Three-dimensional reconstruction and volume quantification method of corrosion area based on multispectral feature fusion

Abstract

The application discloses a corrosion area three-dimensional reconstruction and volume quantification method based on multispectral feature fusion, and particularly relates to the technical field of three-dimensional reconstruction, and obtains a multispectral image of a to-be-measured component, carries out corrosion identification on the multispectral image, obtains a two-dimensional corrosion label and a corrosion probability distribution, projects triangular facets in a three-dimensional grid to the multispectral image, and obtains a two-dimensional projection area, generates a facet inner label fragmentation index based on the discrete degree of the corrosion label, the corrosion connected domain quantity and the corrosion probability wave in the two-dimensional projection area, generates a grid micro-pit engulfing index based on the triangular facet scale, the neighborhood curvature change, the depth residual error and the local sampling sparsity degree, fuses the two indexes to obtain a resolution mismatch risk value and divides a risk level, carries out multi-point sampling, local subdivision or local reconstruction on the triangular facet according to the risk level, and calculates a corrosion area, a depth and a volume, and the application can reduce the averaging or missed detection of micro-point corrosion, and improve the corrosion volume quantification precision.

Description

Technical Field

[0001] This invention relates to the field of corrosion detection and three-dimensional reconstruction technology, specifically to a method for three-dimensional reconstruction and volume quantization of corrosion regions based on multispectral feature fusion. Background Technology

[0002] In existing corrosion detection technologies, multispectral imaging can utilize the differences in spectral response of corrosion products, matrix materials, and surface contaminants at different wavelengths to identify and segment corrosion areas. Three-dimensional reconstruction technology can acquire the spatial morphology of corrosion areas and be used for area, depth, and volume quantification. However, in the process of mapping two-dimensional multispectral corrosion labels to three-dimensional mesh models, existing methods usually directly assign image labels to corresponding triangular facets or vertices based on camera calibration parameters. This lacks an effective assessment of the matching relationship between two-dimensional image resolution and three-dimensional mesh facet scale. When there are dense pinholes, micro-corrosion spots, or broken-boundary corrosion areas in the multispectral image, and the three-dimensional mesh has a large facet scale due to sparse point cloud, mesh simplification, or insufficient reconstruction accuracy, a single triangular facet may cover multiple corrosion and non-corrosion states. This can lead to multiple micro-corrosion defects being averaged, merged, or missed, resulting in significant deviations in the corrosion area boundary, number of corrosion pits, maximum corrosion depth, and volume quantification results. This makes it difficult to meet the needs of high-precision corrosion assessment and structural safety monitoring. Summary of the Invention

[0003] The purpose of this invention is to provide a method for three-dimensional reconstruction and volume quantization of eroded regions based on multispectral feature fusion, so as to overcome the shortcomings of the prior art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a method for three-dimensional reconstruction and volume quantization of corrosion regions based on multispectral feature fusion, comprising:

[0005] Acquire multispectral images and a three-dimensional mesh model of the component under test, perform corrosion identification on the multispectral images, and obtain two-dimensional corrosion labels and corrosion probability distributions;

[0006] Based on the camera calibration parameters, each triangular facet in the 3D mesh model is projected onto the multispectral image to obtain the 2D projection region corresponding to each triangular facet.

[0007] Based on the discreteness of the etched labels, the number of etched connected components, and the fluctuation of the etched probability within the two-dimensional projection area, a label fragmentation index within the patch is generated.

[0008] Based on the scale, neighborhood curvature variation, depth residual, and local sampling sparsity of the triangular facets, a mesh micro-pit swallowing index is generated;

[0009] By combining the label fragmentation index and the mesh micro-pit agglomeration index within the patch, a resolution mismatch risk value is obtained, and the risk level is classified accordingly.

[0010] Based on the risk level, multi-point sampling, local subdivision, or local reconstruction are performed on the triangular facets to obtain an adjusted three-dimensional mesh. The two-dimensional corrosion labels are then remapped onto the adjusted three-dimensional mesh, and the area, depth, and volume of the corrosion region are calculated based on the mapping results.

[0011] Preferably, the corrosion identification of the multispectral image to obtain a two-dimensional corrosion label and corrosion probability distribution includes: dividing the multispectral image into several geometrically consistent regions based on the local normal changes and depth continuity after backprojecting the three-dimensional mesh model onto the multispectral image; selecting pixels with stable multi-band reflection characteristics as regional reference spectra in each geometrically consistent region; performing relative spectral correction on the multispectral pixels in the same region based on the regional reference spectra; generating a corrosion response map based on the difference between the corrected multispectral pixels and the regional reference spectra; and performing connectivity constraints and probabilistic smoothing on the corrosion response map to obtain a two-dimensional corrosion label and corrosion probability distribution.

[0012] Preferably, obtaining the two-dimensional projection region corresponding to each triangular facet includes:

[0013] Based on the camera calibration parameters, the vertices of each triangular facet are transformed to the multispectral image coordinate system to obtain the initial image coverage area of ​​each triangular facet. According to the depth continuity and edge projection curvature of each triangular facet after transformation, local virtual subdivision is performed on the triangular facets that do not meet the projection consistency to obtain a set of subdivided facets. According to the depth order and normal visibility of the subdivided facet set in the multispectral image, the occluded projection pixels are removed, and the effective projection pixels that are directly visible to the multispectral camera are retained. The two-dimensional projection area is formed by the effective projection pixels.

[0014] Preferably, the two-dimensional projection area is determined by the sub-pixel coverage ratio, including:

[0015] The pixels in the multispectral image are divided into multiple sub-pixel sampling points, and it is determined whether each sub-pixel sampling point falls within the projection range of the subdivided patch set. When multiple subdivided patches cover the same sub-pixel sampling point, the sub-pixel sampling point corresponding to the subdivided patch with the smallest interpolation depth and the normal visibility is retained. The sub-pixel coverage ratio is determined according to the number of sub-pixel sampling points belonging to the same original triangular patch in the same image pixel, and the association between the two-dimensional projection region and the original triangular patch is established based on the sub-pixel coverage ratio.

[0016] Preferably, the generation of the label fragmentation index within the patch includes:

[0017] The two-dimensional projection area corresponding to each triangular facet is converted into an equal-area sampling unit within the facet according to the sub-pixel coverage ratio, and the two-dimensional corrosion label and corrosion probability distribution are assigned to each equal-area sampling unit.

[0018] Based on the adjacency relationship between equal-area sampling units, the erosion labels within the same triangular facet are traced to obtain the number of erosion connected regions within the facet and the spatial interval of each erosion connected region.

[0019] Based on the spatial interval, area ratio difference, centroid dispersion state, and corrosion probability fluctuation state of the corrosion connected domains, a normalized fusion is performed to obtain the in-patch label fragmentation index, which is used to characterize the risk of small corrosion defects being averaged by the same triangular patch.

[0020] Preferably, the generation of the grid micro-pit swallowing index includes:

[0021] Each triangular facet is selected as the center, and adjacent triangular facests that share edges or vertices are selected to construct a facet neighborhood. The facet scale characterization value is determined based on the edge length distribution of the triangular facests in the facet neighborhood and the equivalent diameter of the eroded connected domain in the two-dimensional projection region.

[0022] Based on the direction and magnitude of the normal change of each triangular facet relative to the central triangular facet within the neighborhood of the facet, the curvature change features of the neighborhood are extracted.

[0023] The depth residual characteristics are determined based on the signed distance between the center point of each triangular facet in the neighborhood of the facet and the local reference facet. The local sampling sparsity characteristics are determined by combining the spacing between local sampling points. The mesh micro-pit swallowing index is obtained by fusing the facet scale characterization value, neighborhood curvature variation characteristics, depth residual characteristics and local sampling sparsity characteristics.

[0024] Preferably, the resolution mismatch risk value is obtained by fusing the label fragmentation index and the mesh micro-pit swallowing index within the patch, including:

[0025] Based on the label fragmentation index and mesh micro-pit absorption index within the same triangular facet, the triangular facet is determined to be in a low mismatch state, a 2D erosion label fragmentation-dominated state, a 3D mesh absorption-dominated state, or a 2D-3D joint mismatch state. The 2D-3D joint mismatch state is assigned a higher risk enhancement weight than the 2D erosion label fragmentation-dominated state and the 3D mesh absorption-dominated state. Based on the risk enhancement weight, the label fragmentation index and mesh micro-pit absorption index within the facet are fused to obtain the resolution mismatch risk value of each triangular facet.

[0026] Preferably, risk levels are classified accordingly, including:

[0027] Using the neighborhood of a triangular facet as the calculation range, the neighborhood consistency risk value is determined based on the resolution mismatch risk value of each triangular facet in the neighborhood and the distance to the center point. The adjacent risk consistency is determined based on the difference between the resolution mismatch risk value of the current triangular facet and the neighborhood consistency risk value, and the resolution mismatch risk value is corrected using the adjacent risk consistency.

[0028] Based on the corrected resolution mismatch risk value, each triangular facet is classified into low-risk, medium-risk, high-risk, and severe-risk levels.

[0029] Preferably, based on the risk level, multi-point sampling, local subdivision, or local reconstruction are performed on the triangular facets to obtain an adjusted 3D mesh, including:

[0030] For low-risk triangular faces, the original topology is preserved. For medium-risk triangular faces, multiple sampling points are set at the vertices, edge midpoints, and face centroids. For high-risk triangular faces, the boundary pixels between eroded and non-eroded labels within the 2D projection area are extracted. These boundary pixels are then back-calculated into the triangular face and used as dividing points for local subdivision. For severely risky triangular faces, local reconstruction is performed based on the original 3D sampling points in their neighborhood and the corresponding projection range of the eroded connected domains, resulting in an adjusted 3D mesh.

[0031] Preferably, the two-dimensional corrosion labels are remapped to the adjusted three-dimensional mesh, and the area, depth, and volume of the corrosion region are calculated based on the mapping results, including:

[0032] The adjusted facets, formed through multi-point sampling, local subdivision, or local reconstruction, are established as inheritance relationships with the original triangular facets. The original two-dimensional projection region, erosion connectivity number, and erosion probability distribution are associated with the corresponding adjusted facets. Based on the inheritance relationship, two-dimensional erosion labels are reassigned to the adjusted facets. The erosion coverage state, erosion probability, and associated erosion connectivity of each adjusted facet are determined. The erosion participation area is determined based on the erosion coverage state and erosion probability of the adjusted facets. The area, depth, and volume of the eroded region are calculated by combining the depth deviation of the adjusted facets relative to the local reference surface.

[0033] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0034] This invention does not directly attach two-dimensional corrosion labels to three-dimensional mesh patches. Instead, it first establishes a two-dimensional projection region for each triangular patch, then generates a label fragmentation index within the patch from the two-dimensional label side and a mesh micro-pit merging index from the three-dimensional geometry side. These are then fused to obtain a resolution mismatch risk value and classify the risk level. This method can proactively identify mismatched regions where a triangular patch covers multiple micro-corrosion spots and where the three-dimensional mesh cannot support the micro-pitting morphology before label mapping. It solves the problem in existing technologies where micro-pitting is averaged, merged, or missed due to the mismatch between image resolution and mesh scale, thereby improving the reliability of identifying corrosion region boundaries, the number of pits, and the location of corrosion risks.

[0035] This invention further performs differentiated dynamic adjustments to triangular facets based on risk levels, rather than uniformly encrypting the entire 3D model or directly using a fixed sampling method. Low-risk areas retain the original facets, medium-risk areas undergo in-face multi-point sampling, high-risk areas are locally subdivided along the boundaries of the 2D corrosion labels, and severely risky areas are locally reconstructed based on the original 3D sampling points. The 2D corrosion labels are then remapped to the adjusted 3D mesh. This method allows mesh adjustments to be directly driven by the fragmentation state of the corrosion labels and the micro-pit merging state, avoiding computational redundancy caused by global encryption and providing a finer geometric foundation for the boundaries and depths of micro-corrosion pits, thereby directly improving the accuracy of corrosion area, maximum depth, and volume quantification. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0037] Figure 1 This is a flowchart of the three-dimensional reconstruction and volume quantization method for corrosion regions based on multispectral feature fusion of the present invention.

[0038] Figure 2 This is a flowchart of the method for generating resolution mismatch risk values ​​according to the present invention. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] Example 1, please refer to Figure 1 As shown in this embodiment, the method for three-dimensional reconstruction and volume quantization of corrosion regions based on multispectral feature fusion includes:

[0041] A multispectral image and a three-dimensional mesh model of the component under test are acquired. Corrosion identification is performed on the multispectral image to obtain a two-dimensional corrosion label and corrosion probability distribution.

[0042] In one implementation, after acquiring the multispectral image and 3D mesh model of the component under test, a pixel correspondence between the 3D mesh model and the multispectral image is first established. The 3D mesh model consists of a set of vertices and a set of triangular faces, with each triangular face determined by three vertices. Based on the intrinsic and extrinsic parameters of the multispectral camera and the coordinates of the 3D vertices, the vertices of the triangular faces are projected onto the multispectral image plane. The projection relationship is as follows: the homogeneous coordinates of the pixels are equal to the camera intrinsic parameter matrix multiplied by the 3D point coordinates transformed by the rotation matrix and translation vector, and then homogeneously normalized to obtain the pixel coordinates. The image region enclosed by the three projected vertices serves as the back-projection region of the triangular facet. The back-projection regions of multiple adjacent triangular faces jointly cover the effective region in the multispectral image corresponding to the component under test.

[0043] After backprojection, the multispectral image is divided into regions with consistent geometric response. For each backprojected region, the unit normal vector of its corresponding triangular facet is calculated, and the angle between the normal vectors of adjacent triangular facets is calculated; the angle between the normal vectors is obtained by the inverse cosine of the dot product of the two unit normal vectors. Then, the depth difference between the center points of adjacent triangular facets along the camera optical axis is calculated, and the depth difference is the absolute value of the difference in depth coordinates between the two center points after extrinsic parameter transformation.

[0044] When the angle between the normal vectors of adjacent triangular facets does not exceed a preset angle threshold and the depth difference does not exceed a preset depth threshold, the corresponding back-projected regions are grouped into the same geometrically consistent region; otherwise, they are treated as region boundaries. The purpose of this is to group surfaces with similar orientations and continuous depth together for correction, avoiding mixing edges, corners, pit sidewalls, and flat areas facing the camera together.

[0045] Within each geometrically consistent region, stability screening is performed on the multi-band reflectance characteristics of each pixel in the multispectral image. Let the reflectance of a pixel in bands 1 to m be an m-dimensional spectral vector. First, the median reflectance of each band within the region is calculated. Then, the spectral deviation between each pixel's spectral vector and the median reflectance vector is calculated, with the spectral deviation taken as the square root of the sum of the squares of the reflectance differences across bands. Pixels with spectral deviations within the top 30% of the region and neighborhood gray-level gradients within the top 50% of the region are identified as stable pixels. The arithmetic mean of the reflectance of the stable pixels in each band is taken to obtain the regional reference spectrum. The regional reference spectrum is used to characterize the reference reflectance state of non-abnormal surfaces within this geometrically consistent region.

[0046] Relative spectral correction is performed on multispectral pixels within the same region using a regional reference spectrum. For any pixel within the region, the reflectance of that pixel in each band is divided by the regional reference spectral value of the corresponding band to obtain the relative reflectance; when the regional reference spectral value is 0, it is replaced by the non-zero minimum value of the stable pixel reflectance in that band.

[0047] After relative spectral correction, the overall reflection shift caused by differences in light intensity, surface orientation, and local shading is compressed, while the inter-band differences caused by corrosion products are preserved.

[0048] Corrosion response maps are calculated based on the differences between the corrected multispectral pixels and the regional reference spectrum. The corrosion response value consists of three parts: the first part is the Euclidean distance between the corrected spectral vector and the all-1 vector, used to characterize overall spectral anomalies; the second part is the ratio difference of a specified combination of corrosion-sensitive bands, used to characterize band response changes caused by iron oxides, salt deposition, or coating degradation; and the third part is the local variance of the corrected spectral response within the pixel's neighborhood, used to characterize pitting edges and rust patch texture. After normalizing each of the three parts to the range of 0 to 1, a weighted sum is performed to obtain the corrosion response value for each pixel. The corrosion response values ​​of all pixels are arranged according to image coordinates to form the corrosion response map.

[0049] Based on the corrosion response map, region connectivity constraints and probabilistic smoothing are performed. First, the corrosion response map is binarized according to a preset corrosion judgment value to obtain initial corrosion labels. Then, 8-neighborhood connectivity analysis is performed on the initial corrosion labels to retain connected regions with an area of ​​no less than 3 pixels and an average corrosion response value no less than the preset response judgment value, while removing isolated noise points. For the retained connected regions, the corrosion response value is used as the initial probability, and edge smoothing is performed within the same geometrically consistent region. During smoothing, the update probability of the target pixel is equal to the weighted average of the probabilities of its neighboring pixels, and the weights are determined by the spatial distance, spectral distance, and whether they are located in the same geometrically consistent region. Pixels that are closer in spatial distance, closer in spectral distance, and located within the same geometrically consistent region receive higher weights. Finally, a two-dimensional corrosion label and corrosion probability distribution consistent with the spatial location of the multispectral image are obtained, providing pixel-level corrosion location and probability basis for subsequent calculation of the label fragmentation index within the triangular patch projection area.

[0050] Based on the camera calibration parameters, each triangular facet in the 3D mesh model is projected onto the multispectral image to obtain the 2D projection area corresponding to each triangular facet.

[0051] In one implementation, the camera calibration parameters include a multispectral camera intrinsic matrix, distortion parameters, and a rotation matrix and translation vector from the 3D mesh coordinate system to the multispectral camera coordinate system. Coordinate transformation is performed on the three vertices of each triangular facet in the 3D mesh model. The coordinates of the vertices in the camera coordinate system are equal to the rotation matrix multiplied by the 3D mesh coordinates plus the translation vector. Then, pixel coordinates are calculated based on the pinhole imaging relationship: the pixel x-coordinate is equal to the focal length x-component multiplied by the ratio of the camera coordinate x-component to the depth component, plus the principal point x-coordinate; the pixel y-coordinate is equal to the focal length x-component multiplied by the camera coordinate x-component to the depth component, plus the principal point y-coordinate. Subsequently, radial and tangential distortion corrections are performed according to the distortion parameters to obtain the projection points of the triangular facet vertices in the multispectral image coordinate system. The initial image coverage area is formed by the three projection points, and the depth values ​​corresponding to the three vertices are recorded.

[0052] After obtaining the initial image coverage area, the projection consistency of the triangular facets is determined. Depth continuity is obtained by dividing the difference between the maximum and minimum depths of the three vertices of the triangular facet by the average depth of the three vertices. The preferred depth continuity value is 0.08, with the following rules: 0.05 when the working distance of the measured component is less than 1 m, 0.08 when the working distance is 1 m to 3 m, and 0.12 when the working distance is greater than 3 m. The edge projection curvature is obtained by the pixel distance between the midpoint projection position of the three three-dimensional edges of the triangular facet and the midpoint of the corresponding two-dimensional projection edge. The maximum pixel distance of the three edges is taken as the edge projection curvature, with a preferred value of 0.5 pixels. The following rules are set: 0.3 pixels when the object size corresponding to a single pixel in the multispectral image is not greater than 0.1 mm, 0.5 pixels when it is greater than 0.1 mm but not greater than 0.5 mm, and 0.8 pixels when it is greater than 0.5 mm. If the depth continuity exceeds the corresponding value, or the edge projection curvature exceeds the corresponding value, the triangular facet is determined to not meet the projection consistency requirement.

[0053] For triangular patches that do not meet the projection consistency requirements, local virtual subdivision is performed. Local virtual subdivision does not change the topology of the original 3D mesh model; it only temporarily adds sampled patches during the projection calculation. During subdivision, the midpoints of the three edges of the triangular patch are first taken, and then the original triangular patch is divided into four sub-triangular patches. For each sub-triangular patch, the depth continuity and edge projection curvature are recalculated. If projection consistency is still not met, the same four-part subdivision is performed again. The maximum number of subdivision layers is preferably three, corresponding to a maximum of 64 subdivided patches. The rule for setting the maximum number of subdivision layers is: two layers are used when the average side length of the original triangular patch is less than twice the equivalent diameter of the smallest eroded connected region in the erosion probability distribution; otherwise, three layers are used. All subdivided patches generated from the same original triangular patch form a subdivided patch set, and the original triangular patch number is retained for each subdivided patch.

[0054] After obtaining the set of subdivided patches, the occlusion relationship of the subdivided patches in the multispectral image is determined. Each pixel in the multispectral image is divided into 4×4 subpixel sampling points. For each subpixel sampling point, it is determined whether it is within the projection range of a certain subdivided patch. If it is within the projection range, the interpolation depth corresponding to the subpixel sampling point is calculated based on the barycentric coordinates. The barycentric coordinates are obtained by the area ratio of the subpixel sampling point relative to the three projection vertices of the subdivided patch, and the interpolation depth is equal to the weighted sum of the depths of the three vertices according to the corresponding barycentric coordinates. When multiple subdivided patches cover the same subpixel sampling point, the subdivided patch with the smallest interpolation depth is retained, and the subpixel sampling points corresponding to subdivided patches with non-minimum interpolation depths are determined to be part of the occluded projected pixels.

[0055] Based on the depth priority determination, a normal visibility determination is further performed. The subdivision patch normal vector is obtained by cross product of two edge vectors in 3D coordinates and normalized, and the viewing direction is obtained by pointing from the center point of the subdivision patch to the optical center of the multispectral camera and normalized. The dot product of the normal vector and the viewing direction is calculated. If the dot product is greater than 0 and the viewing angle does not exceed 85°, the subdivision patch is determined to be directly visible to the multispectral camera; if the dot product is not greater than 0, or the viewing angle exceeds 85°, the corresponding projected pixels are removed from the initial image coverage area. After this processing, the edges of components, the sidewalls of pits, and the surfaces behind weld obstructions will not be incorrectly assigned the same image label. In short, only the part of the surface that the camera actually sees is retained.

[0056] After determining the depth order and normal visibility, the effective projected pixels corresponding to each original triangle are counted. For any image pixel, if k of the 16 sub-pixel sampling points belong to the same original triangle, then the sub-pixel coverage ratio of that pixel relative to that original triangle is k divided by 16. Pixels with a non-zero sub-pixel coverage ratio are included in the two-dimensional projection region corresponding to that original triangle, and this ratio is used as the pixel area weight during subsequent erosion label mapping. Finally, each original triangle obtains a two-dimensional projection region composed of effective projected pixels and their sub-pixel coverage ratios. Simultaneously, the association between the two-dimensional projection region, the subdivided patch set, and the original triangle number is established, providing an accurate visible projection range for subsequent calculation of the label fragmentation index within the patch based on the two-dimensional projection region.

[0057] Based on the discreteness of the etched labels, the number of etched connected domains, and the fluctuation of the etched probability within the two-dimensional projection area, a label fragmentation index within the patch is generated.

[0058] After obtaining the two-dimensional projection area corresponding to each triangular facet, the original triangular facet is used as the processing object, and the effective projection pixels within the two-dimensional projection area are converted into equal-area sampling units within the facet.

[0059] For any triangular facet, each effective projected pixel in the two-dimensional projection region has a sub-pixel coverage ratio, denoted as ri, where ri ranges from 0 to 1. Let the area of ​​the triangular facet be Af, and the sum of the sub-pixel coverage ratios of all effective projected pixels in the two-dimensional projection region be Rf. Then, the area corresponding to the i-th effective projected pixel on the triangular facet is: the area corresponding to the i-th pixel equals Af multiplied by ri and then divided by Rf. Subsequently, the pixel's corresponding area is divided into several equal-area sampling units within the facet according to a preset sampling area As. As is preferably 1 / 100 of the triangular facet area; when the triangular facet area is less than 0.5 square millimeters, As is 0.005 square millimeters. This avoids treating pixels with different coverage ratios as having the same area.

[0060] Each equal-area sampling unit inherits the two-dimensional erosion label and erosion probability distribution of its source effective projected pixels. If an equal-area sampling unit is contributed by multiple effective projected pixels, the erosion probability of the sampling unit is the average of the erosion probabilities of each source pixel weighted according to the sub-pixel coverage ratio. The two-dimensional erosion label of the sampling unit is determined by the weighted erosion probability. When the weighted erosion probability is not less than 0.5, it is assigned an erosion label; when the weighted erosion probability is less than 0.5, it is assigned a non-erosion label. The rule for the value of 0.5 is as follows: the erosion probability distribution output in the erosion recognition stage has been normalized to 0 to 1, and 0.5 corresponds to the equal probability boundary between erosion and non-erosion. If manually labeled samples are used for calibration, the probability value that minimizes the absolute value of the difference between the precision and recall of erosion recognition in the sample is taken.

[0061] After label assignment, connected component tracing is performed based on the adjacency relationships between sampling units. Each sampling unit has its center point coordinates in the local two-dimensional parametric coordinates of the triangular facet. Any two erosion label sampling units are considered adjacent if they meet the following conditions: boundary contact or the distance between their center points does not exceed 1.5 times the equivalent side length of the sampling unit with the same area. Connected component tracing uses a breadth-first traversal: first, an unvisited erosion label sampling unit is selected as the starting unit, and adjacent erosion label sampling units are sequentially added to the same connected component until no new adjacent erosion label sampling units exist; then, the next unvisited erosion label sampling unit is selected, and the process is repeated. After the traversal, the number of eroded connected components Nf within the facet is obtained, along with the area, centroid, and average erosion probability of each eroded connected component.

[0062] The spatial intervals between each eroded connected region are then calculated. The spatial interval between the a-th and b-th eroded connected regions is the minimum Euclidean distance between the center points of all their sampling cells; if there is only one eroded connected region within a patch, the spatial interval is recorded as 0. The mean spatial interval Df within a patch is equal to the arithmetic mean of the pairwise spatial intervals of all connected regions, divided by the length of the longest side of the triangular patch to complete scale normalization. The resulting Df reflects whether multiple pitting erosion spots are separated from each other within the same triangular patch.

[0063] The dispersion of the erosion labels is determined by the location of the connected regions and the difference in area proportion. First, the proportion of each eroded connected region's area to the area of ​​the triangular facet is calculated, resulting in an area proportion sequence. Then, the coefficient of variation of the area proportion sequence is calculated, which is the standard deviation of the area proportion divided by the average area proportion. Next, the average distance from the centroid of each eroded connected region to the weighted centroid of all eroded sampling units is calculated and normalized using the longest side length of the triangular facet to obtain the location dispersion value. The dispersion Lf is the average of the area proportion coefficient of variation and the location dispersion value, limited to the range of 0 to 1; when the calculated result exceeds 1, it is taken as 1.

[0064] The fluctuations in the corrosion probability distribution are obtained through the probability fluctuations of the corrosion tag sampling units. Let the number of corrosion tag sampling units be Mf, and the corrosion probability of the j-th corrosion tag sampling unit be pj. The mean corrosion probability is the arithmetic mean of all pj. The probability fluctuation value Bf is equal to the square root of the average of the squares of the differences between each pj and the mean corrosion probability. To allow for comparison between different triangular facets, Bf is directly limited to the range of 0 to 1, since the corrosion probability itself already lies within the 0 to 1 range.

[0065] The number of connected corrosion domains, spatial spacing, dispersion, and fluctuations in the corrosion probability distribution within a facet are normalized and fused to obtain the facet label fragmentation index. The normalized value of the number of connected domains, Cf, is equal to Nf divided by a preset reference number of 6; if the result exceeds 1, it is taken as 1. The rule for setting the reference number of 6 is: in the scenario of micro-pitting corrosion detection, when there are 6 independent connected corrosion domains within a triangular facet, a single facet label can no longer represent the independent pitting corrosion distribution. The facet label fragmentation index Ff is equal to the sum of Cf, Df, Lf, and Bf multiplied by 0.30, 0.25, 0.25, and 0.20, respectively. The closer the index is to 1, the more likely it is that there are multiple separate corrosion labels with uneven areas and significant probability fluctuations within the same triangular facet, and the higher the risk that micro-corrosion defects are averaged out by the same triangular facet; the closer the index is to 0, the more likely it is that the two-dimensional corrosion labels are continuously and stably distributed within the triangular facet, and conventional mapping methods can be used subsequently.

[0066] Based on the scale of the triangular facet, the change in neighborhood curvature, the depth residual, and the local sampling sparsity, a mesh micro-pit swallowing index is generated.

[0067] After obtaining the fragmentation index of the label within the facet, a facet neighborhood is constructed centered on the same original triangular facet. The facet neighborhood includes the central triangular facet, triangular facets sharing edges with the central triangular facet, and triangular facets sharing vertices with the central triangular facet. If a duplicate triangular facet exists within the same neighborhood, it is retained only once. Let the central triangular facet be f, the facet neighborhood be Nf, and the number of triangular facets within the neighborhood be m. The three side lengths of each triangular facet are obtained from the Euclidean lengths of the differences in the three-dimensional coordinates of its two endpoints, i.e., the side length is equal to the square root of the sum of the squares of the horizontal, vertical, and height differences of the two endpoints in the three-dimensional coordinates.

[0068] The facet scale is used to evaluate whether the triangular facet scale is sufficient to represent micro-corrosion pits. First, the average side length of all triangular facets in the facet's neighborhood, Elf, is calculated. Then, the average equivalent diameter, Edf, of the corrosion connected domains falling within the two-dimensional projection area of ​​the triangular facet is extracted from the aforementioned two-dimensional corrosion labels. The equivalent diameter of the corrosion connected domain is obtained by calculating the diameter after the area of ​​the connected domain is equivalent to the area of ​​a circle; that is, the equivalent diameter equals the square root of four times the area of ​​the connected domain divided by pi. When there are no corrosion connected domains within the two-dimensional projection area, Edf is taken as the median of the equivalent diameters of all corrosion connected domains in the component under test where the triangular facet is located. The facet scale value, Sf, is the result of Elf divided by 2Edf, limited to the range of 0 to 1; if the result is higher than 1, it is taken as 1. This processing indicates that when the average side length reaches twice the equivalent diameter of the micro-corrosion connected domain, a single facet may already have the potential to cross a complete pitting corrosion pit.

[0069] The neighborhood curvature variation characteristic is jointly determined by the direction and magnitude of the normal variation within the neighborhood of a facet. The normal vector of any triangular facet is obtained by normalizing the cross product of any two edge vectors of that facet. The normal vector of the central triangular facet is denoted as nf, and the normal vector of the i-th triangular facet in the neighborhood is denoted as ni. The included normal angle θi is equal to the angle obtained by taking the inverse cosine of the dot product of nf and ni. Then, the direction vi of the line connecting the center points of the triangular facets is calculated, and the sign of the dot product of vi and nf is calculated; when the sign points in the opposite direction to the normal of the central triangular facet and θi continuously increases, it is denoted as a concave variation; otherwise, it is denoted as a non-concave variation. The neighborhood curvature variation characteristic Kf is equal to the average included normal angle of the concave-varying triangular facets divided by 45°; if the result is greater than 1, it is taken as 1. The rule for setting 45° is: when the local fold corresponding to the edge of the corrosion pit reaches 45°, coarse mesh smoothing will significantly weaken the pit wall boundary; therefore, 45° is taken as the normalized angle reference.

[0070] Depth residual features are used to evaluate whether there are concave morphologies within the neighborhood that are masked by the smoothing of facets. First, a local reference surface is fitted using the center points of all triangular facets within the neighborhood of a facet. The local reference surface is expressed as ax + by + cz + d = 0, where a, b, c, and d are obtained using the least squares method, minimizing the sum of the squared perpendicular distances from all center points to this plane. The signed distance from the i-th center point to the local reference surface is... The calculation expression is: ;in, Let ei be the three-dimensional coordinates of the center point of the i-th triangular facet. A positive ei value indicates that the center point is located on one side of the reference surface normal, while a negative ei value indicates that the center point is located on the opposite side of the reference surface normal. The average absolute value of all negative distances is taken as the mean of the depression residual Eef, and then divided by the aforementioned Edf to obtain the depth residual characteristic Rf, with the result limited to the range of 0 to 1. This allows the subtle depth anomalies after the pit bottom is flattened to be included in the evaluation, avoiding reliance solely on facet size.

[0071] The local sampling sparsity feature is used to evaluate the support density of the source points of the 3D mesh in a local region. If the triangular facet retains reconstructed sampling points, the average distance between adjacent sampling points, Eqf, is calculated; if only the triangular mesh is retained, the average of all edge lengths in the neighborhood is used as Eqf. The local sampling sparsity feature Pf is the result of Eqf divided by Edf, and is limited to the range of 0 to 1; if the calculated result is higher than 1, it is taken as 1. This feature reflects the correspondence between the sampling spacing and the pitting scale. When the sampling spacing reaches the equivalent diameter of the pitting, the boundary of an independent micro-pit is easily covered by a single facet.

[0072] After completing the above calculations, the patch-scale characterization values, neighborhood curvature variation features, depth residual features, and local sampling sparsity features are fused at the same scale to obtain the mesh micro-pit merging index Gf. The preferred calculation is: Gf = 0.35Sf + 0.25Kf + 0.25Rf + 0.15Pf. The weights are set as follows: the patch scale directly determines whether micro-corrosion pits can be geometrically represented, so it is set to 0.35; curvature variation and depth residual jointly reflect the fidelity of pit morphology, each set to 0.25; sampling sparsity affects the reliability of local reconstruction, so it is set to 0.15. The closer Gf is to 1, the higher the degree of 3D mesh merging of the geometric morphology of micro-corrosion pits; the closer Gf is to 0, the more continuously the patch neighborhood can represent micro-corrosion pits. This mesh micro-pit merging index, together with the patch-in-label fragmentation index obtained in the previous step, is used for subsequent resolution mismatch risk value calculation.

[0073] Please see Figure 2 As shown, the resolution mismatch risk value is obtained by fusing the label fragmentation index and the mesh micro-pit swallowing index within the patch, and the risk level is classified accordingly.

[0074] After obtaining the label fragmentation index Ff and the mesh micro-pit absorption index Gf for the same triangular facet, the mismatch state of the triangular facet is determined. The state determination value Ts is preferably 0.55, and the rule is set as follows: in the verification samples, the number of samples where micro-corrosion defects are merged, averaged, or missed is counted. When the number of erroneous samples accounts for 10% of the total number of verification samples, the corresponding critical statistical values ​​of Ff and Gf are taken as Ts; when there are no verification samples, Ts is 0.55. If Ff is not less than Ts and Gf is less than Ts, the triangular facet is determined to be in a two-dimensional corrosion label fragmentation-dominated state; if Gf is not less than Ts and Ff is less than Ts, the triangular facet is determined to be in a three-dimensional mesh absorption-dominated state; if both Ff and Gf are not less than Ts, the triangular facet is determined to be in a two-dimensional and three-dimensional mismatch state; if both Ff and Gf are less than Ts, the triangular facet is determined to be in a low mismatch state. By using this status determination, we can first distinguish the sources of risk and avoid mixing up the problems caused by the discretization of two-dimensional labels and the problems caused by the insufficient representation of three-dimensional meshes.

[0075] After determining the mismatch state, a risk enhancement weight Wf is assigned to different mismatch states. The Wf value is 1.00 for low mismatch states, 1.12 for states dominated by 2D corrosion label fragmentation, 1.18 for states dominated by 3D mesh absorption, and 1.35 for states with both 2D and 3D mismatch. The above values ​​are based on the following: 2D corrosion label fragmentation mainly affects the boundary of the corrosion region and the number of connected components; 3D mesh absorption simultaneously affects corrosion depth, corrosion area, and corrosion volume; and 2D and 3D mismatch indicates that the tiny corrosion labels have become discrete, and the triangular facets cannot support the corresponding geometric details, thus requiring the highest risk enhancement weight. In simpler terms, when there is fragmentation in 2D and coarseness in 3D, tiny pitting corrosion is most easily "eaten up" by a single facet.

[0076] Subsequently, the initial resolution mismatch risk value Qf0 is calculated based on Ff and Gf, with the formula: Qf0 = 1 − (1 − Ff)(1 − Gf); where the values ​​of Ff and Gf are both between 0 and 1. This calculation method ensures that Qf0 increases synchronously when either Ff or Gf increases; when both Ff and Gf increase simultaneously, Qf0 exhibits a joint amplification effect, used to characterize the risk superposition relationship when two-dimensional corrosion label discretization and three-dimensional mesh merging coexist. Then, Qf0 is corrected using the risk enhancement weight Wf to obtain the corrected risk value Qf1, calculated as: Qf1 = min(1, Wf × Qf0), where min represents taking the lower of the two values. When the product of Wf and Qf0 exceeds 1, Qf1 is set to 1, thus ensuring that the corrected risk value remains within the range of 0 to 1.

[0077] To reduce the risk level jumps of individual triangular faces caused by local noise, a risk consistency correction between adjacent triangular faces is introduced. The calculation range is defined as the neighborhood Nf of triangular face f, which includes the central triangular face, shared edge triangular face, and shared vertex triangular face. The corrected risk value of the i-th triangular face in the neighborhood is denoted as Qi1, the distance between the center point of the central triangular face and the center point of the i-th adjacent triangular face is denoted as di, and the average side length of the neighborhood of the central triangular face is denoted as lbarf. The neighborhood weight ui of the i-th adjacent triangular face is calculated according to the following formula: Where exp represents the exponential operation with the natural constant as the base, and dj represents the distance between the center triangle and the center point of the j-th adjacent triangle. The closer the distance is to the center triangle, the higher the corresponding weight. The neighborhood consistency risk value Qf2 is calculated based on the neighborhood weights, using the following formula: Further calculate the adjacent risk consistency Af, the formula is: Wherein, 0.30 is the risk difference normalization threshold, set as the minimum risk difference corresponding to adjacent triangular facets crossing one risk level. If the difference between Qf1 and Qf2 is 0, Af is 1, indicating that the central triangular facet and its neighbors have completely consistent risks; if the difference between the two reaches or exceeds 0.30, Af is 0, indicating that the central triangular facet has a risk abrupt change relative to its neighbors.

[0078] The final resolution mismatch risk value Qf is calculated using the following formula: Qf = 0.70Qf1 + 0.30Af × Qf2; where 0.70 is used to retain the risk contribution of the central triangular facet itself, and 0.30 is used to introduce the continuous risk contribution from the neighborhood. Af serves as a neighborhood consistency adjustment term; when the risk of the central triangular facet is consistent with that of the surrounding facets, the neighborhood risk participates in the correction; when the central triangular facet exhibits an isolated abrupt change relative to the surrounding facets, the neighborhood correction contribution is weakened. This process can enable dense pitting areas to form continuous risk regions, while suppressing isolated noisy facets from mistakenly triggering subsequent local subdivision or local reconstruction.

[0079] Risk levels are assigned based on the final resolution mismatch risk value Qf. Low risk corresponds to 0 ≤ Qf < 0.30, medium risk to 0.30 ≤ Qf < 0.55, high risk to 0.55 ≤ Qf < 0.75, and severe risk to 0.75 ≤ Qf ≤ 1.00. The threshold values ​​of 0.30, 0.55, and 0.75 are set as follows: 0.30 corresponds to the upper limit of risk that can be absorbed by area-weighted correction in conventional label mapping; 0.55 corresponds to the risk starting point requiring local subdivision to recover micro-corrosion boundaries; and 0.75 corresponds to the risk starting point where the current 3D mesh cannot be directly used for micro-pitting volume quantization. Low risk corresponds to subsequent conventional label mapping, medium risk to subsequent multi-point sampling, high risk to subsequent local subdivision, and severe risk to subsequent local reconstruction. Thus, each triangular facet obtains a resolution mismatch risk level directly corresponding to the subsequent dynamic adjustment method.

[0080] Based on the risk level, multi-point sampling, local subdivision, or local reconstruction are performed on the triangular facets to obtain an adjusted three-dimensional mesh. The two-dimensional corrosion labels are then remapped onto the adjusted three-dimensional mesh, and the area, depth, and volume of the corrosion region are calculated based on the mapping results.

[0081] After obtaining the resolution mismatch risk level for each triangular facet, the corresponding adjustment method is determined according to the risk level. Low-risk triangular facets retain their original topology, without changing vertex coordinates or facet connectivity. Medium-risk triangular facets undergo in-plane multi-point sampling, setting seven sampling positions at the three vertices, the midpoints of the three edges, and the facet's centroid. These seven sampling positions are then projected onto their corresponding two-dimensional projection regions, and the two-dimensional corrosion labels and corrosion probability distributions are read. High-risk triangular facets undergo local subdivision, first extracting the boundary pixels between corrosion and non-corrosion labels within the two-dimensional projection region, then calculating the centroid coordinates of the triangular facet, and finally using these positions as in-plane dividing points to divide the original triangular facet into multiple sub-facets. Severe-risk triangular facets undergo local reconstruction, calling the original three-dimensional sampling points from their neighborhood, and re-triangulating the local sampling points using the projection range corresponding to the corrosion connected components as constraints, forming reconstructed facets capable of expressing the boundaries and depths of minute pitting corrosion.

[0082] For multi-point sampling at medium-risk levels, any sampling location is represented using centroid coordinates. Let the three vertices of the triangular facet be P1, P2, and P3. The sampling location Ps is determined by three centroid weights w1, w2, and w3, satisfying the condition that the sum of w1, w2, and w3 equals 1, and Ps equals the sum of w1 multiplied by P1, w2 multiplied by P2, and w3 multiplied by P3. The centroid weights for the seven sampling locations are taken from the three vertex positions, the midpoints of the three edges, and one centroid position, with each centroid position having a weight of 1 / 3. After projecting each sampling location onto the multispectral image, the erosion probability of the corresponding pixel is read. If the projection falls within the coverage area of ​​multiple effective projected pixels, the erosion probability is weighted and averaged according to the sub-pixel coverage ratio. The medium-risk triangular facet does not change the geometric topology but forms an original facet with seven erosion sampling results.

[0083] For high-risk local subdivisions, the erosion label boundaries are first searched within the 2D projection region using an 8-neighborhood approach. A boundary pixel is defined as having an erosion label and having at least one non-erosion label pixel within its 8-neighborhood. The center point of the boundary pixel is converted to the centroid coordinates within the triangular facet using the original projection relationship. If the distance between adjacent boundary points within the triangular facet is less than 1 / 20 of the longest side of the facet, they are merged into a single segmentation point. The merged segmentation point is the arithmetic mean of the coordinates of adjacent boundary points. Subsequently, a local point set is constructed using the three vertices of the original triangular facet, the segmentation point corresponding to the erosion label boundary, and the facet centroid. Constrained triangulation is then performed within the triangular facet parameter plane, with the segmentation point corresponding to the erosion label boundary serving as the endpoint of the constraint edge, ensuring that the newly formed sub-facet boundary closely matches the 2D erosion label boundary. This prevents a high-risk facet from simultaneously spanning multiple micro-pitted areas and normal areas.

[0084] For local reconstruction at the severe risk level, original 3D sampling points are first extracted within the neighborhood of the severe risk triangular facet and its shared edges and vertices. These original 3D sampling points include the point cloud points retained during the 3D reconstruction stage; if only the triangular mesh is retained, facet vertices, edge midpoints, and centroids are used as substitute sampling points. The sampling points are projected onto a local reference surface, which is fitted by the center points of the non-corrosion-covered facets within the neighborhood of the severe risk triangular facet. The local reference surface is expressed as ax + by + cz + d = 0, where a, b, c, and d are obtained using the least squares method to minimize the sum of the squared perpendicular distances from all center points involved in the fitting to the plane. The projected local sampling points are triangulated in the 2D parametric plane, and the triangulation results are then mapped back to 3D space to obtain the locally reconstructed facet. If the number of non-corrosion-covered facets in the neighborhood is less than four, all neighborhood facet center points are used for fitting, and this region is marked as a severe risk adjustment result in the subsequent volume calculation results.

[0085] All patches formed through multi-point sampling, local subdivision, or local reconstruction establish an inheritance relationship with the original triangular patches. This inheritance relationship includes the original triangular patch number, the adjusted patch number, the 2D projection region number, the eroded connected component number, and the pixel representing the source of the erosion probability. For low-risk triangular patches, one original triangular patch corresponds to one adjusted patch; for medium-risk triangular patches, one original triangular patch corresponds to one adjusted patch with multi-point sampling attributes; for high-risk and severe-risk triangular patches, one original triangular patch corresponds to multiple adjusted patches. Each adjusted patch inherits the set of visible pixels from the original 2D projection region and recalculates the corresponding 2D pixel coverage relationship according to its own vertices or sampling positions.

[0086] When remapping the 2D erosion labels, the erosion coverage state and erosion probability are calculated for each adjusted facet. Let the area of ​​the adjusted facet be Aa, the area corresponding to the i-th effective sampling unit within the facet be Ai, the corresponding erosion probability be Pi, and the corresponding 2D erosion label be Li, where Li = 1 indicates erosion and Li = 0 indicates non-erosion. The erosion coverage Ca of the adjusted facet is calculated using the following formula: The corrosion probability Pa of the adjusted film is calculated using the following formula: When Ca is not less than 0.50 and Pa is not less than 0.50, the adjusted film is determined to be in a corroded state; when Ca is less than 0.50 or Pa is less than 0.50, the adjusted film is determined to be in a non-corroded state. The rule for setting 0.50 is the boundary value when the corroded and non-corroded areas are equal and the corrosion probability is equal to the non-corrosion probability.

[0087] After adjustment, the eroded connected region of the image is assigned the number with the highest area weighting among its valid sampling units; if no eroded labeled sampling unit exists, no eroded connected region number is assigned. The area, depth, and volume of the eroded region are calculated based on the mapping results of the adjusted image. The erosion participation area Acor is calculated using the following formula: Only the adjusted facet with erosion coverage is considered. The depth deviation of the adjusted facet relative to the local reference plane is obtained by the signed distance from the facet centroid to the local reference plane. The coordinates of the facet centroid Pc are the arithmetic mean of the coordinates of the three vertices, and the signed distance ha from Pc to the local reference plane is calculated by the following formula: Where xc, yc, and zc are the three-dimensional coordinates of Pc. If ha is negative, it indicates that the adjusted area forms a depression relative to the local reference surface, and the corrosion depth Da is the absolute value of ha; if ha is non-negative, Da is 0. The average depth Davg and maximum depth Dmax of the corrosion region are calculated by the following formulas: , The volume of the corroded region, Vcor, is calculated using the following formula: Through the above processing, computational efficiency is maintained in low-risk areas, corrosion coverage is corrected by in-plane multi-point sampling in medium-risk areas, micro-pitting boundaries are restored in high-risk areas by local subdivision along the two-dimensional corrosion label boundary, and three-dimensional concave morphology is restored in severely risky areas by local reconstruction. Finally, the two-dimensional corrosion label is remapped to the adjusted three-dimensional mesh, and the corrosion area, depth, and volume are obtained from the adjusted patch-level corrosion coverage state and depth deviation state.

[0088] Example 2, to verify the effectiveness of the method in this application in averaging micro-pitting corrosion and suppressing the risk of missed detection, is set up based on the method described in Example 1 to verify the improvement effect on the quantification accuracy of corrosion area, depth, and volume after dynamic adjustment based on the label fragmentation index Ff, the mesh micro-pit engulfment index Gf, and the resolution mismatch risk value Qf. Example 1 has already provided a complete processing flow for multispectral image corrosion recognition, two-dimensional projection of triangular patches, calculation of label fragmentation index within patches, calculation of mesh micro-pit engulfment index, risk level classification, and multi-point sampling, local subdivision, or local reconstruction based on the risk level. This example supplements the process with experimental data and comparative results.

[0089] In this embodiment, a Q345 bridge steel plate with dimensions of 300 mm × 200 mm × 8 mm was selected as the test component. Three types of corrosion areas were prepared on its surface: the first type was a dense pinhole corrosion area with a pitting diameter of 0.20 mm to 0.80 mm and a pitting depth of 0.05 mm to 0.42 mm; the second type was a sheet-like shallow corrosion area with a corrosion depth of 0.03 mm to 0.18 mm; and the third type was a boundary-fragmented corrosion area with discontinuous distribution of corrosion spots. Four groups of each type of corrosion area were set up, for a total of 12 groups of sample areas. Images of the surface of the test component were acquired using a 6-band multispectral camera with center bands of 450 nm, 550 nm, 650 nm, 720 nm, 850 nm, and 940 nm. The image resolution was 2448 × 2048 pixels, and the object size corresponding to a single pixel was 0.05 mm. A structured light 3D scanning device was used to acquire 3D point clouds and generate a triangular mesh model. The average side length of the triangular facets was adjusted to 0.62 mm. A 3D profilometer was used to re-measure the same area, and the profilometer measurement results were used as reference values ​​for the corrosion area, maximum depth, and volume.

[0090] To verify the effectiveness of the method in this application, Comparative Example 1 is provided. Comparative Example 1 uses a conventional direct mapping method, which projects two-dimensional corrosion labels onto a three-dimensional mesh based on camera calibration parameters, and determines whether the entire triangular facet belongs to the corrosion region based on the label value at the center point of the triangular facet. It does not calculate Ff, Gf, and Qf, nor does it perform multi-point sampling, local subdivision, or local reconstruction based on risk level. The method in this application, following the steps of Example 1, calculates Ff and Gf for each triangular facet, merges them to obtain Qf, and then classifies the risk level. It then performs functions such as preserving the original facet, in-face multi-point sampling, local subdivision along the corrosion label boundary, and local reconstruction based on the original three-dimensional sampling points. For high-risk and severe-risk areas, the average side length of the adjusted triangular facet is reduced from 0.62 mm to 0.18 mm to 0.27 mm, enabling the adjusted facet scale to match the spatial distribution of micro-pitting corrosion labels.

[0091] Table 1

[0092]

[0093] Table 1 shows that the Ff and Gf values ​​in the dense pinhole pitting corrosion region are both high, indicating that this region suffers from both 2D corrosion label discrepancies and excessively large 3D mesh patch sizes. If the 2D corrosion labels are directly assigned to the original triangular patches, multiple independent pitting corrosions are easily merged into the same patch. The method in this application classifies such regions as high-risk or severe-risk and triggers local subdivision or local reconstruction, reducing the adjusted average patch edge length to 0.19 mm to 0.24 mm. The Ff value in the boundary-fractured corrosion region is high while the Gf value is at a medium level, indicating that the main problem originates from the fragmentation of the 2D label boundary. Therefore, local subdivision along the corrosion label boundary can restore the corrosion boundary. The Ff and Gf values ​​in the flaky shallow corrosion region do not reach the severe-risk level. The corrosion coverage can be corrected by sampling 7 points within the surface without changing the mesh topology.

[0094] Table 2

[0095]

[0096] As shown in Table 2, Comparative Example 1, due to the direct assignment of values ​​to the center points of triangular facets, easily misclassifies micro-pitting as non-corrosion when it does not cover the center point. Furthermore, multiple pittings falling into the same facet are easily merged into a single corrosion region, resulting in a total false negative rate of 29.87%. The method in this application identifies the dispersion of corrosion labels within the facet using Ff, identifies the degree to which the 3D mesh absorbs micro-corrosion pits using Gf, and then dynamically adjusts based on Qf. This reduces the total false negative rate to 7.23%, demonstrating that the method in this application can effectively reduce the averaging or false negative occurrence of micro-pitting.

[0097] Table 3

[0098]

[0099] As shown in Table 3, the method of this application outperforms Comparative Example 1 in terms of corrosion area, maximum corrosion depth, average corrosion depth, corrosion volume, and number of pitting corrosions. Specifically, the average error in corrosion volume is reduced from 28.5% to 9.7%, a reduction of 66.0%; the error in the number of pitting corrosions is reduced from 29.9% to 7.2%, a reduction of 75.9%. This is because Comparative Example 1 uses a triangular facet as the smallest tag-bearing unit, which cannot represent the spatial differences between multiple corrosion spots within the facet. The method of this application incorporates both the discrete state of the two-dimensional corrosion tag and the geometric carrying capacity of the three-dimensional mesh into the evaluation, and adopts different adjustment methods for different risk levels, enabling the two-dimensional corrosion tag to be remapped to the adjusted three-dimensional mesh in a manner closer to the actual corrosion boundary.

[0100] Table 4

[0101]

[0102] As shown in Table 4, although the corrosion volume obtained in Comparative Example 1 in region A1 appears to be similar to the reference value, its corrosion area is significantly larger and its maximum depth is significantly smaller. This indicates that multiple deep and small pinhole corrosions are averaged into a larger but shallower corrosion area. This result masks the risk of local maximum depth corrosion, which is not conducive to the assessment of the remaining wall thickness of the load-bearing component. After local reconstruction, the method in this application increases the number of identified pits from 29 to 43 and the maximum depth from 0.29 mm to 0.39 mm, which is closer to the 3D profilometer reference value of 0.42 mm. This indicates that the adjusted 3D mesh can more realistically represent the pit boundary and bottom morphology.

[0103] In summary, this embodiment demonstrates that by calculating the label fragmentation index Ff and the mesh micro-pit merging index Gf within the patch, and fusing them to obtain the resolution mismatch risk value Qf for risk level classification, regions where the scale of the two-dimensional multispectral corrosion label and the three-dimensional mesh patch do not match can be accurately identified. Furthermore, by performing multi-point sampling, local subdivision, or local reconstruction based on the risk level, the probability of small corrosion defects being averaged or missed by the same triangular patch can be effectively reduced, improving the accuracy of corrosion area, maximum corrosion depth, and corrosion volume quantification results. The above data indicate that the method of this application has higher pitting corrosion identification completeness and volume quantification reliability compared to conventional direct mapping methods, and can meet the high-precision corrosion assessment needs of scenarios with dense small corrosion defects, such as bridge steel plates, rail transit steel components, and pipeline outer walls.

[0104] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for three-dimensional reconstruction and volume quantization of corroded regions based on multispectral feature fusion, characterized in that, include: Acquire multispectral images and a three-dimensional mesh model of the component under test, perform corrosion identification on the multispectral images, and obtain two-dimensional corrosion labels and corrosion probability distributions; Based on the camera calibration parameters, each triangular facet in the 3D mesh model is projected onto the multispectral image to obtain the 2D projection region corresponding to each triangular facet. Based on the discreteness of the etched labels, the number of etched connected components, and the fluctuation of the etched probability within the two-dimensional projection area, a label fragmentation index within the patch is generated. Based on the scale, neighborhood curvature variation, depth residual, and local sampling sparsity of the triangular facets, a mesh micro-pit swallowing index is generated; By combining the label fragmentation index and the mesh micro-pit agglomeration index within the patch, a resolution mismatch risk value is obtained, and the risk level is classified accordingly. Based on the risk level, multi-point sampling, local subdivision, or local reconstruction are performed on the triangular facets to obtain an adjusted three-dimensional mesh. The two-dimensional corrosion labels are then remapped onto the adjusted three-dimensional mesh, and the area, depth, and volume of the corrosion region are calculated based on the mapping results.

2. The method for three-dimensional reconstruction and volume quantization of eroded regions based on multispectral feature fusion according to claim 1, characterized in that, The process of performing corrosion identification on the multispectral image to obtain a two-dimensional corrosion label and corrosion probability distribution includes: dividing the multispectral image into several geometrically consistent regions based on the local normal variation and depth continuity after backprojecting the three-dimensional mesh model onto the multispectral image; selecting pixels with stable multi-band reflection characteristics as regional reference spectra within each geometrically consistent region; performing relative spectral correction on the multispectral pixels in the same region based on the regional reference spectra; generating a corrosion response map based on the difference between the corrected multispectral pixels and the regional reference spectra; and applying connectivity constraints and probabilistic smoothing to the corrosion response map to obtain a two-dimensional corrosion label and corrosion probability distribution.

3. The method for three-dimensional reconstruction and volume quantization of corrosion regions based on multispectral feature fusion according to claim 2, characterized in that, Obtaining the two-dimensional projection region corresponding to each triangular facet includes: transforming the vertices of each triangular facet to the multispectral image coordinate system according to the camera calibration parameters to obtain the initial image coverage area of ​​each triangular facet; performing local virtual subdivision on triangular facets that do not meet the projection consistency requirements based on the depth continuity and edge projection curvature of each triangular facet after transformation to obtain a set of subdivided facets; removing occluded projection pixels according to the depth order and normal visibility of the subdivided facet set in the multispectral image; retaining the effective projection pixels that are directly visible to the multispectral camera; and forming a two-dimensional projection region with the effective projection pixels.

4. The method for three-dimensional reconstruction and volume quantization of corroded regions based on multispectral feature fusion according to claim 3, characterized in that, The two-dimensional projection region is determined by the sub-pixel coverage ratio, including: dividing the pixels in the multispectral image into multiple sub-pixel sampling points, and determining whether each sub-pixel sampling point falls within the projection range of the subdivided patch set. When multiple subdivided patches cover the same sub-pixel sampling point, the sub-pixel sampling point corresponding to the subdivided patch with the smallest interpolation depth and satisfying normal visibility is retained. The sub-pixel coverage ratio is determined based on the number of sub-pixel sampling points belonging to the same original triangular patch in the same image pixel, and the association between the two-dimensional projection region and the original triangular patch is established based on the sub-pixel coverage ratio.

5. The method for three-dimensional reconstruction and volume quantization of eroded regions based on multispectral feature fusion according to claim 1, characterized in that, Generate the label fragmentation index within the patch, including: The two-dimensional projection area corresponding to each triangular facet is converted into an equal-area sampling unit within the facet according to the sub-pixel coverage ratio, and the two-dimensional corrosion label and corrosion probability distribution are assigned to each equal-area sampling unit. Based on the adjacency relationship between equal-area sampling units, the erosion labels within the same triangular facet are traced to obtain the number of erosion connected regions within the facet and the spatial interval of each erosion connected region. Based on the spatial interval, area ratio difference, centroid dispersion state, and corrosion probability fluctuation state of the corrosion connected domains, a normalized fusion is performed to obtain the in-patch label fragmentation index, which is used to characterize the risk of small corrosion defects being averaged by the same triangular patch.

6. The method for three-dimensional reconstruction and volume quantization of eroded regions based on multispectral feature fusion according to claim 1, characterized in that, Generate the mesh micro-pit swallowing index, including: Each triangular facet is selected as the center, and adjacent triangular facests that share edges or vertices are selected to construct a facet neighborhood. The facet scale characterization value is determined based on the edge length distribution of the triangular facests in the facet neighborhood and the equivalent diameter of the eroded connected domain in the two-dimensional projection region. Based on the direction and magnitude of the normal change of each triangular facet relative to the central triangular facet within the neighborhood of the facet, the curvature change features of the neighborhood are extracted. The depth residual characteristics are determined based on the signed distance between the center point of each triangular facet in the neighborhood of the facet and the local reference facet. The local sampling sparsity characteristics are determined by combining the spacing between local sampling points. The mesh micro-pit swallowing index is obtained by fusing the facet scale characterization value, neighborhood curvature variation characteristics, depth residual characteristics and local sampling sparsity characteristics.

7. The method for three-dimensional reconstruction and volume quantization of eroded regions based on multispectral feature fusion according to claim 6, characterized in that, By fusing the label fragmentation index and the mesh micro-pit agglomeration index within the patch, a resolution mismatch risk value is obtained, including: Based on the label fragmentation index and mesh micro-pit absorption index within the same triangular facet, the triangular facet is determined to be in a low mismatch state, a 2D erosion label fragmentation-dominated state, a 3D mesh absorption-dominated state, or a 2D-3D joint mismatch state. The 2D-3D joint mismatch state is assigned a higher risk enhancement weight than the 2D erosion label fragmentation-dominated state and the 3D mesh absorption-dominated state. Based on the risk enhancement weight, the label fragmentation index and mesh micro-pit absorption index within the facet are fused to obtain the resolution mismatch risk value of each triangular facet.

8. The method for three-dimensional reconstruction and volume quantization of eroded regions based on multispectral feature fusion according to claim 7, characterized in that, Based on this, risk levels are classified, including: Using the neighborhood of a triangular facet as the calculation range, the neighborhood consistency risk value is determined based on the resolution mismatch risk value of each triangular facet in the neighborhood and the distance to the center point. The adjacent risk consistency is determined based on the difference between the resolution mismatch risk value of the current triangular facet and the neighborhood consistency risk value, and the resolution mismatch risk value is corrected using the adjacent risk consistency. Based on the corrected resolution mismatch risk value, each triangular facet is classified into low-risk, medium-risk, high-risk, and severe-risk levels.

9. The method for three-dimensional reconstruction and volume quantization of corroded regions based on multispectral feature fusion according to claim 8, characterized in that, Based on the risk level, multi-point sampling, local subdivision, or local reconstruction are performed on the triangular facets to obtain an adjusted 3D mesh, including: For low-risk triangular faces, the original topology is preserved. For medium-risk triangular faces, multiple sampling points are set at the vertices, edge midpoints, and face centroids. For high-risk triangular faces, the boundary pixels between eroded and non-eroded labels within the 2D projection area are extracted. These boundary pixels are then back-calculated into the triangular face and used as dividing points for local subdivision. For severely risky triangular faces, local reconstruction is performed based on the original 3D sampling points in their neighborhood and the corresponding projection range of the eroded connected domains, resulting in an adjusted 3D mesh.

10. The method for three-dimensional reconstruction and volume quantization of corroded regions based on multispectral feature fusion according to claim 9, characterized in that, The process involves remapping two-dimensional corrosion labels onto the adjusted three-dimensional mesh, and calculating the area, depth, and volume of the corrosion region based on the mapping results. This includes: establishing an inheritance relationship between the adjusted facets formed through multi-point sampling, local subdivision, or local reconstruction and the original triangular facets; associating the original two-dimensional projection region, corrosion connectivity number, and corrosion probability distribution with the corresponding adjusted facets; reassigning two-dimensional corrosion labels to the adjusted facets according to the inheritance relationship; determining the corrosion coverage state, corrosion probability, and associated corrosion connectivity of each adjusted facet; determining the corrosion participation area based on the corrosion coverage state and corrosion probability of the adjusted facets; and calculating the area, depth, and volume of the corrosion region by combining the depth deviation of the adjusted facets relative to the local reference surface.

Images