Method for optimizing micro-crack segmentation based on deep learning and super-resolution reconstruction

Abstract

The invention discloses a method for optimizing micro-crack segmentation based on deep learning and super-resolution reconstruction, and the method comprises the steps: cutting an image in real time, obtaining an image block which takes a component as a target main body, and synchronously recording a homography matrix for geometric mapping; selecting an amplification strategy to improve the resolution, and recording a scale mapping relation; inputting the enhanced image block into a double-flow network; adaptive fusion and reconstruction are carried out on the two branch features, and a high-resolution texture image is output; generating a geometrically corrected ortho-image, fusing the geometrically corrected ortho-image with original illumination information, and outputting a corrected image with a known pixel size; identifying cracks, spalling and honeycomb diseases in parallel; generating a unified defect confidence map; calculating real geometric parameters of the BIM in a BIM global coordinate system through coordinate back projection; and generating quantitative defect reports and maintenance suggestions. The method has the advantage that seamless connection between the detection result and the BIM global coordinates is realized.

Description

Technical Field

[0001] This invention relates to the technical field of concrete defect identification, and in particular to a method for optimizing microcrack segmentation based on deep learning and super-resolution reconstruction. Background Technology

[0002] With the rapid development of computer vision technology, the identification and inspection of defects in bridge concrete structures has gradually shifted from manual inspection to machine inspection, which offers numerous advantages such as high recognition accuracy and high inspection efficiency. Currently, bridge inspections are mostly conducted using a BIM + drone + computer vision framework. Generally, bridge inspection includes the following steps:

[0003] Typically, inspection surfaces such as main beams, bridge decks, and piers are first extracted from BIM models or laser point clouds. Then, a full-coverage flight path is planned based on bounding box or meshing strategies. A drone, controlled by GPS-IMU fusion, maintains a constant tilt angle and overlap for continuous shooting to obtain multi-view images suitable for stereo reconstruction. After shooting, correspondences are established between multiple images using SIFT or ORB features. Sparse point clouds are then calculated using bundle adjustment and input into patch-match or semi-global matching algorithms to obtain dense point clouds and texture maps, laying the geometric foundation for subsequent inspection.

[0004] For images with insufficient resolution, two common compensation methods are employed: one is single-image super-resolution technology, which uses convolutional networks such as SRCNN, EDSR, and ESRGAN to directly upscale the image by 2–4 times, relying on texture priors learned from massive datasets to reconstruct details; the other is multi-view fusion methods, which constrain the photometric consistency of overlapping photos of the same component under unified extrinsic parameters, and generate high-pixel textures through gradient preservation or Laplacian pyramid weighting. For perspective distortion caused by overhead shots, RANSAC is often used to calculate the homography matrix, straightening the region of interest to a local plane; pixel length is then converted to physical scale using focal length and shooting distance for subsequent defect quantification.

[0005] For defect identification, mainstream solutions divide the image into blocks of different scales using image pyramids or sliding windows, and then feed them into U-Net, DeepLab, or Swin-Transformer semantic segmentation networks for multi-class detection. Cracks are often first enhanced by Canny or Frangi filtering, and then the center line is extracted using Hough transform and skeletonization. Honeycombing and spalling are jointly discriminated by combining texture operators such as gray-level co-occurrence matrix, local variance, or LBP with depth features, and outputting a multi-channel probability map. In the denoising stage, thresholding and maximum connected component selection are generally used to suppress isolated pixels, and morphological thinning is used to measure the crack width or spalling area.

[0006] For the final positioning stage, pixel coordinates are mapped back to BIM through SfM point cloud and RTK extrinsic parameters. Defect categories and quantitative results are superimposed under global coordinates, and then pushed to the operation and maintenance platform in the format of component number, shooting time and risk level to realize digital file management and inspection closed loop.

[0007] However, current BIM+UAV+computer vision inspection processes mostly employ bounding box trajectories and constant tilt angle photography. Waypoint selection lacks dynamic adaptation to the normal and safe lateral distance of individual components, resulting in occlusion or field-of-view deviation in some images. When shooting distances vary significantly, resolution is compensated for through fixed magnification or general super-resolution networks, making it difficult to guarantee pixel consistency at the physical scale, often leading to resolution jumps between adjacent components. Existing single-image super-resolution methods mainly rely on data-driven texture priors, neglecting structural information such as the direction of rebar placement, resulting in insufficient directionality of crack edges after reconstruction. Multi-view fusion methods require high overlap and accurate extrinsic parameters, which is uneconomical for long-distance inspection of long-span bridges. Perspective straightening typically only applies homography matrices at the image level, failing to fully bind camera extrinsic parameters to the BIM coordinate system, leaving arcsecond-level errors between images and models, limiting the precise quantification of defects. Multi-scale inspection commonly uses pyramids or sliding windows for sorting, but the results at each scale lack uniformity in physical resolution, confidence level, and source index. Post-fusion requires manual threshold adjustment, making it difficult to make maintenance decisions directly under the same coordinate reference. Finally, defect re-projection often relies on sparse SfM point clouds, and the positioning accuracy is limited by feature matching errors and RTK drift, which cannot meet the measurement requirements of millimeter-level crack width and length. Summary of the Invention

[0008] The purpose of this invention is to address the shortcomings of the existing technology by providing a microcrack segmentation method based on deep learning and super-resolution reconstruction. This method establishes a complete process including continuous shooting planning, resolution consistency, orientation-preserving super-resolution, precise geometric alignment, and unified fusion of multi-scale detection results, enabling the final generated defect information to achieve millimeter-level quantitative identification and traceable positioning in the BIM global coordinate system.

[0009] The objective of this invention is achieved through the following technical solutions:

[0010] A microcrack segmentation optimization method based on deep learning and super-resolution reconstruction is applied to an inspection device equipped with a camera. The method includes the following steps:

[0011] S1: Based on the geometric attributes and spatial pose of the components in the BIM model, combined with the inspection equipment trajectory, automatically plan and execute the optimal shooting pose for each component; crop the image in real time to obtain image blocks with the component as the target subject, and simultaneously record the homography matrix used for geometric mapping.

[0012] S2: Calculate the physical resolution of the image block based on the shooting parameters, and adaptively select a magnification strategy to improve the resolution based on the shooting distance and the importance of the components, and record the scale mapping relationship;

[0013] S3: The enhanced image patch is input into the dual-stream network, where the mainstream branch learns general texture features and the directional branch fuses the prior of rebar layout to enhance directional disease features; the features of the two branches are adaptively fused and reconstructed to output a high-resolution texture map;

[0014] S4: Using camera pose and BIM model, the high-resolution texture map perspective transformation is applied to the local planar coordinate system of the component itself to generate a geometrically straightened orthophoto image, which is then fused with the original lighting information to output a corrected image with known pixel size.

[0015] S5: Detection windows of different sizes are defined according to the physical pixel size, and cracks, peeling and honeycomb defects are identified in parallel; a sparse hash indexing mechanism is adopted to perform pixel-level weighted fusion of multi-scale detection results according to their physical scale and feature saliency to generate a unified defect confidence map.

[0016] S6: Refine and vectorize the fused defect results, calculate their true geometric parameters in the BIM global coordinate system through coordinate back projection; combine component structural attributes and maintenance rules to generate quantitative defect reports and maintenance suggestions.

[0017] Step S1 specifically includes the following steps:

[0018] S11: Import the bridge BIM drawings and structural database simultaneously, and for each component g identified in the structural database... j Perform the following actions in sequence:

[0019] Traverse the 3D geometric model and analyze the center coordinate vector c. j , Outer dimension vector (d a ) j and normal vector n j ;

[0020] Access the navigation logs of the inspection equipment and analyze each motion trajectory curve. k Find its relationship with n j The discrete location point s with the smallest included angle and satisfying the safest lateral distance threshold. j,k ; Regarding the pixel side length δ of the camera and the focal length f of the lens j The direction of discrete position points o is calculated using the plane angle inverse calculation method. j,k With the expected field of view angle θ j,k ; the shortest distance h from the bottom surface of the component to the ground or horizontal plane j Record the data; output a component-level photo list L = {s}j,k ,o j,k ,θ j,k ,h j};

[0021] S12: Perform a confirmation check on each record in the shooting list L, when the inspection equipment is detected to have reached the designated position s. j,k And the orientation is the same as o j,k If the deviation does not exceed the threshold, read the current pose p of the camera. t ;

[0022] The corresponding component g in the BIM model j 3D vertex V j via p t Mapped onto the image plane, generating a projected polygon P. j ; For P j All areas outside the concrete surface are immediately masked, resulting in a temporary field of view that includes only the concrete surface. right The image is cropped to form a simplified image block I. j ;

[0023] S13: Simplify image block I j With component number id j Real-time measured shooting distance d j The camera intrinsic parameter matrix K and extrinsic parameter matrix E are bound and written into task Q;

[0024] For any simplified image block I j If the effective pixel area is less than the threshold A min If the cropping step returns a failure flag, then the missing frame flag will be automatically set in Q;

[0025] Each record in Q is accompanied by a set of homography matrices T. j , used to describe I j The mapping relationship between any pixel in the component and the local coordinate system of the component.

[0026] Step S2 specifically includes the following steps:

[0027] S21: Convert the simplified image block I generated in step S1 into... j Corresponding shooting distance d j and lens focal length f j Perform amplification scheduling sequentially;

[0028] By reading the camera pixel size δ and field of view φ, and using the principle of similar triangles, the actual length λ of a single pixel on the concrete surface is derived. j ; will λ j Compare each line with the reference resolution r0 to calculate the resolution gap Δr that needs to be compensated. j =λ j-r0;

[0029] Synchronous retrieval of component category tags j And the tolerance error threshold η preset for this category in the business rules. j ;

[0030] Press Δr j , tag j With η j Generate scale-up requirement table T up ;

[0031] S22: Enlarge the demand table T up Perform adaptive distance partitioning for each item:

[0032] The partitioning process sequentially performs pixel physical length verification, progressive grading, and magnification selection. Each level is based on three factors: distance, focal length, and safety weight, ensuring that the magnification strategy can simultaneously take into account long-distance perspective distortion, focal length differences, and the priority of key components.

[0033] The first level is pixel physical length verification:

[0034] f j Pixel size δ and shooting distance d j Substitution

[0035] The second level is a progressive classification:

[0036] Based on the baseline resolution r0 as the threshold, a piecewise function is used: Among them, s j For the grade number, s j =1 represents the long distance segment, s j =2 represents the mid-range segment, s j =3 represents the near segment, and λ j Classify into long-distance, medium-distance, and short-distance segments;

[0037] The third level is for selecting the multiplier:

[0038] For each tier, a set of pre-set candidate multipliers m Combining component safety weight ω j Construct the cost function: Where, m j This represents the final magnification; |λ j / m-r0|Measures the deviation between the magnified pixel length and the target resolution; ωj / m is the safety priority;Ensures that the magnified pixel length stably falls within the range of [r0, 1.5r0], and automatically shifts high-safety-level components to the lower boundary of the range within the same range;

[0039] s j m jWrite the adjustment flags back to the amplification requirement table and cache them in order from farthest to nearest.

[0040] S23: Employs a distance-focal length linked magnification strategy to increase the focal length of the image. j Block by block applied to I j ;

[0041] During the magnification process, directional interpolation is performed on the high-gradient neighborhood of the crack to maintain line texture continuity, while smooth interpolation is used in low-texture areas to suppress over-sharpening, and automatic switching is achieved through local gradient thresholding. After all pixels have been magnified, an image frame with uniform resolution and controlled geometric proportions is generated. Synchronously update the mapping table, and add λ j With m j Related;

[0042] S24: Will Together with m j d j f j δ and φ are written together into the metadata table C of the amplification stage. meta This forms a complete enlarged record row. j .

[0043] Step S3 specifically includes the following steps:

[0044] S31: Convert the output of step S2 to m j I j and ID j Write to the texture reconstruction queue Q in sequence tex ;

[0045] Synchronously read the reinforcement layout direction α of the same numbered component from the BIM model j , and then d j f j The key-value pairs with δ are encapsulated into the metadata table. j According to m j ×δ performs normalization calibration on the pixel length, with α j A local directional coordinate system is established by using an initial reference angle. Based on the dual-stream super-resolution, the mainstream branch and directional branch are processed under the same physical scale and the same angular reference to handle crack details.

[0046] S32: I j With α j Jointly estimate the prior fusion direction of the reinforcing bars. dir The process involves sequentially performing multi-scale directional convolution, local principal axis filtering, and prior coupling weighting to generate a single-channel directional field F that matches the crack growth trend. j ;

[0047] Multi-scale directional convolution is: for I jThree spatial scales are defined {s = 1, 2, 3}. For each scale, eight directions Φ = {φ1, ..., φ8} of the Sobel-Rot spin convolution kernel are applied, where the i-th direction of the Sobel-Rot spin convolution kernel is φ. i The convolution operation outputs the magnitude of the gradient direction at each pixel (x, y).

[0048] The local principal axis selection is as follows: for eight-directional gradient maps within the same spatial scale *s*, linear Hough clustering is used to statistically analyze local voting peaks; then, the maps are sorted in ascending order of voting intensity, and only the top *p* intensity values ​​are retained. s The direction of % is used as the candidate set for the current pixel. Where, p s % is the preset value, and (x,y) are the pixel coordinates;

[0049] Prior coupling weighting is: for each pixel, Perform a union, and for each direction φ in the set... i Calculate the prior coupling weight of the reinforcing bars; based on two factors, direction difference penalty and scale weight, scale fusion and prior adjustment are completed in one step using the following formula:

[0050]

[0051] Where, γ s κ is the scale weight; F is the direction difference attenuation coefficient; j (x, y) represents the φ corresponding to the maximum weighted sum. i , as the final orientation of the pixel;

[0052] If I j The existence of texture-poor areas leads to The response is below the threshold, with α j Assign a value and in F j (x,y) is marked as low confidence;

[0053] S33: I j Perform residual mainstream N main ;

[0054] The network outputs a full-frame reference feature map M at the end. j ;

[0055] S34: Simultaneously I j With F j Input direction branch N ori Directional branch N oriRotatable convolution is used for convolution alignment at the guide angle, fusing only strip textures parallel or perpendicular to αj, and maintaining fine interpolation to repair edges in the high-gradient neighborhood of the crack. Local contrast enhancement is performed after each convolution stage to dynamically amplify the grayscale difference between the crack and the background, ultimately producing a high-contrast texture feature map O. j ;

[0056] S35: M j O j With F j Jointly conduct grid-guided fusion G mix ; Grid-guided fusion G mix High-resolution texture fusion is achieved by physical mesh segmentation, mesh confidence evaluation, pixel-level weight refinement, and single deconvolution restoration.

[0057] The physical mesh is divided as follows: based on the magnified pixel size and a certain physical side length reference in the metadata, the three feature maps are simultaneously divided into a set of regular meshes G on the image plane. j ={g1,g2,…};

[0058] The grid confidence assessment is as follows: for each grid g k The three indicators are calculated sequentially as follows: directional consistency ξ k Crack significance χ k and brightness stability ζ k , where ξ k By F j With α j The angle difference statistics were obtained, χ k By O j With M j The average of the absolute values ​​of the gradient differences is obtained, ζ k By M j The internal brightness variance is obtained; a rule table is used to map the three factors to the initial value η of the pattern weights. k ∈[0,1]: when ξ k Small and χ k Increase η during large time k When ξ k Big or ζ k Reduce η when large k In all other cases, the median value remains.

[0059] Pixel-level weighting is refined as follows: within each grid cell, the orientation field angle θ is read for pixel (x,y). xy And calculate the normalized angle difference δ xy =|θ xy -α j | / π;Then put η k Multiply by (1-δ) xy Generate pixel weights ω xy =ηk (1-δ xy The weighted fusion of the two features is performed using the following formula:

[0060] Z j (x,y)=[η k (1-δ xy )]O j (x,y)+[1-η k (1-δ xy )]M j (x,y);

[0061] Among them, Z j (x,y) is the fused output of pixel (x,y); O j (x,y),M j (x, y) are the pixel values ​​of the directional pattern and the mainstream pattern, respectively;

[0062] A single deconvolution is used to restore the fused feature map Z. j Input a 2× deconvolution layer to restore the resolution to the magnified level of step S2, and obtain the final high-resolution texture. Simultaneously record η for each grid. k With ω xy ;

[0063] S36: Will and ID j m j With F j Write it together into texture reconstruction result table C tex Simultaneously record the generation timestamp ts j ...

[0064] Step S4 specifically includes the following steps:

[0065] S41: Convert n, the output of step S1, into n. j V j With the generation in step S3 Simultaneously perform geometric correction; adjust the original camera pose P... j Coordinate alignment: The extrinsic parameter composite method is used to align P... j The camera's coordinate system is rotated and translated to the BIM global coordinate system, so that the camera's optical axis vector v j Maintain the same reference as BIM coordinates; for v j With n j Calculate the included angle and derive the optimal single-plane mapping plane Π using the minimum included angle criterion. j ; and then to Π j Perform bounding box analysis on the outer contour, define a four-point reference box {p1,p2,p3,p4}, and then perform bounding box analysis based on V. j Complete the imaging correspondence, thereby enabling the... With Π j Establish corresponding perspective transformation matrix H between them j ;

[0066] H j λ recorded in S2 j Bind to form a geometric index

[0067] S42: For H j Perform pixel-by-pixel mapping, Pull the whole thing straight to Π j Completely overlapping frontal view textures

[0068] During the mapping process, a bilinear grid sampling method is used to resample and fill in the edge holes caused by angular deformation; the new pixel size is recorded in real time for the straightened image. And its associated index ID in the BIM coordinate system map ;

[0069] S43: Will with I j Perform luminance coupling; for I j Low-frequency luminance channel L is extracted using a mean Gaussian kernel. j According to the channel order, L j and Vertical overlay is performed to achieve simultaneous fusion of the on-site lighting field and the super-resolution texture;

[0070] After the superposition is completed, cosine smoothing is performed at the junction of the two channels to remove brightness discontinuities;

[0071] Output geometric correction-luminance coupling diagram and with id j H j as well as Write to geometry index table C geo ...

[0072] Step S5 specifically includes the following steps:

[0073] S51: Will d j and component type tags j Scale prediction P sca :

[0074] a. Canonical mapping: The predictor is based on tags j Obtain the physical width r of the three types of defects—peeling, honeycombing, and cracking—from the industry standard table. k interval Read Mapping it to the image plane scale yields the desired pixel width w.k interval in,

[0075] b. Window generation: For all three types of defects... Find max, mean, and min, and multiply them by the empirical factor β. L =6, β M =4, β S =2 generates three window sizes: large, medium, and small, respectively (H L W L ), (H M W M ), (H S W S );

[0076] c. Raster scanning: For A row-first, column-incrementing, fixed-step sampling method is used, with the step size corresponding to the side length of the window to ensure no overlap. Each slide records the coordinates u of the top-left corner of the window, the window resolution ρ, and the predicted defect label l, outputting a task list T. j =<u,ρ,l> ;

[0077] S52: Move the task list T j Perform scale scheduling D sca The process involves three sequential steps: scale classification, mapping supplementation, and region deduplication.

[0078] The scale classification is as follows: tasks are automatically assigned to macro contour branches B1, coarse texture branches B2, or fine texture branches B3 based on ρ.

[0079] The mapping is supplemented as follows: for each task, a mapping table ψ is written from the original image pixels to the window coordinates. u ;

[0080] Also includes pixel size adjustment

[0081] Region deduplication involves maintaining an occupancy bitmap Ω. j When a new task window overlaps with an already registered window, only the smaller one will be retained.

[0082] S53: After completing the splitting, input the three tasks into B1, B2 and B3 respectively;

[0083] Macroscopic contour branch B1: A sparse convolution sequence with dilation=4 and kernel=7 is used for large window blocks. After two steps, the receptive field is approximately equal to the window width W. LThe entire outline of the peeling or hole is covered in one step; the convolution output is activated by Soft-Sigmoid to obtain the peeling candidate mask, and then the maximum connected component selector is used to retain only the area-dominant connected components to suppress fragmented pseudo-peeling; finally, the peeling probability map P is returned. sp and window index;

[0084] Coarse texture branch B2: For the center window block, first use a convolution stack with dilation=2 and kernel=5 to extract the mid-frequency features of the honeycomb texture, and then use the local roughness operator R at the end of the channel; R reads the 3×3 neighborhood N of each pixel (u,v) of the convolution feature map C. uv Calculate roughness strength:

[0085]

[0086] Among them, R uv For pixel roughness, Let |N| be the neighborhood mean. uv | represents the number of neighboring pixels; R uv With threshold r ho Comparison: The probability of writing to the cellular map P at locations above the threshold. ho ;

[0087] Fine-line branch B3: For small window blocks according to F j A narrow-aperture 3×3 convolution kernel is rotated to achieve orientation alignment, thereby enhancing cracks parallel to the reinforcing bars; three layers of dilated=1 convolution are concatenated to capture fine lines and streaks; finally, a double-threshold hysteresis method is used to distinguish between main cracks and noise streaks, generating a crack probability map P. cr ;

[0088] S54: Perform reversible scaling index fusion of the probability map, window coordinates, and branch numbers of the three-branch return path. rev The multi-scale results are seamlessly stitched into a single, traceable defect confidence tensor by constructing a sparse hash index, performing pixel aggregation to remove collisions, calculating weighted segments, and merging and re-attaching confidence scores.

[0089] Sparse hash indexing: For each window, create a hash bucket of the same size on the original image plane, with the key being the absolute pixel coordinates p = (x, y) and the value being a triple. Where b∈{1,2,3} represents the branch number. P represents the window scale number corresponding to this branch. b (p) represents the three-channel probability vector of the branch output;

[0090] Pixel aggregation for collision resolution: When the same pixel p appears in multiple records in the hash bucket, it is aggregated into the collision list C(p); the list contains different... And different P b(p); If a pixel is not covered by any window, then fill C(p) with a background placeholder;

[0091] Segmented weighting: For each record in C(p), the gradient of the probability graph is calculated on the fly. Then, the pixel-level fusion weights are obtained based on the following piecewise function:

[0092]

[0093] Where, ω b (p) represents the fusion weight of branch b at pixel p; τ is the gradient magnitude of the branch probability map at pixel p; τ is the adaptive gradient threshold; P b,c (p) represents the probability of branch b for class c∈{crack, peeling, background}; Q c (p) represents the final category c confidence level of pixel p after fusion;

[0094] Confidence-based merge reposting: For each pixel p, based on Q... c (p) constitutes the three-channel defect tensor T j Simultaneously record the maximum ω b (p) corresponds to Mark the source as π(p); after all pixels have been processed, output a T of uniform size. j and source mapping π;

[0095] S55: T j Perform single-unit decoding U dec Decoding directly outputs the final mask M for three categories: cracks, peeling, and background, through consecutive 3×3 convolutions superimposed with a single 2× deconvolution. j The crack mask M is respectively cr Peeling mask M sp With background mask M bg And the defect confidence distribution is statistically analyzed at the window dimension and written into the metadata table. def ; M j With pixels-window-original Figure 3 The reindexing is synchronously pushed to the geometry positioning module.

[0096] Step S6 specifically includes the following steps:

[0097] S61: Convert the crack mask M output by S5 cr With peeling mask M sp Post-processing:

[0098] For M cr For each crack pixel (x, y), call the F generated in step S3. j Reading the direction field angle θ xy ; at the direction field angle θxy Take length L in the parallel direction ∥ =5, Width W ∥ A strip window with a grayscale value of 1 undergoes grayscale extension, writing the maximum grayscale value of the window back to the center pixel to connect the main crack; a length L is taken in the vertical direction. ⊥ =3, Width W ⊥ Suppression is performed on a narrow-band window with a value of 1, and the minimum gray value of the window is averaged with the center pixel.

[0099] The filtered crack map is refined in a single pass, retaining only the skeleton; eight-neighbor redundant points are removed from the skeleton to keep the centerline width constant at 1px, and the original pixel to physical length mapping is appended to the pixel attribute table.

[0100] Crack width w l Calculation: After obtaining the crack skeleton, for each pixel on the skeleton, use the orientation field output in step S3 to obtain its orientation field angle θ. xy The geometric correction-luminosity coupling map generated in step S4 Above, construct a short line segment along the normal direction, intersecting with the crack mask M. cr The intersection of the two boundaries is used to obtain the pixel positions of the left and right boundaries, and the pixel distance between them is denoted as d. p The actual length λ of a single pixel on the concrete surface is recorded in step S4. j The local crack width w at that point can be obtained. p =d p λ j Along the entire centerline for all w p By performing statistical analysis, the overall width w of the crack was obtained. l ;

[0101] Watershed erosion refinement: for M sp Calculate the grayscale morphological gradient, and then use the gradient minimum as a seed to perform watershed flood expansion; after expansion, calculate the concavity of each peeled connected component, and mark subdomains with concavity above a threshold as void regions, outputting the outer frame vector Γ. k ;

[0102] Synchronously generate crack centerline vector set L j With the empty outer frame set Γ j It returns a pixel-physical mapping table.

[0103] S62: L j ,Γ j Based on the pixel coordinates, retrieve the H saved in step S1 j and Perform coordinate backprojection on all vector objects, mapping them back to the component's local coordinate system Σ. j ;

[0104] For each centerline l∈L j The true length L of the crack is obtained by accumulating the pixel spacing segment by segment. l Calculate the azimuth angle φ using the first and last points. l For each outer frame γ∈Γ j Use a geometric algorithm based on the minimum bounding rectangle to calculate the area A. γ Aspect Ratio η γ With center of gravity c γ ; will {L l ,φ l}、{A γ ,η γ The attribute field is attached to the corresponding vector, and the id is also linked. j With ts j Generate an independent layer D j ;

[0105] S63: D j L in l w l and A γ Input maintenance decision rule base R mnt ;

[0106] The rule base first matches the stress zone marker of the component: if L l Length threshold L crit If the crack is located in the tension zone, then mark the traffic closure priority in the results;

[0107] Otherwise if w l Crack width critical value w crit And L l ≤ Length critical value L crit Then mark the glue injection suggestion; then perform dual-threshold filtering on the peeling records by area and aspect ratio, for all A γ >Critical value of peeling area A crit And η γ Aspect Ratio Limit η lim Add a recasting symbol to the entry.

[0108] The advantages of this invention are:

[0109] 1) Firstly, a component-waypoint coupling planning strategy is proposed in the image acquisition stage. Unlike existing trajectory layouts based on bounding boxes or uniform grids, this strategy uses the center coordinates, outer dimensions, and normal of each concrete component in the BIM model as a reference. It then searches for finite waypoints with the smallest angle to the normal and satisfying the lateral safety distance threshold on the continuous trajectory of known UAV or vehicle platform navigation logs. Subsequently, combining the camera focal length and sensor size, the desired field of view angle is calculated inversely using the planar angle, and the 3D vertices of the component are mapped to the image plane in real time, cropping areas outside the field of view. This process locks the minimum usable field of view and outputs the homography matrix during the shooting plan generation stage. This provides a native reference for subsequent geometric uniformity, avoiding the reliance on highly overlapping images for feature registration as in traditional methods. It calculates the physical length of a single pixel by linking the shooting distance, focal length, and pixel size corresponding to waypoints, and drives the magnification selection with three consecutive levels: far, medium, and near. The cost function also introduces component safety weights, causing key components such as the main beam to automatically shift to higher resolutions, thus maintaining a smooth transition in physical resolution across spatial dimensions. This distance-focal length-safety three-factor uniformity mechanism eliminates the reliance on the fixed magnification of a general super-resolution network for resolution adjustment, solving the problem of pixel scale jumps caused by differences in shooting distance at different parts of the same component.

[0110] 2) In the image enhancement and texture restoration stage, a dual-stream super-resolution network structure is proposed and adopted. The mainstream branch is based on residual units to fuse high and low frequency information and output geometric and brightness baseline feature maps. The orientation branch is coupled with the prior of the rebar arrangement orientation through multi-scale rotatable convolution kernels. First, a single-channel orientation field is generated. Then, the filter is aligned in orientation within each convolutional layer to enhance only the strip texture parallel or perpendicular to the local dominant orientation. After the two features are divided into physically equal area grids in the image plane, the grid weights are calculated based on three indicators: orientation consistency, crack saliency, and brightness stability. The weights are further refined and weighted within the grid based on the pixel-level angular difference of the orientation field. The subsequent single deconvolution restores the fused features to the magnified target resolution and outputs pixel-level weights simultaneously, forming a high-resolution texture that retains the grayscale consistency of the concrete substrate and maintains the continuity of the crack orientation. Compared with the existing texture compensation methods that rely on global GAN ​​or single-image SRCNN, the dual-stream structure limits the orientation search space through structural priors, reduces the uncertainty of data-driven texture hallucination, and ensures the complete transmission of the relationship between crack and rebar orientation at the sub-pixel level.

[0111] 3) For multi-scale defect identification and result fusion, this application does not adopt the independent inference of traditional image pyramids or fixed sliding windows. Instead, it generates three levels of square windows (large, medium, and small) on the correction-coupling map with the expected pixel width obtained by the normalized mapping as the scale benchmark. The windows are fed into three parallel branches: macro contour, coarse texture, and fine texture, which output probability maps of spalling, honeycombing, and cracks, respectively. All branch results are written into a sparse hash structure, and then the pixel-level fusion weights are automatically calculated based on the scale inverse and gradient magnitude piecewise function, thereby generating a defect confidence tensor with traceable source and consistent physical resolution on a single image. The tensor is bound to the previously saved perspective matrix and pixel size, and is directly projected into the local coordinate system of the BIM component. The crack centerline and spalling outline can thus be quantified in length, orientation, and area with millimeter precision, and maintenance suggestions are generated in conjunction with the stress area marker. The entire process achieves seamless connection between the detection results and the global coordinates of BIM through continuous physical scale control from shooting planning to detection fusion, which is different from the existing positioning mode that relies on sparse SfM point clouds or manual threshold adjustment. Attached Figure Description

[0112] Figure 1 This is a schematic flowchart of the overall process of the microcrack segmentation method based on deep learning and super-resolution reconstruction of the present invention.

[0113] Figure 2 This is a schematic diagram of the process of dual-stream super-resolution texture reconstruction of the present invention;

[0114] Figure 3 This is a comparative diagram of the present invention and the prior art. Detailed Implementation

[0115] The features and other related features of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments, so as to facilitate understanding by those skilled in the art:

[0116] Example: Figure 1 As shown, this embodiment relates to a microcrack segmentation optimization method based on deep learning and super-resolution reconstruction, applied to camera-equipped inspection equipment (flying or traveling platform), based on bridge BIM drawings and structural database. The method mainly includes the following steps:

[0117] S1. After importing BIM drawings and structural database, traverse the three-dimensional geometric model of each concrete component identified in the database, and analyze the center coordinates, outer dimensions and normals. Combined with navigation logs, find discrete position points on the flight (or driving) trajectory that have the smallest angle with the component normal and a lateral distance not less than a preset threshold. Combined with the camera focal length, use the planar angle inverse method to determine the desired field of view angle. Then, map the corresponding three-dimensional vertices of the component to the image plane to generate projection polygons. Perform real-time cropping on the area outside the polygons to obtain image blocks containing only the target concrete surface and the homography matrix corresponding to the image blocks, forming a component-level shooting list.

[0118] In this embodiment, step S1 specifically includes the following steps:

[0119] S11: Import the bridge BIM drawings and structural database simultaneously, and for each component g identified in the structural database... j (j is the component number) Perform the following actions sequentially:

[0120] Traverse the 3D geometric model and analyze the center coordinate vector c. j , Outer dimension vector (d a ) j and normal vector n j ;

[0121] Access the navigation logs of the inspection equipment and analyze each motion (flight or driving) trajectory curve. k (k is the index of the trajectory curve), find its relationship with n. j The discrete location point s with the smallest included angle and satisfying the safest lateral distance threshold. j,k Regarding camera sensor specifications (camera pixel side length δ) and focal length configuration (lens focal length f) j The direction of discrete position points is calculated using the plane angle inverse calculation method. j,k With the expected field of view angle θ j,k (desired field of view θ) j,k The location point is not greater than the camera's field of view angle φ, thus ensuring that the target component can fall completely into the camera's field of view; the shortest distance h from the bottom surface of the component to the ground or horizontal plane. j Record the data for subsequent safety height verification during shooting; output a component-level shooting list L = {s} j,k ,o j,k ,θ j,k ,h j}

[0122] S12: Perform a confirmation check on each record in the shooting list L, when the inspection equipment is detected to have reached the designated position s. j,k And the orientation is the same as o j,k If the deviation does not exceed the threshold, read the current pose p of the camera. t(t is time);

[0123] The corresponding component g in the BIM model j 3D vertex V j via p t Mapped onto the image plane, generating a projected polygon P. j ; For P j All areas outside the concrete surface are immediately masked, resulting in a temporary field of view that includes only the concrete surface. right The image is cropped to form a simplified image block I. j Block the sky, lanes, guardrails, and other non-target elements;

[0124] S13: Simplify image block I j With component number id j Real-time measured shooting distance d j The camera intrinsic parameter matrix K and extrinsic parameter matrix E are bound and written into task Q;

[0125] For any simplified image block I j If the effective pixel area is less than the threshold A min If the cropping step returns a failure flag, then the missing frame flag will be automatically set in Q;

[0126] Each record in Q is accompanied by a set of homography matrices T. j , used to describe I j The mapping relationship between any pixel in the component and the local coordinate system of the component.

[0127] S2: For the image block, read the camera pixel size, focal length and shooting distance to deduce the physical length of a single pixel on the concrete surface, divide the image into three progressively larger segments: long distance, medium distance and short distance, and construct a cost function based on the segment number, pixel physical length and component safety weight to select the magnification ratio, perform adaptive distance-focal length linkage magnification on the image block, use direction-preserving interpolation in the high gradient neighborhood of cracks to maintain line continuity, use smooth interpolation in low texture areas to suppress over-sharpening, and record the mapping between magnification ratio and physical length.

[0128] In this embodiment, step S2 specifically includes the following steps:

[0129] S21: Convert the simplified image block I generated in step S1 into... j Corresponding shooting distance d j and lens focal length f j Perform amplification scheduling sequentially;

[0130] By reading the camera pixel size δ and field of view φ, and using the principle of similar triangles, the actual length λ of a single pixel on the concrete surface is derived. j ; will λ jCompare each line with the reference resolution r0 to calculate the resolution gap Δr that needs to be compensated. j =λ j -r0;

[0131] Synchronous retrieval of component category tags j And the tolerance error threshold η preset for this category in the business rules. j This allows for the allocation of differentiated magnification precision for different safety levels, such as main beams and secondary beams, during subsequent magnification stages;

[0132] Press Δr j , tag j With η j Generate scale-up requirement table T up .

[0133] S22: Enlarge the demand table T up Perform adaptive distance partitioning for each item:

[0134] The partitioning process sequentially performs pixel physical length verification, progressive grading, and magnification selection. Each level is based on three factors: distance, focal length, and safety weight, ensuring that the magnification strategy can simultaneously take into account long-distance perspective distortion, focal length differences, and the priority of key components.

[0135] The first level is pixel physical length verification:

[0136] f j Pixel size δ and shooting distance d j Substitution

[0137] The second level is a progressive classification:

[0138] Based on the baseline resolution r0 as the threshold, a piecewise function is used: Among them, s j For the grade number, s j =1 represents the long distance segment, s j =2 represents the mid-range segment, s j =3 represents the near segment, and λ j Divide the magnification into long-range, medium-range, and short-range segments; progressive division only shifts to the next closer segment to ensure a continuous spatial transition of magnification and prevent sudden jumps in resolution between adjacent components;

[0139] The third level is for selecting the multiplier:

[0140] For each tier, a set of pre-set candidate multipliers m The cost function is constructed by combining the component safety weights ωj (larger values ​​for main beams and smaller values ​​for secondary beams): Where, m j This represents the final magnification; |λ j / m-r0| Measures the deviation between the magnified pixel length and the target resolution; ω j / m is the safety priority option; the larger the value, the more likely it is to choose a higher magnification. It ensures that the pixel length after magnification is stably within the range of [r0, 1.5r0], and automatically shifts high-safety-level components to the lower boundary of the range (finer pixels) within the same range.

[0141] s j m j Write the adjustment flags back to the amplification requirement table and cache them in order from farthest to nearest.

[0142] S23: Employs a distance-focal length linked magnification strategy to increase the focal length of the image. j Block by block applied to I j ;

[0143] During the magnification process, directional interpolation is performed on the high-gradient neighborhood of the crack to maintain line texture continuity, while smooth interpolation is used in low-texture areas to suppress over-sharpening, and automatic switching is achieved through local gradient thresholding. After all pixels have been magnified, an image frame with uniform resolution and controlled geometric proportions is generated. Synchronously update the mapping table, and add λ j With m j Related.

[0144] S24: Will Together with m j d j f j δ and φ are written together into the metadata table C of the amplification stage. meta (Used to uniformly store the magnified records of each simplified image block), forming a complete magnified record row. j .

[0145] S3: As Figure 2 As shown, the magnified image patch is input into a dual-stream super-resolution texture reconstruction network. The main branch is fused with high and low frequency information through residual units to output a baseline feature map. The directional branch generates a single-channel directional field based on multi-scale directional convolution and prior information on the direction of rebar arrangement, and performs rotatable convolution for directional alignment. The two features are fused in a regular grid according to directional consistency, crack saliency, and brightness stability. Then, the magnified resolution is restored through a single deconvolution to obtain a high-resolution texture map and simultaneously save the pixel-level fusion weights.

[0146] In this embodiment, step S3 specifically includes the following steps:

[0147] S31: Convert the output of step S2 to m j I j and ID j Write to the texture reconstruction queue Q in sequence tex ;

[0148] Synchronously read the reinforcement layout direction α of the same numbered component from the BIM model j , and then d j f j The key-value pairs with δ are encapsulated into the metadata table. j According to m j ×δ performs normalization calibration on the pixel length, with α j A local directional coordinate system is established by setting an initial reference angle. Based on the dual-stream super-resolution, the mainstream branch and directional branch are processed under the same physical scale and the same angular reference to handle crack details.

[0149] S32: I j With α j Jointly estimate the prior fusion direction of the reinforcing bars. dir The process involves sequentially performing multi-scale directional convolution, local principal axis filtering, and prior coupling weighting to generate a single-channel directional field F that matches the crack growth trend. j ;

[0150] Multi-scale directional convolution is: for I j Three spatial scales are defined {s = 1, 2, 3}. For each scale, eight directions Φ = {φ1, ..., φ8} of the Sobel-Rot spin convolution kernel are applied, where the i-th direction of the Sobel-Rot spin convolution kernel is φ. i The convolution operation outputs the magnitude of the gradient direction at each pixel (x, y). The three scales together cover three typical size segments: crack width, rebar spacing, and honeycomb particles, ensuring that fine cracks and coarse textures receive sufficient response at the same time.

[0151] The local principal axis selection is as follows: for eight-directional gradient maps within the same spatial scale *s*, linear Hough clustering is used to statistically analyze local voting peaks; then, the maps are sorted in ascending order of voting intensity, and only the top *p* intensity values ​​are retained. s The direction of % is used as the candidate set for the current pixel. Noise directions are eliminated and the search space is compressed, where p s % is the preset value, and (x,y) are the pixel coordinates;

[0152] Prior coupling weighting is: for each pixel, Perform a union, and for each direction φ in the set... i Calculate the prior coupling weight of the reinforcing bars; based on two factors, direction difference penalty and scale weight, scale fusion and prior adjustment are completed in one step using the following formula:

[0153]

[0154] Where, γ sκ is the scale weight, and its value increases with s to emphasize the contribution of the coarse scale to the strip texture; F is the orientation difference attenuation coefficient; j (x, y) represents the φ corresponding to the maximum weighted sum. i , as the final orientation of the pixel;

[0155] If I j The existence of texture-poor areas leads to The response is below the threshold, with α j Assign a value and in F j The (x,y) value is marked as low confidence, which suggests that the subsequent network should dynamically increase the weight of the directional branch at the corresponding position.

[0156] The three-step process, through scale coverage, voting screening, and prior coupling, allows the directional field to inherit the overall direction of the reinforcing steel in the prior stage while retaining the real texture details of the concrete surface.

[0157] S33: I j Perform residual mainstream N main That is, it can immediately extract the edge and brightness gradient through shallow convolution, complete the high and low frequency information fusion based on skip connections in multi-level residual units, and add a channel normalization layer at the end of each level to unify the cross-component lighting reference.

[0158] The network outputs a full-frame reference feature map M at the end. j The figure maintains the geometric integrity and grayscale consistency of the concrete base.

[0159] S34: Simultaneously I j With F j Input direction branch N ori Directional branch N ori Rotatable convolution is used for convolution alignment at the orientation field angle, fusing only strip textures parallel or perpendicular to αj. Fine-grained interpolation is applied in the high-gradient neighborhood of the crack to repair edges with directional precision. Local contrast enhancement is performed after each convolution stage to dynamically amplify the grayscale difference between the crack and the background, ultimately producing a high-contrast texture feature map O. j Enhance input during the fusion stage.

[0160] S35: M j O j With F j Jointly conduct grid-guided fusion G mix ; Grid-guided fusion G mix High-resolution texture fusion is achieved by physical mesh segmentation, mesh confidence evaluation, pixel-level weight refinement, and single deconvolution restoration.

[0161] The physical mesh is divided as follows: based on the magnified pixel size and a certain physical side length reference in the metadata, the three feature maps are simultaneously divided into a set of regular meshes G on the image plane.j ={g1,g2,…}; the division ensures that each grid covers a similar actual concrete area;

[0162] The grid confidence assessment is as follows: for each grid g k The three indicators are calculated sequentially as follows: directional consistency ξ k Crack significance χ k and brightness stability ζ k , where ξ k By F j With α j The angle difference statistics were obtained, χ k By O j With M j The average of the absolute values ​​of the gradient differences is obtained, ζ k By M j The internal brightness variance is obtained; a rule table is used to map the three factors to the initial value η of the pattern weights. k ∈[0,1]: when ξ k Small and χ k Increase η during large time k When ξ k Big or ζ k Reduce η when large k In all other cases, the median value remains.

[0163] Pixel-level weighting is refined as follows: within each grid cell, the orientation field angle (guide angle) θ is read for the pixel (x, y). xy And calculate the normalized angle difference δ xy =|θ xy -α j | / π;Then put η k Multiply by (1-δ) xy Generate pixel weights ω xy =η k (1-δ xy The weighted fusion of the two features is performed using the following formula:

[0164] Z j (x,y)=[η k (1-δ xy )]O j (x,y)+[1-η k (1-δ xy )]M j (x,y);

[0165] Among them, Z j (x,y) is the fused output of pixel (x,y); O j (x,y),M j (x, y) are the pixel values ​​of the directional pattern and the mainstream pattern, respectively;

[0166] A single deconvolution is used to restore the fused feature map Z. j Input a 2× deconvolution layer to restore the resolution to the magnified level of step S2, and obtain the final high-resolution texture map. Simultaneously record η for each grid. k With ω xy This provides a traceable source of weights for subsequent defect probability fusion.

[0167] S36: Will and ID j m j With F j Write it together into texture reconstruction result table C tex (Used to centrally store the high-resolution textures and related information of each simplified image block), and simultaneously record the generation timestamp ts j .

[0168] S4: Rotate and translate the camera coordinate system to the BIM global coordinate system using the extrinsic composite method for the high-resolution texture map. Derive the optimal mapping plane based on the minimum included angle criterion and calculate the perspective transformation matrix. Adjust the texture map to a frontal view image that coincides with the plane of the component according to the matrix. Then extract the low-frequency brightness channel of the original image and overlap it longitudinally with the adjusted texture. Eliminate brightness discontinuities through cosine smoothing to generate a geometric correction-brightness coupling map and record the pixel size.

[0169] In this embodiment, step S4 specifically includes the following steps:

[0170] S41: Convert n, the output of step S1, into n. j V j With the generation in step S3 Simultaneously perform geometric correction; adjust the original camera pose P... j Coordinate alignment: The extrinsic parameter composite method is used to align P... j The coordinate system of the machine body is rotated and translated to the BIM global coordinate system, and the external parameter composite method (or other methods) is used to transform P. j The camera's coordinate system is rotated and translated to the BIM global coordinate system, so that the camera's optical axis vector v j Maintain the same reference as BIM coordinates; for v j With n j Calculate the included angle and derive the optimal single-plane mapping plane Π using the minimum included angle criterion. j ; and then to Π j Perform bounding box analysis on the outer contour, define a four-point reference box {p1,p2,p3,p4}, and then perform bounding box analysis based on V. j Complete the imaging correspondence, thereby enabling the... With Π j Establish corresponding perspective transformation matrix H between them j ;

[0171] H j λ recorded in S2 j Bind to form a geometric index

[0172] S42: For H j Perform pixel-by-pixel mapping, Pull the whole thing straight to Π j Completely overlapping frontal view textures

[0173] During the mapping process, a bilinear grid sampling method is used to resample and fill in the edge holes caused by angular deformation; the new pixel size is recorded in real time for the straightened image. And its associated index ID in the BIM coordinate system map Through geometric consistency correction, the upward and downward distortions caused by shooting at a distance are eliminated at the image level.

[0174] S43: Will with I j Perform luminance coupling; for I j Low-frequency luminance channel L is extracted using a mean Gaussian kernel. j According to the channel order, L j and Vertical overlay is performed to achieve simultaneous fusion of the on-site lighting field and the super-resolution texture;

[0175] After the superposition is completed, cosine smoothing is performed at the junction of the two channels to remove brightness discontinuities;

[0176] Output geometric correction-luminance coupling diagram and with id j H j as well as Write to geometry index table C geo (Used to centrally store key information of each simplified image block after geometric correction), providing a unified input that preserves crack details and reflects on-site lighting for large-scale peeling depth identification and multi-scale defect stitching.

[0177] S5: Based on the industry standard mapping of defect width range to pixel size, generate three levels of square windows (large, medium, and small) and scan them in a row-first manner to form a task list. The windows are then assigned to the macro contour branch, coarse texture branch, and fine texture branch respectively. The macro contour branch uses a hole receptive field convolution to output a peeling candidate mask. The coarse texture branch extracts a honeycomb probability map through a local roughness operator. The fine texture branch uses a direction field rotating narrow aperture convolution kernel combined with a double threshold hysteresis method to output a crack probability map. The results of each branch are back-pasted to the original image coordinates in a sparse hashing manner. The pixel-level fusion weights are calculated segmentally based on the scale inverse and gradient magnitude and synthesized into a defect confidence tensor of uniform size.

[0178] In this embodiment, step S5 specifically includes the following steps:

[0179] S51: Will d j and component type tags j Scale prediction P sca :

[0180] a. Canonical mapping: The predictor is based on tags j Obtain the physical width r of the three types of defects—peeling, honeycombing, and cracking—from the industry standard table. k interval Read Mapping it to the image plane scale yields the desired pixel width w. k interval in,

[0181] b. Window generation: For all three types of defects... Find max, mean, and min, and multiply them by the empirical factor β. L =6, β M =4, β S =2 generates three window sizes: large, medium, and small, respectively (H L W L ), (H M W M ), (H S W S The height H = width W ensures that the window is square so that the convolution kernel maintains the same receptive field shape in the three branches.

[0182] c. Raster scanning: For A row-first, column-incrementing, fixed-step sampling method is used, with the step size corresponding to the side length of the window to ensure no overlap. Each slide records the coordinates u of the top-left corner of the window, the window resolution ρ, and the predicted defect label l, outputting a task list T. j =<u,ρ,l> .

[0183] S52: Move the task list T j Perform scale scheduling D sca The process involves three sequential steps: scale classification, mapping supplementation, and region deduplication.

[0184] The scale classification is as follows: based on ρ, the task is automatically assigned to the macroscopic contour branch B1 (=(H)). L W L )), coarse-textured branch B2 (=(H M W M )) or fine line branch B3(=(H S W S ));

[0185] The mapping is supplemented as follows: for each task, a mapping table ψ is written from the original image pixels to the window coordinates. u This makes it easier to reply directly with the results later;

[0186] Also includes pixel size adjustment Ensure that each branch can obtain the physical width of the defect;

[0187] Region deduplication involves maintaining an occupancy bitmap Ω. j When a new task window overlaps with an already registered window, only the smaller one is retained. The overlapping window is indexed in the metadata record, prioritizing smaller sizes and avoiding redundant calculations. After processing, task flows are issued according to branch numbers, with the three branches running in parallel without overlap.

[0188] S53: After completing the splitting, input the three tasks into B1, B2 and B3 respectively;

[0189] Macroscopic contour branch B1: A sparse convolution sequence with dilation=4 and kernel=7 is used for large window blocks. After two steps, the receptive field is approximately equal to the window width W. L The entire outline of the peeling or hole is covered in one step; the convolution output is activated by Soft-Sigmoid to obtain the peeling candidate mask, and then the maximum connected component selector is used to retain only the area-dominant connected components to suppress fragmented pseudo-peeling; finally, the peeling probability map P is returned. sp and window index;

[0190] Coarse texture branch B2: For the center window block, first use a convolution stack with dilation=2 and kernel=5 to extract the mid-frequency features of the honeycomb texture, and then use the local roughness operator R at the end of the channel; R reads the 3×3 neighborhood N of each pixel (u,v) of the convolution feature map C. uv Calculate roughness strength:

[0191]

[0192] Among them, R uv For pixel roughness, Let |N| be the neighborhood mean. uv | represents the number of neighboring pixels; R uv With threshold r ho Comparison: The probability of writing to the cellular map P at locations above the threshold. ho The variance metric in the formula can effectively distinguish between honeycomb (high roughness) and normal substrate (low roughness), and is not sensitive to changes in illumination.

[0193] Fine-line branch B3: For small window blocks according to F jA narrow-aperture 3×3 convolution kernel is rotated to achieve orientation alignment, thereby enhancing cracks parallel to the reinforcing bars; three layers of dilated=1 convolution are concatenated to capture fine lines and streaks; finally, a double-threshold hysteresis method is used to distinguish between main cracks and noise streaks, generating a crack probability map P. cr .

[0194] S54: Perform reversible scaling index fusion of the probability map, window coordinates, and branch numbers of the three-branch return path. rev The multi-scale results are seamlessly stitched into a single, traceable defect confidence tensor by constructing a sparse hash index, performing pixel aggregation to remove collisions, calculating weighted segments, and merging and re-attaching confidence scores.

[0195] Sparse hash indexing: For each window, create a hash bucket of the same size on the original image plane, with the key being the absolute pixel coordinates p = (x, y) and the value being a triple. Where b∈{1,2,3} represents the branch number. P represents the window scale index corresponding to this branch (the smaller the value, the smaller the window). b (p) is the three-channel probability vector of the branch output; the pixel → branch → window link can be queried in reverse at any time;

[0196] Pixel aggregation for collision resolution: When the same pixel p appears in multiple records in the hash bucket, it is aggregated into the collision list C(p); the list contains different... And different P b (p); If a pixel is not covered by any window, then fill C(p) with a background placeholder;

[0197] Segmented weighting: For each record in C(p), the gradient of the probability graph is calculated on the fly. (A high crack gradient means a clear edge), and then the pixel-level fusion weights are obtained based on the following piecewise function:

[0198]

[0199] Where, ω b (p) represents the fusion weight of branch b at pixel p; τ is the gradient magnitude of the branch probability map at pixel p; τ is the adaptive gradient threshold; P b,c (p) represents the probability of branch b for class c∈{crack, peeling, background}; Q c (p) represents the final category c confidence level of pixel p after fusion;

[0200] Based on the above, a priority principle is proposed: those with insufficient gradients are directly cleared to zero, and those with qualified gradients are then weighted according to the reciprocal of the scale, so as to ensure that fine-scale and sharp-edge crack information dominates the fusion process.

[0201] Confidence-based merge reposting: For each pixel p, based on Q... c (p) constitutes the three-channel defect tensor T j Simultaneously record the maximum ω b (p) corresponds to Mark the source as π(p); after all pixels have been processed, output a T of uniform size. j And source mapping π.

[0202] S55: T j Perform single-unit decoding U dec Decoding directly outputs the final mask M for three categories: cracks, peeling, and background, through consecutive 3×3 convolutions superimposed with a single 2× deconvolution. j The crack mask M is respectively cr Peeling mask M sp With background mask M bg And the defect confidence distribution is statistically analyzed at the window dimension and written into the metadata table. def ; M j With pixels-window-original Figure 3 The reindex is synchronously pushed to the geometry positioning module.

[0203] S6: Perform directional guided filtering on the defect confidence tensor and refine the crack skeleton pixel by pixel. Use the watershed algorithm to refine the spalling connected domain and separate the void region. Back-project the crack centerline and spalling outline onto the component local coordinate system according to the perspective transformation matrix and pixel size. Calculate the true length, azimuth, and spalling area of ​​the crack. Combine the stress zone marking and maintenance rule library of the structural database to generate maintenance suggestion results with corresponding inspection timestamps.

[0204] In this embodiment, step S6 specifically includes the following steps:

[0205] S61: Convert the crack mask M output by S5 cr With peeling mask M sp Post-processing:

[0206] For M cr For each crack pixel (x, y), call the F generated in step S3. j Reading the direction field angle θ xy ; at the direction field angle θ xy Take length L in the parallel direction ∥ =5, Width W ∥ A strip window with a grayscale value of 1 undergoes grayscale extension, writing the maximum grayscale value of the window back to the center pixel to connect the main crack; a length L is taken in the vertical direction. ⊥ =3, Width W ⊥ Suppression is performed on a narrow-band window with a value of 1, and the minimum gray value of the window is averaged with the center pixel.

[0207] The filtered crack map is refined in a single pass, retaining only the skeleton; eight-neighbor redundant points are removed from the skeleton to keep the centerline width constant at 1px, and the original pixel to physical length mapping is appended to the pixel attribute table.

[0208] Crack width w l Calculation: After obtaining the crack skeleton, for each pixel on the skeleton, use the orientation field output in step S3 to obtain its orientation field angle θ. xy The geometric correction-luminosity coupling map generated in step S4 Up, along the normal direction (with θ) xy Construct a short line segment (vertically) and connect it to the crack mask M. cr The intersection of the two boundaries is used to obtain the pixel positions of the left and right boundaries, and the pixel distance between them is denoted as d. p The actual length λ of a single pixel on the concrete surface is recorded in step S4. j The local crack width w at that point can be obtained. p =d p λ j Along the entire centerline for all w p Perform statistical analysis (e.g., take the average or upper quantile value) to obtain the overall width w of the crack. l (Subsequent crack width w) l and the actual length L of the crack l Write it together to a separate layer D j );

[0209] Watershed erosion refinement: for M sp Calculate the grayscale morphological gradient, and then use the gradient minimum as a seed to perform watershed flood expansion; after expansion, calculate the concavity of each peeled connected component, and mark subdomains with concavity above a threshold as void regions, outputting the outer frame vector Γ. k ;

[0210] Synchronously generate crack centerline vector set L j With the empty outer frame set Γ j It returns a pixel-physical mapping table.

[0211] S62: L j ,Γ j Based on the pixel coordinates, retrieve the H saved in step S1 j and Perform coordinate backprojection on all vector objects, mapping them back to the component's local coordinate system Σ. j ;

[0212] For each centerline l∈L j The true length L of the crack is obtained by accumulating the pixel spacing segment by segment. l Calculate the azimuth angle φ using the first and last points. lFor each outer frame γ∈Γ j Use a geometric algorithm based on the minimum bounding rectangle to calculate the area A. γ Aspect Ratio η γ With center of gravity c γ ; will {L l ,φ l}、{A γ ,η γ The attribute field is attached to the corresponding vector, and the id is also linked. j With ts j Generate an independent layer D j .

[0213] S63: D j L in l w l and A γ Input maintenance decision rule base R mnt ;

[0214] The rule base first matches the stress zone marker of the component: if L l Length threshold L crit If the crack is located in the tension zone, then mark the traffic closure priority in the results;

[0215] Otherwise if w l Crack width critical value w crit And L l ≤ Length critical value L crit Then mark the glue injection suggestion; then perform dual-threshold filtering on the peeling records by area and aspect ratio, for all A γ >Critical value of peeling area A crit And η γ Aspect Ratio Limit η lim Add a recasting symbol to the entry.

[0216] like Figure 3 As shown, this method adopts BIM prior knowledge + safety view, dual-flow direction super-resolution, multi-scale hash fusion detection, and automatic vectorization + coordinate mapping design to establish a whole-process method that integrates shooting planning, resolution consistency, direction-preserving super-resolution, precise geometric alignment, and unified fusion of multi-scale detection results. This enables the final generated defect information to achieve millimeter-level quantitative identification and traceable positioning in the BIM global coordinate system.

[0217] The beneficial technical effects of this embodiment are as follows:

[0218] 1) Firstly, a component-waypoint coupling planning strategy is proposed in the image acquisition stage. Unlike existing trajectory layouts based on bounding boxes or uniform grids, this strategy uses the center coordinates, outer dimensions, and normal of each concrete component in the BIM model as a reference. It then searches for finite waypoints with the smallest angle to the normal and satisfying the lateral safety distance threshold on the continuous trajectory of known UAV or vehicle platform navigation logs. Subsequently, combining the camera focal length and sensor size, the desired field of view angle is calculated inversely using the planar angle, and the 3D vertices of the component are mapped to the image plane in real time, cropping areas outside the field of view. This process locks the minimum usable field of view and outputs the homography matrix during the shooting plan generation stage. This provides a native reference for subsequent geometric uniformity, avoiding the reliance on highly overlapping images for feature registration as in traditional methods. It calculates the physical length of a single pixel by linking the shooting distance, focal length, and pixel size corresponding to waypoints, and drives the magnification selection with three consecutive levels: far, medium, and near. The cost function also introduces component safety weights, causing key components such as the main beam to automatically shift to higher resolutions, thus maintaining a smooth transition in physical resolution across spatial dimensions. This distance-focal length-safety three-factor uniformity mechanism eliminates the reliance on the fixed magnification of a general super-resolution network for resolution adjustment, solving the problem of pixel scale jumps caused by differences in shooting distance at different parts of the same component.

[0219] 2) In the image enhancement and texture restoration stage, a dual-stream super-resolution network structure is proposed and adopted. The mainstream branch is based on residual units to fuse high and low frequency information and output geometric and brightness baseline feature maps. The orientation branch is coupled with the prior of the rebar arrangement orientation through multi-scale rotatable convolution kernels. First, a single-channel orientation field is generated. Then, the filter is aligned in orientation within each convolutional layer to enhance only the strip texture parallel or perpendicular to the local dominant orientation. After the two features are divided into physically equal area grids in the image plane, the grid weights are calculated based on three indicators: orientation consistency, crack saliency, and brightness stability. The weights are further refined and weighted within the grid based on the pixel-level angular difference of the orientation field. The subsequent single deconvolution restores the fused features to the magnified target resolution and outputs pixel-level weights simultaneously, forming a high-resolution texture that retains the grayscale consistency of the concrete substrate and maintains the continuity of the crack orientation. Compared with the existing texture compensation methods that rely on global GAN ​​or single-image SRCNN, the dual-stream structure limits the orientation search space through structural priors, reduces the uncertainty of data-driven texture hallucination, and ensures the complete transmission of the relationship between crack and rebar orientation at the sub-pixel level.

[0220] 3) For multi-scale defect identification and result fusion, this application does not adopt the independent inference of traditional image pyramids or fixed sliding windows. Instead, it generates three levels of square windows (large, medium, and small) on the correction-coupling map with the expected pixel width obtained by the normalized mapping as the scale benchmark. The windows are fed into three parallel branches: macro contour, coarse texture, and fine texture, which output probability maps of spalling, honeycombing, and cracks, respectively. All branch results are written into a sparse hash structure, and then the pixel-level fusion weights are automatically calculated based on the scale inverse and gradient magnitude piecewise function, thereby generating a defect confidence tensor with traceable source and consistent physical resolution on a single image. The tensor is bound to the previously saved perspective matrix and pixel size, and is directly projected into the local coordinate system of the BIM component. The crack centerline and spalling outline can thus be quantified in length, orientation, and area with millimeter precision, and maintenance suggestions are generated in conjunction with the stress area marker. The entire process achieves seamless connection between the detection results and the global coordinates of BIM through continuous physical scale control from shooting planning to detection fusion, which is different from the existing positioning mode that relies on sparse SfM point clouds or manual threshold adjustment.

Claims

1. A microcrack segmentation optimization method based on deep learning and super-resolution reconstruction, applied to inspection equipment equipped with a camera, characterized in that... The method includes the following steps: S1: Based on the geometric attributes and spatial pose of the components in the BIM model, combined with the inspection equipment trajectory, automatically plan and execute the optimal shooting pose for each component; crop the image in real time to obtain image blocks with the component as the target subject, and simultaneously record the homography matrix used for geometric mapping. S2: Calculate the physical resolution of the image block based on the shooting parameters, and adaptively select a magnification strategy to improve the resolution based on the shooting distance and the importance of the components, and record the scale mapping relationship; S3: The enhanced image patch is input into the dual-stream network, where the mainstream branch learns general texture features and the directional branch fuses the prior of rebar layout to enhance directional disease features; the features of the two branches are adaptively fused and reconstructed to output a high-resolution texture map; S4: Using camera pose and BIM model, the high-resolution texture map perspective transformation is applied to the local planar coordinate system of the component itself to generate a geometrically straightened orthophoto image, which is then fused with the original lighting information to output a corrected image with known pixel size. S5: Detection windows of different sizes are defined according to the physical pixel size, and cracks, peeling and honeycomb defects are identified in parallel; a sparse hash indexing mechanism is adopted to perform pixel-level weighted fusion of multi-scale detection results according to their physical scale and feature saliency to generate a unified defect confidence map. S6: Refine and vectorize the fused defect results, calculate their true geometric parameters in the BIM global coordinate system through coordinate back projection; combine component structural attributes and maintenance rules to generate quantitative defect reports and maintenance suggestions.

2. The method for optimizing micro-crack segmentation based on deep learning and super-resolution reconstruction as described in claim 1, characterized in that... Step S1 specifically includes the following steps: S11: Bridge BIM drawings and structure database are imported at the same time, and each component g identified in the structure database is targeted j The following actions are performed in sequence: traversing the three-dimensional geometric model, resolving a center coordinate vector c j , an envelope dimension vector (d a ) j , and a normal vector n j ; calling navigation log of the inspection device, for each motion trajectory curve Γ k , find the discrete position point s j with the smallest angle with n j,k and satisfying the threshold of the safest lateral distance; for the pixel side length δ of the camera and the lens focal segment f j , calculate the pointing o j,k of the discrete position point by plane angle backstepping method and the expected view angle θ j,k ; record the shortest distance h j from the component bottom surface to the ground or horizontal plane; output the component-level shooting list L = {s j,k , o j,k , θ j,k , h j}. S12: Perform a confirmation check on each record in the shooting list L, when the inspection equipment is detected to have reached the designated position s. j,k And the orientation is the same as o j,k If the deviation does not exceed the threshold, read the current pose p of the camera. t ; The corresponding component g in the BIM model j 3D vertex V j via p t Mapped onto the image plane, generating a projected polygon P. j ; For P j All areas outside the concrete surface are immediately masked, resulting in a temporary field of view that includes only the concrete surface. right The image is cropped to form a simplified image block I. j ; S13: Reduce image block I j With component number id j Real-time measured shooting distance d j The camera intrinsic parameter matrix K and extrinsic parameter matrix E are bound and written into task Q; For any simplified image block I j If the effective pixel area is less than the threshold A min If the cropping step returns a failure flag, then the missing frame flag will be automatically set in Q; Each record in Q is accompanied by a set of homography matrices T. j , used to describe I j The mapping relationship between any pixel in the graph and the local coordinate system of the component.

3. The method for optimizing micro-crack segmentation based on deep learning and super-resolution reconstruction as described in claim 2, characterized in that... Step S2 specifically includes the following steps: S21: Convert the simplified image block I generated in step S1 into... j Corresponding shooting distance d j and lens focal length f j Perform amplification scheduling sequentially; By reading the camera pixel size δ and field of view φ, and using the principle of similar triangles, the actual length λ of a single pixel on the concrete surface is derived. j ; will λ j Compare each line with the reference resolution r0 to calculate the resolution gap Δr that needs to be compensated. j =λ j -r0; Synchronous retrieval of component category tags j And the tolerance error threshold η preset for this category in the business rules. j ; Press Δr j , tag j With η j Generate scale-up requirement table T up ; S22: Enlarge the demand table T up Perform adaptive distance partitioning for each item: The partitioning process sequentially performs pixel physical length verification, progressive grading, and magnification selection. Each level is based on three factors: distance, focal length, and safety weight, ensuring that the magnification strategy can simultaneously take into account long-distance perspective distortion, focal length differences, and the priority of key components. The first level is pixel physical length verification: f j Pixel size δ and shooting distance d j Substitution The second level is a progressive classification: Based on the baseline resolution r0 as the threshold, a piecewise function is used: Among them, s j For the grade number, s j =1 represents the long distance segment, s j =2 represents the mid-range segment, s j =3 represents the near segment, and λ j Classify into long-distance, medium-distance, and short-distance segments; The third level is for selecting the multiplier: For each tier, a set of pre-set candidate multipliers m Combining component safety weight ω j Construct the cost function: Where, m j This represents the final magnification; |λ j / m-r0| Measures the deviation between the magnified pixel length and the target resolution; ω j / m is the safety priority option; it ensures that the pixel length after magnification is stably within the range of [r0, 1.5r0], and automatically shifts high-safety-level components to the lower boundary of the range within the same range; s j m j Write the adjustment flags back to the amplification requirement table and cache them in order from farthest to nearest. S23: Employs a distance-focal length linked magnification strategy to increase the focal length of the image. j Block by block applied to I j ; During the magnification process, directional interpolation is performed on the high-gradient neighborhood of the crack to maintain line texture continuity, while smooth interpolation is used in low-texture areas to suppress over-sharpening, and automatic switching is achieved through local gradient thresholding. After all pixels have been magnified, an image frame with uniform resolution and controlled geometric proportions is generated. Synchronously update the mapping table, and add λ j With m j Related; S24: Will Together with m j d j f j δ and φ are written together into the metadata table C of the amplification stage. meta This forms a complete enlarged record row. j .

4. The method for optimizing micro-crack segmentation based on deep learning and super-resolution reconstruction as described in claim 3, characterized in that... Step S3 specifically includes the following steps: S31: Convert the output of step S2 to m j I j and ID j Write to the texture reconstruction queue Q in sequence tex ; Synchronously read the reinforcement layout direction α of the same numbered component from the BIM model j , and then d j f j The key-value pairs with δ are encapsulated into the metadata table. j According to m j ×δ performs normalization calibration on the pixel length, with α j A local directional coordinate system is established by using an initial reference angle. Based on the dual-stream super-resolution, the mainstream branch and directional branch are processed under the same physical scale and the same angular reference to handle crack details. S32: I j With α j Jointly estimate the prior fusion direction of the reinforcing bars. dir The process involves sequentially performing multi-scale directional convolution, local principal axis filtering, and prior coupling weighting to generate a single-channel directional field F that matches the crack growth trend. j ; Multi-scale directional convolution is: for I j Three spatial scales are defined {s = 1, 2, 3}. For each scale, eight directions Φ = {φ1, ..., φ8} of the Sobel-Rot spin convolution kernel are applied, where the i-th direction of the Sobel-Rot spin convolution kernel is φ. i The convolution operation outputs the magnitude of the gradient direction at each pixel (x, y). The local principal axis selection is as follows: for eight-directional gradient maps within the same spatial scale *s*, linear Hough clustering is used to statistically analyze local voting peaks; then, the maps are sorted in ascending order of voting intensity, and only the top *p* intensity values ​​are retained. s The direction of % is used as the candidate set for the current pixel. Where, p s % is the preset value, and (x,y) are the pixel coordinates; Prior coupling weighting is: for each pixel, Perform a union, and for each direction φ in the set... i Calculate the prior coupling weight of the reinforcing bars; based on two factors, direction difference penalty and scale weight, scale fusion and prior adjustment are completed in one step using the following formula: Where, γ s κ is the scale weight; F is the direction difference attenuation coefficient; j (x, y) represents the φ corresponding to the maximum weighted sum. i , as the final orientation of the pixel; If I j The existence of texture-poor areas leads to The response is below the threshold, with α j Assign a value and in F j (x,y) is marked as low confidence; S33: I j Perform residual mainstream N main ; The network outputs a full-frame reference feature map M at the end. j ; S34: Simultaneously I j With F j Input direction branch N ori Directional branch N ori Rotatable convolution is used for convolution alignment at the guide angle, fusing only strip textures parallel or perpendicular to αj, and maintaining fine interpolation to repair edges in the high-gradient neighborhood of the crack. Local contrast enhancement is performed after each convolution stage to dynamically amplify the grayscale difference between the crack and the background, ultimately producing a high-contrast texture feature map O. j ; S35: M j O j With F j Jointly conduct grid-guided fusion G mix ; Grid-guided fusion G mix High-resolution texture fusion is achieved by physical mesh segmentation, mesh confidence evaluation, pixel-level weight refinement, and single deconvolution restoration. The physical mesh is divided as follows: based on the magnified pixel size and a certain physical side length reference in the metadata, the three feature maps are simultaneously divided into a set of regular meshes G on the image plane. j ={g1,g2,…}; The grid confidence assessment is as follows: for each grid g k The three indicators are calculated sequentially as follows: directional consistency ξ k Crack significance χ k and brightness stability ζ k , where ξ k By F j With α j The angle difference statistics were obtained, χ k By O j With M j The average of the absolute values ​​of the gradient differences is obtained, ζ k By M j The internal brightness variance is obtained; a rule table is used to map the three factors to the initial value η of the pattern weights. k ∈[0,1]: when ξ k Small and χ k Increase η during large time k When ξ k Big or ζ k Reduce η when large k In all other cases, the median value remains. Pixel-level weighting is refined as follows: within each grid cell, the orientation field angle θ is read for pixel (x,y). xy And calculate the normalized angle difference δ xy =|θ xy -α j | / π;Then η k Multiply by (1-δ) xy Generate pixel weights ω xy =η k (1-δ xy The weighted fusion of the two features is performed using the following formula: Z j (x,y)=[η k (1-d xy )]O j (x,y)+[1-η k (1-d xy )]M j (x,y); Among them, Z j (x,y) is the fused output of pixel (x,y); O j (x,y),M j (x, y) are the pixel values ​​of the directional pattern and the mainstream pattern, respectively; A single deconvolution is used to restore the fused feature map Z. j Input a 2× deconvolution layer to restore the resolution to the magnified level of step S2, and obtain the final high-resolution texture. Simultaneously record η for each grid. k With ω xy ; S36: Will and ID j m j With F j Write it together into texture reconstruction result table C tex Simultaneously record the generation timestamp ts j .

5. The method for optimizing micro-crack segmentation based on deep learning and super-resolution reconstruction as described in claim 4, characterized in that... Step S4 specifically includes the following steps: S41: Convert n, the output of step S1, into n. j V j With the generation in step S3 Simultaneously perform geometric correction; adjust the original camera pose P... j Coordinate alignment: The extrinsic parameter composite method is used to align P... j The camera's coordinate system is rotated and translated to the BIM global coordinate system, so that the camera's optical axis vector v j Maintain the same reference as BIM coordinates; for v j With n j Calculate the included angle and derive the optimal single-plane mapping plane Π using the minimum included angle criterion. j ; and then to Π j Perform bounding box analysis on the outer contour, define a four-point reference box {p1,p2,p3,p4}, and then perform bounding box analysis based on V. j Complete the imaging correspondence, thereby enabling the... With Π j Establish corresponding perspective transformation matrix H between them j ; H j λ recorded in S2 j Bind to form a geometric index S42: For H j Perform pixel-by-pixel mapping, Pull the whole thing straight to Π j Completely overlapping frontal view textures During the mapping process, a bilinear grid sampling method is used to resample and fill in the edge holes caused by angular deformation; the new pixel size is recorded in real time for the straightened image. And its associated index ID in the BIM coordinate system map ; S43: Will with I j Perform luminance coupling; for I j Low-frequency luminance channel L is extracted using a mean Gaussian kernel. j According to the channel order, L j and Vertical overlay is performed to achieve simultaneous fusion of the on-site lighting field and the super-resolution texture; After the superposition is completed, cosine smoothing is performed at the junction of the two channels to remove brightness discontinuities. Output geometric correction-luminance coupling diagram and with id j H j as well as Write to geometry index table C geo .

6. The method for optimizing micro-crack segmentation based on deep learning and super-resolution reconstruction as described in claim 5, characterized in that... Step S5 specifically includes the following steps: S51: Will d j and component type tags j Scale prediction P sca : a. Canonical mapping: The predictor is based on tags j Obtain the physical width r of the three types of defects—peeling, honeycombing, and cracking—from the industry standard table. k interval Read Mapping it to the image plane scale yields the desired pixel width w. k interval in, b. Window generation: For all three types of defects... Find max, mean, and min, and multiply them by the empirical factor β. L =6, β M =4, β S =2 generates three window sizes: large, medium, and small, respectively (H L W L ), (H M W M ), (H S W S ); c. Raster scanning: For A row-first, column-incrementing, fixed-step sampling method is used, with the step size corresponding to the side length of the window to ensure no overlap. Each slide records the coordinates u of the top-left corner of the window, the window resolution ρ, and the predicted defect label l, outputting a task list T. j =<u,ρ,l> ; S52: Move the task list T j Perform scale scheduling D sca The process involves three sequential steps: scale classification, mapping supplementation, and region deduplication. The scale classification is as follows: tasks are automatically assigned to macro contour branches B1, coarse texture branches B2, or fine texture branches B3 based on ρ. The mapping is supplemented as follows: for each task, a mapping table ψ is written from the original image pixels to the window coordinates. u ; Also includes pixel size adjustment Region deduplication involves maintaining an occupancy bitmap Ω. j When a new task window overlaps with an already registered window, only the smaller one will be retained. S53: After completing the splitting, input the three tasks into B1, B2 and B3 respectively; Macroscopic contour branch B1: A sparse convolution sequence with dilation=4 and kernel=7 is used for large window blocks. After two steps, the receptive field is approximately equal to the window width W. L The entire outline of the peeling or hole is covered in one step; the convolution output is activated by Soft-Sigmoid to obtain the peeling candidate mask, and then the maximum connected component selector is used to retain only the area-dominant connected components to suppress fragmented pseudo-peeling; finally, the peeling probability map P is returned. sp and window index; Coarse texture branch B2: For the center window block, first use a convolution stack with dilation=2 and kernel=5 to extract the mid-frequency features of the honeycomb texture, and then use the local roughness operator R at the end of the channel; R reads the 3×3 neighborhood N of each pixel (u,v) of the convolution feature map C. uv Calculate roughness strength: Among them, R uv For pixel roughness, Let |N| be the neighborhood mean. uv | represents the number of neighboring pixels; R uv With threshold r ho Comparison: The probability of writing to the cellular map P at locations above the threshold. ho ; Fine-line branch B3: For small window blocks according to F j A narrow-aperture 3×3 convolution kernel is rotated to achieve orientation alignment, thereby enhancing cracks parallel to the reinforcing bars; three layers of dilated=1 convolution are concatenated to capture fine line streaks; finally, a double-threshold hysteresis method is used to distinguish between main cracks and noise streaks, generating a crack probability map P. cr ; S54: Perform reversible scaling index fusion of the probability map, window coordinates, and branch numbers of the three-branch return path. rev The multi-scale results are seamlessly stitched into a single, traceable defect confidence tensor by using sparse hashing for indexing, pixel aggregation for collision removal, weighted segmentation for calculation, and confidence merging and re-attaching. Sparse hash indexing: For each window, a hash bucket of the same size is created on the original image plane, with the key being the absolute pixel coordinates p = (x, y) and the value being a triple. Where b∈{1,2,3} represents the branch number. P represents the window scale number corresponding to this branch. b (p) represents the three-channel probability vector of the branch output; Pixel aggregation for collision resolution: When the same pixel p appears in multiple records in the hash bucket, it is aggregated into the collision list C(p); the list contains different... And different P b (p); If a pixel is not covered by any window, then fill C(p) with a background placeholder; Segmented weighting: For each record in C(p), the gradient of the probability graph is calculated on the fly. Then, the pixel-level fusion weights are obtained based on the following piecewise function: Where, ω b (p) represents the fusion weight of branch b at pixel p; τ is the gradient magnitude of the branch probability map at pixel p; τ is the adaptive gradient threshold; P b,c (p) represents the probability of branch b for class c∈{crack, peeling, background}; Q c (p) represents the final category c confidence score of pixel p after fusion; confidence score merging and backtracking: for each pixel p, based on Q... c (p) constitutes the three-channel defect tensor T j Simultaneously record the maximum ω b (p) corresponds to Mark the source as π(p); after all pixels have been processed, output a T of uniform size. j and source mapping π; S55: T j Perform single-unit decoding U dec Decoding directly outputs the final mask M for three categories: cracks, peeling, and background, through consecutive 3×3 convolutions superimposed with a single 2× deconvolution. j The crack mask M is respectively cr Peeling mask M sp With background mask M bg And the defect confidence distribution is statistically analyzed at the window dimension and written into the metadata table. def ; M j The pixel-window-original image triple index is pushed synchronously to the geometric positioning module.

7. The method for optimizing micro-crack segmentation based on deep learning and super-resolution reconstruction as described in claim 6, characterized in that... Step S6 specifically includes the following steps: S61: Convert the crack mask M output by S5 cr With peeling mask M sp Post-processing: For M cr For each crack pixel (x, y), call the F generated in step S3. j Reading the direction field angle θ xy ; at the direction field angle θ xy Take length L in the parallel direction ∥ =5, Width W ∥ A strip window with a grayscale value of 1 undergoes grayscale extension, writing the maximum grayscale value of the window back to the center pixel to connect the main crack; a length L is taken in the vertical direction. ⊥ =3, Width W ⊥ Suppression is performed on a narrow-band window with a resolution of 1, and a weighted average is calculated between the minimum grayscale value of the window and the center pixel. A single-pass thinning process is applied to the filtered crack image, retaining only the skeleton. Redundant points in the eight-neighborhood are removed from the skeleton, ensuring the centerline width remains constant at 1px, and the original pixel's mapping to the physical length is appended to the pixel attribute table. Crack width w l Calculation: After obtaining the crack skeleton, for each pixel on the skeleton, use the orientation field output in step S3 to obtain its orientation field angle θ. xy The geometric correction-luminosity coupling map generated in step S4 Above, construct a short line segment along the normal direction, intersecting with the crack mask M. cr The intersection of the two boundaries is used to obtain the pixel positions of the left and right boundaries, and the pixel distance between them is denoted as d. p The actual length λ of a single pixel on the concrete surface is recorded in step S4. j The local crack width w at that point can be obtained. p =d p λ j Along the entire centerline for all w p By performing statistical analysis, the overall width w of the crack was obtained. l ; Watershed erosion refinement: for M sp Calculate the grayscale morphological gradient, and then use the gradient minimum as a seed to perform watershed flood expansion; after expansion, calculate the concavity of each peeled connected component, and mark subdomains with concavity above a threshold as void regions, outputting the outer frame vector Γ. k ; Synchronously generate crack centerline vector set L j With the empty outer frame set Γ j It returns a pixel-physical mapping table. S62: L j ,Γ j Based on the pixel coordinates, retrieve the H saved in step S1 j and Perform coordinate backprojection on all vector objects, mapping them back to the component's local coordinate system Σ. j ; For each centerline l∈L j The true length L of the crack is obtained by accumulating the pixel spacing segment by segment. l Calculate the azimuth angle φ using the first and last points. l For each outer frame γ∈Γ j Use a geometric algorithm based on the minimum bounding rectangle to calculate the area A. γ Aspect Ratio η γ With center of gravity c γ ; will {L l ,φ l }、{A γ ,η γ The attribute field is attached to the corresponding vector, and the id is also linked. j With ts j Generate an independent layer D j ; S63: D j L in l w l and A γ Input maintenance decision rule base R mnt ; The rule base first matches the stress zone marker of the component: if L l Length threshold L crit If the crack is located in the tension zone, then mark the traffic closure priority in the results; Otherwise if w l Crack width critical value w crit And L l ≤ Length critical value L crit Then mark the glue injection suggestion; then perform dual-threshold filtering on the peeling records by area and aspect ratio, for all A γ >Critical value of peeling area A crit And η γ Aspect Ratio Limit η lim Add a recasting symbol to the entry.

Images