statistics , the reconstruction confidence and the acquisition condition set are combined to form the crack feature code , that is:

[0381] ;

[0382] wherein is the structured coding data used for subsequent threshold judgment and time sequence comparison analysis.

[0383] In the specific embodiment, the S7 is specifically:

[0384] The crack feature code is recorded as:

[0385] ;

[0386] wherein is the unique number of the current crack, is the crack spatial positioning information set, is the opening width distribution statistics set, wherein is the average value of the opening width, is the standard deviation of the opening width, is the maximum value of the opening width in all cross sections, is the set of statistics of the misaligned height distribution, where is the average value of the misaligned height, is the standard deviation of the misaligned height, is the maximum value of the misaligned height in all cross sections, is the strike angle of the crack in the horizontal plane, is the set of statistics of the crack curvature, where is the average curvature along the center line of the crack, is the maximum curvature along the center line of the crack, is the reconstruction confidence scalar mapped from the reprojection residual, whose value range is the set of acquisition conditions of this detection, including the camera extrinsic parameters, polarization angle parameters, and optional acquisition time period and illumination condition markers, etc.

[0387] When there is a historical crack feature code, take the crack number as the index, retrieve the associated historical code set from the historical database, and record it as:

[0388] ;

[0389] where is the set of historical feature codes of this crack, is the number of historical records, and when there is no historical record, let be the crack feature code obtained by the th historical detection, whose internal structure is consistent with the current code ;

[0390] Secondly, according to the preset threshold value given by the engineering specification and the optional adaptive correction strategy, the over-limit judgment is carried out on the current opening width distribution statistics and misaligned height distribution statistics, and the reference limit value of the opening width is , and the reference limit value of the misaligned height is , where and are given according to the relevant specification or design requirement, in order to appropriately tighten the threshold value when the reconstruction confidence is low or the acquisition condition is poor, the adaptive correction function based on and can be defined to obtain the effective threshold value and :

[0391] ;

[0392] where and The adaptive correction functions of the opening width and the misalignment height thresholds, respectively, can appropriately relax or tighten the reference thresholds according to the reconstruction confidence and the acquisition conditions;

[0393] After obtaining the effective threshold, the current statistical quantity is compared with the effective threshold to determine whether the crack is developing The over-limit indicator variables of the opening width and the misalignment height are constructed, and the following equations are established:

[0394] ;

[0395] wherein is a Boolean variable indicating whether the opening width is over the limit, and the value 1 indicates that the opening width exceeds the limit, and the value 0 indicates that the opening width is not over the limit, is a Boolean variable indicating whether the misalignment height is over the limit, and the meaning is similar, is an indicator function, and the value is 1 when the condition in the parentheses is true, and the value is 0 otherwise;

[0396] Further, the two are combined to obtain the overall over-limit determination result:

[0397] ;

[0398] wherein is a total determination Boolean variable indicating whether the current crack is over the limit, and the value is 1 when either the opening width or the misalignment height is over the limit;

[0399] Then, the reconstruction confidence and the acquisition conditions are combined to modify the over-limit determination result, and the lower limit of the confidence is set to When , it is considered that the reconstruction result is unreliable, and the acquisition condition effectiveness determination function is defined, wherein indicates that the acquisition condition meets the preset requirements (for example, the angle coverage, the stable illumination, and the reasonable polarization angle setting), indicates that the acquisition condition does not meet the requirements, and the recheck marker is constructed according to the above two conditions:

[0400] ;

[0401] wherein is a Boolean variable of the recheck marker, and when , it indicates that the detection result needs to be arranged for review, and even can also prompt the engineering personnel to carefully interpret the result through the marker;

[0402] When there is a historical crack feature code, i.e. , to determine whether the crack is developing, the latest historical record from the historical set is selected, and is recorded as and extract the corresponding opening width distribution statistics from it and the misalignment height distribution statistics where is the maximum opening width at the historical time, is the maximum misalignment height at the historical time, the superscript (prev) indicates the statistics corresponding to the historical record;

[0403] To measure the degree of evolution of the crack between two detections, the present embodiment calculates the increment of the current statistics relative to the historical statistics in the form of "change amount", and defines the change amount of the opening width and the misalignment height as:

[0404] ;

[0405] where is the change amount of the maximum opening width, is the change amount of the maximum misalignment height, a positive value indicates an increase relative to the historical record, and a negative value indicates a decrease;

[0406] A growth threshold is given in combination with engineering experience and where is the growth limit value of the maximum opening width, is the growth limit value of the maximum misalignment height, and a developing crack judgment variable can be constructed:

[0407] ;

[0408] where is a Boolean variable indicating whether the crack is a developing crack, when it indicates that there is a significant increase in the opening or misalignment between two detections that exceeds the growth threshold, and it can be judged as a developing crack, when that is, when there is no historical record, it can be defaulted to , indicating that it is not yet identified as a developing crack;

[0409] After obtaining the over-limit judgment and the developing crack judgment , they are combined and mapped to the risk level according to the preset level mapping rule, and the risk level is denoted as , which can take several discrete levels, for example corresponding to "low, lower, medium, high" risk respectively, and the present embodiment can define the mapping function as:

[0410] ;

[0411] where is the risk level mapping function, for example when and At that time, it can be made Take the lowest level, when or When at least one of the conditions is true, the risk level is increased according to the preset rules. When both conditions are true, the highest risk level can be applied. The specific mapping relationship can be configured in engineering applications in combination with specifications and experience.

[0412] Finally, determine whether the limit has been exceeded. Determination of whether it is a developing crack Risk level and re-inspection mark The combined results are denoted as:

[0413] ;

[0414] in The final output data structure for this step clearly indicates whether the current crack exceeds the limit, whether it is a developing crack, the risk level, and whether a re-inspection is required.

[0415] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0416] This invention addresses the technical challenge of accurately acquiring the crack opening width and misalignment height in deep foundation pit support structures. It constructs an integrated algorithm chain encompassing "short-sequence and multi-polarization imaging acquisition - crack masking and centerline and left / right lateral domain division - local 3D reconstruction and normal field optimization of the crack neighborhood - absolute scale calibration of the reference object - millimeter-level cross-sectional measurement along the centerline - crack feature encoding and temporal determination." By transforming 2D image observations step-by-step into left / right implicit surfaces, normal fields, and cross-sectional measurement results under a unified imaging model and absolute scale, and ultimately solidifying them into structured crack feature codes, this invention enables stable output of key 3D engineering quantities such as crack opening width and misalignment height with millimeter-level accuracy. Combined with reprojection residual-driven reconstruction confidence and re-inspection marking, it achieves the expected technical effects of high-precision quantitative characterization of deep foundation pit crack morphology, risk level determination, and traceable management.

[0417] In terms of algorithm structure design, this invention makes targeted improvements to address the characteristics of cracks that are "slender, have weak texture, are easy to bridge across cracks, and are highly sensitive to scale":

[0418] In one aspect, the bilateral geometry representation of the left and right implicit surfaces is introduced in the vicinity of the crack mask, and the low occupancy prior and the cross-crack repulsive potential are explicitly applied along the normal direction of the crack center line, so that the local reconstruction avoids the cross-crack filling and surface sticking problems which are prone to occur under the traditional photometric consistency driving at the model level;

[0419] In another aspect, the normal observation obtained by the multi-polarization image solution is embedded in the joint optimization target of the implicit geometry with a differentiable consistency term, instead of a post-processing loose fusion, so that the reality and stability of the surface geometry are improved from the normal field level.

[0420] Meanwhile, the present application unifies the reference absolute scale calibration, the local coordinate system and the cross-section sequence construction along the center line, and the nearest point pairing measurement in the same implicit geometry framework, so that the consistency in the data domain, the coordinate system and the measurement scale is ensured from the pixel-level crack segmentation, the local three-dimensional reconstruction to the millimeter-level opening width and the fault height output, so that the technical cooperation of the overall algorithm link is more conducive to be realized, and the accuracy and reliability of the crack three-dimensional measurement and risk assessment are significantly improved.

Claims

1. A method for detecting deep foundation pit crack features based on image recognition, characterized in that, include: S1. Acquire short-time image sequences and polarization image groups with different polarization angles of the surface of the deep foundation pit support structure, and complete the camera internal parameters and camera external parameters calibration. S2. Using the image sequence as input, perform crack semantic segmentation to generate a crack mask, and determine and output the crack centerline and its left and right sides accordingly. S3. Input the crack mask, crack centerline, left domain, right domain, image sequence, camera intrinsic and extrinsic parameters, and polarization image group. Perform local 3D reconstruction in the crack mask neighborhood, establish the left implicit surface and the right implicit surface, and jointly optimize by combining polarization information to obtain the normal field and reprojection residual statistics. S4. Using the left and right implicit surfaces, image sequence, camera intrinsic parameters, and camera extrinsic parameters as input, identify reference objects with known dimensions and obtain absolute scale parameters accordingly. Perform scale consistency correction on the two implicit surfaces and output the absolute scale parameters and the scale-corrected left and right implicit surfaces. S5. Using the crack centerline, the scale-corrected left and right implicit surfaces, the normal field, and the absolute scale parameters as input, generate a sequence of sections along the centerline, and calculate the opening width and misalignment height of each section based on the two implicit surfaces and the normal field. S6. Using the cross-section sequence, the opening width and misalignment height of each cross-section, the crack centerline and the statistical value of the reprojection residual as input, generate crack feature codes according to engineering rules. S7. Input the crack feature code and retrieve the historical code. Based on the threshold and time sequence comparison, determine and generate the detection conclusion. S5 includes establishing a local coordinate system in a local three-dimensional space along the crack centerline with the tangent of the centerline as the axis and the normal of the centerline as the cross-section normal, and generating a cross-section sequence orthogonal to the normal of the centerline in this coordinate system; Within each section, the opening width is calculated based on the nearest point pairing of the scale-corrected left implicit surface and the scale-corrected right implicit surface, and the misalignment height is obtained by calculating the difference in coordinate components of the two implicit surfaces in the height direction defined by the normal field within the section. The opening width and misalignment height are converted to millimeter-level values ​​based on absolute dimensional parameters.

2. The method for detecting deep foundation pit crack features based on image recognition according to claim 1, characterized in that, S1 specifically refers to: At least two frames of images with short temporal attributes are continuously acquired within the same field of view on the surface of the deep foundation pit support structure to form an image sequence, and corresponding images are acquired under at least two different polarization angle polarization transmission axis settings to form a polarization image group, while recording the polarization angle parameters of each polarization image. Complete camera calibration to obtain camera intrinsic parameters including focal length, principal point coordinates, radial distortion coefficient and tangential distortion coefficient, and estimate camera extrinsic parameters including rotation and translation for each frame in the image sequence; Distortion correction and resolution unification are performed on image sequences and polarization image groups to match them with an imaging model that aligns with camera intrinsic and extrinsic parameters. Output image sequences, polarization image groups, camera intrinsic parameters, and camera extrinsic parameters.

3. The method for detecting deep foundation pit crack features based on image recognition according to claim 1, characterized in that, S2 specifically refers to: The image sequence is input into an image recognition network to perform crack semantic segmentation and generate a pixel-level crack probability map. The pixel-level crack probability map is converted into a crack mask according to preset rules, and the crack mask is subjected to connectivity screening and morphological refinement to form a continuous crack region. The crack mask is skeletonized to obtain a single-pixel wide connected path. When there are branches in the path, the main trunk is selected based on the path continuity and length to generate the crack center line. At each discrete point of the crack centerline, the normal direction is determined on the image plane according to the local tangential direction of the centerline. The two sides of the crack mask are divided by the normal direction and merged to obtain the left and right domains.

4. The method for detecting deep foundation pit crack features based on image recognition according to claim 1, characterized in that, S3 specifically refers to: Within the neighborhood defined by the crack mask, image sequences, camera intrinsic parameters and camera extrinsic parameters are used as inputs. Neural radiation field or three-dimensional Gaussian sputtering is used for local three-dimensional reconstruction. The reprojection residual is calculated with semantic mask weighted photometric consistency as the basic optimization objective. Create a left implicit surface and a right implicit surface in the left and right domains respectively to form a two-sided representation; The normals are observed using a polarization image group solution algorithm, and a consistency constraint is imposed on the normal outputs of the left and right implicit surfaces to match the normal observations. The gap region is determined based on the normal zone of the crack centerline. An occupation prior is applied to the volume density or occupation value in the gap region to make it approach a low value. The cross-crack repulsion potential is defined to limit the minimum distance between the left and right implicit surfaces in the gap region. The photometric consistency, polarization consistency, semantic mask weighting, occupied prior, and cross-slit repulsion potential are jointly optimized in a differentiable manner to output the left implicit surface, right implicit surface, and normal field, and update the reprojection residual statistics.

5. The method for detecting deep foundation pit crack features based on image recognition according to claim 1, characterized in that, S4 specifically refers to: Using the left and right implicit surfaces, as well as the image sequence, camera intrinsic parameters, and camera extrinsic parameters as input, identify at least one reference object with a known size in the image sequence; Based on the camera's intrinsic and extrinsic parameters, the pixel boundaries of the reference object are projected onto the left and right implicit planes to obtain the corresponding geometric position and current size of the reference object on the left and right implicit planes. The absolute scale parameter is obtained by using the proportional relationship between the known size of the reference object and the corresponding current size, and a uniform absolute scale parameter is obtained by using the consistency criterion when there are multiple reference objects. The coordinate scale consistency of the left and right implicit surfaces is corrected using an absolute scale parameter so that the two implicit surfaces are expressed under the same metric scale. Output the absolute scale parameters and the scale-corrected left implicit surface and the scale-corrected right implicit surface.

6. The method for detecting deep foundation pit crack features based on image recognition according to claim 1, characterized in that, S6 specifically refers to: The inputs include the cross-section sequence, the opening width and misalignment height of each cross-section, the crack centerline, the reprojection residual statistics, and the external polarization angle parameters of the camera. The opening width and misalignment height of each section are statistically calculated to obtain the opening width distribution statistics and misalignment height distribution statistics; The orientation of the crack is calculated based on the tangential direction of the crack centerline, and curvature statistics are performed using the curvature sequence of the crack centerline. The centerline of the crack is mapped to the scene coordinate system using camera extrinsic parameters to generate spatial positioning information; The reconstruction confidence level is obtained by mapping the reprojection residual statistics with a monotonically decreasing method. Crack numbers are assigned according to preset unique identifier generation rules; The crack feature code is composed of crack number, spatial location information, opening width distribution statistics, misalignment height distribution statistics, direction and curvature, reconstruction confidence and acquisition conditions, and then outputs the crack feature code.

7. The method for detecting deep foundation pit crack features based on image recognition according to claim 1, characterized in that, S7 specifically refers to: The crack feature code is used as input, and when a historical crack feature code exists, it is retrieved by association using the crack number. The distribution statistics of opening width and misalignment height in the crack feature encoding are compared based on a preset threshold or an adaptive threshold to determine whether the limit is exceeded. The validity of the determination of whether the limit is exceeded is corrected based on the reconstruction confidence and the acquisition conditions. A re-inspection mark is generated when the reconstruction confidence is lower than the threshold or the acquisition conditions do not meet the requirements. When historical crack features are encoded, the distribution statistics of opening width and the distribution statistics of misalignment height are compared over time. The amount of change or the rate of change is calculated and compared with the growth threshold to determine whether it is a developing crack. Based on the level mapping rules, whether the limit is exceeded and whether it is a developing crack are mapped to risk levels. Detection conclusions are generated and output, including whether the limit is exceeded, whether it is a developing crack, risk level, and re-inspection mark.

Images