A method for fusing colmap and deep learning for three-dimensional reconstruction of cuttings

Abstract

The application discloses a kind of fusion Colmap and the method for 3D reconstruction of cutting, it is related to computer vision and geological image processing technical field.Utilize Colmap to the pre-processing of multi-view cutting image, obtain camera parameter, dense point cloud and depth map;Image pixel is converted into world coordinate system three-dimensional point and completes ray sampling, and three-dimensional sampling point is encoded using NeuS method with multi-frequency sine cosine position;With high-dimensional feature vector input multilayer perceptron, obtain reconstruction network by depth, color and signed distance function multi-loss joint optimization;Finally, isosurface extraction algorithm is used to extract isosurface and generate cutting 3D model.The application is high in precision, detail is good, and can realize efficient reconstruction only by relying on two-dimensional image, provides reliable technical support for geological interpretation, reservoir evaluation and drilling engineering.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and geological image processing technology, specifically to a method for three-dimensional reconstruction of rock debris that integrates Colmap and deep learning. Background Technology

[0002] With the continuous development of geological exploration and drilling technologies, the requirements for the accuracy of stratigraphic information acquisition are increasing. Rock cuttings are important materials that directly reflect stratigraphic characteristics, and their analysis is crucial for stratigraphic division and oil and gas reservoir evaluation. Traditional rock cuttings analysis methods suffer from insufficient information extraction, poor timeliness, and difficulty in quantifying parameters such as porosity and permeability, thus failing to meet the needs of high-precision geological analysis.

[0003] With the widespread application of digital and intelligent technologies in the petroleum industry, 3D reconstruction of rock cuttings has become an important part of the construction of digital oilfields. However, existing reconstruction technologies have obvious shortcomings: traditional multi-view stereo vision reconstruction is easily affected by the surface features of rock cuttings, resulting in sparse point clouds and missing surface details; pure deep learning implicit reconstruction lacks accurate camera pose and depth priors, resulting in poor global consistency and low detail fidelity, making it difficult to support the accurate description of complex geological structures.

[0004] Meanwhile, drilling waste contains rich geological information, and traditional treatment methods pollute the environment and waste resources. Therefore, there is an urgent need for a high-precision and high-efficiency three-dimensional reconstruction method for drilling waste to achieve accurate acquisition of stratigraphic information, environmental protection and emission reduction, and efficient utilization of resources, providing reliable technical support for oil and gas exploration and development. Summary of the Invention

[0005] To address the aforementioned shortcomings in existing technologies, this invention provides a 3D reconstruction method for rock debris that integrates Colmap and deep learning. This method solves the problems of missing details in traditional methods, poor global consistency in pure deep learning methods, insufficient camera pose and depth priors, and difficulty in accurately reconstructing the surface texture and fine structure of rock debris.

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

[0007] A method for 3D reconstruction of rock cuttings integrating Colmap and deep learning includes the following steps: S1. Multi-view rock debris images are processed using Colmap to obtain training data for the 3D reconstruction network; S2. Combining camera intrinsic and extrinsic parameters with Colmap data, the pixels of each rock debris image are converted into three-dimensional points in the world coordinate system, and a three-dimensional sampling point set is obtained through ray sampling. S3. The NeuS method is used to encode the position of the three-dimensional sampling point set to obtain a high-dimensional feature vector; S4. Based on the multilayer perceptron, and using high-dimensional feature vectors combined with Colmap data, a three-dimensional reconstruction network model is obtained through joint optimization of multiple losses. S5. Using a three-dimensional reconstruction network model and an isosurface extraction algorithm, a three-dimensional model of rock debris is obtained.

[0008] Further, step S1 includes the following sub-steps: S11. Prepare a set of rock debris images in RGB encoding format from different perspectives, covering all angles of the rock debris scene; S12. Use Colmap to extract feature points from each rock cutting image and match the feature points of different rock cutting images to find the correspondence between the rock cutting images. S13. Based on the feature point matching results of S12, the camera pose is obtained by triangulation and a sparse point cloud is generated. S14. Based on the sparse point cloud, the MVS algorithm is used to generate a dense point cloud; S15. Use Colmap to generate corresponding depth maps for each rock cuttings image.

[0009] Furthermore, step S2 includes the following sub-steps: S21. Normalize and calibrate the original pixel two-dimensional coordinates of each rock cuttings image using the following formula to obtain a two-dimensional normalized physical coordinate sequence. : , , in, For the first The original x-coordinate of each pixel. For the first The original ordinate of each pixel. This represents the lateral offset of the camera's intrinsic parameters. This represents the vertical offset of the camera's intrinsic parameters. This refers to the camera's intrinsic lateral focal length. The vertical focal length is the camera's intrinsic parameter. For the first Normalized physical x-coordinate of each pixel For the first Normalized physical ordinate of each pixel It is a two-dimensional space. The number of pixels; S22. Merge the depth values ​​from the Colmap output depth map into a two-dimensional normalized physical coordinate sequence to obtain a three-dimensional transition sequence. : in, For the first The depth value of each pixel. It is a three-dimensional space; S23. The three-dimensional transition sequence is mapped to three-dimensional spatial coordinates in the camera coordinate system using the following formula. : , , , in, The x-coordinate in the camera coordinate system The vertical coordinate in the camera coordinate system These are the depth coordinates in the camera coordinate system. S24. Convert the 3D spatial coordinates in the camera coordinate system to 3D points in the world coordinate system using the following formula: , in, For the world coordinate system The vector form of a pixel's three-dimensional points. The first in camera coordinate system Pixel three-dimensional spatial coordinate vector form, Let the rotation matrix be the camera's extrinsic parameters. for transpose, The translation vector of the camera's extrinsic parameters; S25. Using the origin of the world coordinate system as the origin of all rays, and taking the direction pointing to each three-dimensional point in the world coordinate system as different ray directions, construct each ray, and sample on each ray to obtain a three-dimensional sampling point set.

[0010] Furthermore, step S3 includes the following sub-steps: S31. For each three-dimensional sampling point in the three-dimensional sampling point set, the three coordinate components are independently encoded using multi-frequency sine and cosine functions. S32. Concatenate the encoded result vectors of the three coordinate components to obtain the high-dimensional feature vectors of each three-dimensional sampling point.

[0011] Furthermore, the expression for the multi-frequency sine and cosine function encoding method for each coordinate component in S31 is as follows: , in, It is a sine function. It is a cosine function. Pi For coordinate components, Coordinate components The encoded result vector, For frequency levels.

[0012] Furthermore, in step S32, the encoded result vectors of the three coordinate components of each three-dimensional sampling point are concatenated in the following manner: , in, The three-dimensional sampling points obtained by stitching High-dimensional feature vectors, Three-dimensional sampling points The encoded vector of the horizontal coordinate component. Three-dimensional sampling points The encoded vector of the ordinate component. Three-dimensional sampling points The encoded vector of depth coordinate components.

[0013] Furthermore, step S4 includes the following sub-steps: S41. Based on a multilayer perceptron, a 3D reconstruction network model is constructed, which includes: Deep subnetworks are used to calculate depth values ​​based on high-dimensional feature vectors; The symbolic distance field subnetwork is used to calculate symbolic distance values ​​based on high-dimensional feature vectors; The color sub-network is used to calculate color values ​​based on high-dimensional feature vectors; S42. Construct loss functions corresponding to depth, symbol distance, and color values ​​respectively. Based on the high-dimensional feature vector and combined with Colmap data, optimize the 3D reconstruction network model through backpropagation joint training.

[0014] Furthermore, step S5 includes the following sub-steps: S51. Construct a cube-shaped bounding box with a side length of 1.01, so that it completely encloses the rock debris target object inside the unit sphere; S52. Generate a dense and uniform 3D mesh inside the bounding box, input the mesh vertices into the 3D reconstruction network model, and calculate the symbolic distance value. S53. Using the isosurface extraction algorithm, a three-dimensional model of rock debris is obtained based on the symbolic distance value.

[0015] Furthermore, the method of S53 is as follows: traverse all cube elements, and when an isosurface with a symbolic distance value of 0 passes through a cube element, calculate the intersection point between the isosurface and the cube edge by linear interpolation, and generate triangular facets based on the intersection points to form a continuous surface as a three-dimensional model of rock debris.

[0016] The beneficial effects of this invention are as follows: 1) This invention achieves feature extraction, matching, camera pose calculation and depth map generation of multi-view rock debris images through Colmap preprocessing, providing accurate and stable camera intrinsic and extrinsic parameters and depth priors for subsequent 3D reconstruction, and solving the problems of lack of geometric constraints and poor global consistency in pure deep learning reconstruction.

[0017] 2) This invention achieves accurate mapping from two-dimensional images to three-dimensional spatial points through pixel normalization calibration, depth fusion, and step-by-step transformation between camera coordinate system and world coordinate system, ensuring the physical consistency and spatial accuracy of coordinate transformation and improving the reconstruction accuracy of three-dimensional points of rock debris.

[0018] 3) This invention draws on the NeuS method and uses multi-frequency sine and cosine position encoding, which can map low-dimensional coordinates to a high-dimensional feature space, enabling the network to simultaneously capture the low-frequency overall outline of rock fragments and the high-frequency fine geometric structure, effectively restoring the minute textures, edges and uneven details of the rock fragment surface, and significantly improving surface fidelity.

[0019] 4) This invention employs a multilayer perceptron and sets up three branches: a depth subnetwork, a symbolic distance field subnetwork, and a color subnetwork. It can simultaneously learn the geometric structure, spatial distance, and surface color information of rock fragments. Through multi-loss joint constraints, it achieves stable optimization of the network, resulting in more complete and robust reconstruction results.

[0020] 5) This invention achieves a balance between geometric smoothness, surface detail and color fidelity in the 3D reconstruction network by jointly optimizing depth loss, color loss and signed distance function loss, thus avoiding overfitting or loss of detail caused by a single loss.

[0021] 6) This invention uses the Marching Cubes isosurface extraction algorithm, which can directly generate a continuous and closed triangular mesh model from the trained implicit network. It can obtain a three-dimensional rock cutting model that can be directly used for geological analysis without post-processing. The process is simple and highly practical.

[0022] 7) This invention can complete high-precision three-dimensional reconstruction using only multi-view rock debris images, without the need for additional high-precision scanning equipment, reducing the cost of geological sample modeling and enabling rapid acquisition of three-dimensional morphology, pores, cracks and other structural information of rock debris.

[0023] 8) The three-dimensional rock cutting model obtained by this invention can intuitively reflect the internal structure and composition of the formation rocks, and can provide a reliable basis for geological interpretation, reservoir evaluation, porosity and permeability analysis, and drilling scheme optimization, and has strong engineering application value. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating a method for 3D reconstruction of rock cuttings that integrates Colmap and deep learning, according to an embodiment of the present invention. Figure 2 This is a structural diagram of the three-dimensional reconstruction network model according to an embodiment of the present invention. Detailed Implementation

[0025] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0026] like Figure 1 As shown, a method for 3D reconstruction of rock cuttings integrating Colmap and deep learning includes the following steps: S1. Process multi-view debris images using Colmap to obtain training data for the 3D reconstruction network. Colmap is an open-source tool used to reconstruct 3D scenes from multiple 2D images. This embodiment sets the following initialization parameters: batch size = 4, epochs = 200, learning rate = 0.0005, seed = 42, hidden layer dimension = 256, number of samples = 1024. S1 includes the following sub-steps: S11. Prepare a set of rock debris images in RGB encoding format from different perspectives, covering all angles of the rock debris scene; S12. Use Colmap to extract feature points from each rock cutting image, such as SIFT feature points, and match the feature points of different rock cutting images to find the correspondence between each rock cutting image. S13. Based on the feature point matching results of S12, Colmap obtains the camera pose through triangulation and generates a sparse point cloud. S14. Based on the sparse point cloud, Colmap uses the MVS (Multi-View Stereo) algorithm to generate a dense point cloud. S15. Use Colmap to generate corresponding depth maps for each rock cutting image. The depth map records the distance from each pixel to the camera.

[0027] This invention achieves feature extraction, matching, camera pose calculation, and depth map generation of multi-view rock debris images through Colmap preprocessing, providing accurate and stable camera intrinsic and extrinsic parameters and depth priors for subsequent 3D reconstruction, and solving the problems of lack of geometric constraints and poor global consistency in pure deep learning reconstruction.

[0028] In this embodiment, a Python script is used to organize the camera intrinsic parameters (including focal length and principal point) and camera extrinsic parameters (including pose, i.e., rotation matrix and translation vector) into a .npz file for use in subsequent steps.

[0029] S2. Combining camera intrinsic and extrinsic parameters with Colmap data, convert each rock cutting image pixel into a three-dimensional point in the world coordinate system, and obtain a three-dimensional sampling point set through ray sampling, including the following sub-steps: S21. Normalize and calibrate the original pixel two-dimensional coordinates of each rock cuttings image using the following formula to obtain a two-dimensional normalized physical coordinate sequence. : , , in, For the first The original x-coordinate of each pixel. For the first The original ordinate of each pixel. This represents the lateral offset of the camera's intrinsic parameters. This represents the vertical offset of the camera's intrinsic parameters. This refers to the camera's intrinsic lateral focal length. The vertical focal length is the camera's intrinsic parameter. For the first The normalized physical x-coordinate of each pixel. For the first Normalized physical ordinate of each pixel It is a two-dimensional space. This represents the number of pixels.

[0030] , Image spatial height, This represents the image space width.

[0031] This step preprocesses and calibrates the original pixel coordinates, eliminating the inherent interference from pixel discretization and principal point offset, and providing a standardized basic coordinate sequence for subsequent 2D coordinate upsizing.

[0032] S22. Merge the depth values ​​from the Colmap output depth map into a two-dimensional normalized physical coordinate sequence to obtain a three-dimensional transition sequence. : in, For the first The depth value of each pixel. It is a three-dimensional space.

[0033] This step achieves the fusion of depth information in two-dimensional normalized coordinates, providing complete two-dimensional depth feature support for subsequent three-dimensional coordinate transformation, and ensuring an effective transition of coordinates from a plane to space.

[0034] S23. The three-dimensional transition sequence is mapped to three-dimensional spatial coordinates in the camera coordinate system using the following formula. : , , , in, The x-coordinate in the camera coordinate system The vertical coordinate in the camera coordinate system These are the depth coordinates in the camera coordinate system.

[0035] This step ensures that during the 2D to 3D conversion process, the coordinate mapping always follows the physical rules of the camera's intrinsic parameters, maintaining the accuracy and consistency of the spatial positions of all pixels in the camera's local coordinate system.

[0036] S24. Convert the 3D spatial coordinates in the camera coordinate system to 3D points in the world coordinate system using the following formula: , in, For the world coordinate system The vector form of a pixel's three-dimensional points. The first in camera coordinate system Pixel three-dimensional spatial coordinate vector form, Let the rotation matrix be the camera's extrinsic parameters. for transpose, This is the translation vector of the camera's extrinsic parameters.

[0037] This step ensures that all pixels are mapped to a unified world coordinate system in subsequent global scene analysis and multi-image point cloud fusion, fully preserving the global spatial location information of the scene and guaranteeing the accuracy and consistency of the 3D coordinate results.

[0038] S25. Using the origin of the world coordinate system as the origin of all rays, and taking the direction pointing to each three-dimensional point in the world coordinate system as different ray directions, construct each ray, and sample on each ray to obtain a three-dimensional sampling point set.

[0039] This invention achieves precise mapping from two-dimensional images to three-dimensional spatial points through pixel normalization calibration, depth fusion, and step-by-step transformation between the camera coordinate system and the world coordinate system. This ensures the physical consistency and spatial accuracy of the coordinate transformation and improves the reconstruction accuracy of three-dimensional points of rock debris.

[0040] In order for the MLP (Multilayer Perceptron) to effectively learn high-frequency geometric details, the three coordinate components of each 3D sampling point need to be independently encoded to map the original low-dimensional input to a high-dimensional feature space.

[0041] S3. The NeuS (Neural Implicit Surfaces by Volume Rendering) method is used to encode the position of the 3D sampling point set to obtain a high-dimensional feature vector, including the following steps: S31. For each of the three coordinate components of the three-dimensional sampling point in the three-dimensional sampling point set, perform multi-frequency sine and cosine function encoding independently: The expression for the multi-frequency sine and cosine function encoding method for each coordinate component is as follows: , in, It is a sine function. It is a cosine function. Pi For coordinate components, Coordinate components The encoded result vector, For frequency levels.

[0042] This embodiment The value is set to 6. Each frequency corresponds to a pair of sine and cosine functions, thus representing each one-dimensional coordinate component. (For example, the horizontal, vertical, and depth coordinate components) are mapped to a 12-dimensional feature vector. Simultaneously, the frequencies in the formula increase by powers of 2, enabling the network to simultaneously capture low-frequency global shape contours and high-frequency local geometric features (such as sharp edges and small protrusions). This design balances the global smoothness of the SDF (Signed Distance Function) with the fidelity of local details.

[0043] S32. Concatenate the encoded vectors of the three coordinate components to obtain the high-dimensional feature vectors of each three-dimensional sampling point. Concatenate as follows: , in, The three-dimensional sampling points obtained by stitching High-dimensional feature vectors, Three-dimensional sampling points The encoded vector of the horizontal coordinate component. Three-dimensional sampling points The encoded vector of the ordinate component. Three-dimensional sampling points The encoded vector of depth coordinate components.

[0044] Since each component generates a 12-dimensional vector, the high-dimensional feature vector of each 3D sampling point is 36-dimensional. Through this positional encoding, MLP can more accurately fit the signed distance field of complex geometric surfaces, effectively recovering fine geometric structures while maintaining smoothness, thus laying the foundation for high-quality 3D reconstruction.

[0045] This invention draws on the NeuS method and employs multi-frequency sine and cosine position encoding, which can map low-dimensional coordinates to a high-dimensional feature space. This allows the network to simultaneously capture the low-frequency overall contour of rock fragments and the high-frequency fine geometric structure, effectively restoring the minute textures, edges, and uneven details of the rock fragment surface and significantly improving surface fidelity.

[0046] S4. Based on a multilayer perceptron, and using high-dimensional feature vectors combined with Colmap data, a 3D reconstruction network model is obtained through joint optimization using multiple losses. This includes the following steps: S41. Based on a multilayer perceptron, construct a 3D reconstruction network model, such as... Figure 2 As shown, it includes: Deep subnetworks are used to calculate depth values ​​based on high-dimensional feature vectors; The symbolic distance field subnetwork is used to calculate symbolic distance values ​​based on high-dimensional feature vectors; The color sub-network is used to calculate color values ​​based on high-dimensional feature vectors.

[0047] In the 3D reconstruction network model constructed in this embodiment, the volume density and SDF symbolic distance function are calculated respectively through the MLP multilayer perceptron, and the color value and depth value are synthesized through the volume rendering equation, thus realizing the construction of the three major sub-networks.

[0048] S42. Construct loss functions corresponding to depth, symbol distance, and color values ​​respectively. Based on the high-dimensional feature vector and combined with Colmap data, optimize the 3D reconstruction network model through backpropagation joint training.

[0049] In this embodiment, when using the Colmap tool, sufficient depth and symbolic distance values ​​of sample points are obtained. Combined with the known color values, these constitute the sample data for training. The loss function is constructed based on the error, and through backpropagation and joint optimization, a 3D reconstruction network model is finally trained.

[0050] This invention employs a multilayer perceptron with three branches: a depth subnetwork, a symbolic distance field subnetwork, and a color subnetwork. It can simultaneously learn the geometric structure, spatial distance, and surface color information of rock fragments. Through joint constraints using multiple losses, network stability optimization is achieved, resulting in more complete and robust reconstruction results. Joint optimization using depth loss, color loss, and symbolic distance function loss balances geometric smoothness, surface detail, and color realism in the 3D reconstruction network, avoiding overfitting or detail loss caused by a single loss function.

[0051] S5. Using the 3D reconstruction network model and the Marching Cubes algorithm, a 3D model of rock cuttings is obtained, including the following steps: S51. Construct a cube-shaped bounding box with a side length of 1.01, so that it completely encloses the rock debris target object located within the unit sphere.

[0052] S52. Generate a dense and uniform 3D mesh inside the bounding box, input the mesh vertices into the 3D reconstruction network model, and calculate the symbolic distance value.

[0053] Therefore, in this embodiment, a discretized symbolic distance voxel field is formed, where the values ​​of internal points are positive and the values ​​of external points are negative.

[0054] S53. Using the isosurface extraction algorithm, a three-dimensional model of rock debris is obtained based on the sign distance value. The specific process is as follows: traverse all cube elements. When an isosurface with a sign distance value of 0 passes through a cube element (the sign of the sign distance value of the eight vertices of the cube element determines whether the SDF=0 isosurface passes through the element), calculate the intersection point between the isosurface and the cube edge through linear interpolation. Connect the intersection points to generate triangular facets, forming a continuous surface. Finally, output a surface model composed of triangular meshes as the three-dimensional model of rock debris.

[0055] This invention employs the Marching Cubes isosurface extraction algorithm, which can directly generate continuous, closed triangular mesh models from trained implicit networks. It can obtain three-dimensional rock cutting models that can be directly used for geological analysis without post-processing, and the process is simple and highly practical.

[0056] In summary, this invention achieves high-precision 3D reconstruction solely using multi-view rock cuttings images, eliminating the need for additional high-precision scanning equipment, thus reducing the cost of geological sample modeling. It can quickly acquire 3D morphology, porosity, fracture, and other structural information of rock cuttings. The resulting 3D rock cuttings model can intuitively reflect the internal structure and composition of formation rocks, providing a reliable basis for geological interpretation, reservoir evaluation, porosity and permeability analysis, and drilling program optimization, and possesses significant engineering application value.

[0057] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for 3D reconstruction of rock cuttings integrating Colmap and deep learning, characterized in that, Includes the following steps: S1. Multi-view rock debris images are processed using Colmap to obtain training data for the 3D reconstruction network; S2. Combining camera intrinsic and extrinsic parameters with Colmap data, the pixels of each rock debris image are converted into three-dimensional points in the world coordinate system, and a three-dimensional sampling point set is obtained through ray sampling. S3. The NeuS method is used to encode the position of the three-dimensional sampling point set to obtain a high-dimensional feature vector; S4. Based on the multilayer perceptron, and using high-dimensional feature vectors combined with Colmap data, a three-dimensional reconstruction network model is obtained through joint optimization of multiple losses. S5. Using a three-dimensional reconstruction network model and an isosurface extraction algorithm, a three-dimensional model of rock debris is obtained.

2. The method for 3D reconstruction of rock cuttings integrating Colmap and deep learning according to claim 1, characterized in that, S1 includes the following steps: S11. Prepare a set of rock debris images in RGB encoding format from different perspectives, covering all angles of the rock debris scene; S12. Use Colmap to extract feature points from each rock cutting image and match the feature points of different rock cutting images to find the correspondence between the rock cutting images. S13. Based on the feature point matching results of S12, the camera pose is obtained by triangulation and a sparse point cloud is generated. S14. Based on the sparse point cloud, the MVS algorithm is used to generate a dense point cloud; S15. Use Colmap to generate corresponding depth maps for each rock cuttings image.

3. The method for 3D reconstruction of rock cuttings integrating Colmap and deep learning according to claim 1, characterized in that, S2 includes the following steps: S21. Normalize and calibrate the original pixel two-dimensional coordinates of each rock cuttings image using the following formula to obtain a two-dimensional normalized physical coordinate sequence. : ， ， in, For the first The original x-coordinate of each pixel. For the first The original ordinate of each pixel. This represents the lateral offset of the camera's intrinsic parameters. This represents the vertical offset of the camera's intrinsic parameters. This refers to the camera's intrinsic lateral focal length. The vertical focal length is the camera's intrinsic parameter. For the first Normalized physical x-coordinate of each pixel For the first Normalized physical ordinate of each pixel It is a two-dimensional space. The number of pixels; S22. Merge the depth values ​​from the Colmap output depth map into a two-dimensional normalized physical coordinate sequence to obtain a three-dimensional transition sequence. : in, For the first The depth value of each pixel. It is a three-dimensional space; S23. The three-dimensional transition sequence is mapped to three-dimensional spatial coordinates in the camera coordinate system using the following formula. : ， ， ， in, The x-coordinate in the camera coordinate system The vertical coordinate in the camera coordinate system These are the depth coordinates in the camera coordinate system. S24. Convert the 3D spatial coordinates in the camera coordinate system to 3D points in the world coordinate system using the following formula: ， in, For the world coordinate system The vector form of a pixel's three-dimensional points. The first in camera coordinate system Pixel three-dimensional spatial coordinate vector form, Let be the camera extrinsic rotation matrix. for transpose, The translation vector of the camera's extrinsic parameters; S25. Using the origin of the world coordinate system as the origin of all rays, and taking the direction pointing to each three-dimensional point in the world coordinate system as different ray directions, construct each ray, and sample on each ray to obtain a three-dimensional sampling point set.

4. The method for 3D reconstruction of rock cuttings integrating Colmap and deep learning according to claim 1, characterized in that, S3 includes the following steps: S31. For each three-dimensional sampling point in the three-dimensional sampling point set, the three coordinate components are independently encoded using multi-frequency sine and cosine functions. S32. Concatenate the encoded result vectors of the three coordinate components to obtain the high-dimensional feature vectors of each three-dimensional sampling point.

5. The method for 3D reconstruction of rock cuttings integrating Colmap and deep learning according to claim 4, characterized in that, The expression for the multi-frequency sine and cosine function encoding method for each coordinate component in S31 is as follows: ， in, It is a sine function. It is a cosine function. Pi For coordinate components, Coordinate components The encoded result vector, For frequency levels.

6. The method for 3D reconstruction of rock cuttings integrating Colmap and deep learning according to claim 4, characterized in that, The encoding result vector of the three coordinate components of each three-dimensional sampling point in S32 is concatenated in the following manner: ， in, The three-dimensional sampling points obtained by stitching The high-dimensional feature vector, Three-dimensional sampling points The encoded vector of the horizontal coordinate component. Three-dimensional sampling points The encoded vector of the ordinate component. Three-dimensional sampling points The encoded vector of depth coordinate components.

7. The method for 3D reconstruction of rock cuttings integrating Colmap and deep learning according to claim 1, characterized in that, S4 includes the following sub-steps: S41. Based on a multilayer perceptron, a 3D reconstruction network model is constructed, which includes: Deep subnetworks are used to calculate depth values ​​based on high-dimensional feature vectors; The symbolic distance field subnetwork is used to calculate symbolic distance values ​​based on high-dimensional feature vectors; The color sub-network is used to calculate color values ​​based on high-dimensional feature vectors; S42. Construct loss functions corresponding to depth, symbol distance, and color values ​​respectively. Based on the high-dimensional feature vector and combined with Colmap data, optimize the 3D reconstruction network model through backpropagation joint training.

8. The method for 3D reconstruction of rock cuttings integrating Colmap and deep learning according to claim 7, characterized in that, S5 includes the following steps: S51. Construct a cube-shaped bounding box with a side length of 1.01, so that it completely encloses the rock debris target object inside the unit sphere; S52. Generate a dense and uniform 3D mesh inside the bounding box, input the mesh vertices into the 3D reconstruction network model, and calculate the symbolic distance value. S53. Using the isosurface extraction algorithm, a three-dimensional model of rock debris is obtained based on the symbolic distance value.

9. The method for 3D reconstruction of rock cuttings integrating Colmap and deep learning according to claim 8, characterized in that, The method of S53 is as follows: traverse all cube elements, and when an isosurface with a symbolic distance value of 0 passes through a cube element, calculate the intersection point between the isosurface and the cube edge by linear interpolation, and generate triangular facets based on the intersection points to form a continuous surface as a three-dimensional model of rock debris.

Images