A GPU-based gold mine particle size distribution three-dimensional reconstruction method, device and medium

Abstract

The application discloses a kind of GPU-based gold mine particle size distribution three-dimensional reconstruction method, equipment and medium, it is related to three-dimensional image processing technical field, including, acquisition gold mine stockpile multi-view image and camera pose, complete image normalization and calibration alignment processing on GPU, and calculate perspective quality weight, obtain synchronous frame group, pose matrix and quality weight table;According to pose matrix, candidate depth mutual projection is executed in parallel to synchronous frame group, and according to quality weight table, similarity is gathered, and matching cost body is generated, and matching cost body generates dense depth map and depth confidence map by semi-global matching, and the three-dimensional sampling point is obtained by executing back projection to dense depth map, three-dimensional sampling point is parameterized as Gaussian primitive and forms three-dimensional model state.The application realizes the inhibition of false matching under the condition of occlusion and weak texture and improves the stability and reliability of depth description by fusing multi-view mutual projection similarity and combining semi-global matching to output dense depth and depth confidence.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional image processing technology, and in particular to a GPU-based method, device, and medium for three-dimensional reconstruction of gold ore particle size distribution. Background Technology

[0002] With the increasing demand for digital and intelligent mineral processing in mines, the 3D acquisition of particle size distribution in gold ore stockpiles is gradually shifting from contact screening and single-point measurement to 3D reconstruction and statistical evaluation based on multi-view imaging. Multi-camera arrays, structured light, and photogrammetry are being used more widely in open-pit and plant environments. The improved parallel computing capabilities of graphics processors have accelerated the realization of depth estimation, point cloud generation, and 3D representation, enabling 3D reconstruction under conditions of large field of view, high resolution, and high frame rate acquisition to gradually possess near real-time characteristics, providing basic data support for online particle size assessment, process control, and production line closed-loop. Among these technologies, the generation of 3D models has become a key step in assessing particle size distribution and analyzing mineral particle characteristics, providing more accurate data support for decision support in the mine production process.

[0003] Existing methods have shortcomings. Multi-view depth solving is sensitive to differences in viewpoint quality, occlusion, and weak textures, which can easily lead to mismatches and depth instability. Furthermore, the lack of a unified characterization of depth reliability results in insufficient stability and reproducibility of granular statistical values ​​in densely packed regions. In addition, the consistency between instance partitioning and 3D correspondence under conditions of particle adhesion and blurred boundaries is poor, which can easily lead to cross-particle mismatches, fragmentation of the same particle, and unstable attribution. This affects the credibility and interpretability of the calculation of particle-level geometric features and the results of particle-level determination, thereby affecting the accuracy and reliability of the final generated 3D model. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a GPU-based three-dimensional reconstruction method for gold ore particle size distribution to solve the problems of unstable depth estimation and inconsistent three-dimensional separation of cohesive particles.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a GPU-based method for three-dimensional reconstruction of gold ore particle size distribution, which includes acquiring multi-view images of gold ore piles and camera poses, performing image normalization and calibration alignment processing on the GPU, and calculating viewpoint quality weights to obtain a synchronized frame group, pose matrix and quality weight table.

[0008] Candidate depth cross-projection is performed in parallel on the synchronous frame group according to the pose matrix, and similarity is aggregated according to the quality weight table to generate a matching cost body. The matching cost body generates a dense depth map and a depth confidence map through semi-global matching. Back-projection is performed on the dense depth map to obtain 3D sampling points. The 3D sampling points are parameterized into Gaussian primitives to form the 3D model state. Particle instance segmentation is performed on the dense depth map and the depth confidence map to obtain particle boundary cues. The particle boundary cues are mapped to the 3D model state according to the pose matrix to form 3D boundary evidence. The adjacency relationship of Gaussian primitives is weighted and graph segmentation clustered using the 3D boundary evidence to obtain particle instance clusters and conflict lists. Particle instance clusters and conflict lists are combined to calculate granular geometric features and determine granularity to generate granular results. The granular results are labeled on the 3D model state to generate a granularity distribution 3D model and a granularity list.

[0009] As a preferred embodiment of the GPU-based 3D reconstruction method for gold ore grain size distribution described in this invention, the specific steps for obtaining the synchronization frame group, pose matrix, and mass weight table are as follows:

[0010] Collect multi-view images of gold ore piles and camera poses, and associate and bind them with timestamps and camera numbers to generate multi-view image sequences and camera pose sequences.

[0011] The camera pose sequence is converted and aligned into a pose matrix. The multi-view image sequence is normalized and calibrated on the GPU to obtain a normalized multi-view image corresponding to the pose matrix.

[0012] Spatiotemporal alignment is performed based on the timestamps of normalized multi-view images. A set of multi-camera frames corresponding to the same time point is extracted to generate a synchronization frame group, and the synchronization frame group is aligned and associated with the pose matrix.

[0013] Calculate the view quality weights for the normalized multi-view images within the synchronization frame group, and organize them into a quality weight table according to the synchronization frame group and camera number.

[0014] As a preferred embodiment of the GPU-based 3D reconstruction method for gold ore grain size distribution described in this invention, the steps of performing parallel candidate depth cross-projection on synchronous frame groups according to the pose matrix and aggregating similarities according to a quality weight table to generate a matching cost volume are as follows.

[0015] By utilizing the normalized multi-view images and pose matrices within the synchronized frame group, a candidate depth set is generated at the image location of the normalized multi-view images;

[0016] Parallel computation of the cross-projection similarity of image locations at different candidate depths in normalized multi-view images;

[0017] By combining the quality weight table, the similarity of multi-view mutual projection is weighted and fused to construct the matching cost body.

[0018] As a preferred embodiment of the GPU-based 3D reconstruction method for gold ore grain size distribution described in this invention, the matching cost volume generates a dense depth map and a depth confidence map through semi-global matching. Back projection is performed on the dense depth map to obtain 3D sampling points. These 3D sampling points are then parameterized into Gaussian primitives to form the 3D model state. The specific steps are as follows:

[0019] Perform semi-global matching optimization on the matching cost volume to solve for the optimal depth value corresponding to the image location, and generate a dense depth map and a depth confidence map;

[0020] Based on the dense depth map and the corresponding pose matrix, a set of 3D sampling points is obtained by back projection.

[0021] Each 3D sampling point is parameterized into a corresponding Gaussian primitive, and the state of the 3D model is constructed through all Gaussian primitives.

[0022] As a preferred embodiment of the GPU-based 3D reconstruction method for gold ore grain size distribution described in this invention, the specific steps for obtaining grain boundary clues by performing grain instance segmentation on the dense depth map and depth confidence map are as follows:

[0023] The effective image locations in the dense depth map are filtered based on the depth confidence map to generate a set of effective image locations;

[0024] The effective image location set and the corresponding depth values ​​of the dense depth map are organized into an effective depth set;

[0025] The effective depth set includes a pairing relationship between the effective image location set and the corresponding depth value;

[0026] Using the effective depth set as the processing domain, particle instance segmentation is performed on the dense depth map and the depth confidence map to generate particle instance segmentation results, and the boundaries are extracted from the particle instance segmentation results to form particle boundary cues.

[0027] As a preferred embodiment of the GPU-based 3D reconstruction method for gold ore grain size distribution described in this invention, the following steps are taken: Particle boundary cues are mapped to the 3D model state according to the pose matrix to form 3D boundary evidence. Weighting and graph segmentation clustering are then performed on the adjacency relationships of Gaussian primitives using this 3D boundary evidence to obtain particle instance clusters and a conflict list.

[0028] Particle boundary cues are combined with dense depth maps and mapped to the state of a 3D model according to the pose matrix to form 3D boundary evidence.

[0029] Weighting and graph segmentation clustering are performed on the adjacency relations of Gaussian primitives based on three-dimensional boundary evidence, and the weighting results and graph segmentation clustering results are obtained.

[0030] Based on the weighting results and graph segmentation clustering results, granular instance clusters and conflict lists are generated.

[0031] As a preferred embodiment of the GPU-based 3D reconstruction method for gold ore grain size distribution described in this invention, the steps of calculating and determining grain size geometric features by combining particle instance clusters and conflict lists to generate grain size results, and then labeling the grain size results onto the 3D model state to generate a 3D model of grain size distribution and a grain size list, are as follows:

[0032] By performing conflict entry processing on the Gaussian primitives contained in the granular instance cluster using the conflict list, the set of valid Gaussian primitives corresponding to the granular instance cluster is obtained.

[0033] The conflict entry handling refers to removing the Gaussian primitives corresponding to the conflict list from the Gaussian primitives contained in the granular instance cluster;

[0034] granular geometric features are calculated based on the effective Gaussian primitive set, and the granular geometric features are associated with the particle instance cluster to form a granular geometric feature set;

[0035] Granularity determination is performed based on the set of granular geometric features, and granularity results are generated. The granularity results maintain a correspondence with the particle instance clusters.

[0036] The particle size distribution 3D model is generated by labeling the particle size distribution results with the Gaussian primitives corresponding to the particle instance clusters in the 3D model state, and a particle size list is generated based on the particle size distribution results.

[0037] As a preferred embodiment of the GPU-based three-dimensional reconstruction method for gold ore particle size distribution described in this invention, the particle size geometric features include volume, surface area, shape factor, and equivalent particle size.

[0038] The particle size determination includes performing conservative determination marking, equivalent particle size threshold range determination, and shape factor boundary verification on the particle instance clusters marked on the conflict list.

[0039] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the GPU-based three-dimensional reconstruction method for gold ore particle size distribution as described in the first aspect of the present invention.

[0040] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the GPU-based three-dimensional reconstruction method for gold ore grain size distribution as described in the first aspect of the present invention.

[0041] The beneficial effects of this invention are as follows: By fusing multi-view mutual projection similarity according to view quality weights and combining it with semi-global matching to output dense depth and depth confidence, it achieves the suppression of mismatches and the improvement of depth stability and reliability characterization under occlusion and weak texture conditions, thereby improving the stability and reproducibility of 3D reconstruction and providing a more consistent depth basis for the construction of 3D models; by mapping particle boundary cues to 3D boundary evidence and weighting Gaussian primitive adjacency relations before graph segmentation and clustering and generating a conflict list, it achieves consistent 3D separation and stable control of the affiliation of adhering particles, thereby reducing cross-particle misconnection and fragmentation and improving the accuracy and interpretability of particle size determination and particle size statistics, thus improving the consistency and usability of the particle size distribution 3D model. Attached Figure Description

[0042] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a flowchart of a GPU-based method for 3D reconstruction of gold ore grain size distribution.

[0044] Figure 2 A flowchart for generating the matching cost body.

[0045] Figure 3 A flowchart for generating the state and particle boundary cues of the 3D model.

[0046] Figure 4 A flowchart for generating a 3D model of granularity distribution and a list of granularity levels.

[0047] Figure 5 A comparison of deep RMSE data for multi-view mutual projection similarity fusion and semi-global matching.

[0048] Figure 6 A comparative data graph showing the relationship between the binned mean of depth confidence and the mean absolute depth error. Detailed Implementation

[0049] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0050] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0051] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0052] Reference Figures 1-6 This is one embodiment of the present invention, which provides a GPU-based three-dimensional reconstruction method for gold ore grain size distribution, including the following steps:

[0053] S1. Acquire multi-view images of the gold ore pile and camera pose, perform image normalization and calibration alignment on the GPU, and calculate view quality weights to obtain synchronized frame groups, pose matrices and quality weight tables.

[0054] S1.1 Acquire multi-view images and camera poses of the gold ore pile, and associate and bind them with the timestamp and camera number to generate multi-view image sequences and camera pose sequences.

[0055] It should be noted that an image acquisition array consisting of multiple industrial cameras is deployed around the gold ore pile. Each industrial camera is clock-synchronized and numbered. During operation, a unified synchronization trigger signal is sent to the image acquisition array so that all industrial cameras complete exposure acquisition at the same trigger moment, obtaining a multi-view image set covering the entire surface of the gold ore pile at the same time.

[0056] Each time an exposure is triggered, the acquisition control terminal generates a global timestamp for this trigger and writes the global timestamp into the image metadata of the multi-view image group corresponding to this trigger. At the same time, it synchronously acquires the camera attitude data corresponding to each industrial camera. The camera attitude data includes the position and orientation parameters of the camera in three-dimensional space, and the global timestamp is written into the camera attitude data to form an attitude record. When the sampling frequency of the camera attitude data is higher than the image frame rate, the attitude data is subjected to nearest neighbor selection or interpolation alignment based on the global timestamp to obtain the attitude record corresponding to the global timestamp. The same global timestamp is assigned to the multi-view image group acquired at the same trigger time, and a unique number of the camera to which each multi-view image in the multi-view image group belongs is written. The pairing relationship between multi-view images and attitude records is established based on "global timestamp + camera number". All pairing results are sorted and organized according to the global timestamp to form a multi-view image sequence and a camera attitude sequence arranged in chronological order.

[0057] S1.2. Convert and align the camera pose sequence into a pose matrix. Perform image normalization and calibration alignment processing on the GPU to obtain normalized multi-view images corresponding to the pose matrix.

[0058] It should be noted that the position and orientation parameters in each pose record are read using the "global timestamp + camera number" index, and a mapping relationship is established with the corresponding multi-view images. When the orientation parameter is represented by quaternions, the quaternions are normalized to eliminate scale errors, and the rotation matrix is ​​calculated according to the analytical rules of the quaternion-to-rotation relationship, expressed as follows:

[0059] ;

[0060] ;

[0061] in, Represents the rotation matrix; Represent quaternions; Represents the four components of a quaternion; Represents the scalar components of a quaternion; , , This represents the three components of the vector components of a quaternion along the three-dimensional coordinate axes.

[0062] When the orientation parameter is expressed using Euler angles, the roll, pitch, and yaw angles are combined into a rotation matrix, expressed as follows:

[0063] ;

[0064] in, Indicates circling Axis rotation rotation matrix; Indicates circling Axis rotation rotation matrix; Indicates circling Axis rotation rotation matrix; Indicates the roll angle; Indicates the pitch angle; Indicates the yaw angle; It indicates a combination of yaw, pitch, and roll angles in that order.

[0065] Position parameters directly constitute the 3D coordinates of the camera center in the world coordinate system. Under a unified external parameter direction convention (world coordinate system to camera coordinate system or camera coordinate system to world coordinate system), the 3D coordinates and rotation matrix are converted into translation components, expressed as follows:

[0066] ;

[0067] in, Indicates the translation component; This represents the rotation matrix that rotates a vector from the world coordinate system to the camera coordinate system; subscript Represents the world coordinate system; Indicates the camera coordinate system; This represents the position vector of the camera center in the world coordinate system; An abbreviation for the center of a camera.

[0068] Will and Assembled into a pose matrix, the expression is:

[0069] ;

[0070] in, This represents the pose matrix.

[0071] The multi-view image sequence is written to the GPU memory in batches according to the index. On the GPU, the size is resampled, the pixel value range is normalized, and the channel order is unified for each frame to obtain a normalized multi-view image. Based on the camera intrinsic parameters (focal length and principal point parameters) and distortion parameters, the pixel coordinates of the normalized multi-view image are reversed and the distortion correction is completed by interpolation sampling to obtain a distortion-corrected normalized multi-view image. The distortion-corrected normalized multi-view image is calibrated and aligned, and the pixel coordinate origin and axial direction, and the positive depth direction and right-hand rule are unified to ensure that the camera intrinsic parameter definition and the coordinate rule of the pose matrix are consistent. This establishes a frame-by-frame correspondence between the normalized multi-view image and the pose matrix, and outputs a normalized multi-view image corresponding to the pose matrix.

[0072] S1.3. Perform spatiotemporal alignment based on the timestamps of standardized multi-view images, extract the set of multi-camera frames corresponding to the same time point to generate a synchronization frame group, and align and associate the synchronization frame group with the pose matrix.

[0073] It should be noted that, based on standardized multi-view images and the "global timestamp + camera number" index information, all standardized multi-view images are binned and aggregated according to the global timestamp, and a time alignment tolerance window is set. Standardized image frames from different camera numbers are extracted within the same time bucket or at the same time point falling within the same tolerance window. These frames are then sorted by camera number and organized into a multi-camera frame set at the same time point, defined as a synchronization frame group. The synchronization frame group records the shared time point identifier, the set of participating camera numbers, and the corresponding frame index. For each standardized multi-view image within each synchronization frame group, the corresponding pose matrix is ​​retrieved in the pose matrix sequence using the "global timestamp + camera number" key, and consistency verification is performed. If there are missing frames or missing pose records, the corresponding camera frame or the corresponding time point synchronization frame group is removed to ensure that the frames within the group correspond to the pose matrix frame by frame. The synchronization frame group and the retrieved pose matrix are bound with the same index to form a "synchronization frame group - pose matrix" alignment association result.

[0074] S1.4 Calculate the view quality weights for the normalized multi-view images within the synchronization frame group, and organize them into a quality weight table according to the synchronization frame group and camera number.

[0075] It should be noted that within each synchronized frame group, the corresponding normalized multi-view images are traversed according to the camera number. For each frame of normalized multi-view images, the view quality sub-indices are calculated in parallel on the GPU and view quality weights are generated. Specifically, the sharpness index is obtained based on the image pixel gradient energy; the exposure contrast index is obtained based on the mean brightness, standard deviation of brightness, and the proportion of overexposed / underexposed pixels; the effective coverage index is obtained based on the proportion of effective pixel masks or material pile area masks; the angle between the camera optical axis direction and the material pile reference direction is calculated based on the pose matrix bound to the corresponding frame; and the camera center distance between the camera and the other cameras in the synchronized frame group is calculated. The mean or median of the camera center distance is taken as the baseline statistic. The geometric view index is obtained by combining the angle and the baseline statistic.

[0076] The expression for calculating the sharpness index is as follows:

[0077] ;

[0078] in, Indicates synchronization frame group Inside, camera Clarity index; superscript Indicates clarity; Indicates the synchronization frame group number; Indicates the camera number; Indicates synchronization frame group Inside, camera Normalized multi-view images at the pixel level Brightness at that location; Represents pixel coordinates; This represents the set of pixels that participated in the statistics; Indicates the number of pixels; This represents the sharpness normalization scale constant, used to map sharpness to the range of 0 to 1.

[0079] The expression for calculating the exposure contrast index is:

[0080] ;

[0081] in, Indicates synchronization frame group Inside, camera Exposure contrast index; superscript Indicates exposure contrast; Indicates the average brightness; Indicates the standard deviation of brightness; This represents the average target brightness, indicating the desired exposure level. This represents the standard deviation of the target brightness, indicating the desired contrast level. , This represents the scale parameter, used to control the intensity of the penalty when the target value is deviated from. Indicates the percentage of overexposed pixels; This indicates the percentage of underexposed pixels.

[0082] The expression for calculating the effective coverage metric is as follows:

[0083] ;

[0084] in, Indicates synchronization frame group Inside, camera Effective coverage indicators; superscript Indicates coverage; This represents the effective pixel mask, with a value of 0 or 1. 1 indicates that the pixel belongs to the effective area, and 0 indicates that the pixel belongs to the invalid area.

[0085] The expression for the computational geometric perspective index is as follows:

[0086] ;

[0087] in, Indicates synchronization frame group Inside, camera Geometric perspective indicators; superscript Represents a geometric perspective; Indicates the angle of view; Indicates the included angle of the target; Indicates the angle penalty scale parameter; Indicates baseline statistics; This represents the baseline saturation value.

[0088] Each viewpoint quality sub-index is mapped to a uniform dimension between 0 and 1, and the viewpoint quality sub-indexes are combined to obtain the original quality score of the corresponding camera frame. Non-negative clipping and normalization are performed on the original quality scores of all camera frames within the same synchronization frame group to obtain the viewpoint quality weights of each camera frame in the group, and the sum of the weights is 1. The viewpoint quality weights are organized into a quality weight table according to "synchronization frame group number + camera number". The quality weight table contains at least the synchronization frame group number, camera number and viewpoint quality weight fields.

[0089] S2. Perform candidate depth cross-projection on the synchronous frame group in parallel according to the pose matrix, and aggregate similarity according to the quality weight table to generate matching cost volume. The matching cost volume generates dense depth map and depth confidence map through semi-global matching. Perform back projection on the dense depth map to obtain three-dimensional sampling points, parameterize the three-dimensional sampling points into Gaussian primitives and form the three-dimensional model state.

[0090] S2.1. Using the normalized multi-view images and pose matrices within the synchronized frame group, a set of candidate depths is generated at the image positions of the normalized multi-view images, and the multi-view mutual projection similarity of the image positions at different candidate depths is calculated in parallel.

[0091] It should be noted that within each synchronized frame group, the camera frame with the highest viewpoint quality weight is selected as the reference viewpoint frame. The normalized multi-view image corresponding to the reference viewpoint frame is used as the reference frame. A candidate depth set is generated at each image position of the reference frame. The multi-view cross-projection similarity of each image position at different candidate depths is calculated in parallel. The candidate depth set is determined by a preset depth range and sampling rules. The preset depth range is determined during the calibration stage in combination with the installation position of the camera array, the field of view coverage, and the typical working distance of the gold ore pile. Several representative pile surface positions (e.g., near end, middle, far end, and edge areas of the pile) are selected within the observable area of ​​the camera array and multi-view matching is performed. The coarse depth distribution from the surface of the material pile to the camera is obtained by triangulation. Based on the coarse depth distribution, the "near-end depth boundary" is defined as the minimum working distance that can be stably observed and still ensure effective matching, and the "far-end depth boundary" is defined as the maximum working distance that the material pile may reach while still meeting the imaging resolution and matching reliability. To resist changes in the shape of the material pile, minor equipment vibrations and measurement noise during operation, safety margins are reserved on both sides of the near-end depth boundary and the far-end depth boundary, so that the probability of the true depth falling within the preset range during subsequent reconstruction is higher. The sampling rule generates a candidate depth set with a fixed number of candidates within the preset depth range, and reverse depth uniform sampling is used to make the near-distance sampling denser and the far-distance sampling sparser.

[0092] For any image position and any candidate depth in the reference frame, a 3D point corresponding to the image position at the corresponding candidate depth is constructed along the viewing direction of the reference view frame. Using the pose matrices of the reference view frame and other cameras, the 3D point is transformed from the coordinate convention of the reference view frame to the coordinate convention of other cameras. Based on this, it is projected onto the normalized multi-view images corresponding to other cameras to obtain the mutual projection pixel positions. Out-of-bounds and invalid regions are removed from the mutual projection pixel positions. Bilinear interpolation is used to sample the brightness values ​​of the mutual projection pixel positions from the normalized multi-view images corresponding to other cameras. A similarity score is calculated between the brightness values ​​of the reference view frame and the brightness values ​​sampled from each other camera. Before calculating the similarity score, the image brightness values ​​of each camera within the synchronization frame group are uniformly scaled: after converting each frame image into a brightness map, it is linearly normalized according to the pixel dynamic range output by the camera, so that the brightness value falls within the range of 0 to 1; where the image is 8-bit grayscale, it is normalized according to the range of 0 to 255. Normalization is performed on the brightness values. For higher bit depths or other formats, normalization is performed according to the corresponding minimum and maximum value ranges. Through unified scaling, brightness values ​​from different cameras and under different exposure conditions have a consistent numerical scale, ensuring that the similarity results calculated based on brightness differences can be compared and fused within the same range. At the same candidate depth, the brightness difference between the reference camera and other cameras at their mutual projection positions is calculated and averaged over the view dimension. Based on the normalized brightness values ​​to the range of 0 to 1, the brightness difference is converted into multi-view mutual projection similarity. The value of multi-view mutual projection similarity also ranges from 0 to 1, with a larger value indicating better consistency across multiple views at the candidate depth. After removing out-of-bounds and invalid regions from the mutual projection pixel positions, the camera numbers that still have valid mutual projection pixel positions are grouped into a valid set. Subsequent similarity calculations and summaries are performed only for cameras within this valid set, and normalization is performed based on the number of cameras in the valid set. The expression for generating multi-view mutual projection similarity is:

[0093] ;

[0094] in, Indicates synchronization frame group In the image position of the middle reference view frame Candidate depth Multi-view cross-projection similarity; Indicates the image location; Indicates candidate depth; Indicates synchronization frame group In the image location With candidate depth The set of camera numbers that still participate in the calculation after being removed from out-of-bounds and invalid regions; Indicates synchronization frame group Chinese reference camera At image location The brightness value; Indicates the reference camera; Indicates the position of the mutually projected pixels, referring to the reference camera. Image location In candidate depth The corresponding 3D points are transformed by pose relationship and projected onto other cameras. The pixel coordinates obtained from the normalized multi-view image.

[0095] When a certain , All other cameras either exceeded the limit or were invalid. An undefined condition will occur, and the multi-view cross-projection similarity of the corresponding image position at the corresponding candidate depth will be set to 0, indicating that an effective multi-view consistency constraint cannot be formed.

[0096] S2.2. Combine the quality weight table to perform weighted fusion of the multi-view mutual projection similarity to construct the matching cost body. Perform semi-global matching optimization on the matching cost body to solve the optimal depth value corresponding to the image position and generate a dense depth map and a depth confidence map.

[0097] It should be noted that within each synchronized frame group, a reference viewpoint frame is used, and multi-view cross-projection similarities are aggregated at each image position and each candidate depth of the reference viewpoint frame. Viewpoint quality weights for each camera frame are extracted from the quality weight table according to the synchronized frame group number and camera number. For the same image position and the same candidate depth, only entries with valid cross-projection (not exceeding boundaries and with valid sampled pixels) from different camera frames are retained, and weighted fusion is performed according to the corresponding quality weights to obtain the fused similarity. This automatically removes invalid contributions when some camera projections are invalid and ensures the stability of the fused score scale. A preset cost mapping is defined. The mapping rule converts the fusion similarity into a matching cost value. The preset cost mapping rule is to linearly adjust the fusion similarity to the range of 0 to 1 (if the original similarity is already in the range of 0 to 1, it is used directly; if the original similarity is in the range of -1 to 1, it is first linearly translated and scaled to the range of 0 to 1). Then, the difference between 1 and the fusion similarity is calculated to obtain the matching cost value. This monotonically inverts the value to satisfy the optimization objective of "the greater the similarity, the smaller the cost". The matching cost values ​​corresponding to all image positions and all candidate depths are organized into a matching cost body using a three-dimensional index method of "image row coordinates × image column coordinates × candidate depth index".

[0098] For a given image location at a certain candidate depth, if mutual projection of all camera frames is invalid, the matching cost of the index unit is uniformly set to the maximum cost constant to avoid subsequent optimization in selecting the current candidate depth. Semi-global matching optimization is performed on the matching cost body, and path accumulation is performed on the candidate depths along multiple directional paths, with small depth change penalties and large depth jump penalties applied. The accumulated results in each direction are aggregated to obtain the aggregate cost. For each image location, the optimal depth value corresponding to the candidate depth with the minimum aggregate cost is selected in the candidate depth dimension and organized to generate a dense depth map. The depth confidence is calculated based on the difference between the aggregate costs of the optimal solution and the suboptimal solution, and a depth confidence map is generated.

[0099] S2.3. Based on the dense depth map and the corresponding pose matrix, perform back projection to obtain a set of three-dimensional sampling points, parameterize each three-dimensional sampling point into a corresponding Gaussian primitive, and construct the three-dimensional model state through all Gaussian primitives.

[0100] It should be noted that, based on the pose matrix corresponding to the dense depth map and the reference view frame, and using the camera intrinsic parameters (focal length and principal point parameters) registered during the calibration and alignment stage, the optimal depth value of the dense depth map is read point by point on the pixel coordinate grid of the reference view frame. The pixel coordinates are corrected for principal point offset according to the camera intrinsic parameters and normalized according to the focal length to obtain the camera viewing direction corresponding to the current pixel. The camera viewing direction is then scaled according to the optimal depth to obtain camera coordinate sampling points. The pose matrix is ​​used to transform the camera coordinate sampling points to a unified world coordinate system to obtain world coordinate sampling points, which are then aggregated to form a three-dimensional sampling point set.

[0101] When the depth value at a certain image location is an invalid depth identifier (corresponding to an invalid optimal depth caused by completely invalid mutual projection), no 3D sampling point is generated. For other locations, the sampling points are subjected to continuous reliability constraints based on the depth confidence given by the depth confidence map. This ensures that the lower the depth confidence, the smaller the contribution in the subsequent representation and the greater the spatial uncertainty. Each world coordinate sampling point is parameterized as a Gaussian primitive. Specifically, the mean of the Gaussian primitive is taken as the coordinate of the world coordinate sampling point, the covariance is determined by the spatial scale corresponding to the optimal depth at the unit pixel scale for lateral diffusion, and the depth confidence is used to monotonically adjust the diffusion in the depth direction to characterize the uncertainty in the depth direction. The color attribute is taken as the brightness or color component of the reference view frame at the corresponding pixel position, and the transparency attribute is given by the monotonically mapped by the depth confidence to reflect the effective contribution of the Gaussian primitive to the 3D representation. The 3D model state is composed of all Gaussian primitives and index relationships.

[0102] It should also be noted that existing technologies directly and equally weight the mutual projection similarity of each viewpoint to determine depth and back-project it into 3D points. This is easily affected by low-quality viewpoints and occlusion noise, resulting in increased mismatches, significant depth jumps, difficulty in distinguishing reliability, and a large number of 3D outliers. This solution performs weighted fusion of similarity according to quality weights and converts it into cost. Then, it uses semi-global matching to obtain a more stable dense depth and depth confidence. Furthermore, it expresses the uncertainty and contribution of back-projected points using Gaussian parameters, thereby improving the depth stability and 3D representation usability in occluded and weakly textured scenes, and solving the problems of depth instability and difficulty in controlling outliers caused by equal-weighted fusion.

[0103] Figure 5 Using occlusion rate (horizontal axis) as the independent variable and depth RMSE (vertical axis, scientific notation) as the evaluation index, this paper shows the trend of dense depth result accuracy with increasing occlusion under the difficult condition of "weak texture + low quality viewpoint ratio of 0.66". This is achieved by calculating multi-view cross-projection similarity based on normalized multi-view images and pose matrices of synchronized frame groups and constructing a matching cost volume. The five curves in the figure correspond to "equal-weighted convergence similarity (excluding semi-global matching)", "equal-weighted convergence similarity + simplified semi-global matching", "equal-weighted convergence similarity + semi-global matching", "viewpoint quality weighted fusion (excluding semi-global matching)", and "viewpoint quality weighted fusion + semi-global matching (this invention)". Overall, it is visible that the RMSE of all schemes increases with increasing occlusion rate, but the present invention... The curve consistently remains at its lowest point with shorter error bars, indicating that fusing similarity scores by weighting them according to viewpoint quality and combining them with semi-global matching can more effectively suppress mismatches and improve depth stability and reproducibility under occlusion and weak texture conditions. The error bars represent the 95% confidence interval under repeated simulations, reflecting the fluctuation range of results under different random occlusion / noise perturbations. The red dashed rectangle and double-headed arrow in the overview figure mark the local magnification window. The local magnification further uses double-headed arrows to mark the differences between the schemes, and the point of maximum difference is highlighted with dashed lines and arrows at an occlusion rate of 0.8. It can be seen that the present invention has a 45% reduction in RMSE and stability advantage relative to the baseline in the high occlusion interval, thereby providing a more consistent depth basis for subsequent 3D reconstruction and making the reliability characterization of the depth confidence map more usable.

[0104] Figure 6This diagram illustrates the changes in consensus among different schemes under varying occlusion rates. The horizontal axis represents the occlusion rate (0 to 1), and the vertical axis represents the scheme method, including equal-weighted fusion, quality-weighted fusion, scheme group, strong-weighted scheme group, and weak-weighted scheme group. The color intensity of each cell reflects the consensus level under the corresponding occlusion rate and scheme; darker colors indicate higher consensus. As the occlusion rate increases, the consensus of most schemes gradually decreases. The strong-weighted scheme group exhibits higher consensus at lower occlusion rates, indicating that the scheme can maintain high stability under high-quality perspective conditions. In contrast, the equal-weighted fusion scheme shows a more significant decrease in consensus at higher occlusion rates, demonstrating lower stability. This approach clearly shows the stability differences among different schemes under varying occlusion conditions. The color differences in local detail areas further amplify the performance differences among different schemes under the same occlusion rate, helping to analyze the performance of various methods under different occlusion conditions.

[0105] S3. Perform particle instance segmentation on the dense depth map and depth confidence map to obtain particle boundary clues. Map the particle boundary clues to the 3D model state according to the pose matrix to form 3D boundary evidence. Perform weighting and graph segmentation clustering on the adjacency relationship of Gaussian primitives through the 3D boundary evidence to obtain particle instance clusters and conflict lists.

[0106] S3.1. Based on the depth confidence map, the effective image locations of the dense depth map are filtered to generate a set of effective image locations, and the set of effective image locations and the corresponding depth values ​​of the dense depth map are organized into a set of effective depths.

[0107] It should be noted that the dense depth map and depth confidence map are aligned pixel by pixel. The depth confidence and corresponding depth value of each image location are read point by point on the pixel coordinates of the reference view frame. Image locations with invalid depth values ​​are first removed to prevent invalid depths from entering subsequent particle instance segmentation. The number of remaining usable image locations is then counted. A pre-defined "target effective image location number" is used to constrain the computational scale of subsequent particle instance segmentation. The target effective image location number is pre-defined based on the reference view frame resolution, the parallel computational load of subsequent particle instance segmentation, and the upper limit of GPU memory usage. Set a preset retention ratio and determine the ratio of the number of target valid image locations to the number of available image locations (when the ratio is greater than 1, take 1 to ensure that no additional locations are discarded when there are insufficient available locations). Sort the available image locations from high to low depth confidence and extract the first few image locations corresponding to the preset retention ratio to form a set of valid image locations. When multiple image locations have the same depth confidence, determine the selection based on the fixed order of pixel coordinates (row first, column second). Pair each image location in the set of valid image locations with the corresponding depth value in the dense depth map and collect and organize them into a set of valid depths.

[0108] S3.2. Using the effective depth set as the processing domain, perform particle instance segmentation on the dense depth map and the depth confidence map to generate particle instance segmentation results, and extract the boundaries from the particle instance segmentation results to form particle boundary cues.

[0109] It should be noted that, within the image location range defined by the effective depth set, the depth values ​​of the dense depth map and the confidence scores of the depth confidence map are paired according to pixel coordinates to form a joint "depth-confidence" representation. The back-projection relationship is then used to convert the depth values ​​corresponding to each pixel within the processing domain into 3D sampling point coordinates. The brightness or color components, depth confidence scores, and morphological features calculated from local neighborhood depth changes at the same location in the reference view frame are extracted and organized into a point feature set. An improved deep learning instance segmentation network is used for parallel inference on the point feature set, outputting the particle instance number and boundary probability for each 3D sampling point. The particle instance number is then backfilled according to pixel coordinates to obtain the particle instance segmentation result. The differences in instance number within the neighborhood are checked point by point on the particle instance segmentation result, and the boundary pixel set at the instance boundary is selected by combining the boundary probability. The boundary pixel set, along with adjacent instance number pairs and corresponding boundary probabilities, is then compiled to form particle boundary clues.

[0110] It should also be noted that when training the improved deep learning instance segmentation network, multi-view samples covering different stacking densities, lighting, and occlusion conditions are collected. Point sets corresponding to the multi-view samples are generated, and each point in the point set is labeled with its corresponding particle instance number and boundary position, forming supervised learning data pairs. Fixed-point resampling and coordinate normalization are performed on each point set, and data augmentation such as random rotation, scaling, jitter, and local missing data are performed to improve robustness to viewpoint changes and occlusion noise. The three-dimensional coordinates, brightness or color components, depth confidence, and morphological features calculated from neighborhood depth changes are concatenated as point feature inputs. The particle instance number is used as the instance segmentation supervision signal, enabling the improved deep learning instance segmentation network to learn the aggregation of the same particle point and the separation of different particle points. Boundary labels are used as boundary supervision signals to enable the improved deep learning instance segmentation network to output boundary probabilities to strengthen the boundary representation of adhered particles. The training loss is calculated iteratively in batches on the GPU. When the training loss is less than the convergence threshold, the training is considered complete. After training, the obtained network parameters are saved as improved deep learning instance segmentation network parameters that can be directly loaded and called.

[0111] The convergence threshold is defined based on the normalized smooth training loss. An exemplary value range is 0.01 to 0.05. If the convergence threshold is set higher than 0.05, the training is more likely to stop prematurely, and the segmentation boundary and adhesion separation ability have not been fully learned. If the convergence threshold is set lower than 0.01, it usually takes longer training to occasionally reach the threshold, which will waste computing resources.

[0112] S3.3 Combine the particle boundary cues with the dense depth map and map them to the state of the 3D model according to the pose matrix to form 3D boundary evidence. Based on the 3D boundary evidence, perform weighting and graph segmentation clustering on the adjacency relationship of Gaussian primitives to obtain the weighting results and graph segmentation clustering results and generate particle instance clusters. Based on the weighting results and graph segmentation clustering results, organize and generate a conflict list.

[0113] It should be noted that the boundary pixels in the particle boundary cues are aligned with the dense depth map according to pixel coordinates to read the corresponding depth values, and invalid depth entries are removed. For the remaining boundary pixels, the world coordinate boundary points are calculated using the back projection and pose matrix transformation relationship. A correlation is established between the pixel position and the Gaussian primitives in the 3D model state, transforming the 2D boundary into 3D boundary evidence. Within the 3D model state, spatial adjacency relationships are constructed using the mean coordinates of the Gaussian primitives, and the basic adjacency strength is formed using the mean distance, covariance overlap, and color similarity. This is then used to establish the 3D boundary evidence. Boundary evidence reduces the strength of adjacent edges crossing the boundary and increases the strength of adjacent edges connected on the same side to complete adjacency weighting. A spatial clustering algorithm is then executed on the weighted adjacency relationships. The spatial clustering algorithm uses graph partitioning clustering to group Gaussian primitives to obtain granular instance clusters. Consistency checks are performed on the weighted edges near the boundary and their cluster affiliations. Gaussian primitives and their adjacency relationships that have "strong boundary evidence but are still classified into the same cluster" or "maintain similar adjacency strength with multiple clusters at the same time, resulting in unstable affiliation" are collected and organized into conflict entries, and the conflict entries are summarized to obtain a conflict list.

[0114] It should also be noted that existing technologies mostly rely on abrupt changes in dense depth to divide particles and directly group them in 3D. This is easily affected by noise, occlusion, and weak textures, resulting in unclear boundaries, particle adhesion or fragmentation, unstable 3D grouping, and difficulty in tracing disputed points. This solution uses depth reliability as a constraint and links the 2D boundary with the 3D grouping for correction, outputting a conflict list to locate uncertain areas. This makes particle boundaries clearer and more coherent, 3D grouping more stable and consistent, and significantly reduces misconnections and fragmentation. It improves the usability of 3D reconstruction results and the reliability of particle size determination, and solves the problems of boundary aliasing and 3D clustering errors caused by unstable depth.

[0115] S4. Combine particle instance clusters and conflict list to calculate particle-level geometric features and determine particle level, generate particle level results, and label the particle level results to the 3D model state to generate a 3D model of particle size distribution and a particle level list.

[0116] S4.1. Perform conflict entry processing on the Gaussian primitives contained in the granular instance cluster through the conflict list to obtain the set of valid Gaussian primitives corresponding to the granular instance cluster.

[0117] It should be noted that the granular instance cluster provides the Gaussian primitives contained in each granular instance, and the conflict list provides the Gaussian primitive records that need to be processed. First, the Gaussian primitive records in the conflict list are deduplicated and summarized into a conflict Gaussian primitive set. Then, the Gaussian primitives contained in each granular instance cluster are checked one by one. For Gaussian primitives belonging to the conflict Gaussian primitive set, conflict entry processing is performed. Conflict entry processing involves removing the Gaussian primitive from the Gaussian primitives contained in the current granular instance cluster. When the same Gaussian primitive is contained in multiple granular instance clusters at the same time, the removal is performed on each granular instance cluster containing the Gaussian primitive. The remaining Gaussian primitives after removal are summarized to form the valid Gaussian primitive set corresponding to the granular instance cluster, while maintaining the correspondence with the original granular instance cluster.

[0118] S4.2 Calculate the granular geometric features based on the effective Gaussian primitive set, and establish a correlation between the granular geometric features and the particle instance cluster to form a granular geometric feature set.

[0119] It should be noted that, for each cluster of particle instances, the effective set of Gaussian primitives is traversed one by one, the mean coordinates and covariance of each Gaussian primitive are read, and the spatial distribution expression of the current particle instance is obtained by summing them. The mean coordinates are weighted and averaged according to the contribution represented by transparency to obtain the geometric center of the particle instance, which is used for subsequent unified reference of geometric quantities. Each Gaussian primitive is decomposed according to covariance to obtain three principal axis directions and scales. The square root of the covariance eigenvalue is taken as the three axis radius and regarded as a three-dimensional ellipsoid corresponding to the current Gaussian primitive. The approximate values ​​of the ellipsoid volume and ellipsoid surface area are calculated respectively, and then calculated according to transparency. The volume and surface area of ​​a particle instance cluster are obtained by weighted summation of brightness, volume, and surface area. The shape factor is calculated based on the volume and surface area. The shape factor is a combination ratio of volume and surface area, which characterizes the degree to which the particle is close to a sphere and is organized into a range of 0 to 1, so that the larger the value, the closer it is to a sphere. The equivalent particle size is converted from the volume. The equivalent particle size is the diameter of a sphere with the same volume. The volume, surface area, shape factor, and equivalent particle size of the current particle instance cluster are obtained, which are the particle-level geometric features. The particle-level geometric features are paired with the corresponding particle instance clusters and summarized to form a set of particle-level geometric features.

[0120] S4.3 Perform particle size determination based on the particle size geometric feature set, generate particle size results, and maintain a correspondence between the particle size results and the particle instance clusters.

[0121] It should be noted that the equivalent particle size, shape factor, volume, and surface area of ​​each particle instance cluster are read through the set of particle size geometric features one by one. Before performing particle size determination, a set of particle size boundary values ​​and a set of particle size numbers are given in advance as particle size division parameters. The set of particle size boundary values ​​is determined by the operation requirements or screening diameter and is saved along with the task parameters. It is a set of boundary values ​​arranged in ascending order of particle size and is used to divide the equivalent particle size into several adjacent intervals. The equivalent particle size is compared with the set of particle size boundary values ​​level by level to determine the interval to which it falls. The comparison is carried out in ascending order, and the interval that satisfies "the lower limit is not greater than the equivalent particle size and the equivalent particle size is less than the upper limit" is taken as the assigned interval. The corresponding particle size number is output to complete the particle size determination.

[0122] When the equivalent particle size falls at or is very close to the boundary between adjacent intervals, causing instability in interval assignment, the assignment is stabilized using the shape factor. This ensures that particle clusters with a higher shape factor and closer to spherical shape are assigned to the corresponding interval based on their equivalent particle size, while particle clusters with a lower shape factor and less spherical shape are preferentially assigned to a coarser interval near the boundary to reflect greater uncertainty in their external dimensions. The particle size number, along with the corresponding equivalent particle size and the particle size geometric features used for determination, is combined to form the particle size result, which is then associated with the particle cluster.

[0123] S4.4. Label the particle size distribution results to the Gaussian primitives corresponding to the particle instance clusters in the 3D model state, generate a 3D model of particle size distribution, and generate a particle size list based on the particle size distribution results.

[0124] It should be noted that the particle level information, such as the particle level number and equivalent particle size, is read for each particle instance cluster. This particle level information is then appended as particle size annotation information to the corresponding valid Gaussian primitives of the particle instance cluster. At the same time, the particle instance identifier to which the Gaussian primitive belongs is retained, so that all valid Gaussian primitives in the 3D model state carry a consistent annotation of "particle instance identifier + particle level number". This allows Gaussian primitives to be grouped and displayed or presented differently according to particle level, resulting in a 3D model of particle size distribution. The particle level results of all particle instance clusters are statistically summarized using the particle level number as the summation dimension. At least the number of particles corresponding to each particle level, the statistics of the equivalent particle size (mean, minimum, maximum or quantile), and the proportion of the sum of the particle volumes corresponding to each particle level to the total volume are summarized. The summary items of each particle level are then organized in order of particle level number to form a particle level list.

[0125] This embodiment also provides a computer device applicable to the GPU-based three-dimensional reconstruction method for gold ore grain size distribution, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the GPU-based three-dimensional reconstruction method for gold ore grain size distribution as proposed in the above embodiment.

[0126] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0127] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the GPU-based three-dimensional reconstruction method for gold ore grain size distribution as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0128] In summary, this invention achieves improved stability and reliability of 3D reconstruction by fusing multi-view mutual projection similarity according to view quality weights and combining it with semi-global matching to output dense depth and depth confidence. This suppresses mismatches and improves depth stability and reliability under occlusion and weak texture conditions, thereby enhancing the stability and reproducibility of 3D reconstruction and providing a more consistent depth foundation for 3D model construction. Furthermore, by mapping particle boundary cues to 3D boundary evidence and weighting Gaussian primitive adjacency relationships before graph segmentation and clustering to generate a conflict list, this invention achieves consistent 3D separation and stable control of the affiliation of adherent particles. This reduces cross-particle misconnections and fragmentation, improves the accuracy and interpretability of particle size determination and granularity statistics, and enhances the consistency and usability of the granularity distribution 3D model.

[0129] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A GPU-based method for three-dimensional reconstruction of gold ore grain size distribution, characterized in that: include, Multi-view images of gold ore piles and camera poses are acquired. Image normalization and calibration alignment are performed on the GPU, and view quality weights are calculated to obtain synchronized frame groups, pose matrices, and quality weight tables. Candidate depth cross-projection is performed in parallel on the synchronous frame group according to the pose matrix, and similarity is aggregated according to the quality weight table to generate matching cost volume. The matching cost volume generates dense depth map and depth confidence map through semi-global matching. Back projection is performed on the dense depth map to obtain 3D sampling points. The 3D sampling points are parameterized into Gaussian primitives and form the 3D model state. Particle instance segmentation is performed on dense depth maps and depth confidence maps to obtain particle boundary cues. The particle boundary cues are then mapped to the state of the 3D model according to the pose matrix to form 3D boundary evidence. The adjacency relations of Gaussian primitives are weighted and graph segmentation clustered using the 3D boundary evidence to obtain particle instance clusters and conflict lists. By combining particle instance clusters and conflict lists, particle-level geometric features are calculated and particle-level determination is performed to generate particle-level results. The particle-level results are then labeled onto the 3D model state to generate a 3D model of particle size distribution and a particle-level list.

2. The GPU-based three-dimensional reconstruction method for gold ore grain size distribution as described in claim 1, characterized in that: The specific steps for obtaining the synchronization frame group, pose matrix, and quality weight table are as follows. Collect multi-view images of gold ore piles and camera poses, and associate and bind them with timestamps and camera numbers to generate multi-view image sequences and camera pose sequences. The camera pose sequence is converted and aligned into a pose matrix. The multi-view image sequence is normalized and calibrated on the GPU to obtain a normalized multi-view image corresponding to the pose matrix. Spatiotemporal alignment is performed based on the timestamps of normalized multi-view images. A set of multi-camera frames corresponding to the same time point is extracted to generate a synchronization frame group, and the synchronization frame group is aligned and associated with the pose matrix. Calculate the view quality weights for the normalized multi-view images within the synchronization frame group, and organize them into a quality weight table according to the synchronization frame group and camera number.

3. The GPU-based three-dimensional reconstruction method for gold ore grain size distribution as described in claim 2, characterized in that: The process involves parallel candidate depth cross-projection onto the synchronized frame group based on the pose matrix, and similarity aggregation according to the quality weight table to generate a matching cost volume. The specific steps are as follows: By utilizing the normalized multi-view images and pose matrices within the synchronized frame group, a candidate depth set is generated at the image location of the normalized multi-view images; Parallel computation of the cross-projection similarity of image locations at different candidate depths in normalized multi-view images; By combining the quality weight table, the similarity of multi-view mutual projection is weighted and fused to construct the matching cost body.

4. The GPU-based three-dimensional reconstruction method for gold ore grain size distribution as described in claim 3, characterized in that: The matching cost volume generates a dense depth map and a depth confidence map through semi-global matching. Back projection is performed on the dense depth map to obtain 3D sampling points. These 3D sampling points are then parameterized into Gaussian primitives to form the 3D model state. The specific steps are as follows: Perform semi-global matching optimization on the matching cost volume to solve for the optimal depth value corresponding to the image location, and generate a dense depth map and a depth confidence map; Based on the dense depth map and the corresponding pose matrix, a set of 3D sampling points is obtained by back projection. Each 3D sampling point is parameterized into a corresponding Gaussian primitive, and the state of the 3D model is constructed through all Gaussian primitives.

5. The GPU-based three-dimensional reconstruction method for gold ore grain size distribution as described in claim 4, characterized in that: The specific steps for performing particle instance segmentation on the dense depth map and depth confidence map to obtain particle boundary clues are as follows. The effective image locations in the dense depth map are filtered based on the depth confidence map to generate a set of effective image locations; The effective image location set and the corresponding depth values ​​of the dense depth map are organized into an effective depth set; The effective depth set includes a pairing relationship between the effective image location set and the corresponding depth value; Using the effective depth set as the processing domain, particle instance segmentation is performed on the dense depth map and the depth confidence map to generate particle instance segmentation results, and the boundaries are extracted from the particle instance segmentation results to form particle boundary cues.

6. The GPU-based three-dimensional reconstruction method for gold ore grain size distribution as described in claim 5, characterized in that: The process involves mapping particle boundary cues to the 3D model state using the pose matrix, forming 3D boundary evidence. This 3D boundary evidence is then used to perform weighting and graph segmentation clustering on the adjacency relationships of Gaussian primitives to obtain particle instance clusters and a conflict list. The specific steps are as follows: Particle boundary cues are combined with dense depth maps and mapped to the state of a 3D model according to the pose matrix to form 3D boundary evidence. Weighting and graph segmentation clustering are performed on the adjacency relations of Gaussian primitives based on three-dimensional boundary evidence, and the weighting results and graph segmentation clustering results are obtained. Based on the weighting results and graph segmentation clustering results, granular instance clusters and conflict lists are generated.

7. The GPU-based three-dimensional reconstruction method for gold ore grain size distribution as described in claim 6, characterized in that: The process involves calculating and determining the particle size distribution using a combination of particle instance clusters and a conflict list. This generates particle size results, which are then annotated onto the 3D model to create a 3D model of particle size distribution and a particle size list. The specific steps are as follows: By performing conflict entry processing on the Gaussian primitives contained in the granular instance cluster using the conflict list, the set of valid Gaussian primitives corresponding to the granular instance cluster is obtained. The conflict entry handling refers to removing the Gaussian primitives corresponding to the conflict list from the Gaussian primitives contained in the granular instance cluster; granular geometric features are calculated based on the effective Gaussian primitive set, and the granular geometric features are associated with the particle instance cluster to form a granular geometric feature set; Granularity determination is performed based on the set of granular geometric features, and granularity results are generated. The granularity results maintain a correspondence with the particle instance clusters. The particle size distribution 3D model is generated by labeling the particle size distribution results with the Gaussian primitives corresponding to the particle instance clusters in the 3D model state, and a particle size list is generated based on the particle size distribution results.

8. The GPU-based three-dimensional reconstruction method for gold ore grain size distribution as described in claim 7, characterized in that: The particle size geometry features include volume, surface area, shape factor, and equivalent particle size; The particle size determination includes performing conservative determination marking, equivalent particle size threshold range determination, and shape factor boundary verification on the particle instance clusters marked on the conflict list.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the GPU-based three-dimensional reconstruction method for gold ore grain size distribution as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the GPU-based three-dimensional reconstruction method for gold ore grain size distribution as described in any one of claims 1 to 8.

Images