A three-dimensional reconstruction method of bridge underwater pile pier structure combining structure priori and feature fusion

Abstract

The application provides a three-dimensional reconstruction method of a bridge underwater pile pier structure combining structure priori and feature fusion, comprising: obtaining multi-view surrounding underwater images of the bridge underwater pile pier; generating a sparse three-dimensional point cloud of the pile pier based on the underwater images, and fitting to obtain cylindrical geometric parameters of a pile pier body; converting the sparse three-dimensional point cloud into a three-dimensional Gaussian point cloud, and the three-dimensional Gaussian point being represented by a spatial position, a covariance matrix and a transparency parameter; constructing and optimizing a total loss function, and the total loss function fusing a pixel reconstruction loss term, a cylindrical geometric priori loss term and an underwater radiation transmission loss term, wherein a constraint weight applied to a pile pier disease area by the cylindrical geometric priori loss term is lower than a constraint weight applied to a pile pier body area, and geometric deviation of the disease area relative to an ideal cylindrical shape is allowed; and iteratively optimizing parameters of the three-dimensional Gaussian point cloud by minimizing the total loss function, and generating a three-dimensional model of the bridge underwater pile pier structure.

Description

Technical Field

[0001] This invention belongs to the field of underwater structure surface detection technology, specifically involving a three-dimensional reconstruction method for underwater bridge pile pier structures that combines structural priors and feature fusion. Background Technology

[0002] In existing technologies, deep learning-based semantic segmentation methods have been widely applied to target extraction in remote sensing scenes. These methods typically employ an encoder-decoder structure, obtaining high-level semantic features through downsampling and gradually restoring spatial resolution during the decoding stage to output pixel-level classification results. However, in complex photovoltaic (PV) scenarios, existing methods still suffer from the following shortcomings:

[0003] First, photovoltaic (PV) modules often exhibit significant multi-scale variations and dense arrangement in remote sensing imagery. Simultaneously, influenced by installation methods, terrain materials, weather conditions, and lighting conditions, the spectral and textural characteristics of PV modules are highly variable and easily confused with similar background features such as roads, rooftops, water bodies, and shadows, leading to both false positives and false negatives. While conventional convolutional neural network (CNN) encoders, such as those used in the U-Net and DeepLab series, excel in extracting local texture and edge features, their inherent local receptive field limits their ability to model long-range contextual dependencies. This makes it challenging for models to distinguish between terrain features with similar spectral characteristics but vastly different spatial layouts (such as large areas of PV panels versus large areas of concrete pavement or rooftops of specific materials), hindering their ability to understand the scene layout from a global perspective and thus suppress background interference.

[0004] Secondly, in pixel-level segmentation tasks, deep features possess strong semantic expressive power, but they are prone to losing small targets and edge details during continuous downsampling and nonlinear transformations. Shallow features, while retaining more texture and edge information, lack sufficient semantic representation capabilities. Existing feature fusion strategies, such as simple skip connections or feature map concatenation, often struggle to simultaneously consider both high-level semantics and low-level details. This simple information superposition fails to fully consider the essential differences and complementary relationships between features at different levels in terms of semantic granularity and spatial details, and the fusion process lacks effective guidance or modulation mechanisms. Therefore, when decoding to restore resolution, feature confusion is easily caused, leading to blurred boundaries of photovoltaic modules, jagged artifacts, or adhesion or breakage between targets in densely arranged areas, severely affecting the structural integrity and geometric accuracy of the segmentation results.

[0005] Furthermore, although some improvement methods attempt to introduce attention mechanisms (such as channel attention and spatial attention) to enhance feature representation or use dilated convolution to expand the receptive field, these improvements are mostly general designs and do not optimize for the specific, regular rectangular or strip-shaped geometry of photovoltaic modules. The strong directionality (horizontal or vertical arrangement) and regular grid-like layout of photovoltaic arrays in remote sensing images is an important prior knowledge. However, existing general segmentation models lack mechanisms to effectively utilize such strong structural priors, resulting in insufficient ability of the models to capture and maintain the regular boundaries of photovoltaic panels in complex backgrounds. Their robustness in dealing with scenes of partial shading, uneven illumination, or irregular arrangement needs to be improved.

[0006] In summary, existing technologies still have limitations in processing photovoltaic module segmentation in complex remote sensing scenarios, particularly in global context modeling, effective fusion and guidance of multi-level features, and utilization of target-specific geometric priors, which restrict further improvements in segmentation accuracy and practicality. Summary of the Invention

[0007] To address the shortcomings and deficiencies of existing technologies, this invention provides a method and system for three-dimensional reconstruction of underwater bridge pier structures based on surround-view images. The method acquires multi-view surround-view underwater images of the bridge piers and generates a sparse three-dimensional point cloud of the piers based on these images. This sparse point cloud is then fitted to obtain the cylindrical geometric parameters of the pier body. The sparse three-dimensional point cloud is transformed into a three-dimensional Gaussian point cloud characterized by spatial location, covariance matrix, and transparency parameters. A total loss function is constructed and optimized, fusing the cylindrical geometry prior loss term, the underwater radiative transfer loss term, and the pixel reconstruction loss term. During the optimization process, to accurately characterize surface defects on the piers, the constraint weight of the cylindrical geometry prior loss term is reduced in defective areas, allowing for geometric deviations from the ideal cylindrical shape. The parameters of the three-dimensional Gaussian point cloud are iteratively optimized by minimizing the total loss function, ultimately generating a high-precision three-dimensional model of the underwater bridge pier structure. This invention enhances the morphological fidelity of the main structure of the pile pier by introducing prior constraints of cylindrical geometry, compensates for light intensity attenuation and scattering effects caused by water quality using an underwater radiation transmission physical model, and combines an adaptive weight adjustment mechanism for the diseased area. While ensuring the overall structural accuracy, it effectively preserves the detailed features of surface diseases, overcoming the technical difficulties of low image quality caused by the underwater environment and the difficulty of traditional reconstruction methods in taking into account both the overall shape and local details.

[0008] The specific technical solution adopted by this invention to solve its technical problem is as follows:

[0009] A three-dimensional reconstruction method for an underwater bridge pier structure includes:

[0010] Acquire underwater images of bridge piers from multiple perspectives using panoramic underwater photography.

[0011] Based on the underwater image, a sparse three-dimensional point cloud of the pile pier is generated, and the cylindrical geometric parameters of the pile pier body are obtained by fitting.

[0012] The sparse 3D point cloud is transformed into a 3D Gaussian point cloud, wherein the 3D Gaussian points are characterized by spatial location, covariance matrix and transparency parameter.

[0013] A total loss function is constructed and optimized, which integrates a pixel reconstruction loss term, a cylindrical geometry prior loss term, and an underwater radiation transmission loss term. The cylindrical geometry prior loss term applies a lower constraint weight to the defective area of ​​the pile pier than to the constraint weight applied to the main body area of ​​the pile pier, allowing the defective area to produce a geometric deviation from the ideal cylindrical shape.

[0014] By iteratively optimizing the parameters of the three-dimensional Gaussian point cloud by minimizing the total loss function, a three-dimensional model of the underwater pile pier structure of the bridge is generated.

[0015] Furthermore, the method for acquiring underwater images by multi-view surround shooting is as follows: a professional underwater camera is used to perform 360° full-circle surround shooting and vertical layer shooting of the damaged pile pier, so as to achieve full-dimensional image coverage of the pile pier. The height of the imaged pier body accounts for a preset proportion of the original pile pier height, and the images of adjacent layers maintain a preset proportion of overlapping area. During surround shooting, images are acquired at fixed angle intervals.

[0016] Furthermore, before generating the sparse 3D point cloud, the process includes a step of enhancing the details of the multi-view panoramic underwater image by adjusting the local sharpness, contrast, and color balance of the image to improve the recognizability of image feature points.

[0017] Furthermore, the sparse 3D point cloud is generated by extracting stable feature points from the image, matching the same feature points from different images, and combining them with triangulation. The cylindrical geometric parameters include the direction of the cylinder axis, the position of the axis, and the cylinder radius. These parameters are obtained by fitting the radial residual from each point in the sparse point cloud to the cylinder axis, and a cylindrical coordinate system for the pile pier is established based on the direction of the cylinder axis.

[0018] Furthermore, when converting the sparse 3D point cloud into a 3D Gaussian point cloud, initial constraints are applied to the 3D Gaussian point cloud. The spatial positions of all 3D Gaussian points are within the preset deviation range of the fitted cylinder radius, and the major axis of the covariance matrix of the 3D Gaussian points on the cylinder surface is parallel to the cylinder tangent direction, while the minor axis is perpendicular to the cylinder surface.

[0019] Furthermore, the cylindrical geometric prior loss term is calculated based on the deviation between the radial distance from the center of the three-dimensional Gaussian point to the axis of the cylinder and the cylinder radius. The main body area and the defect area of ​​the pile pier are distinguished by a weighting coefficient. The weighting coefficient of the main body area is greater than that of the defect area, so as to strengthen the cylindrical shape constraint of the main body area and relax the constraint intensity of the defect area.

[0020] Furthermore, the underwater radiative transmission loss term is constructed based on a depth-related underwater illumination attenuation and scattering physical model, which constrains the color, brightness, and contrast changes of the rendered image to conform to the physical laws of underwater imaging, compensates for light intensity attenuation and scattering effects caused by water quality, and reduces image color drift and detail distortion caused by underwater turbidity.

[0021] Furthermore, the total loss function is weighted and summed by preset weighting coefficients for the pixel reconstruction loss term, the cylindrical geometry prior loss term, and the underwater radiative transfer loss term; the iterative optimization projects the 3D Gaussian point cloud from the world coordinate system to the camera coordinate system through a differentiable rasterizer and performs pixel-level rendering, calculates the total loss function value, and then adjusts the parameters of the 3D Gaussian point cloud through gradient backpropagation until the loss function value meets the preset threshold.

[0022] Furthermore, after generating the three-dimensional model, the method also includes a step of evaluating the accuracy of the three-dimensional model, using structural similarity index and peak signal-to-noise ratio as quantitative evaluation indicators to characterize the structural similarity and pixel-level difference between the reconstructed rendered image and the real underwater image, respectively.

[0023] Furthermore, a system for implementing the above-described method for three-dimensional reconstruction of underwater bridge pier structures includes:

[0024] The image acquisition module is used to acquire multi-view panoramic underwater images of the bridge's underwater piers;

[0025] The point cloud and parameter fitting module is used to generate a sparse three-dimensional point cloud of the pile pier based on the underwater image, and to fit the cylindrical geometric parameters of the pile pier body.

[0026] A Gaussian point cloud conversion module is used to convert the sparse 3D point cloud into a 3D Gaussian point cloud, wherein the 3D Gaussian points are characterized by spatial location, covariance matrix and transparency parameter.

[0027] The loss function optimization module is used to construct and optimize the total loss function, which integrates pixel reconstruction loss term, cylindrical geometry prior loss term and underwater radiation transmission loss term. The constraint weight of the cylindrical geometry prior loss term is reduced for the pile pier defect area, and the parameters of the three-dimensional Gaussian point cloud are iteratively optimized by minimizing the total loss function.

[0028] The model generation module is used to generate a 3D model of the underwater pile pier structure of the bridge based on the optimized 3D Gaussian point cloud parameters.

[0029] And a computer device including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the method described above.

[0030] A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0031] Compared to existing technologies, this invention and its preferred embodiments offer at least the following advantages: By introducing a cylindrical geometric prior loss term, the regular geometric shape of the pile pier body is incorporated as a strong constraint into the optimization process of the 3D Gaussian point cloud, effectively overcoming the problem of overall morphological distortion in the reconstructed model caused by low underwater image quality, and significantly improving the geometric fidelity of the main structure. Simultaneously, by introducing an underwater radiation transmission loss term, the physical process of light attenuation and scattering in water is explicitly modeled, compensating for the unique radiation distortion of underwater imaging, making the surface material and texture of the reconstructed model closer to reality. Crucially, this invention creatively proposes a differentiated processing mechanism to reduce the weight of the cylindrical geometric prior constraint in the pile pier defect area. This mechanism allows reasonable geometric deviations in the defect area during the optimization process, thereby accurately capturing and preserving the geometric details of surface defects such as cracks, spalling, and holes while ensuring the accuracy of the main structure, achieving a balance between overall morphological fidelity and local detail representation. Furthermore, by synergistically optimizing pixel reconstruction loss, cylindrical geometric prior loss, and underwater physical loss, and further defining the preferred steps such as image enhancement, point cloud initialization, and weight allocation strategies in the dependent claims, a complete and robust reconstruction process is formed, which can adapt to different water quality conditions and disease types, providing a high-precision and quantifiable three-dimensional model data foundation for the structural health detection of underwater bridge piers. Attached Figure Description

[0032] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0033] Figure 1 This is a diagram illustrating the process of three-dimensional reconstruction based on image enhancement and Gaussian splashing in an embodiment of the present invention;

[0034] Figure 2 A diagram of the raw image dataset captured by the optical camera in an embodiment of the present invention;

[0035] Figure 3 This is an image of the dataset enhanced by the point sharpness weighted image enhancement method according to an embodiment of the present invention;

[0036] Figure 4 This is a diagram showing the three-dimensional reconstruction result of the underwater bridge structure according to an embodiment of the present invention;

[0037] Figure 5 This is a flowchart illustrating the overall process of an embodiment of the present invention. Detailed Implementation

[0038] To make the features and advantages of the present invention more apparent and understandable, specific embodiments are described below in detail:

[0039] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0040] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0041] To address the problem of insufficient accuracy in existing 3D reconstruction techniques for digitally archiving damaged pile structures, this invention proposes a 3D reconstruction method for underwater structures based on surround-view imaging. The aim is to solve the problems of massive data volume, the significant impact of underwater image quality on existing algorithms, and slow 3D modeling and structural evaluation efficiency. Compared with traditional manual judgment methods, the method in this application achieves high accuracy and fast processing speed in generating 3D models of underwater structures.

[0042] like Figure 5 As shown, the implementation process of the three-dimensional reconstruction method for underwater bridge structures based on surround-view images provided by the present invention includes the following steps:

[0043] Step 1, Image Dataset Acquisition and Prior Information Acquisition: Use a professional underwater camera to take multi-angle photos around the damaged bridge pier to acquire an image dataset of the underwater structure. The image dataset includes images that cover the pier body at least 360 degrees and that capture images of the pier body at a height of 70% of the original pier body, to ensure a comprehensive view of the underwater structure.

[0044] While acquiring underwater image datasets, a structured light scanning device is simultaneously used to scan the same area of ​​the pile pier to obtain initial sparse point cloud data of the local area of ​​the pile pier. This data is used to perform spatial location compensation for subsequent feature points extracted from images, thereby improving the accuracy of the spatial coordinates of the feature points.

[0045] Preferably, the above image dataset acquisition steps and equipment include using an underwater camera with 1080p 60fps shooting capability. The camera parameters include a wide-angle lens, a horizontal field of view of 155 degrees, a 1 / 1.7-inch CMOS sensor, an autofocus system, and built-in image stabilization technology, which can effectively reduce image blur caused by water flow and camera vibration during underwater shooting. During the shooting process, a professional underwater camera is used to perform 360-degree surround shooting of the starting position on the upper part of the damaged pier. Then, the camera position is moved vertically downwards, and the surround shooting is repeated in a loop. It is necessary to ensure that the lower and upper images overlap by at least 30% until the vertical imaging height accounts for approximately 70% of the original pier height, ensuring that all directions of the pier are fully covered. At least one image is taken every 30 degrees to ensure high density and uniformity of the image data.

[0046] Step 2, Image Enhancement Processing: Enhancement processing is performed on the acquired underwater image dataset. By adjusting local sharpness and contrast, the image quality is improved by enhancing the details in the image and improving the contrast.

[0047] As a preferred approach, the above original image enhancement method adopts a customized enhancement method based on point sharpness weights, which specifically includes: 1) performing Gaussian filtering on the underwater original image to denoise and suppress water scattering noise; 2) calculating the point sharpness value of each pixel in the image to characterize the richness of details; 3) assigning weights according to the point sharpness value to improve the contrast and sharpness of high-sharpness detail areas and to make appropriate adjustments to low-sharpness background areas; 4) performing color balance optimization to correct underwater color drift and restore the true color characteristics of the pile pier.

[0048] Step 3: Generating a sparse point cloud: By analyzing sparse feature points in the image, combining the compensated feature points with structured light scanning data, the pose changes experienced by the camera during shooting are estimated, and a 3D geometric point cloud of the scene is reconstructed. Specifically, stable and identifiable feature points are extracted from the enhanced image dataset, and these feature points are described. Then, by matching identical feature point pairs in different images, the relative positional relationships between images are determined. Next, based on the known camera parameters and the corresponding feature point positions in the images, triangulation is used to calculate the position of each feature point in 3D space. Triangulation uses the known image coordinates of the same feature point from two viewpoints, combined with the camera's intrinsic and extrinsic parameters, to calculate the accurate position of the point in 3D space and generate a sparse point cloud.

[0049] Step four, 3D reconstruction with prior constraints: The cylindrical structural parameters of the pile pier are fitted using the initial sparse point cloud to obtain the cylinder axis direction, position, and radius parameters. A cylindrical coordinate representation of the pile pier is then established based on the cylinder axis direction, allowing 3D points to be represented by axial coordinates, circumferential angles, and radial distances to the axis. These point clouds are then transformed into a 3D Gaussian ellipsoid set. Each point cloud is described by its position, covariance matrix, and transparency, forming a preliminary 3D Gaussian point cloud.

[0050] Specifically, this includes: converting sparse point clouds into Gaussian point clouds, where each 3D Gaussian point is determined by its spatial location. Covariance matrix and transparency The representation is made, and each 3D Gaussian point needs to be combined with the prior constraints of the pile pier cylindrical structure, specifically including: (1) Gaussian point cloud initialization constraints: the spatial position of all three-dimensional Gaussian points must satisfy: the radial distance from the center of the Gaussian point to the axis ∈ [radius mean - 5cm, radius mean + 5cm] (radius mean is the optimal estimate of the radius of the pile pier cylindrical structure obtained by fitting the sparse point cloud by the least squares method, which is a single determined value and serves as the benchmark value for the spatial position constraint of the Gaussian point cloud); the major axis of the covariance matrix of the Gaussian points on the cylindrical surface is parallel to the tangent direction of the cylinder, and the minor axis is perpendicular to the cylindrical surface; (2) Gaussian point cloud optimization constraints: a cylindrical fitting error control mechanism is introduced in the three-dimensional reconstruction process to force the Gaussian point cloud to fit the cylindrical structure and ensure that the reconstruction model conforms to the actual geometric shape of the pile pier.

[0051] As a preferred method, the cylindrical structure parameters of the pile pier are fitted based on the initial sparse point cloud. The cylindrical structure parameters include at least the direction of the cylindrical axis, the position of the axis, and the radius, and a coordinate expression of the cylindrical structure of the pile pier is established.

[0052] Then, these 3D Gaussian points are projected from the world coordinate system to the camera coordinate system, and then rendered using a differentiable rasterizer to generate the final rendered image, accurately restoring the spatial position and shape of the underwater structure. This includes:

[0053] Initialize a 3D Gaussian point cloud. Each 3D Gaussian point is represented by its spatial location, covariance matrix, and transparency, generating an initial Gaussian point cloud dataset. The generated Gaussian ellipsoid set is then projected from the world coordinate system to the planar camera coordinate system. Using the camera's intrinsic and extrinsic parameters, the coordinates of the 3D Gaussian points are converted to coordinates on a 2D image. Finally, the 3D Gaussian points are rasterized based on the projection results.

[0054] The initialized 3D Gaussian points are optimized by using a loss function that conforms to the physical laws of underwater light attenuation and light scattering to adjust the spatial position, covariance matrix and transparency of each Gaussian point, so as to ensure that these points can accurately represent the geometry of the underwater structure.

[0055] The optimized 3D Gaussian points are projected onto the 2D image space through viewpoint transformation, using... -mix( Image rendering is performed using a blending technique to generate a fused image, ensuring that the surface details and color information of underwater structures are accurately reproduced.

[0056] Step 5, Model Optimization and Accuracy Evaluation: The parameters of the 3D Gaussian point cloud are optimized using a loss function constrained by physical conditions; the accuracy of the 3D reconstruction model is evaluated using SSIM (Structural Similarity Index) and PSNR (Peak Signal-to-Noise Ratio). Specifically, this includes:

[0057] Based on the complexity of the reconstructed scene, the density of 3D Gaussian points is adjusted. More Gaussian points are added in complex or detailed areas, while the density of Gaussian points is reduced in simplified areas. This ultimately generates a high-precision 3D reconstruction model that accurately represents the geometry and details of the underwater bridge damage structure.

[0058] Depending on the complexity of the reconstructed scene, SSIM is used to evaluate the structural similarity between the generated rendered image and the real image. SSIM reflects the similarity of brightness, contrast, and structure; a value closer to 1 indicates a greater structural similarity between the reconstructed image and the real image. PSNR is used to evaluate the pixel-level differences between images; a higher PSNR value indicates better image quality and smaller pixel-level differences between the generated 3D model and the actual underwater structure.

[0059] As a preferred approach, a differentiable rasterizer and loss function are used for optimization. The loss function is calculated by comparing the rendered result with the real image, and the parameters of all three-dimensional Gaussian points are optimized through gradient backpropagation. In addition to the image reconstruction-related loss, the loss function further incorporates prior constraints on the pile pier cylindrical structure and constraints that satisfy the physical conditions for underwater imaging. Finally, by minimizing the loss function, the rendered image is made as close as possible to the real image, thereby generating a more accurate three-dimensional model.

[0060] As a further preferred option, the specific process based on the enhanced image data and the point cloud data generated by structured light scanning includes:

[0061] First, some initial sparse point cloud data is obtained through structured light scanning. This point cloud data provides the preliminary spatial location of the underwater structure, with each point representing a position coordinate of the underwater structure. The initial sparse point cloud obtained by structured light scanning is then fused with the sparse point cloud generated in step three based on image feature matching to obtain a high-precision fused sparse point cloud, which serves as the basis for subsequent parameter fitting of the pile pier cylindrical structure.

[0062] Based on the initial sparse point cloud, the cylindrical structure parameters of the pile pier are fitted. These parameters include at least the direction, position, and radius of the cylindrical axis. A coordinate representation of the cylindrical structure is established for subsequent calculation of prior constraints in the optimization process. The cylindrical axis can be represented as:

[0063]

[0064] in, For a point on the axis, The unit vector along the axis. Let be a scalar parameter, and let the cylinder radius be . For any three-dimensional point in a sparse point cloud Its radial distance to the cylinder axis is defined as:

[0065]

[0066] in, For point The closest point on the axis For point Perpendicular distance to the axis. Cylinder parameters. This can be obtained by minimizing the radial residual, for example:

[0067]

[0068] in, For any and not given Parallel vectors, then the point Cylindrical coordinates can be represented as:

[0069]

[0070] in, For axial coordinates, For circumferential angle, This represents the radial distance.

[0071] As a further preferred option, the specific process for initializing the 3D Gaussian point cloud includes:

[0072] Initialize a 3D Gaussian point cloud based on the fitted point cloud data. The Gaussian representation of each point cloud is as follows:

[0073]

[0074] in, The mean of the Gaussian points. Let covariance matrix be the variance matrix. It is the determinant of the covariance matrix. It is its inverse matrix.

[0075] The generated Gaussian ellipsoid set is projected from the world coordinate system to the planar camera coordinate system. Using the camera's intrinsic and extrinsic parameter matrices, the coordinates of the 3D Gaussian points are converted to coordinates on the 2D image. The projection formula is:

[0076]

[0077]

[0078] In the formula, It is the Jacobian matrix of the affine approximation of the projection transformation. It is the corresponding transpose matrix. It is the camera pose matrix transpose, It is a location The coordinates on the z-axis.

[0079] The 3D Gaussian points are rasterized based on the projection results. The rasterization process converts each projected Gaussian point into a pixel in the image space and assigns transparency and color information to each pixel.

[0080] As a further preferred option, the above adopts -The specific process of image rendering using hybrid techniques includes:

[0081] The color of each Gaussian point is calculated using the spherical harmonic coefficients. The formula for calculating the spherical harmonic function, used to describe the lighting model, is:

[0082]

[0083] in, The color of the Gaussian point. yes The first order A spherical harmonic function, and It is the angular representation of the viewpoint direction vector.

[0084] Use the transparency and position of Gaussian points for blended rendering. - The blending technique synthesizes multiple Gaussian points at different depths. The transparency of each Gaussian point... and color The mixing formula used in the generation of the final image is:

[0085]

[0086] in, It is the weight of each Gaussian point. It is the color of each Gaussian point. It is the number of three-dimensional Gaussian points projected onto the pixel.

[0087] As a further preferred option, the specific process of the above further optimization includes:

[0088] Pixel-level rendering is performed using a differentiable rasterizer. A loss function is calculated by comparing the color of each pixel with the color of the real image. The loss function is defined as:

[0089]

[0090] in, It is a predicted image obtained through rendering. It is a real image. It refers to the number of pixels.

[0091] Further, a priori constraint loss term for pile-pier cylindrical structures is introduced. and the constraint loss term that satisfies the physical conditions for underwater imaging .in, The constraint can be based on the consistency between the radial distance from the center of the Gaussian point to the axis of the cylinder and the radius of the cylinder, thus constraining the point set of the main area of ​​the pile pier. definition:

[0092]

[0093] in, For point Radial distance to the axis, To fit the cylinder radius, A tolerance threshold is set; and the constraint strength is reduced or relaxed for the diseased area to allow reasonable deviations of the diseased part relative to the cylindrical structure, thus enabling accurate representation. Weights are introduced. definition:

[0094]

[0095] Among them, the main area Take the larger value, the diseased area Take the smaller value. Loss term due to physical constraints of underwater imaging. To constrain the color, brightness, and contrast variations of rendered images under different viewpoints and distances to maintain consistency with depth-related illumination attenuation and scattering-induced variation patterns, thereby reducing color drift and detail distortion caused by underwater turbidity, the total loss function can be expressed as:

[0096]

[0097] in, and These are preset weighting coefficients.

[0098] As a further preferred option, the specific process of evaluating the accuracy of the 3D reconstruction model using SSIM (Structural Similarity Index) and PSNR (Peak Signal-to-Noise Ratio) metrics includes:

[0099] The SSIM metric is calculated using the following formula:

[0100]

[0101] in: and These are the pixel values ​​of the two images. and These are the average brightness values ​​of the two images, respectively. and It is the variance of the two images. It is the covariance of the two images. and It is a constant, where , This is the commonly used default value.

[0102] The PSNR index is calculated using the following formula:

[0103]

[0104] in: It is the maximum possible pixel value in the image, usually 255. It is the mean squared error, calculated using the following formula:

[0105]

[0106] in: and These are two images at different positions. pixel values, and It refers to the size of the image.

[0107] The following, in conjunction with the accompanying drawings, illustrates and describes more details of the implementation of the technical solution of the present invention through embodiments:

[0108] like Figure 1 As shown, this embodiment of the invention provides a method for three-dimensional reconstruction of underwater bridge pier defects based on image enhancement and 3D Gaussian Splatting, specifically including the following steps:

[0109] The original images of several underwater damaged pile piers were obtained by taking pictures with an underwater optical camera. First, after determining the location of the underwater pile pier damage, the device equipped with the underwater camera was operated so that the camera was facing the damaged area and taking pictures around the pier in 360 degrees. One image was taken every 30 degrees during the circling. Then the camera position was moved vertically downwards and the circling pictures were repeated. It is necessary to ensure that the lower image overlaps with the upper image by at least 30% until the vertical imaging height accounts for about 70% of the original pier height, so as to ensure that all directions of the pier are fully covered.

[0110] Figure 2 This dataset contains raw images of underwater bridge piers with defects, collected using a professional underwater camera and following standardized surround-view imaging techniques. The images cover a 360° circumferential area of ​​the piers, with the vertical imaging height covering 70% of the original pier height. Upper and lower images maintain at least 30% overlap. One image is captured every 30° circumferentially, comprehensively recording the original appearance features of the underwater piers, including the main body of the pier, the defective areas, and the underwater background. However, due to environmental factors such as insufficient underwater lighting, turbid water, and water flow scattering, the images suffer from issues such as blurred details, low contrast, and color shift in some areas. This dataset serves as the primary data for subsequent image enhancement processing.

[0111] The acquired original image is further enhanced using an image enhancement method based on point sharpness weights.

[0112] Figure 3 To Figure 2 The dataset consists of original underwater pile pier images and the resulting image enhancement method based on point sharpness weighting. This dataset addresses the issues of lost details and blurriness in the original images through local sharpness adjustment, contrast-weighted fusion, color balance optimization, and edge sharpening. It significantly improves the recognition of key features such as pile pier surface texture and diseased area boundaries, while effectively suppressing color distortion caused by the underwater environment. This preserves the true visual information of the pile pier structure, providing a high-quality image foundation for subsequent feature point extraction and matching of sparse point clouds, and greatly reducing the false matching rate of feature point detection.

[0113] The enhanced image data and point cloud data generated by structured light scanning are further fitted to the bridge pier cylinder. In this embodiment, to make the 3D Gaussian point cloud more suitable for the underwater bridge pier scene, the cylindrical structural parameters of the pier body are fitted based on the initial sparse point cloud generated by SFM, and a coordinate representation of the pier cylindrical structure is established. The cylindrical axis is represented as:

[0114]

[0115] in, For a point on the axis, Let be the unit vector along the axis, and let the radius of the cylinder be . For any point in a sparse point cloud Its projection parameters on the axis, the nearest point, and the radial distance are as follows:

[0116]

[0117] Cylinder parameters This can be obtained by minimizing the radial residual:

[0118]

[0119] Furthermore, in the direction of the axis Construct orthogonal basis in a perpendicular plane ,For example:

[0120]

[0121] in, For any and not given Parallel vectors, then the point Cylindrical coordinates can be represented as:

[0122]

[0123] Initialize a 3D Gaussian point cloud. Each 3D Gaussian point is represented by its spatial location, covariance matrix, and transparency, generating a preliminary Gaussian point cloud dataset. The expression for each Gaussian point is:

[0124]

[0125] in, The mean of the Gaussian points. Let covariance matrix be the variance matrix. It is the determinant of the covariance matrix. It is its inverse matrix.

[0126] Then, using the projection formula:

[0127]

[0128]

[0129] In the formula, It is the Jacobian matrix of the affine approximation of the projection transformation. It is the corresponding transpose matrix. It is the camera pose matrix transpose, It is a location exist Coordinates on the axis.

[0130] Each optimized 3D Gaussian point is converted into two-dimensional image coordinates.

[0131] After projection, the mixing weights can be calculated based on the Gaussian distribution and the opacity of the point. Specifically:

[0132]

[0133] In the formula, For opacity, It is the position vector of that point. It is the position vector of the corresponding transpose.

[0134] The color of each Gaussian point is calculated using the spherical harmonic coefficients. The formula for calculating the spherical harmonic function, used to describe the lighting model, is:

[0135]

[0136] in, The color of the Gaussian point. yes The first order A spherical harmonic function, and It is the angular representation of the viewpoint direction vector.

[0137] Furthermore, the transparency and position of the Gaussian points are blended and rendered. Multiple Gaussian points at different depths are composited using alpha blending. The transparency of each Gaussian point is... and color The mixing formula used in the generation of the final image is:

[0138]

[0139] in, It is the weight of each Gaussian point. It is the color of each Gaussian point. It is the number of three-dimensional Gaussian points projected onto the pixel.

[0140] Finally, the images generated by the model are compared with the actual training set images, the loss function is calculated, and the position, opacity, covariance matrix, and spherical harmonic coefficients of all 3D Gaussian points are optimized through gradient backpropagation. The loss function is:

[0141]

[0142] in, It is a predicted image obtained through rendering. It is a real image. It refers to the number of pixels.

[0143] Based on this, a prior constraint loss term for cylindrical structures is introduced for the prior constraints of pile-pier cylindrical structures. The loss function is used to constrain the consistency between the radial distance from the center of the Gaussian point in the main body area of ​​the pile pier to the cylinder axis and the fitted radius. The loss function is defined as:

[0144]

[0145] in, For the point set of the main area of ​​the pile pier, Calculated based on the aforementioned axis model To fit the cylinder radius, The tolerance threshold is set according to the engineering requirements for the surface flatness of underwater bridge piers, and is 5% to 10% of the estimated cylinder radius. In this embodiment, 2 cm is preferred to allow for reasonable geometric deviations in the Gaussian points of the main pier area, avoiding excessive constraints that could lead to model distortion. The constraint strength is reduced or relaxed for the defective areas to allow for reasonable deviations from the cylindrical structure, thus ensuring accurate representation. Weights are introduced accordingly. definition:

[0146]

[0147] Among them, the main area Take the larger value, the diseased area Take the smaller value. Specifically: Automatic identification of diseased areas is performed based on enhanced underwater image features and 3D point cloud geometric features. Diseased areas in the image are identified through image edge detection and texture mutation analysis, and verified by combining point cloud features such as local density mutations and excessive radial distance deviations. The weighting coefficient of the diseased area... The weighting coefficient for the main area of ​​the pile pier is set to 0.1~0.3, weakening the cylindrical constraint to preserve the true geometric shape of the defect. A value of 0.7~1.0 is used to strengthen the cylindrical constraint to ensure the accuracy of the structural shape. In this embodiment, the defective area is preferably selected. Take 0.2, main area Take 0.9.

[0148] Furthermore, to meet the physical conditions for underwater imaging and reduce color drift and detail distortion caused by underwater turbidity, physical condition constraints are introduced. In this embodiment, a depth-related attenuation model can be used to constrain the physical consistency of the rendered image. For example, depth can be utilized... With attenuation coefficient Constructing an underwater imaging model:

[0149]

[0150] in, The colors of the rendered image, To meet the requirements of underwater attenuated observation images, This is the water attenuation coefficient, characterizing the degree of attenuation of light propagation in water. It is related to water turbidity and the concentration of suspended particles in the water. For turbid water, a value of 0.05~0.2m is used. -1 Take 0.01~0.05m of clean water. -1 ; Represents pixels The corresponding depth can be determined by the depth of the Gaussian points that contribute to the pixel during the rasterization process. The weighted average is used to obtain A; A is the ambient light constant, which characterizes the background light intensity of the underwater environment and is related to water depth and surface lighting conditions. Its value ranges from 0 to 255 (matching image pixel values). In this embodiment... Take 0.1m -1 A is 50.

[0151] The loss due to physical constraints can be defined as:

[0152]

[0153] in, For pixel index, This is a real image.

[0154] Therefore, the total loss function can be expressed as:

[0155]

[0156] in, and These are preset weighting coefficients. Among them, The weighting coefficients for the prior geometric loss term of the cylinder are used to control the constraint strength of the cylindrical structure morphology of the pile pier. The weighting coefficients for the underwater radiation transmission loss term are used to control the constraint strength of the physical laws governing underwater imaging. Both values ​​are adjusted based on the underwater image quality and the structural characteristics of the pile pier. The preferred value is 0.1~0.5. The preferred value is 0.05~0.2; in this embodiment, the underwater turbidity environment and the cylindrical structure characteristics of the pile pier are considered. Take 0.3, Take 0.1.

[0157] By optimizing the parameters of all Gaussian points through gradient backpropagation, the optimization of these parameters will enable the Gaussian points to fit the target scene more accurately, ultimately generating a high-quality 3D model that matches the real image.

[0158] Figure 4The final 3D reconstruction model of an underwater bridge pier structure, achieved using the full-process method described in this invention, is a high-precision 3D model obtained through sparse point cloud generation, cylindrical structure parameter fitting, 3D Gaussian point cloud initialization and optimization, differentiable rasterization rendering, and gradient backpropagation optimization using multi-constraint loss functions. This model accurately reproduces the cylindrical geometry of the underwater pier, with axial, circumferential, and radial dimensions matching the actual pier height. It also clearly presents the spatial location, morphological characteristics, and detailed information of the defective areas in the pier body, achieving regular reconstruction of the pier's main structure and accurate representation of defective areas. The model's accuracy, as evaluated by SSIM and PSNR indices, meets the high-precision requirements for underwater structure digital archiving in terms of structural similarity and pixel-level fidelity. It can be directly applied to defect detection, structural assessment, and digital management of underwater bridge piers.

[0159] Based on the same inventive concept, this invention also provides a computer device, comprising: one or more processors, and a memory for storing one or more computer programs; the programs include program instructions, and the processor executes the program instructions stored in the memory. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, used to implement one or more instructions, specifically for loading and executing one or more instructions stored in a computer storage medium to implement the above-described method.

[0160] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, performs the above-described method. This storage medium can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0161] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0162] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

[0163] This invention is not limited to the above-described preferred embodiments. Anyone inspired by this invention can derive other forms of three-dimensional reconstruction methods for underwater bridge pier structures that combine structural a priori knowledge with feature fusion. All equivalent variations and modifications made within the scope of the claims of this invention shall fall within the scope of this invention.

Claims

1. A three-dimensional reconstruction method for underwater pile pier structures of bridges, characterized in that, include: Acquire underwater images of bridge piers from multiple perspectives using panoramic underwater photography. Based on the underwater image, a sparse three-dimensional point cloud of the pile pier is generated, and the cylindrical geometric parameters of the pile pier body are obtained by fitting. The sparse 3D point cloud is transformed into a 3D Gaussian point cloud, wherein the 3D Gaussian points are characterized by spatial location, covariance matrix and transparency parameter. A total loss function is constructed and optimized, which integrates a pixel reconstruction loss term, a cylindrical geometry prior loss term, and an underwater radiation transmission loss term. The cylindrical geometry prior loss term applies a lower constraint weight to the defective area of ​​the pile pier than to the constraint weight applied to the main body area of ​​the pile pier, allowing the defective area to produce a geometric deviation from the ideal cylindrical shape. By iteratively optimizing the parameters of the three-dimensional Gaussian point cloud by minimizing the total loss function, a three-dimensional model of the underwater pile pier structure of the bridge is generated.

2. The three-dimensional reconstruction method for an underwater bridge pier structure according to claim 1, characterized in that: The method for acquiring underwater images by multi-view surround shooting is as follows: a professional underwater camera is used to perform 360° full-circle surround shooting and vertical layer shooting of the damaged pile pier to achieve full-dimensional image coverage of the pile pier. The height of the imaged pier body accounts for a preset proportion of the original pile pier height, and the images of adjacent layers maintain a preset proportion of overlapping area. During surround shooting, images are acquired at fixed angle intervals.

3. The three-dimensional reconstruction method for an underwater bridge pier structure according to claim 1, characterized in that: Before generating the sparse 3D point cloud, the process also includes a step of enhancing the details of the multi-view panoramic underwater images by adjusting the local sharpness, contrast and color balance of the images to improve the recognizability of image feature points.

4. The three-dimensional reconstruction method for an underwater bridge pier structure according to claim 1, characterized in that: The sparse 3D point cloud is generated by extracting stable feature points from the image, matching the same feature points from different images, and combining them with triangulation. The cylindrical geometric parameters include the direction of the cylinder axis, the position of the axis, and the cylinder radius. These parameters are obtained by fitting the radial residual from each point in the sparse point cloud to the cylinder axis, and a cylindrical coordinate system for the pile pier is established based on the direction of the cylinder axis.

5. The three-dimensional reconstruction method for an underwater bridge pier structure according to claim 1, characterized in that: When converting the sparse 3D point cloud into a 3D Gaussian point cloud, initial constraints are applied to the 3D Gaussian point cloud. The spatial positions of all 3D Gaussian points are within the preset deviation range of the fitted cylinder radius, and the major axis of the covariance matrix of the 3D Gaussian points on the cylinder surface is parallel to the cylinder tangent direction, while the minor axis is perpendicular to the cylinder surface.

6. The three-dimensional reconstruction method for an underwater bridge pier structure according to claim 1, characterized in that: The cylindrical geometric prior loss term is calculated based on the deviation of the radial distance from the center of the three-dimensional Gaussian point to the axis of the cylinder from the cylinder radius. The weighting coefficient distinguishes between the main area and the defect area of ​​the pile pier. The weighting coefficient of the main area is greater than that of the defect area, so as to strengthen the cylindrical shape constraint of the main area and relax the constraint intensity of the defect area.

7. The three-dimensional reconstruction method for an underwater bridge pier structure according to claim 1, characterized in that: The underwater radiative transmission loss term is constructed based on a depth-related underwater light attenuation and scattering physical model. It constrains the color, brightness, and contrast changes of the rendered image to conform to the physical laws of underwater imaging, compensates for light intensity attenuation and scattering effects caused by water quality, and reduces image color drift and detail distortion caused by underwater turbidity.

8. The three-dimensional reconstruction method for an underwater bridge pier structure according to claim 1, characterized in that: The total loss function is a weighted sum of the pixel reconstruction loss term, the cylindrical geometry prior loss term, and the underwater radiative transfer loss term using preset weighting coefficients. The iterative optimization projects the 3D Gaussian point cloud from the world coordinate system to the camera coordinate system using a differentiable rasterizer and performs pixel-level rendering. After calculating the total loss function value, the parameters of the 3D Gaussian point cloud are adjusted through gradient backpropagation until the loss function value meets the preset threshold.

9. The three-dimensional reconstruction method for an underwater bridge pier structure according to claim 1, characterized in that: After generating the three-dimensional model, the method further includes a step of evaluating the accuracy of the three-dimensional model, using structural similarity index and peak signal-to-noise ratio as quantitative evaluation indicators to characterize the structural similarity and pixel-level difference between the reconstructed rendered image and the real underwater image, respectively.

10. A system for implementing the three-dimensional reconstruction method for underwater bridge pier structures as described in claim 1, characterized in that, include: The image acquisition module is used to acquire multi-view panoramic underwater images of the bridge's underwater piers; The point cloud and parameter fitting module is used to generate a sparse three-dimensional point cloud of the pile pier based on the underwater image, and to fit the cylindrical geometric parameters of the pile pier body. A Gaussian point cloud conversion module is used to convert the sparse 3D point cloud into a 3D Gaussian point cloud, wherein the 3D Gaussian points are characterized by spatial location, covariance matrix and transparency parameter. The loss function optimization module is used to construct and optimize the total loss function, which integrates pixel reconstruction loss term, cylindrical geometry prior loss term and underwater radiation transmission loss term. The constraint weight of the cylindrical geometry prior loss term is reduced for the pile pier defect area, and the parameters of the three-dimensional Gaussian point cloud are iteratively optimized by minimizing the total loss function. The model generation module is used to generate a 3D model of the underwater pile pier structure of the bridge based on the optimized 3D Gaussian point cloud parameters.

Images