A tunnel fault image analysis method based on image processing

Abstract

The application discloses a tunnel fault image analysis method based on image processing, and aims at solving the problems that the surface image cannot be corresponded with the spatial profile of the transient seismic wave method, the geological radar and the resistivity imaging profile, and the abnormality in a certain distance in front cannot be mapped to the working face, wherein the anisotropic regularization is performed through the Fourier neural operator, the cross-modal Transformer registration based on the differentiable attention mask and the curve mileage and the cross-section polar coordinate position coding of the sector geometry is combined, the analytical geometry mapping and the uncertainty propagation are matched, the working face structure trace and the water seepage texture evidence are fused, the accurate positioning of the abnormality in the working face, the position confidence range estimation and the risk grading early warning are realized, the stripe artifact is inhibited, the abnormal boundary is reserved, and the technical effects of outputting the structured results such as the mileage post number, the azimuth angle and the distance interval are achieved.

Description

Technical Field

[0001] This invention relates to the field of image processing, and more particularly to a method for analyzing tunnel fault images based on image processing. Background Technology

[0002] With the rapid development of tunnel engineering, advanced geological prediction is widely used in predicting fractured zones, water-rich areas, and adverse geological conditions. Among existing technologies, tomographic methods such as transient seismic wave method, ground-penetrating radar, and resistivity imaging can provide profile information such as abnormal reflections, low-velocity zones, or low- or high-resistivity zones at a certain distance in front of the tunnel face. At the same time, digital imaging of the tunnel face is used to record structural traces, seepage, and surrounding rock structural characteristics to assist in construction and risk assessment.

[0003] In recent years, research has also emerged on image processing and learning methods for stripe artifact suppression, super-resolution reconstruction, and cross-modal feature registration, attempting to improve anomaly detection and spatial localization capabilities. In engineering practice, a common approach is to simplify the projection of the profile according to the mileage markers and centerline geometry, and then manually annotate and issue warnings on the tunnel face based on experience.

[0004] However, existing technologies still have the following shortcomings:

[0005] (1) The cross-modal spatial correspondence is unclear: the imaging geometry, transmission and reception sectors and fields of view of different devices are significantly different. In addition, the curvature and pose of the tunnel centerline change. Existing methods mostly use rigid / affine or straight-line mileage projection, which makes it difficult to accurately analyze and map the profile anomaly "a certain distance ahead" to the tunnel face coordinates. It lacks pixel-level correspondence and error quantification.

[0006] (2) Resolution anisotropy and artifacts affect positioning accuracy: The tomographic profile has inconsistent resolution in the longitudinal and transverse directions and is easily affected by artifacts such as stripes and bands. Traditional filtering or general noise reduction can easily cause abnormal boundary blurring and loss of detail, making it difficult to achieve reconstitution and correction of orientation perception while maintaining clear edges.

[0007] (3) Insufficient fusion and uncertainty characterization: Existing cross-modal registration usually relies on empirical similarity measurement, lacking constraints based on sensor sector geometry and measurement coding oriented towards curve mileage and cross-sectional coordinates; the uncertainty of anomaly locations is difficult to propagate and quantify along the analytical mapping chain; the fusion of face image evidence (construction traces, seepage textures) and profile anomalies lacks explicit modeling of credibility and conflict, making it difficult to form standardized coordinates and risk classification outputs that can be used for BIM / GIS publishing.

[0008] Therefore, a tunnel fault image analysis method that can overcome the shortcomings of the existing technology is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0009] One objective of this invention is to propose a tunnel fault image analysis method based on image processing. Addressing the problems of unclear spatial correspondence between existing surface images and transient seismic wave methods, ground-penetrating radar, and resistivity imaging profiles, difficulty in analytically mapping anomalies to the tunnel face, and lack of uncertainty and reliability fusion, this invention proposes technical solutions including acquisition and joint calibration, Fourier neural operator anisotropic renormalization, introduction of sector geometrically differentiable attention masks and curve mileage, cross-modal transformer registration with cross-section polar coordinate position encoding, analytical geometric mapping and Jacobian error propagation based on B-spline curve mileage and cross-section shape parameterization, evidence deep learning fusion, and structured coordinate output. This invention achieves the technical effects of accurately locating anomalies to the tunnel face, providing location confidence ranges and risk levels, suppressing fringe artifacts while preserving anomaly boundaries, improving cross-modal registration and temporal stability, and supporting standardized release of mileage markers, azimuth angles, and distance intervals.

[0010] A tunnel fault image analysis method based on image processing according to an embodiment of the present invention is characterized by comprising the following steps:

[0011] S1. Acquire tunnel face images and tomographic profiles, and simultaneously record tunnel centerline mileage, sensor pose, sensor intrinsic and extrinsic parameters, and transmit and receive sector parameters. Output initial data including tunnel face images, tomographic profiles, mileage and pose parameters, sensor intrinsic and extrinsic parameters, and sector parameters.

[0012] S2. Input the tomographic profile, use the Fourier neural operator with anisotropic resolution renormalization to correct the longitudinal and transverse resolution and fringe artifacts and retain abnormal boundaries, and output the renormalized profile.

[0013] S3. Using the re-reconstructed profile and face image, mileage and pose parameters, sensor intrinsic and extrinsic parameters, and sector parameters as inputs, a differentiable attention mask based on sensor sector geometry is introduced in the cross-modal Transformer registration module to limit the geometrically accessible area in the cross-attention layer. The geometric location code composed of tunnel centerline mileage and cross-section polar coordinates is injected into the query and key-value embedding. The output is the dense correspondence field and pixel-level uncertainty map corresponding to the pixels of the re-reconstructed profile and the pixels of the face image, and the output is the anomaly saliency map aligned with the coordinates of the re-reconstructed profile.

[0014] S4. Using dense correspondence field, pixel-level uncertainty map, anomaly salience map, mileage and pose parameters, sensor intrinsic and extrinsic parameters as input, in the analytical geometry mapping module, based on the curve mileage parameterization and cross-sectional shape parameterization of the tunnel centerline, the coordinates of the abnormal pixels indicated by the anomaly salience map are mapped to the tunnel face coordinates through the analytical transformation from sensor coordinates to tunnel coordinates. Error propagation is performed according to the pixel-level uncertainty map, and the tunnel face coordinate positioning result and position confidence range are output.

[0015] S5. Using the face coordinate positioning result, the position confidence range, and the face image in the initial data as input, first extract the structural trace features and seepage texture features from the face image, and then use evidence deep learning in the evidence fusion module to fuse the face coordinate positioning result with the above features, and output the fused positioning result and risk level.

[0016] S6. Using the fused positioning results and risk level as input, generate structured coordinate output, including mileage station number, azimuth angle, distance range to the working face, and working face unfolded icon annotation, for the location and early warning of fault anomaly landing points.

[0017] Optionally, step S1 specifically includes:

[0018] The tomographic profile is at least one of transient seismic wave image, ground-penetrating radar image, and resistivity imaging image;

[0019] A hardware clock with a unified time base is used to synchronize and trigger the tunnel face image and tomographic profile, and record the timestamp, so that the tunnel face image and tomographic profile correspond within the same acquisition cycle;

[0020] The tunnel centerline mileage is obtained through the mileage marker identification and mileage measurement unit, and the relationship between mileage and cross-section position is recorded using the fitted tunnel centerline curve.

[0021] The intrinsic parameters of the face imaging device are calibrated using a calibration board and a marker point array. The extrinsic parameters and pose parameters of the sensor are jointly calibrated using an external reference coordinate system and field control points. The calibrated intrinsic and extrinsic parameters and pose parameters of the sensor are obtained.

[0022] The transmission and reception geometry of the tomography equipment is recorded. The sector parameters include sector angle, incident angle, transmission and reception distance, sampling time window and scanning frequency band, and are stored in association with mileage and pose parameters.

[0023] The output includes timestamps, face images, tomographic profiles, mileage and pose parameters, sensor intrinsic and extrinsic parameters, and initial data for sector parameters.

[0024] Optionally, step S2 specifically includes:

[0025] Using the tomographic profile as input and sector parameters as conditional input, a Fourier neural operator with a frequency domain mixing layer is employed. In the frequency domain mixing layer, directional encoding for the longitudinal and transverse directions of the tomographic profile is introduced, and the frequency domain weights are gated based on the sector angle, incident angle, and sampling time window.

[0026] When training the Fourier neural operator, edge fidelity loss and stripe artifact suppression loss are used in combination, and consistency regularization is applied to the outputs of adjacent positions within the same sector.

[0027] After processing by the Fourier neural operator, a renormalized profile is output. The longitudinal and lateral resolutions of the renormalized profile are corrected and the abnormal boundaries remain clear.

[0028] Optionally, step S3 specifically includes:

[0029] Using the renormal profile, face image, mileage and pose parameters, sensor intrinsic and extrinsic parameters, and sector parameters as inputs, a differentiable attention mask is generated in the cross-attention layer of the cross-modal Transformer registration module based on the sector parameters. The differentiable attention mask defines the geometrically reachable area based on the sector angle, incident angle, transmit and receive distance, sampling time window, and scanning frequency band, and softens the mask boundary. It is also dynamically updated with mileage and pose parameters.

[0030] The Earth metric location code composed of the tunnel centerline mileage and cross-sectional polar coordinates is injected into the query and key-value embedding to guide attention along the tunnel centerline and cross-section.

[0031] An anomaly saliency decoding head is set up in the cross-modal Transformer registration module. It takes the cross-attention fusion features as input and the response domain of the differentiable attention mask and the geometric location encoding as constraints. An anomaly saliency map aligned with the coordinates of the renormal profile map is generated through multi-scale upsampling and convolution.

[0032] When training the cross-modal Transformer registration module, mileage monotonicity constraints and local smoothness constraints are imposed on the dense correspondence field, and forward and backward cyclic consistency loss is introduced.

[0033] After processing by the cross-modal Transformer registration module, the output includes the dense correspondence field and pixel-level uncertainty map corresponding to the pixels of the renormal profile and the pixels of the face image, and also outputs the anomaly significance map.

[0034] Optionally, step S4 specifically includes:

[0035] Using dense correspondence field, pixel-level uncertainty map, anomaly saliency map, renormal profile map, mileage and pose parameters, and sensor intrinsic and extrinsic parameters as inputs, the analytical geometry mapping module first performs threshold segmentation and connected component analysis on the anomaly saliency map to obtain the coordinates of the abnormal pixels. Then, based on the mileage and pose parameters and the sensor intrinsic and extrinsic parameters, an analytical transformation function from sensor coordinates to tunnel coordinates is established. B-splines are used to parameterize the tunnel centerline curve mileage and simultaneously parameterize the cross-sectional shape to generate a local tangent plane of the tunnel face. The coordinates of the abnormal pixels in the renormal profile map are projected onto the local tangent plane of the tunnel face for the first time after the analytical transformation to output the initial tunnel face coordinates. Then, the initial tunnel face coordinates are compensated for a second time based on the curvature and torsion of the tunnel centerline to output the compensated tunnel face coordinates.

[0036] Subsequently, in the error propagation layer, the sensitivity of the analytical transformation to mileage and pose parameters and sensor intrinsic and extrinsic parameters is calculated based on the pixel-level uncertainty map to form a Jacobian matrix, and the pixel-level uncertainty and the Jacobian matrix are combined to form the position confidence range.

[0037] The compensated face coordinates are back-projected to the reconstructed profile coordinates through the analytical transformation. The reprojection error is calculated and the B-spline parameters and the control points of the cross-sectional shape parameterization are iteratively updated until the reprojection error meets the set threshold. The face coordinate positioning result and position confidence range are then output.

[0038] Optionally, step S5 specifically includes:

[0039] Using the face coordinate positioning results, position confidence range, and face image as input, the face image is first subjected to specular reflection suppression and brightness normalization to obtain a preprocessed face image. Then, using the preprocessed face image as input, multi-scale linear structure enhancement and connected component analysis are used to extract construction trace features, and color and brightness joint segmentation and texture consistency analysis are used to extract seepage texture features.

[0040] Subsequently, in the evidence fusion module, evidence deep learning is used to fuse the face coordinate positioning results with the position confidence range, structural trace features, and seepage texture features. During the fusion process, evidence parameters including confidence level, conflict level, random uncertainty, and model uncertainty are output, and the fusion weights are adaptively adjusted according to the conflict level. At the same time, mileage and pose parameters are used as input conditions to perform temporal smoothing and drift penalty on the fusion positioning results corresponding to adjacent mileages. Finally, the fusion positioning results and risk level are output.

[0041] Optionally, step S6 specifically includes:

[0042] Using the fused positioning result, risk level, and location confidence range as input, the fused positioning result is represented as cross-sectional polar coordinates in the face coordinate system, and the angle is read to obtain the azimuth angle;

[0043] The fused positioning result is then associated with mileage and pose parameters, and the longitudinal position of the fused positioning result is converted into a mileage station number based on the curve mileage parameterization of the tunnel centerline.

[0044] Then, the component of the position confidence range along the normal direction of the working face is calculated as the distance interval to the working face;

[0045] Then, a working face unfolding diagram is generated using the mileage marker and azimuth as coordinates, and color-coded according to the risk level and line width coded according to the distance range to the working face to form the working face unfolding diagram annotation;

[0046] The final output includes structured coordinates with mileage markers, azimuth angles, distance ranges to the working face, risk levels, and unfolded diagrams of the working face. These coordinates are organized and published according to the data fields of the Building Information Modeling (BIM) interface and the Geographic Information System (GIS) interface.

[0047] The beneficial effects of this invention are:

[0048] (1) Achieve dense registration and analytical mapping with cross-modal geometric consistency: By using a differentiable attention mask based on sector geometry and geodesic location coding of curve mileage and cross-sectional polar coordinates, combined with mileage monotonicity and forward and backward cyclic consistency constraints, and analytical mapping of B-spline curve mileage and cross-sectional shape parameterization and curvature and torsion compensation, the anomaly of "a certain distance ahead" can be uniquely and continuously mapped to the working face and the position confidence range obtained by pixel-level uncertainty propagation can be output, which significantly improves the reliability of spatial alignment and positioning.

[0049] (2) Improve the quality of tomographic profile imaging and anomaly detectability: Fourier neural operators with frequency domain mixing layer and directional coding and sector parameter gating are used to achieve anisotropic renormalization of longitudinal and transverse resolution and suppression of stripe artifacts. The edge fidelity loss and adjacent sector consistency regularization are used to maintain clear anomaly boundaries and spatiotemporal stability, providing high-quality input for subsequent registration and mapping.

[0050] (3) Achieve evidence-based multi-source fusion decision-making and standardized release: Utilize evidence-based deep learning to fuse the tunnel face construction traces, seepage textures and positioning results, explicitly output the confidence level, conflict level and randomness, model uncertainty, adaptively adjust the fusion weights, and combine mileage and pose for temporal smoothing and drift penalty, and finally generate mileage station number, azimuth angle, distance range to the tunnel face and risk level and tunnel face unfolding icon annotation, which is convenient for BIM / GIS integration and early warning application. Attached Figure Description

[0051] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0052] Figure 1 This is a flowchart of a tunnel fault image analysis method based on image processing proposed in this invention. Detailed Implementation

[0053] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0054] refer to Figure 1 A method for analyzing tunnel fault images based on image processing, characterized by the following steps:

[0055] S1. Acquire tunnel face images and tomographic profiles, and simultaneously record tunnel centerline mileage, sensor pose, sensor intrinsic and extrinsic parameters, and transmit and receive sector parameters. Output initial data including tunnel face images, tomographic profiles, mileage and pose parameters, sensor intrinsic and extrinsic parameters, and sector parameters.

[0056] S2. Input the tomographic profile, use the Fourier neural operator with anisotropic resolution renormalization to correct the longitudinal and transverse resolution and fringe artifacts and retain abnormal boundaries, and output the renormalized profile.

[0057] S3. Using the re-reconstructed profile and face image, mileage and pose parameters, sensor intrinsic and extrinsic parameters, and sector parameters as inputs, a differentiable attention mask based on sensor sector geometry is introduced in the cross-modal Transformer registration module to limit the geometrically accessible area in the cross-attention layer. The geometric location code composed of tunnel centerline mileage and cross-section polar coordinates is injected into the query and key-value embedding. The output is the dense correspondence field and pixel-level uncertainty map corresponding to the pixels of the re-reconstructed profile and the pixels of the face image, and the output is the anomaly saliency map aligned with the coordinates of the re-reconstructed profile.

[0058] S4. Using dense correspondence field, pixel-level uncertainty map, anomaly salience map, mileage and pose parameters, sensor intrinsic and extrinsic parameters as input, in the analytical geometry mapping module, based on the curve mileage parameterization and cross-sectional shape parameterization of the tunnel centerline, the coordinates of the abnormal pixels indicated by the anomaly salience map are mapped to the tunnel face coordinates through the analytical transformation from sensor coordinates to tunnel coordinates. Error propagation is performed according to the pixel-level uncertainty map, and the tunnel face coordinate positioning result and position confidence range are output.

[0059] S5. Using the face coordinate positioning result, the position confidence range, and the face image in the initial data as input, first extract the structural trace features and seepage texture features from the face image, and then use evidence deep learning in the evidence fusion module to fuse the face coordinate positioning result with the above features, and output the fused positioning result and risk level.

[0060] S6. Using the fused positioning results and risk level as input, generate structured coordinate output, including mileage station number, azimuth angle, distance range to the working face, and working face unfolded icon annotation, for the location and early warning of fault anomaly landing points.

[0061] In this specific embodiment, S1 specifically refers to:

[0062] A unified time-base hardware clock is used to synchronize and trigger the face imaging equipment and each tomographic device, and timestamps are recorded. The acquisition cycle index is defined as follows. and Establish time synchronization constraints for the total number of data acquisition cycles:

[0063] ;

[0064] in, For the first The timestamp of the palm face imaging device for each acquisition cycle. For the first One acquisition cycle tomography mode timestamp, The maximum allowable synchronization error, This is a tomographic modal index, with values ​​taken from the set. Transient seismic wave method, ground-penetrating radar, resistivity imaging ;

[0065] To achieve a geometric correspondence between mileage and space, a curve fitting was performed on the tunnel centerline based on mileage identification and measurement to establish a centerline and time-mileage mapping with mileage as a parameter:

[0066] ;

[0067] in, To use the curve mileage Let be the three-dimensional centerline position vector of the independent variable. For three-dimensional real space, To timestamp A function that maps to the corresponding mileage. To provide hardware time readings under a unified time base;

[0068] To achieve geometrically consistent calibration of multiple sensors, a calibration board and a marker point array were used to complete the intrinsic parameter calibration of the face imaging equipment. Furthermore, under the constraints of an external reference coordinate system and field control points, external parameters were used for joint pose calibration of each sensor. The pose is represented by a homogeneous transformation as follows:

[0069] ;

[0070] in, For the first Each acquisition cycle is from the sensor coordinate system To the world coordinate system Rigid body transformation, Let be a rotation matrix. It is a translation vector. It is a three-dimensional zero vector. It is a three-dimensional special Euclidean group. It is a special orthogonal group in three dimensions. and Indicates a palm face imaging device;

[0071] The intrinsic parameter matrix of the face imaging device is denoted as:

[0072] ;

[0073] in, This is the intrinsic parameter matrix. and These are the equivalent focal length pixel values ​​in the horizontal and vertical directions. and The coordinates of the principal imaging point;

[0074] For tomographic equipment, the transmit and receive geometry is recorded and summarized into sector parameter vectors:

[0075] ;

[0076] in, For modality sector parameters, For sector angle, Angle of incidence The distance between the transmitter and receiver, The endpoint vector of the sampling time window. The endpoint vector of the scanning frequency band;

[0077] The above information is associated and stored as an initial data packet according to the collection cycle:

[0078] ;

[0079] in, For the first Data sets from each collection period Image of the facet of the machine. For modality Chromatographic profile, For the corresponding mileage, The pose of each sensor to the world coordinate system. This is an internal reference for the face imaging equipment. For sector parameters, For timestamps;

[0080] Through the above-mentioned synchronization, mileage-geometric modeling and joint calibration, strict alignment and traceable correlation of face images, tomographic profiles, mileage and pose, sensor intrinsic and extrinsic parameters and sector parameters are achieved within the same acquisition cycle.

[0081] In this specific embodiment, S2 specifically refers to:

[0082] Using tomographic profiles as input and sector parameters as conditions, a Fourier neural operator containing a frequency domain mixing layer and directional coding is constructed to achieve anisotropic renormalization and stripe artifact suppression at both longitudinal and transverse resolutions.

[0083] First, input the profile. It is fed into the frequency domain mixing layer, where for Two-dimensional Fourier transform, For modality The original chromatographic profile, Spatial pixel coordinates, For frequency coordinates, Using the Fourier transform operator, direction-aware weighting and sector-conditional gating are performed in the frequency domain to obtain:

[0084] ;

[0085] in For the frequency domain representation of the renormalization result, For parameters Controlled frequency domain mixing weights and explicit receive direction coding To model anisotropy in both the longitudinal and transverse directions, This is a gate function based on sector parameters. For modality sector parameter vector, For sector corner, For the angle of incidence, For the distance between the transmitter and receiver, For sampling time window, For scanning frequency band, It is a directional encoding vector composed of frequency direction and scale;

[0086] The renormal profile of the spatial domain is then obtained through inverse Fourier transform:

[0087] ;

[0088] in For the output renormalized profile, This is the inverse Fourier transform operator;

[0089] To ensure edge fidelity and suppress stripe artifacts, while improving the consistency of outputs at adjacent positions within the same sector, a composite loss is used during the training phase:

[0090] ;

[0091] in For the total loss, For edge fidelity loss, The reference output is obtained from the edge-guided prior. To reduce the loss of fringe artifacts, Output consistency regularization for adjacent mileage positions within the same sector for The output set at adjacent positions, and These are non-negative weighting coefficients;

[0092] Through the above frequency domain mixing, directional coding and gating, and joint optimization of edge fidelity and consistency, the following results are obtained: With longitudinal and transverse resolution corrected, fringe artifacts significantly suppressed, and anomalous boundaries remaining clear, high-quality input is provided for subsequent cross-modal registration and analytical geometric mapping.

[0093] In this specific embodiment, S3 specifically refers to:

[0094] Using the renormalized profile and face image as input, and combining odometer and pose parameters, sensor intrinsic and extrinsic parameters, and sector parameters, a differentiable attention mask based on sensor sector geometry is explicitly introduced in the cross-attention layer of the cross-modal transformer to define the geometrically accessible region and dynamically update it with odometer and pose. Let the renormalized profile pixel be denoted as... The number of pixels on the palm face is The query vector on the cross-section side and the key vector on the face side are denoted as follows: and The mask weight is Then, the cross-attention with a mask is written as:

[0095] ;

[0096] in For the first The position of the first section is related to the first Attention weights for each palm face position. For embedded dimensions, The mask gate temperature coefficient, It is the numerical stability constant. Indicates index Normalized exponent calculation;

[0097] To constrain cross-modal alignment on curve mileage metrics and cross-sectional coordinates, geodesic location encoding is introduced and injected into the query and key, denoted as mileage. The polar coordinates of the cross section are ,but:

[0098] ;

[0099] in, For position encoding vectors, For multi-frequency embedding functions, For the first The tunnel centerline curve mileage corresponding to each profile location and These represent the radial distance and azimuth angle of the location in the cross-sectional coordinate system, respectively. For the position of the working face Corresponding isomorphic position encoding;

[0100] After cross-attention fusion, the dense correspondence field and pixel-level uncertainty are obtained through decoding. The mapping function is denoted as... Uncertainty is ,have:

[0101] ;

[0102] in Sectional position The corresponding pixel coordinates on the facet image. This corresponds to the pixel-level uncertainty scalar;

[0103] Simultaneously, an anomaly saliency decoding head is set up, taking the cross-attention fusion features as input and constrained by reachability mask and geodesic location encoding. After multi-scale upsampling and convolution, an anomaly saliency map aligned with the renormalized profile coordinates is generated. During the training phase, ingress odometer monotonicity, local smoothing, and forward / backward cyclic consistency constraints are applied to the dense correspondence field. The composite regularization is as follows:

[0104] ;

[0105] in The total loss for the regularization term. To lose weight, The ReLU function is used to ensure the monotonicity of the mileage. To measure local smoothness, the discrete gradient of the corresponding field on the profile grid is used. and These are the L2 norm and L1 norm, respectively. To achieve forward and backward cyclic consistency through the reverse mapping from the face to the profile, the final output includes a dense correspondence field between the pixels of the re-reconstructed profile image and the pixels of the face image, a pixel-level uncertainty map, and an anomaly saliency map aligned with the coordinates of the re-reconstructed profile.

[0106] In this specific embodiment, S4 specifically includes:

[0107] Based on dense correspondence fields, pixel-level uncertainty maps, anomaly saliency maps, and renormalization profile maps, and combined with mileage and pose parameters, sensor intrinsic and extrinsic parameters, the analytical geometry mapping module first performs threshold segmentation and connected component analysis on the anomaly saliency map to obtain a set of anomalous pixels. Then, the correspondence field is used to associate the mileage and local cross-sectional information of each anomalous pixel to determine the location interval of the face to be projected. The threshold segmentation can be expressed as follows:

[0108] ;

[0109] in A set of abnormal pixels, These are the pixel coordinates of the cross-sectional image. This is an abnormally significant value. The significance threshold;

[0110] Subsequently, an analytical transformation chain from sensor coordinates to tunnel coordinates is established, and the first projection is performed on the local tangent plane at the tunnel face. The first connected abnormal pixel is recorded. The corresponding mileage is ,have:

[0111] ;

[0112] in, For pixels The inverse solution is obtained and expressed in the world coordinate system. Three-dimensional points in For mileage From the sensor coordinate system To the world coordinate system homogeneous rigid body transformation, Let be the analytical back-projection function from the pixel to the sensor system in three dimensions, determined by the sensor imaging geometry and medium propagation model. For sensors Internal reference, For tomographic modes The sector parameter vector, where For sector corner, For the angle of incidence, For the distance between the transmitter and receiver, For sampling time window, For scanning frequency bands;

[0113] The tunnel centerline is parameterized using B-splines to obtain the curve mileage parameterization. And construct the Frenet frame Therefore, in Define a local tangent plane on the tunnel face:

[0114] ;

[0115] in, For the tangent plane, It is the normal vector. For any three-dimensional point, For mileage Located at the center line position;

[0116] Will orthogonal projection to Obtain the initial working face coordinates:

[0117] ;

[0118] in The three-dimensional coordinates of the tunnel face obtained from the first projection. It is a 2-norm;

[0119] Considering the non-planar errors caused by the curvature and twist of the centerline, for Secondary compensation will be provided.

[0120] ;

[0121] in The coordinates of the working face after compensation. and mileage The curvature and torsion at that point and These are the primary normal direction and the secondary normal direction (dual normal direction), respectively. and These are compensation coefficients related to the parameterization of the cross-sectional shape;

[0122] In the error propagation layer, a position confidence range is formed based on pixel-level uncertainty and parameter sensitivity, and an input uncertainty vector is constructed. And calculate the Jacobian matrix and covariance propagation:

[0123] ;

[0124] in, This is a sensitivity matrix relating pixel coordinates, odometer, and parameters. The input is the uncertainty covariance, where the pixel uncertainty comes from the pixel-level uncertainty map. Let be the covariance of the working face coordinates;

[0125] The normal vector is used for interval representation to obtain the confidence radius of the distance to the working face:

[0126] ;

[0127] in, To quantify the uncertainty along the normal direction;

[0128] To verify and achieve self-consistent optimization, the compensated working face coordinates are reprojected back to the renormalized profile coordinates via an inverse analytical transformation. This inverse projection is denoted as... And calculate the reprojection error:

[0129] ;

[0130] in, For reprojected pixel coordinates, The L2 norm error is minimized. Iteratively update the B-spline control points and the control points of the cross-sectional shape parameterization until... Up to now, among them The final output includes the tunnel face coordinates, based on the set convergence threshold. Location confidence range The location results are used for subsequent evidence fusion and early warning issuance.

[0131] In this specific embodiment, S5 specifically includes:

[0132] Using the face coordinate positioning results, position confidence range, and face image as input, the preprocessed image is first obtained by specular reflection suppression and brightness normalization. Then, two types of evidence features, namely construction traces and seepage textures, are extracted. In the evidence fusion module, evidence deep learning is used to generate confidence, conflict and uncertainty parameters. Finally, the mileage and pose are combined for temporal smoothing and the fused positioning results and risk level are output.

[0133] In the preprocessing stage, the brightness is first locally normalized and specular highlights are suppressed, and the pixel coordinates are defined as follows: The original image's luminance channel is Local window is Then the normalized brightness is:

[0134] ;

[0135] in, The normalized brightness. and They are respectively The mean and standard deviation of the brightness within the range, It is the numerical stability constant;

[0136] The specular reflection mask is:

[0137] ;

[0138] in This is a function indicating the mirror region. For channel pixel values, Highlight threshold The strength after pretreatment This is the specular suppression coefficient. It is an indicator function;

[0139] The trace feature construction employs a Hessian-based multi-scale linear structure enhancement, where the scale set is denoted as . , In order to scale Top The calculated Hessian matrix has eigenvalues ​​arranged according to... The linear structure response is then sorted as follows:

[0140] ;

[0141] in For trace significance, and for eigenvalues, To suppress the scale parameters of nonlinear structures;

[0142] Water seepage texture features are modeled jointly using brightness thresholding and structure tensor consistency, with the structure tensor set to... Its eigenvalues Define consistency And with brightness threshold To suppress interference, then:

[0143] ;

[0144] in Significant water seepage, The brightness threshold;

[0145] Geometric and image evidence are fed into an evidence deep learning module to form Dirichlet distribution parameters, assuming the number of categories is... (Where, category 1 represents fault anomalies, and the rest are non-fault / background), the uncertainty of the normal distance obtained from step S4 is denoted as And mapped to geometric credibility ,in Define the two input features as scaling coefficients. Each is passed through an evidence network with non-negative outputs. and Obtain the evidence vector and The corresponding Dirichlet parameter is:

[0146] ;

[0147] in For the first Dirichlet parameters of the path evidence, for A dimensional vector of all 1s As non-negative evidence, For the expected class probability, Concentration;

[0148] The conflict degree is defined based on the difference in probabilities between the two paths, and adaptive weighted fusion is performed.

[0149] ;

[0150] in, For L1 conflict metric, and For geometric and image evidence weights, These are the parameters of the merged Dirichlet;

[0151] Depend on To determine the degree of confidence and the degree of uncertainty, define the total evidence. With total concentration Then the trust vector Random uncertainty Model uncertainty can be determined by... and Introducing random inactivation and targeting The class probabilities are obtained by taking the variance estimate of the next random forward pass;

[0152] Then by mileage sequence Perform timing smoothing, let the first... The fusion probability of the cycle is The smoothed probability is Mileage increment is ,but:

[0153] ;

[0154] in, For smoothing coefficients, This is the drift penalty coefficient. For the first Cycle mileage;

[0155] Finally, a risk score is formed by combining the smoothed probability and uncertainty of the fault category, and then quantified into a risk level. Let the fault category index be... The threshold vector is Then the risk score And according to Segment mapping is used to assign risk levels for subsequent structured releases.

[0156] In this specific embodiment, S6 specifically includes:

[0157] Using the fused positioning results, risk level, and location confidence range as input, the fused positioning results are first represented as cross-sectional polar coordinates in the working face coordinate system, and the angle is read as the azimuth. Let the fused 3D positioning point be... Its corresponding centerline projection mileage is The center line is The Frenet frame is And let the normal vector of the face be The azimuth angle is defined with reference to the local tangent plane:

[0158] ;

[0159] in From the center line point Point to fusion point The vector, Let it be the projection of the cutting plane on the tunnel face. Azimuth (in) (Zero angle, counterclockwise is positive) It is a two-parameter arctangent function;

[0160] Subsequently, the fused positioning results are correlated with mileage and pose parameters, and the longitudinal position is converted into a mileage marker based on curve mileage parameterization. The kilometer segment and meter segment are then denoteed as follows:

[0161] ;

[0162] in, For integer segments of kilometers, The remaining meters within one kilometer. The mileage of the merging point along the tunnel centerline is represented by the chainage string "". ";

[0163] Next, the component of the position confidence range along the normal direction of the working face is calculated as the distance interval to the working face, and the mean distance and standard uncertainty of the normal direction are respectively:

[0164] ;

[0165] in The sign distance from the fusion point to the tangent plane of the tunnel face (if On the tangent plane ), The distance range to the working face. For along The positional uncertainty of the direction (obtained by taking the square root of the normal variance given in step S4). Confidence coefficient;

[0166] Then, a face development diagram is generated using the mileage station and azimuth as coordinates. The coordinates and annotation codes of the development diagram are defined as follows:

[0167] ;

[0168] in, The two-dimensional coordinates of the unfolded diagram are shown (horizontal axis is mileage, vertical axis is azimuth). For the objects labeled on the diagram, The risk level obtained in step S5, For the color mapping function based on risk level, width This is a function that maps line widths to distance intervals; the final structured coordinate output is organized as follows:

[0169] ;

[0170] in, For structured records used for publication, For the mileage marker string " ", and catalog and publish according to the data fields of the Building Information Modeling and Geographic Information System interface.

[0171] 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.

Claims

1. A method for analyzing tunnel fault images based on image processing, characterized in that, Includes the following steps: S1. Acquire tunnel face images and tomographic profiles, and simultaneously record tunnel centerline mileage parameters, sensor pose parameters, sensor intrinsic parameters, sensor extrinsic parameters, and corresponding sector parameters for transmission and reception. Output initial data including tunnel face images, tomographic profiles, mileage parameters, pose parameters, sensor intrinsic parameters, sensor extrinsic parameters, and sector parameters. S2. Input the tomographic profile, use the Fourier neural operator with anisotropic resolution renormalization to correct the longitudinal and transverse resolution and fringe artifacts and retain abnormal boundaries, and output the renormalized profile. S3. Using the re-reform profile, tunnel face image, mileage parameters, pose parameters, sensor intrinsic parameters, sensor extrinsic parameters, and sector parameters as input, a differentiable attention mask based on sensor sector geometry is introduced in the cross-attention layer of the cross-modal Transformer registration module to define the geometrically accessible region. Geodesic location codes composed of tunnel centerline mileage and cross-sectional polar coordinates are injected into query embedding and key-value embedding. The output includes a dense correspondence field, a pixel-level uncertainty map, and an anomaly saliency map aligned with the re-reform profile coordinates. The dense correspondence field represents the correspondence between pixels in the re-reform profile and pixels in the tunnel face image, and the pixel-level uncertainty map represents the pixel-level uncertainty corresponding to the dense correspondence field. S4. Using dense correspondence field, pixel-level uncertainty map, anomaly salience map, mileage parameters, pose parameters, sensor intrinsic parameters, and sensor extrinsic parameters as inputs, the analytical geometry mapping module parameterizes the curve mileage and cross-sectional shape based on the tunnel centerline. It then maps the coordinates of the abnormal pixels indicated by the anomaly salience map to the tunnel face coordinates through an analytical transformation from sensor coordinates to the tunnel coordinate system. Based on the pixel-level uncertainty map, it performs error propagation and outputs the tunnel face coordinate positioning result and position confidence range. S5. Using the face coordinate positioning result, position confidence range, and face image in the initial data as input, first extract the structural trace features and seepage texture features from the face image, and then use evidence deep learning in the evidence fusion module to fuse the face coordinate positioning result, position confidence range, structural trace features, and seepage texture features, and output the fused positioning result and risk level. S6. Using the fused positioning results and risk level as input, generate structured coordinate output, which includes mileage station number, azimuth angle, distance range to the working face, and working face unfolded icon annotation, for the location and early warning of fault anomaly landing points.

2. The tunnel fault image analysis method based on image processing according to claim 1, characterized in that, Step S1 is as follows: The tomographic profile is at least one of transient seismic wave image, ground-penetrating radar image, and resistivity imaging image; A hardware clock with a unified time base is used to synchronously trigger and record timestamps for the face images and tomographic profiles, so that the face images and tomographic profiles correspond within the same acquisition cycle. The tunnel centerline mileage is obtained through mileage marker identification and mileage measurement units, and the relationship between mileage and cross-sectional position is recorded using a fitted tunnel centerline curve. The intrinsic parameters of the face imaging device are calibrated using a calibration board and a marker point array. The extrinsic parameters and pose of the sensor are jointly calibrated using an external reference coordinate system and field control points to obtain the calibrated intrinsic, extrinsic, and pose parameters of the sensor. The transmission and reception geometry of the tomography equipment is recorded. The sector parameters include sector angle, incident angle, transmission and reception distance, sampling time window and scanning frequency band, and are stored in association with mileage parameters and pose parameters. The output includes initial data such as timestamps, face images, tomographic profiles, odometer parameters, pose parameters, sensor intrinsic parameters, sensor extrinsic parameters, and sector parameters.

3. The tunnel fault image analysis method based on image processing according to claim 1, characterized in that, Step S2 is as follows: Using the tomographic profile as input and sector parameters as conditional input, a Fourier neural operator with a frequency domain mixing layer is employed. In the frequency domain mixing layer, directional coding for the longitudinal and transverse directions of the tomographic profile is introduced, and the frequency domain weights are gated based on the sector angle, incident angle, and sampling time window. When training the Fourier neural operator, edge fidelity loss and stripe artifact suppression loss are used in combination, and consistency regularization is applied to the outputs of adjacent positions within the same sector. After processing by the Fourier neural operator, a renormalized profile is output. The longitudinal and lateral resolutions of the renormalized profile are corrected and the abnormal boundaries remain clear.

4. The tunnel fault image analysis method based on image processing according to claim 1, characterized in that, Step S3 is as follows: Using the renormal profile, face image, odometer parameters, pose parameters, sensor intrinsic parameters, sensor extrinsic parameters, and sector parameters as inputs, a differentiable attention mask is generated in the cross-attention layer of the cross-modal Transformer registration module based on the sector parameters. The differentiable attention mask defines the geometrically reachable region based on the sector angle, incident angle, transmit and receive distance, sampling time window, and scanning frequency band, and softens the mask boundary. It is also dynamically updated with the odometer parameters and pose parameters. Geodetic location codes composed of tunnel centerline mileage and cross-sectional polar coordinates are injected into query embedding and key-value embedding to guide attention along the tunnel centerline and cross-section. An anomaly saliency decoding head is set in the cross-modal Transformer registration module. It takes the cross-attention fusion features as input and the response domain of the differentiable attention mask and geodesic location encoding as constraints. An anomaly saliency map aligned with the coordinates of the renormal profile map is generated through multi-scale upsampling and convolution. When training the cross-modal Transformer registration module, mileage monotonicity constraints and local smoothness constraints are imposed on the dense correspondence field, and forward and backward cyclic consistency loss is introduced. After processing by the cross-modal Transformer registration module, a dense correspondence field, a pixel-level uncertainty map, and an anomaly significance map are output; wherein, the dense correspondence field is used to represent the correspondence between pixels in the renormal profile and pixels in the face image, and the pixel-level uncertainty map is used to represent the pixel-level uncertainty corresponding to the dense correspondence field.

5. The tunnel fault image analysis method based on image processing according to claim 1, characterized in that, Step S4 is as follows: Using dense correspondence field, pixel-level uncertainty map, anomaly saliency map, renormal profile map, mileage parameters, pose parameters, sensor intrinsic parameters, and sensor extrinsic parameters as inputs, the analytical geometry mapping module first performs threshold segmentation and connected component analysis on the anomaly saliency map to obtain the coordinates of the abnormal pixels. Then, based on the mileage parameters, pose parameters, sensor intrinsic parameters, and sensor extrinsic parameters, an analytical transformation function from sensor coordinates to tunnel coordinates is established. B-splines are used to parameterize the tunnel centerline curve mileage and simultaneously parameterize the cross-sectional shape to generate a local tangent plane of the tunnel face. The coordinates of the abnormal pixels in the renormal profile map are projected onto the local tangent plane of the tunnel face for the first time after the analytical transformation to output the initial tunnel face coordinates. Then, the initial tunnel face coordinates are compensated for a second time based on the curvature and torsion of the tunnel centerline to output the compensated tunnel face coordinates. Subsequently, in the error propagation layer, the sensitivity of the analytical transformation to the odometer parameter, pose parameter, sensor intrinsic parameter and sensor extrinsic parameter is calculated based on the pixel-level uncertainty map, forming a Jacobian matrix, and the pixel-level uncertainty and the Jacobian matrix are combined into a position confidence range; The compensated face coordinates are back-projected to the reconstructed profile coordinates through the analytical transformation. The reprojection error is calculated, and the B-spline parameters and the control points of the cross-sectional shape parameterization are iteratively updated until the reprojection error meets the set threshold. The face coordinate positioning result and position confidence range are then output.

6. The tunnel fault image analysis method based on image processing according to claim 1, characterized in that, Step S5 is as follows: Using the face coordinate positioning results, position confidence range, and face image as input, the face image is first subjected to specular reflection suppression and brightness normalization to obtain a preprocessed face image. Then, using the preprocessed face image as input, multi-scale linear structure enhancement and connected component analysis are used to extract construction trace features, and color and brightness joint segmentation and texture consistency analysis are used to extract water seepage texture features. Subsequently, in the evidence fusion module, evidence deep learning is used to fuse the face coordinate positioning results, position confidence range, structural trace features, and seepage texture features. During the fusion process, evidence parameters including confidence level, conflict level, random uncertainty, and model uncertainty are output. The fusion weights are adaptively adjusted according to the conflict level. At the same time, the mileage parameters and pose parameters are used as input conditions to perform temporal smoothing and drift penalty on the fused positioning results corresponding to adjacent mileages. Finally, the fused positioning results and risk level are output.

7. The tunnel fault image analysis method based on image processing according to claim 1, characterized in that, Step S6 is as follows: Using the fused positioning result, risk level, and location confidence range as input, the fused positioning result is represented as cross-sectional polar coordinates in the face coordinate system, and the angle is read to obtain the azimuth angle; The fused positioning result is then associated with mileage parameters and pose parameters, and the longitudinal position of the fused positioning result is converted into a mileage station number based on the curve mileage parameterization of the tunnel centerline. Then, the component of the position confidence range along the normal direction of the working face is calculated as the distance interval to the working face; Then, a working face unfolding diagram is generated using the mileage marker and azimuth as coordinates. It is then color-coded according to the risk level and line width-coded according to the distance range to the working face, forming a working face unfolding diagram annotation. The final output is a structured coordinate output, which includes mileage station number, azimuth angle, distance range to the working face, risk level, and working face unfolded icon annotation, and is organized and published according to the data fields of the building information model interface and geographic information system interface.

Images