Realistic three-dimensional reconstruction method for earth surface exposed geologic body based on neural radiation field

Abstract

The invention discloses an earth surface exposed geologic body reality sense three-dimensional reconstruction method based on a neural radiation field. The method comprises the steps that an earth surface geologic body image data set is acquired; performing feature matching on the image data set by using a feature matching strategy based on a nearest neighbor algorithm and a LightGlue neural network to generate feature point matching information and a feature description file; sparse three-dimensional point cloud data and camera pose information are obtained based on incremental SfM; and obtaining depth information of an image data set based on sparse three-dimensional point cloud data and camera pose information, and establishing a geologic body realistic three-dimensional reconstruction model by using a neural radiation field optimization method fusing the depth information and Gaussian position coding. According to the method, the feature matching precision and the camera pose calculation accuracy are improved, the texture detail expression ability of a geological structure can be improved, the reconstruction quality of a surface geologic body model is enhanced, and geological information support can be provided for the fields of geological research and engineering application.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning and digital railway construction, and in particular to a method for realistic three-dimensional reconstruction of surface exposed geological bodies based on neural radiation fields. Background Art

[0002] Large-scale geotechnical projects such as urban rail transit, highways, high-speed rail construction, and water conservancy projects are characterized by large scale, long construction period, complex operations, and high difficulty in engineering safety. In large-scale geotechnical projects, the condition of the surface geological body will affect the stability of the project site, and will have an impact on the safety, economic cost, and normal use of specific construction projects such as foundation projects and slopes. Therefore, the development of large-scale geotechnical projects has put forward an urgent need to truly understand the regional surface geological structure. In the early stage of large-scale projects, geologists need to accurately analyze the topography, landforms, and structural conditions of the geological area. By simulating the characteristics of topographic and geomorphic changes, reconstructing the surface outcrop geological model, and intuitively and realistically displaying the geological structure morphology and rock mass distribution, it is of great significance for construction projects in complex geological areas.

[0003] In recent years, three-dimensional geological models have provided important technical means and methods for understanding regional surface geological structures by displaying geological information in a three-dimensional manner and describing geological features in detail. Image-based three-dimensional reconstruction technology has been widely used in three-dimensional modeling of ground targets. This method uses computer vision technology to restore the three-dimensional model of the target object, such as motion recovery structure technology and multi-view stereo geometry technology. These methods have the advantages of convenient data acquisition and a wide range of applications, but the reconstruction process is complicated and the model reconstruction effect is not good. With the continuous development of computer vision, deep learning technology has provided new ideas for three-dimensional reconstruction due to its powerful learning and representation capabilities. Three-dimensional reconstruction technology based on deep learning has a high degree of automation, can learn complex data patterns and features, and can produce more realistic reconstruction results, especially for complex scenes or detailed parts.

[0004] Therefore, this application proposes a realistic three-dimensional reconstruction method for surface exposed geological bodies based on neural radiation fields, which solves the problems of poor matching of weak geological textures and repeated geological texture features and distortion of the reconstructed model based on computer vision and deep learning methods. Summary of the invention

[0005] In view of this, the present invention provides a method for realistic three-dimensional reconstruction of surface exposed geological bodies based on neural radiation fields, so as to solve the technical problems of weak geological textures and poor matching of repeated geological textures in the current reconstruction of three-dimensional surface models of surface exposed geological bodies.

[0006] In order to achieve the above technical objectives, the present invention adopts the following technical solutions:

[0007] The present invention provides a method for realistic three-dimensional reconstruction of surface exposed geological bodies based on neural radiation fields, comprising:

[0008] Acquire surface geological body image data set;

[0009] Perform feature matching on the image data set using a feature matching strategy based on a nearest neighbor algorithm and a LightGlue neural network to generate feature point matching information and a feature description file;

[0010] According to the image data set, feature point matching information and feature description file, sparse three-dimensional point cloud data and camera pose information are obtained based on incremental SfM;

[0011] The depth information of the image data set is obtained based on the sparse three-dimensional point cloud data and the camera posture information, and a realistic three-dimensional reconstruction model of the geological body is established by using a neural radiation field optimization method that integrates the depth information and Gaussian position coding.

[0012] Furthermore, the surface geological body image dataset includes: a synthetic dataset and a real dataset;

[0013] The synthetic data set is generated by controlling angles and lighting conditions using graphics software;

[0014] The real dataset consists of real-world images taken by actual cameras, including surfaces with texture mapping, front scene images, and panoramas.

[0015] Furthermore, feature matching is performed on the image data set using a feature matching strategy based on the nearest neighbor algorithm and the LightGlue neural network to generate feature point matching information and feature description files, including:

[0016] Extracting candidate feature points of the image data set using the SIFT algorithm, performing feature matching on the candidate feature points using the nearest neighbor algorithm to obtain an initial matching image pair;

[0017] Extracting a first feature point and a corresponding first descriptor from the initial matching image pair using a SIFT algorithm;

[0018] Performing feature matching based on the first feature point and the first descriptor through a LightGlue neural network to obtain a second feature point;

[0019] The corresponding relationship between each image feature point in the image data set is obtained according to the second feature point, and feature point matching information and a feature description file are generated.

[0020] Further, the candidate feature points of the image data set are extracted using the SIFT algorithm, and the candidate feature points are matched using the nearest neighbor algorithm to obtain an initial matching image pair, including:

[0021] Obtaining candidate feature points and corresponding candidate feature point descriptors for each image in the image data set;

[0022] Taking any two different images in the image data set as a set of image pairs, traversing and obtaining all image pairs in the image data set;

[0023] Perform feature matching based on the nearest neighbor algorithm on all candidate feature point descriptors in the two images in each group of image pairs, output the matching results of each group of image pairs, and obtain the initial matching image pairs.

[0024] Furthermore, the feature matching method based on the nearest neighbor algorithm includes:

[0025] Perform L2 normalization and similarity calculation on the candidate feature point descriptors corresponding to the two images in each image pair to obtain a similarity matrix;

[0026] The values ​​in the similarity matrix are sorted from large to small, and the candidate feature points corresponding to the first two values ​​are used as the nearest neighbor feature points and the second nearest neighbor feature points;

[0027] Calculate the ratio of the Euclidean distance of the nearest neighbor feature point to the Euclidean distance of the next nearest neighbor feature point;

[0028] Detect the symmetry of the nearest neighbor relationship, and retain the nearest neighbor feature point pairs based on the symmetry detection results;

[0029] A valid matching feature point pair among the nearest neighbor feature point pairs is determined based on a preset ratio threshold.

[0030] Further, according to the image data set, feature point matching information and feature description file, sparse three-dimensional point cloud data and camera pose information are obtained based on incremental SfM, including:

[0031] Convert the feature point matching information and feature description file into a format to obtain a matching strategy of a Colmap framework;

[0032] Based on the matching strategy and the image data set, the Colmap framework is used to perform SfM sparse reconstruction to obtain sparse three-dimensional point cloud and camera pose information.

[0033] Furthermore, based on the matching strategy and the image dataset, a Colmap framework is used to perform SfM sparse reconstruction, including:

[0034] According to the matching strategy, selecting the best matching image in the image data set;

[0035] Using the epipolar geometry constraint relationship, the initial position and posture of the camera are solved through the best matching image;

[0036] Based on the initial pose and the feature points in the matching strategy, using a triangulation method to obtain initial three-dimensional point cloud information corresponding to the feature points;

[0037] By using the BA optimization technology, the initial position and initial three-dimensional point cloud information of the camera are optimized with the goal of minimizing the reprojection error, so as to obtain accurate camera position and three-dimensional point cloud information.

[0038] Furthermore, based on the initial posture and the feature points in the matching strategy, initial three-dimensional point cloud information corresponding to the feature points is obtained by using a triangulation method, including:

[0039] Through the feature point matching relationship in the matching strategy, the essential matrix or basic matrix between the matching feature points is traversed and calculated;

[0040] The essential matrix or basic matrix is ​​decomposed by a singular value decomposition method to obtain the initial poses and initial three-dimensional point cloud coordinates of all cameras.

[0041] Furthermore, a realistic three-dimensional reconstruction model of a geological body is established by using a neural radiation field optimization method that integrates the depth information and Gaussian position coding, including:

[0042] Sampling the projected light according to the camera posture to obtain light sampling points, inputting the light sampling points into the MLP neural network to generate a radiation field describing the volume density and color of the sampling points in the surface geological body scene;

[0043] Performing volume rendering according to the radiation field to obtain an initial three-dimensional reconstruction model representing the morphology of the surface geological body;

[0044] The MLP network weights are iteratively adjusted based on a preset loss function, and the initial three-dimensional reconstruction model is iteratively optimized until a convergence condition is met, thereby generating a realistic three-dimensional reconstruction model of the geological body.

[0045] Furthermore, the projected light is sampled according to the camera posture to obtain light sampling points, and the light sampling points are input into the MLP neural network to generate a radiation field describing the volume density and color of the sampling points in the surface geological body scene, including:

[0046] Using the depth information, the neural radiation field learns the geometric structure of the surface geological body, and presetting the loss function so that the distribution of the light termination position conforms to the three-dimensional key point;

[0047] A Gaussian position encoding method based on random Fourier features is used to perform high-frequency mapping of the input of the neural radiation field.

[0048] Compared with the prior art, the method for realistic 3D reconstruction of surface exposed geological bodies based on neural radiation fields provided by the present invention has the following advantages:

[0049] 1. By combining the nearest neighbor algorithm and LightGlue neural network for feature matching, the matching accuracy and stability can be effectively improved, which is suitable for processing more complex scenes and reducing mismatches, especially in natural scenes such as geological bodies exposed on the surface, and the correspondence between images can be obtained more accurately.

[0050] 2. Three-dimensional reconstruction based on the incremental SfM method can obtain more accurate camera poses and sparse three-dimensional point clouds. By gradually accumulating information, it avoids the computational bottleneck caused by processing a large amount of data at one time, improves computational efficiency and reconstruction accuracy, and can better process large-scale image data sets.

[0051] 3. Based on the sparse 3D point cloud data and camera pose information, the neural radiation field optimization method that integrates depth information and Gaussian position coding is used to eliminate geological information artifacts and blur in the reconstructed model, and enhance the ability to express texture details of geological structures such as joints, inclined bedding and imbricate gravel. In particular, for the processing of areas with weak or repeated geological textures, a surface geological model that is closer to reality can be constructed, which significantly improves the realism and detail restoration of the model and is suitable for geological surface reconstruction.

[0052] In summary, the present invention solves the problem of poor matching of weak geological textures and repeated geological textures, improves the accuracy of feature matching and the accuracy of camera pose solution; the neural radiation field optimization method that integrates depth information and Gaussian position coding effectively solves the problems of geological information blur and artifacts that appear in the neural radiation field reconstruction of surface geological bodies, while improving the ability to express texture details of geological structures such as joints, inclined bedding, and imbricate gravel, thereby enhancing the reconstruction quality of surface geological body models. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 A schematic diagram of the process of the method for realistic three-dimensional reconstruction of surface exposed geological bodies based on neural radiation fields provided by the present invention;

[0054] Figure 2 A schematic diagram of the method flow of the incremental SfM combined with LightGlue to optimize feature matching provided by the present invention;

[0055] FIG3( a ) is a camera position of an image of a surface geological body A for SfM sparse reconstruction using a matching strategy provided by the present invention through a Colmap framework;

[0056] FIG3( b ) is a camera position of the surface geological body B image for SfM sparse reconstruction using the matching strategy provided by the present invention through the Colmap framework;

[0057] Figure 4 A schematic diagram of the process of the neural radiation field optimization method for fusing depth information and Gaussian position coding provided by the present invention;

[0058] FIG5( a ) is a schematic diagram of a reconstruction model of a surface geological body A provided by the present invention;

[0059] FIG5( b ) is a local detail view of the sedimentary rock joint surface in the A reconstruction model provided by the present invention;

[0060] FIG5( c ) is a local detail view of the weathered soil layer in the A reconstruction model provided by the present invention;

[0061] FIG6( a ) is a schematic diagram of a reconstruction model of a surface geological body B provided by the present invention;

[0062] FIG6( b ) is a detailed view of the imbricate arrangement of gravels and oblique bedding in the B reconstruction model provided by the present invention;

[0063] Figure 7 A flowchart for implementing the method for realistic three-dimensional reconstruction of surface exposed geological bodies based on neural radiation fields provided by the present invention. DETAILED DESCRIPTION

[0064] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0065] The present invention provides a method for realistic three-dimensional reconstruction of surface exposed geological bodies based on neural radiation fields, which is described in detail below.

[0066] See also Figure 1 This embodiment provides a method for realistic 3D reconstruction of a surface exposed geological body based on a neural radiation field, comprising:

[0067] Step S101: Acquire a surface geological body image dataset;

[0068] Step S102: performing feature matching on the image data set using a feature matching strategy based on a nearest neighbor algorithm and a LightGlue neural network to generate feature point matching information and a feature description file;

[0069] Step S103: obtaining sparse three-dimensional point cloud data and camera pose information based on incremental SfM according to the image data set, feature point matching information and feature description file;

[0070] Step S104: obtaining depth information of the image data set based on the sparse three-dimensional point cloud data and the camera pose information, and establishing a realistic three-dimensional reconstruction model of the geological body by using a neural radiation field optimization method that integrates the depth information and Gaussian position coding.

[0071] The method of this embodiment, based on machine vision technology (feature matching and SfM) and deep learning network (LightGlue neural network and neural radiation field), can efficiently and accurately perform three-dimensional reconstruction of surface geological bodies. Through fine feature matching, high-quality three-dimensional point cloud reconstruction and neural radiation field optimization, the final generated three-dimensional model eliminates the blurring of geological information and the generation of artifacts, reduces the distortion of model reconstruction, and enhances the ability to express the texture details of geological structures such as joints, inclined bedding, and imbricate gravel, achieving a high level in terms of realism, detail restoration and computational efficiency, and is suitable for geological body analysis, geological exploration, natural environment modeling and other fields.

[0072] As a preferred embodiment, the surface geological body image data set includes: a synthetic data set and a real data set;

[0073] The synthetic data set is generated by controlling angles and lighting conditions using graphics software;

[0074] The real dataset consists of real-world images taken by actual cameras, including surfaces with texture mapping, front scene images, and panoramas.

[0075] Specifically, synthetic datasets are generated by computer graphics software (such as Blender), which can control different angles and lighting conditions to produce a large number of images.

[0076] The real dataset consists of real-world images taken by actual cameras. These images contain texture-mapped surfaces and 180° front scenes and all-round scenes with low geometric complexity. Real images are affected by environmental factors such as on-site lighting and shadows, which are more challenging but also closer to practical applications.

[0077] As a preferred embodiment, feature matching and extraction of the image data set are performed using a feature matching strategy based on a nearest neighbor algorithm and a LightGlue neural network to generate feature point matching information and a feature description file, including:

[0078] Extracting candidate feature points of the image data set using the SIFT algorithm, performing feature matching on the candidate feature points using the nearest neighbor algorithm to obtain an initial matching image pair;

[0079] Extracting a first feature point and a corresponding first descriptor from the initial matching image pair using a SIFT algorithm;

[0080] Performing feature matching based on the first feature point and the first descriptor through a LightGlue neural network to obtain a second feature point;

[0081] The corresponding relationship between each image feature point in the image data set is obtained according to the second feature point, and feature point matching information and a feature description file are generated.

[0082] In some embodiments, in order to verify the feature point matching accuracy and data set construction accuracy of the algorithm on geological features and geological body textures, two surface exposed geological body image sets containing geological structural features and different geological textures can be self-made.

[0083] First, pre-matching is performed based on the SIFT feature point extraction algorithm and the nearest neighbor algorithm to obtain candidate matching image pairs; the goal of this step is to find approximate matching points based on the preliminary matching of the image pairs.

[0084] Then, the SIFT algorithm is used again to extract feature points and corresponding descriptors from the candidate matching image pairs; the purpose here is to extract more refined feature points from the matched image pairs, in order to further improve the accuracy and quality of feature point matching.

[0085] Next, the LightGlue neural network is used to match the feature points in the image pair to obtain the correspondence between the feature points in the first image and the feature points in the second image, and the matching results are usually stored in the form of indexes; through the multi-layer attention mechanism of the LightGlue neural network, more representative feature points are obtained and feature point matching is performed, which solves the problem of poor matching of weak geological textures and repeated geological texture features.

[0086] Finally, a database table is created and the image information and feature matching results are written into the specified database to prepare for subsequent neural network matching.

[0087] As a preferred embodiment, the candidate feature points of the image data set are extracted using the SIFT algorithm, and the candidate feature points are matched using the nearest neighbor algorithm to obtain an initial matching image pair, including:

[0088] Obtaining candidate feature points and corresponding candidate feature point descriptors for each image in the image data set;

[0089] Taking any two different images in the image data set as a set of image pairs, traversing and obtaining all image pairs in the image data set;

[0090] Perform feature matching based on the nearest neighbor algorithm on all candidate feature point descriptors in the two images in each group of image pairs, output the matching results of each group of image pairs, and obtain the initial matching image pairs.

[0091] As a preferred embodiment, the feature matching method based on the nearest neighbor algorithm includes:

[0092] Perform L2 normalization and similarity calculation on the candidate feature point descriptors corresponding to the two images in each image pair to obtain a similarity matrix;

[0093] The values ​​in the similarity matrix are sorted from large to small, and the candidate feature points corresponding to the first two values ​​are used as the nearest neighbor feature points and the second nearest neighbor feature points;

[0094] Calculate the ratio of the Euclidean distance of the nearest neighbor feature point to the Euclidean distance of the next nearest neighbor feature point;

[0095] Detect the symmetry of the nearest neighbor relationship, and retain the nearest neighbor feature point pairs based on the symmetry detection results;

[0096] A valid matching feature point pair among the nearest neighbor feature point pairs is determined based on a preset ratio threshold.

[0097] In the actual implementation process of some embodiments, pre-matching based on the SIFT feature point extraction algorithm and the nearest neighbor algorithm is generally performed according to the following steps:

[0098] The first step is to read the images in the image list, select two images in order to obtain two sets of feature point information; perform L2 normalization on the descriptors of the two sets of feature points and perform similarity calculation to obtain a similarity matrix;

[0099] The second step is to search for the two largest values ​​in the similarity matrix to obtain the nearest neighbor feature point and the second nearest neighbor feature point;

[0100] The third step is to calculate the ratio of the Euclidean distance of the nearest neighbor feature point to the Euclidean distance of the next nearest neighbor feature point;

[0101] The fourth step is to check the symmetry of the nearest neighbor relationship to ensure that the nearest neighbor relationship holds.

[0102] The fifth step is to use the ratio obtained in the third step to perform a test, that is, for a pair of feature points A and B, if the ratio of the Euclidean distance from A to B to the Euclidean distance from B to A is less than the preset ratio threshold of 0.8, then this is considered to be a valid match;

[0103] Repeat the above steps until all feature points on the image and all image pairs are matched, and only the matching results of the image pairs are output.

[0104] As a preferred embodiment, according to the image data set, feature point matching information and feature description file, sparse three-dimensional point cloud data and camera pose information are obtained based on incremental SfM, including:

[0105] Convert the feature point matching information and feature description file into a format to obtain a matching strategy of a Colmap framework;

[0106] Based on the matching strategy and the image data set, the Colmap framework is used to perform SfM sparse reconstruction to obtain sparse three-dimensional point cloud and camera pose information.

[0107] The accurate camera pose and 3D point cloud are calculated through incremental SfM, the obtained 3D point cloud is retained, and a dataset consisting of surface geological body images and their corresponding pose data is constructed for subsequent use in the neural radiation field optimization method that integrates depth information and Gaussian position encoding.

[0108] As a preferred embodiment, based on the matching strategy and the image dataset, the SfM sparse reconstruction is performed using the Colmap framework, including:

[0109] According to the matching strategy, selecting the best matching image in the image data set;

[0110] Using the epipolar geometry constraint relationship, the initial position and posture of the camera are solved through the best matching image;

[0111] Based on the initial pose and the feature points in the matching strategy, using a triangulation method to obtain initial three-dimensional point cloud information corresponding to the feature points;

[0112] By using the BA optimization technology, the initial position and initial three-dimensional point cloud information of the camera are optimized with the goal of minimizing the reprojection error, so as to obtain accurate camera position and three-dimensional point cloud information.

[0113] like Figure 2 As shown, Figure 2 The method flow of incremental SfM combined with LightGlue to optimize feature matching is demonstrated.

[0114] As a specific embodiment, the image data of the surface geological body A and the surface geological body B are used as the input object of the data set construction method, and the image pair information and matching result information are obtained by using the feature matching strategy of the fusion nearest neighbor algorithm and the LightGlue neural network, which are written into the Colmap database. The camera pose is solved by using the matching result information using the epipolar geometry constraint relationship, and the corresponding three-dimensional point cloud information is triangulated on the image according to the known camera pose and the matched feature points. The camera pose and the three-dimensional point cloud are iteratively optimized using the BA method. After completing the SfM sparse reconstruction, the three-dimensional point cloud obtained by the sparse reconstruction is retained, and the internal and external parameter matrix data corresponding to each surface geological body image pose are recorded and sorted, and a reconstruction data set consisting of the surface geological body image and its corresponding pose data file is constructed for the subsequent surface geological body neural radiation field training and reconstruction, wherein the camera pose visualization results calculated by the surface geological body A and the surface geological body B are shown in Figure 3 (a) and Figure 3 (b).

[0115] As a preferred embodiment, based on the initial posture and feature points in the matching strategy, initial three-dimensional point cloud information corresponding to the feature points is obtained by using a triangulation method, including:

[0116] Through the feature point matching relationship in the matching strategy, the essential matrix or basic matrix between the matching feature points is traversed and calculated;

[0117] The essential matrix or basic matrix is ​​decomposed by a singular value decomposition method to obtain the initial poses and initial three-dimensional point cloud coordinates of all cameras.

[0118] As a preferred embodiment, a realistic three-dimensional reconstruction model of a geological body is established by using a neural radiation field optimization method that integrates the depth information and Gaussian position coding, including:

[0119] Sampling the projected light according to the camera posture to obtain light sampling points, inputting the light sampling points into the MLP neural network to generate a radiation field describing the volume density and color of the sampling points in the surface geological body scene;

[0120] Performing volume rendering according to the radiation field to obtain an initial three-dimensional reconstruction model representing the morphology of the surface geological body;

[0121] The MLP network weights are iteratively adjusted based on a preset loss function, and the initial three-dimensional reconstruction model is iteratively optimized until a convergence condition is met, thereby generating a realistic three-dimensional reconstruction model of the geological body.

[0122] As a specific embodiment, the optimization method of NeRF (radiance neural field) fusing depth information and Gaussian position coding has the following specific steps:

[0123] First, the projected light is sampled and the sampling points are input into the MLP neural network to generate a radiation field describing the volume density and color of all sampling points in the surface geological scene;

[0124] Secondly, volume rendering is performed to obtain a model that represents the morphology of the surface geological body;

[0125] By continuously using the loss function to iteratively adjust the MLP network weights, the model will gradually approach the true shape of the geological body until the convergence condition is met, completing the training of the geological body neural radiation field, thereby obtaining a high-fidelity surface geological body model;

[0126] Introducing depth information to guide the neural radiation field to learn the geometric structure of the surface geological body, and designing a loss function to make the distribution of the light termination position conform to the three-dimensional key points;

[0127] A Gaussian position encoding method based on random Fourier features is used to perform high-frequency mapping of the input of the neural radiation field.

[0128] like Figure 4 As shown, Figure 4 A flowchart of the neural radiation field optimization method that integrates depth information and Gaussian position coding is shown. By utilizing the similarity between the ray termination distribution and the depth information distribution, depth information is added to guide the neural radiation field learning model's geometric information, and the loss function is optimized to make the distribution of ray termination positions consistent with the distribution of three-dimensional model control points, thereby improving the effectiveness of the sampling points, eliminating the geological information blur and artifacts that appear in the neural radiation field reconstruction of the surface geological body, and reducing the distortion of the model reconstruction.

[0129] As a preferred embodiment, the projected light is sampled according to the camera posture to obtain light sampling points, and the light sampling points are input into the MLP neural network to generate a radiation field describing the volume density and color of the sampling points in the surface geological body scene, including:

[0130] Using the depth information, the neural radiation field learns the geometric structure of the surface geological body, and presetting the loss function so that the distribution of the light termination position conforms to the three-dimensional key point;

[0131] A Gaussian position encoding method based on random Fourier features is used to perform high-frequency mapping of the input of the neural radiation field.

[0132] By replacing the position coding in the network with Gaussian position coding based on random Fourier features, the input of the neural radiation field is mapped at high frequency to improve the ability to express texture details of geological structures such as joints, inclined bedding, and imbricate gravel, thereby further improving the reconstruction quality of the surface geological model.

[0133] As a specific embodiment, a self-made data set of surface geological body A and surface geological body B is used for training. During the training process, an Adam optimizer with parameters of 0.9 and 0.999 is used. The number of rays randomly sampled in each iteration is 4096, the initial learning rate is set to 0.0005, and the number of training iterations is set to 300,000 times.

[0134] PSNR, SSIM and LPIPS are used to evaluate the neural radiation field or similar 3D surface reconstruction.

[0135] Based on the prepared surface geological body data A and surface geological body data B, the improved surface geological body neural radiation field is trained using the constructed image data of the additional camera pose;

[0136] During the training process, the loss function is optimized and the parameters are continuously adjusted to minimize the difference between the reconstruction result and the real image, so as to reconstruct a surface model of the surface geological body that is closer to the actual surface. The surface reconstruction models of the surface geological body A and the surface geological body B are shown in the figure below: Figure 5(a) and 6(a) As shown, Figure 5(b) and Figure 5(c) respectively show the local details of the sedimentary rock joint surface and the weathered soil layer in Figure 5(a), and Figure 6(b) shows the details of the imbricate gravel and inclined bedding in Figure 6(a). It can be seen that the present invention can accurately reconstruct the information contained in each image, eliminate the geological information artifacts and blurring problems in the reconstructed model, enhance the ability to express the texture details of geological structures such as joint surfaces, inclined bedding and imbricate gravel, and construct a surface geological body model that is closer to reality, thereby truly displaying the characteristic morphology of the exposed geological body on the surface.

[0137] As a specific example, Figure 7 As shown, Figure 7 The flowchart of the implementation of the method for realistic 3D reconstruction of surface exposed geological bodies based on neural radiation field proposed in this application is shown:

[0138] Firstly, based on the prepared surface geological body image dataset, a feature matching strategy integrating the nearest neighbor algorithm and the LightGlue neural network is designed. The LightGlue neural network is introduced to match the feature points in the image pairs. A multi-layer attention mechanism is used to obtain more representative feature points, and feature point matching is performed to solve the problem of poor feature matching of weak geological textures and repeated geological textures. This strategy is applied to the traditional incremental SfM to solve the accurate camera pose and 3D point cloud, retain the obtained 3D point cloud, and construct a dataset consisting of surface geological body images and their corresponding pose data. Finally, a neural network integrating depth information and Gaussian position coding is proposed. Through the radiation field optimization method, the similarity between the ray termination distribution and the depth information distribution is utilized, and the depth information is added to guide the geometric information of the neural radiation field learning model. The loss function is optimized so that the distribution of the ray termination position conforms to the distribution of the three-dimensional model control points, thereby improving the effectiveness of the sampling points, eliminating the geological information ambiguity and artifacts that appear in the neural radiation field reconstruction of the surface geological body, and reducing the distortion of the model reconstruction; at the same time, Gaussian position coding based on random Fourier features is used instead of the position coding in the network to perform high-frequency mapping on the input of the neural radiation field, thereby improving the ability to express texture details of geological structures such as joints, inclined bedding, and imbricate gravel, and further improving the reconstruction quality of the surface geological body model.

[0139] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for realistic 3D reconstruction of surface exposed geological bodies based on neural radiation fields, characterized in that: include: Acquire surface geological body image data set; Perform feature matching on the image data set using a feature matching strategy based on a nearest neighbor algorithm and a LightGlue neural network to generate feature point matching information and a feature description file; According to the image data set, feature point matching information and feature description file, sparse three-dimensional point cloud data and camera pose information are obtained based on incremental SfM; The depth information of the image data set is obtained based on the sparse three-dimensional point cloud data and the camera posture information, and a realistic three-dimensional reconstruction model of the geological body is established by using a neural radiation field optimization method that integrates the depth information and Gaussian position coding.

2. The method for realistic 3D reconstruction of surface exposed geological bodies based on neural radiation fields according to claim 1 is characterized in that: The feature matching strategy based on the nearest neighbor algorithm and the LightGlue neural network is used to perform feature matching extraction on the image dataset to generate feature point matching information and feature description files, including: Extracting candidate feature points of the image data set using the SIFT algorithm, performing feature matching on the candidate feature points using the nearest neighbor algorithm to obtain an initial matching image pair; Extracting a first feature point and a corresponding first descriptor from the initial matching image pair using a SIFT algorithm; Performing feature matching based on the first feature point and the first descriptor through a LightGlue neural network to obtain a second feature point; The corresponding relationship between each image feature point in the image data set is obtained according to the second feature point, and feature point matching information and a feature description file are generated.

3. The method for realistic 3D reconstruction of surface exposed geological bodies based on neural radiation fields according to claim 2 is characterized in that: The candidate feature points of the image data set are extracted by using the SIFT algorithm, and the candidate feature points are matched by using the nearest neighbor algorithm to obtain an initial matching image pair, including: Obtaining candidate feature points and corresponding candidate feature point descriptors for each image in the image data set; Taking any two different images in the image data set as a set of image pairs, traversing and obtaining all image pairs in the image data set; Perform feature matching based on the nearest neighbor algorithm on all candidate feature point descriptors in the two images in each group of image pairs, output the matching results of each group of image pairs, and obtain the initial matching image pairs.

4. The method for realistic 3D reconstruction of surface exposed geological bodies based on neural radiation fields according to claim 3 is characterized in that: The feature matching method based on the nearest neighbor algorithm includes: Perform L2 normalization and similarity calculation on the candidate feature point descriptors corresponding to the two images in each image pair to obtain a similarity matrix; The values ​​in the similarity matrix are sorted from large to small, and the candidate feature points corresponding to the first two values ​​are used as the nearest neighbor feature points and the second nearest neighbor feature points; Calculate the ratio of the Euclidean distance of the nearest neighbor feature point to the Euclidean distance of the next nearest neighbor feature point; Detect the symmetry of the nearest neighbor relationship, and retain the nearest neighbor feature point pairs based on the symmetry detection results; A valid matching feature point pair among the nearest neighbor feature point pairs is determined based on a preset ratio threshold.

5. The method for realistic 3D reconstruction of surface exposed geological bodies based on neural radiation field according to claim 1, characterized in that: According to the image data set, feature point matching information and feature description file, sparse three-dimensional point cloud data and camera pose information are obtained based on incremental SfM, including: Convert the feature point matching information and feature description file into a format to obtain a matching strategy of a Colmap framework; Based on the matching strategy and the image data set, the Colmap framework is used to perform SfM sparse reconstruction to obtain sparse three-dimensional point cloud and camera pose information.

6. The method for realistic 3D reconstruction of surface exposed geological bodies based on neural radiation fields according to claim 5, characterized in that: Based on the matching strategy and the image dataset, the SfM sparse reconstruction is performed using the Colmap framework, including: According to the matching strategy, selecting the best matching image in the image data set; Using the epipolar geometry constraint relationship, the initial position and posture of the camera are solved through the best matching image; Based on the initial pose and the feature points in the matching strategy, using a triangulation method to obtain initial three-dimensional point cloud information corresponding to the feature points; By using the BA optimization technology, the initial position and initial three-dimensional point cloud information of the camera are optimized with the goal of minimizing the reprojection error, so as to obtain accurate camera position and three-dimensional point cloud information.

7. The method for realistic 3D reconstruction of surface exposed geological bodies based on neural radiation fields according to claim 6 is characterized in that: Based on the initial pose and the feature points in the matching strategy, initial three-dimensional point cloud information corresponding to the feature points is obtained by using a triangulation method, including: Through the feature point matching relationship in the matching strategy, the essential matrix or basic matrix between the matching feature points is traversed and calculated; The essential matrix or basic matrix is ​​decomposed by a singular value decomposition method to obtain the initial poses and initial three-dimensional point cloud coordinates of all cameras.

8. The method for realistic 3D reconstruction of surface exposed geological bodies based on neural radiation fields according to claim 6, characterized in that: A realistic three-dimensional reconstruction model of a geological body is established by using a neural radiation field optimization method that integrates the depth information and Gaussian position coding, including: Sampling the projected light according to the camera posture to obtain light sampling points, inputting the light sampling points into the MLP neural network to generate a radiation field describing the volume density and color of the sampling points in the surface geological body scene; Performing volume rendering according to the radiation field to obtain an initial three-dimensional reconstruction model representing the morphology of the surface geological body; The MLP network weights are iteratively adjusted based on a preset loss function, and the initial three-dimensional reconstruction model is iteratively optimized until a convergence condition is met, thereby generating a realistic three-dimensional reconstruction model of the geological body.

9. The method for realistic 3D reconstruction of surface exposed geological bodies based on neural radiation fields according to claim 8, characterized in that: The projected light is sampled according to the camera posture to obtain light sampling points, and the light sampling points are input into the MLP neural network to generate a radiation field describing the volume density and color of the sampling points in the surface geological body scene, including: Using the depth information, the neural radiation field learns the geometric structure of the surface geological body, and presetting the loss function so that the distribution of the light termination position conforms to the three-dimensional key point; A Gaussian position encoding method based on random Fourier features is used to perform high-frequency mapping of the input of the neural radiation field.

10. The method for realistic 3D reconstruction of surface exposed geological bodies based on neural radiation fields according to claim 1, characterized in that: The surface geological body image data set includes: a synthetic data set and a real data set; The synthetic data set is generated by controlling angles and lighting conditions using graphics software; The real dataset consists of real-world images taken by actual cameras, including surfaces with texture mapping, front scene images, and panoramas.

Images