A method for identifying damage of a structured surface based on Gaussian spatter

Abstract

The application discloses a kind of structural surface damage identification methods based on Gaussian splashing, belong to structural health monitoring and three-dimensional computer vision technical field.The method includes the following steps: S1, obtains the multi-angle image of structural surface, and generates the three-dimensional Gaussian splashing reference model of structural surface by motion recovery structure (SfM) and Gaussian splashing rendering technology;S2, based on the geometry and appearance attribute of each Gaussian element in the reference model, calculate the local abnormality between target Gaussian element and its neighborhood element, and according to the abnormality threshold, the region to be identified is screened out;S3, Gaussian element in the region to be identified is carried out cluster analysis and feature extraction, and the classification and quantization of structural surface damage are completed.The present application utilizes the high fidelity and lightweight characteristics of Gaussian splashing model, without explicit mesh reconstruction, realizes the high-precision, automated identification of small damage on structural surface, significantly improves the detection efficiency and accuracy.

Description

Technical Field

[0001] This invention belongs to the field of structural health monitoring and three-dimensional computer vision technology, specifically involving a structural surface damage identification method based on Gaussian splatting, which is particularly suitable for automated and high-precision damage detection on the surfaces of large infrastructure such as bridges, buildings, and dams. Background Technology

[0002] Surface damage to structures (such as cracks, spalling, and corrosion) is a key indicator for assessing structural health. Traditional damage detection mainly relies on manual visual inspection or contact measurement, which suffers from problems such as low efficiency, high subjectivity, high cost, and safety risks.

[0003] With the development of computer vision technology, detection methods based on UAV imagery and 3D reconstruction have gradually become a research hotspot. However, existing methods still face challenges: 1. 3D models based on traditional meshes or point clouds are prone to geometric noise or smoothing effects when reconstructing minute damages such as fine cracks, resulting in the loss of damage details; 2. Although Neural Radiation Field (NeRF)-based methods can generate high-precision views, their training and inference processes are computationally intensive, making it difficult to meet the needs of rapid detection; 3. Most existing damage recognition algorithms are performed on 2D images, lacking consistency in 3D space, easily affected by changes in lighting and viewing angle, and difficult to accurately quantify the spatial scale of damage such as crack depth and spalling volume.

[0004] Therefore, there is an urgent need for a technical solution that can efficiently and faithfully reconstruct a three-dimensional structural model and directly perform automated, high-precision damage identification and quantification on the model. Summary of the Invention

[0005] The purpose of this invention is to propose a structural surface damage identification method based on Gaussian splashing to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A method for identifying structural surface damage based on Gaussian splashing includes the following steps: S1. Acquire multi-angle images of the surface of the structure to be detected, and generate a three-dimensional Gaussian splash reference model of the surface of the structure based on the multi-angle images. S2. Traverse each Gaussian splatter reference model, calculate the local anomaly of each Gaussian splatter based on its geometric and appearance properties, and filter out the regions to be identified according to a preset anomaly threshold. S3. Perform cluster analysis on the Gaussian elements in the region to be identified, and extract and classify the features of the resulting clusters to achieve the identification and quantification of structural surface damage.

[0008] Further, step S1 includes the following sub-steps: S11. Acquire multi-angle images of the surface of the structure to be detected. Using the multi-angle images, estimate the camera pose of the images through the Structure for Motion (SfM) algorithm, and acquire the sparse point cloud of the structure surface. The sparse point cloud serves as the initial geometric support of the structure surface. S12. Initialize the sparse point cloud into a three-dimensional Gaussian splatter set, and use the camera pose to iteratively train and optimize the position, covariance (representing shape and orientation), color (RGB), and opacity (Alpha) of the Gaussian splatter set through differentiable rendering and stochastic gradient descent (SGD) algorithm, so as to minimize the error between the view rendered by the Gaussian splatter set and the multi-angle image, and finally generate the three-dimensional Gaussian splash benchmark model.

[0009] The beneficial effects of the above-mentioned further solutions are as follows: In this invention, Gaussian splashing technology uses three-dimensional Gaussian primitives as explicit representations, avoiding the implicit queries and mesh topology construction of NeRF, thus achieving extremely fast rendering speed. By using the sparse point cloud provided by SfM as the initial anchor point, the convergence speed and geometric accuracy of model training are guaranteed, and the generated benchmark model can reproduce the geometric and texture details of the structural surface with high fidelity.

[0010] Further, in step S2, the calculation of the local anomaly degree includes: S21. For any target Gaussian element in the benchmark model Using spatial index structures such as KD-trees or octrees, we can quickly query the set of K nearest neighbors Gaussian elements in spatial location. In order to establish the local context of the target Gaussian element; S22. Calculate the target Gaussian elements respectively. The set of K nearest neighbors Gaussian elements Geometric deviation between Appearance deviation What deviation is mentioned? Appearance deviation Used to quantify the consistency of the target Gaussian element with its local neighborhood in terms of geometry and surface color; S23, the geometric deviation Appearance deviation Normalization and weighted fusion are performed to obtain the local anomaly degree of the target Gaussian unit. The local anomaly degree The calculation formula is:

[0011] in, and The preset weights are used to adjust the proportion of geometric and appearance factors in anomaly measurement.

[0012] Furthermore, in S22, the geometric deviation Defined as the covariance matrix of the target Gaussian element Mean of the covariance matrix of the K nearest neighbor Gaussian elements The Frobenius norm between, i.e. The appearance deviation Defined as the RGB value of the target Gaussian element The average RGB value of the K nearest neighbor Gaussian element The Euclidean distance between them, i.e. .

[0013] The beneficial effect of the above-mentioned further solution is that, in this invention, structural surface damage such as cracks and spalling manifests as local geometric and appearance anomalies in the three-dimensional Gaussian model. Gaussian elements on intact structural surfaces exhibit local consistency in geometry (covariance, i.e., shape and orientation) and appearance (RGB values). However, the orientation of Gaussian elements in cracked areas is disordered, leading to geometric deviations. The size increases; abnormal color in the rusted areas leads to visual deviations. Increased. This invention calculates local anomaly degree by fusing geometric and appearance deviations, enabling it to sensitively detect different types of minute damage with strong robustness.

[0014] Further, step S3 includes the following sub-steps: S31. The density-based spatial clustering (DBSCAN) algorithm is used to perform spatial location clustering of the three-dimensional coordinates of the centers of all Gaussian elements within the region to be identified. This algorithm can adaptively aggregate spatially adjacent and density-reachable anomalous Gaussian elements into independent damage regions, forming one or more damage clusters, without requiring pre-specifying the number of clusters, and is suitable for damage with irregular shapes; S32. Extract the geometric features of each damaged cluster. Specifically, perform principal component analysis (PCA) on the center coordinates of all Gaussian elements within the cluster to obtain three principal eigenvalues. And the corresponding main direction. Based on the feature values, calculate the linearity, flatness and other morphological descriptors of the clusters, and use these descriptors as feature vectors, input them into a pre-trained damage classifier based on support vector machine, and output the damage type, which includes cracks, spalling or corrosion; S33. Calculate the quantitative index of the damage based on the geometric morphological characteristics of the damage cluster and the damage type.

[0015] Furthermore, in S32, the geometric features include linearity. and flatness High linearity The value indicates that the clusters are linearly distributed and have high flatness. The value indicates that the clusters are distributed in a sheet-like pattern.

[0016] The beneficial effects of the above-mentioned further solutions are: In this invention, the DBSCAN algorithm can effectively aggregate damage regions of arbitrary morphology and filter sparse noise points. By analyzing the geometric morphology of the clusters using PCA, different types of damage can be effectively distinguished: cracks are usually linear (high... Value), peeling is usually in flakes (high) (Value). Using this as a feature input to the classifier improves the accuracy of damage classification. Simultaneously, based on information such as the principal axis length of the clusters, the crack length of the structure is determined by the range along the first principal direction, allowing direct quantification of the damage degree in three-dimensional space and avoiding projection errors from two-dimensional image estimation.

[0017] The beneficial effects of this invention are: (1) This invention proposes a method for damage identification directly on a three-dimensional Gaussian splash (GS) model. The GS model has both high-fidelity rendering quality and real-time rendering speed, overcoming the shortcomings of insufficient detail in traditional mesh models and large computational load in NeRF models.

[0018] (2) By calculating the geometric and appearance anomalies in the neighborhood of Gaussian elements, this invention achieves sensitive detection of minute damages such as fine cracks and early corrosion, avoiding over-smoothing or omission of damage.

[0019] (3) This invention achieves automated classification and quantification of damage by performing three-dimensional spatial clustering and morphological analysis on abnormal high-score elements. The detection results such as crack length and spalling area are directly measured in three-dimensional space with high accuracy, and avoids the perspective and scale problems of two-dimensional image analysis. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings involved in the embodiments are now briefly described. Obviously, the drawings in the following description are merely illustrative of some embodiments of the present invention. For those skilled in the art, other forms of drawings can be constructed based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of the structural surface damage identification method based on Gaussian splashing according to the present invention; Figure 2 This invention provides a Gaussian splashing process for a structural surface damage identification method based on Gaussian splashing. Figure 3 This is the damage identification process of the structural surface damage identification method based on Gaussian splashing in this invention; Figure 4 This is the damage classification process of the structural surface damage identification method based on Gaussian splashing in this invention; Figure 5 This is the baseline model for embodiments of the present invention; Figure 6 This is a local anomaly thermal diagram according to an embodiment of the present invention; Figure 7 This refers to the region to be identified in this embodiment of the invention. Figure 8 The damage clustering and classification results are shown in the embodiments of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the application is described and illustrated below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. All other embodiments obtained by those skilled in the art based on the embodiments provided in this application without inventive effort are within the scope of protection of this application.

[0023] It should be understood that the step numbers used in the text are for ease of description only and are not intended to limit the order in which the steps are performed.

[0024] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0025] The terms “comprising” and “including” indicate the presence of the described feature, whole, step, operation, element and / or component, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or collections thereof.

[0026] The following description, in conjunction with the accompanying drawings, further illustrates the structural surface damage identification method based on Gaussian splashing proposed in this invention.

[0027] Example 1: like Figure 1 As shown, this invention provides a method for identifying structural surface damage based on Gaussian splashing, comprising the following steps: S1. Acquire multi-angle images of the surface of the structure to be inspected, and generate a three-dimensional Gaussian splash reference model of the surface of the structure based on the multi-angle images. Figure 2 The specific process of step S1 is shown: S11. A drone equipped with a high-definition camera can be used to fly around the structure to be inspected, acquiring multi-angle images from different heights and angles, ensuring that the forward and lateral overlap between the images is greater than 80%. Using the multi-angle images, the sparse point cloud of the structure surface and the camera pose are obtained through the Structure of Motion (SfM) algorithm.

[0028] S12. Using the sparse point cloud as the initial position, initialize the 3D Gaussian primitive set, and adjust the position of the Gaussian primitives using differential rendering and stochastic gradient descent (SGD). ), covariance ( Characterizing shape and orientation, color ( (represented by the coefficients of the spherical harmonic function) and opacity ( Training optimization is performed. This optimization process includes not only gradient descent adjustments to the aforementioned parameters, but also adaptive density control, which involves periodically densifying Gaussian primitives based on the rendering gradient. This includes cloning or splitting primitives in regions with large gradients and removing opacity. Primitives close to zero. Training can be performed for 30,000 iterations until the rendering loss function, a combination of L1 loss and SSIM loss, converges, ultimately generating a 3D Gaussian splash baseline model. Figure 5 This is the baseline model for this embodiment, and it includes a simulated damage area, including a linear crack and a circular spall.

[0029] S2. Traverse each Gaussian splatter base model, calculate the local anomaly degree of each Gaussian splatter based on its geometric and appearance properties, and filter out the regions to be identified according to the preset anomaly degree threshold. Figure 3 The specific process of step S2 is shown: S21. In this embodiment, for any target Gaussian element in the model... First, the set of K nearest neighbors Gaussian elements in spatial location is quickly retrieved using spatial indexing structures such as KD-trees. .

[0030] S22. Calculate the geometric deviations respectively. Appearance deviation Geometric deviation Covariance matrix based on target primitives Mean of the covariance matrix of the K nearest neighbor primitives The calculation of the Frobenius norm between them, i.e. This is used to measure the local consistency of its shape and orientation. Appearance deviation. Based on the RGB values ​​of the target primitive , The fundamental color is calculated from a spherical harmonic function at a specific viewpoint, and is the RGB mean of the K nearest neighbor primitives. The Euclidean distance between them is calculated as follows: .

[0031] S23. Before weighted fusion, in order to eliminate the influence of dimensions, the calculated values ​​are first... and The value is min-max normalized over the entire model or a large local window, scaling it to the [0,1] interval. A weighted summation is then performed to obtain the local anomaly degree of the target primitive. In this embodiment, weights can be set. This approach emphasizes geometric morphology to accurately identify abrupt changes in surface discontinuities caused by cracks. Finally, an anomaly threshold is set. , Anomalies can be statistically analyzed from the distribution of known intact sample regions, and the 95th percentile can be used, or the Otsu method can be adaptively employed to determine all anomalies. Greater than this threshold The Gaussian elements are filtered out to form the region to be identified. For example... Figure 6 As shown, primitives from intact surfaces are displayed against a dark background, indicating low local anomaly. Conversely, primitives from damaged areas are displayed against a light background, indicating significantly higher anomaly, forming a clear local anomaly heatmap. Anomaly thresholds are then set. All Gaussian elements with an anomaly score greater than the threshold are filtered out to form the region to be identified. For example... Figure 7 As shown, the filtered areas to be identified are represented by red outliers, which spatially correspond precisely to... Figure 6 The simulated damage location.

[0032] S3. Perform cluster analysis on Gaussian elements within the region to be identified, and extract and classify the features of the resulting clusters to achieve the identification and quantification of structural surface damage. Figure 4 The specific process of step S3 is shown: S31. The density-based spatial clustering (DBSCAN) algorithm is used to cluster the x, y, and z coordinates of the center positions of all Gaussian elements within the region to be identified. The neighborhood radius of DBSCAN can be set. The minimum sample size is 5 cm, which can be adjusted based on the average density and typical damage scale of the baseline model, setting the minimum sample size for the core object. A value of 20 is used to filter out sparse noise points. DBSCAN can automatically divide spatially connected anomalous primitives into clusters, effectively separating different damage areas.

[0033] S32. For each damaged cluster, extract its geometric morphological features. Calculate the eigenvalues ​​of the three principal components of the cluster using Principal Component Analysis (PCA). Linearity is defined accordingly. and flatness .Will , Information such as the volume of clusters and the number of primitives is used as feature vectors and input into a pre-trained damage classifier based on support vector machines. A set of criteria can be set: when and At that time, it was classified as "crack"; when and When the appearance deviates, it is classified as "peeling"; Significantly higher than geometric deviation Furthermore, when the shape is irregular, it can be classified as "rust". For example... Figure 8 As shown, the DBSCAN algorithm successfully clustered two spatially separated damage regions into "damage cluster 1 (crack)" and "damage cluster 2 (stripping)" and distinguished them with different colors.

[0034] S33. Calculate the quantitative index based on the geometric morphological characteristics of the clusters. If classified as "crack," its three-dimensional length can be calculated by the cluster along the first principal direction (corresponding to...). The bounding box length of the cluster is determined, and its average width can be determined by the cluster along the second principal direction (corresponding to the average width of the cluster). The standard deviation or bounding box width estimate.

[0035] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for identifying structural surface damage based on Gaussian splashing, characterized in that, Includes the following steps: S1. Acquire multi-angle images of the surface of the structure to be detected, and generate a three-dimensional Gaussian splash reference model of the surface of the structure based on the multi-angle images. S2. Traverse each Gaussian splatter reference model, calculate the local anomaly of each Gaussian splatter based on its geometric and appearance properties, and filter out the regions to be identified according to a preset anomaly threshold. S3. Perform cluster analysis on the Gaussian elements in the region to be identified, and extract and classify the features of the resulting clusters to achieve the identification and quantification of structural surface damage.

2. The structural surface damage identification method based on Gaussian splashing according to claim 1, characterized in that, Step S1 specifically includes the following: S11. Obtain multi-angle images of the surface of the structure to be detected, and use the multi-angle images to obtain sparse point clouds of the surface of the structure through a motion recovery structure algorithm. S12. Initialize the sparse point cloud into a three-dimensional Gaussian primitive set, and train and optimize the position, covariance, color and opacity of the Gaussian primitives through differential rendering and stochastic gradient descent to generate a three-dimensional Gaussian splash baseline model.

3. The structural surface damage identification method based on Gaussian splashing according to claim 1, characterized in that, The calculation of local anomaly degree in step S2 specifically includes: S21. For any target Gaussian element, query its K nearest neighbor Gaussian element set; S22. Calculate the geometric deviation between the target Gaussian element and the set of K nearest neighbor Gaussian elements respectively. Appearance deviation ; S23, the geometric deviation Appearance deviation Weighted fusion is performed to obtain the local anomaly degree of the target Gaussian element. The calculation formula is as follows: in, and Preset weights.

4. The structural surface damage identification method based on Gaussian splashing according to claim 3, characterized in that, The geometric deviation Covariance matrix based on target Gaussian elements Mean of the covariance matrix of the K nearest neighbors Gaussian elements Calculation of the Frobenius norm between: The appearance deviation Based on the RGB values ​​of the target Gaussian unit mean RGB value of the nearest neighbor Gaussian element to K Euclidean distance between them: 。 5. The structural surface damage identification method based on Gaussian splashing according to claim 1, characterized in that, Step S3 specifically includes the following: S31. A density-based spatial clustering algorithm is used to perform spatial location clustering of Gaussian elements in the region to be identified, forming one or more damage clusters. S32. Extract the geometric features of each damage cluster. It is then input into a pre-trained damage classifier, which outputs the damage type; S33. Calculate the quantitative index of the damage based on the geometric morphological characteristics of the damage cluster.

6. The structural surface damage identification method based on Gaussian splashing according to claim 5, characterized in that, The geometric morphological features include the three principal eigenvalues ​​of the damaged clusters obtained through principal component analysis. And calculate the linearity accordingly. and flatness ; The types of damage include cracks, peeling, or corrosion.

7. The structural surface damage identification method based on Gaussian splashing according to claim 5, characterized in that, The quantitative indicators include, when the damage type is a crack, the quantitative indicators are the length, average width, and maximum width of the crack.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-7.

Images