Method for completing a point cloud of a workpiece for a polishing system and apparatus therefor

Abstract

The application provides a point cloud completion method for a workpiece of a polishing system and a device thereof, and the completion method comprises the following steps: obtaining a defective point cloud D of a workpiece to be processed and a corresponding RGB image; inputting the point cloud into a pre-trained ResNet neural inference model to obtain a field function; generating a corresponding SDF field and a three-dimensional Gaussian expression, optimizing the SDF field based on the three-dimensional Gaussian expression to obtain an SDF field, extracting a point cloud of the SDF field, and optimizing the point cloud by taking the information of the point cloud as a constraint to obtain a point cloud. The completion method can complete the defective point cloud of the workpiece.

Description

Technical Field

[0001] This invention relates to the field of grinding system technology, and in particular to a method and apparatus for completing the point cloud of a workpiece in a grinding system. Background Technology

[0002] In the fields of industrial manufacturing, precision machining, and workpiece quality inspection, obtaining complete and accurate three-dimensional geometric data of the workpiece surface before and after grinding is a core prerequisite for conducting surface quality assessment, identifying minute defects, and iteratively optimizing grinding process parameters. Existing traditional three-dimensional reconstruction methods are prone to problems such as three-dimensional data gaps, surface distortion, and loss of micro-texture when dealing with complex conditions such as highly reflective metal grinding surfaces, partially obscured areas, and missing workpiece edge data. This results in insufficient geometric accuracy in reconstruction, ultimately leading to distorted workpiece quality inspection results and insufficient basis for process optimization.

[0003] In recent years, new 3D reconstruction technologies based on neural implicit SDF representation and 3D Gaussian sputtering have gradually emerged. This type of solution relies on a technical chain of multi-view image acquisition, neural network feature learning, and multi-dimensional geometric constraint optimization to complete high-precision 3D morphology restoration of complex scenes. It provides a new technical approach for the automated inspection of industrial workpiece surfaces and can restore the complete shape of the workpiece and the micro-grinding texture of the surface based on 2D visual data.

[0004] However, when existing neural implicit reconstruction and 3D Gaussian reconstruction schemes are directly applied to the inspection of metal grinding workpieces, significant adaptation defects still exist: the metal grinding surface is prone to strong specular reflection, causing large-area loss of depth acquisition information from the camera, forming voids; workpiece clamping and structural occlusion exacerbate local data loss; at the same time, the workpiece edge region itself is geometrically steep, and after superimposed measurement noise interference, conventional reconstruction algorithms cannot stably reproduce microscopic features such as scratches and grinding marks. These defects significantly reduce the effective reliability of the reconstructed point cloud, directly interfering with subsequent grinding defect identification, quantitative analysis of grinding amount, and process optimization, making it difficult to meet the stringent accuracy requirements of batch online quality inspection of precision metal grinding workpieces.

[0005] Therefore, how to complete the point cloud of the workpiece has become an urgent problem to be solved. Summary of the Invention

[0006] The purpose of this invention is to provide a method and apparatus for completing the point cloud of a workpiece in a grinding system.

[0007] To achieve one of the aforementioned objectives, an embodiment of the present invention provides a method for completing the point cloud of a workpiece in a grinding system, comprising the following steps: obtaining a point cloud D of the workpiece to be processed with defects and an RGB image corresponding to the point cloud D. The point cloud D and its corresponding RGB image All data were obtained from the same viewpoint using a preset camera; all hollow pixels in the point cloud D were acquired, and a set of missing regions was formed from all the acquired hollow pixels. Using the optical axis of the camera as the z-axis, and with a preset origin, the point cloud D is transformed to obtain a point cloud in xyz format under a normalized viewpoint. Hollow regions after transformation of missing region set ; for the RGB image Normalization is performed to obtain the RGB image. The point cloud The input is fed into a pre-trained ResNet neural inference model to obtain a point cloud. Field function; generating point cloud SDF field, generating point cloud The three-dimensional Gaussian representation of the SDF field is obtained, and the SDF field is optimized based on the three-dimensional Gaussian representation. Extract the SDF field Corresponding point cloud , with point clouds Information as constraints on point clouds Optimize and obtain point cloud. .

[0008] As a further improvement to one embodiment of the present invention, the step of processing the RGB image... Normalization is performed to obtain the RGB image. Specifically, this includes: RGB images Pixel normalization is performed by dividing each pixel value by 255 to obtain a normalized RGB image. .

[0009] As a further improvement to one embodiment of the present invention, the training method of the ResNet neural inference model includes: acquiring multiple training data, each training data including: point clouds in xyz format. The point cloud Corresponding RGB image and the point cloud The corresponding ground truth SDF field is used to remove some points from the point cloud in the training data; a ResNet neural inference model is created, in which the gradient descent optimizer is the Adam optimizer and the loss function is the RMSE function; the ResNet neural inference model is then solidified into a function. ,in, For the normalized view of xyz format point cloud, It is the corresponding RGB image. It is the sampling space. It is the field function used to describe the surface of the point cloud. For any point in space, For point The shortest distance to the point cloud surface; the ResNet neural inference model is trained based on all training data.

[0010] As a further improvement to one embodiment of the present invention, the point cloud generation... The "SDF field" specifically includes: generating point clouds minimum bounding cube The minimum enclosing cube frame The internal space is defined as a three-dimensional sampling space. ; to point clouds RGB images and sampling space Input the ResNet neural inference model to generate an SDF field.

[0011] As a further improvement to one embodiment of the present invention, the point cloud generation... The "3D Gaussian representation" specifically includes: using the K-means++ algorithm to represent point clouds Clustering is performed to obtain an initial 3D Gaussian cloud; then, based on the point cloud... From the measuring point to the cavity area The distance generates the hole weight coefficient; the Gaussian center is fixed, and the covariance is corrected by weighting with the hole weight coefficient and regularizing along the Z-axis of the camera; the Gaussian geometric confidence is calculated by combining the hole weight coefficient, the point cluster size and the Gaussian volume; based on the Gaussian center, the corrected covariance and geometric confidence, an optimized three-dimensional Gaussian set is output.

[0012] As a further improvement to one embodiment of the present invention, the SDF field is obtained by optimizing the SDF field based on the three-dimensional Gaussian expression. Specifically, this includes: calculating the local geometric constraints of the three-dimensional Gaussian representation. and the constraints of the inherent properties of the SDF field. Calculate the comprehensive constraints Based on comprehensive constraints The SDF field is obtained by optimizing the SDF field. .

[0013] As a further improvement to one embodiment of the present invention, the extraction of the SDF field... Corresponding point cloud Specifically, this includes: generating point clouds. The point cloud Empty; preserve depth map At the same sampling density, Extract the two-dimensional coordinate set of the sampling points, and for each sampling point... The following operations are performed: Generate sampling points Sampling points satisfy: ,and ,in, , The coordinates are on the X-axis. The coordinates on the Y-axis. , and All coordinates are Z-axis coordinates, and the sampling points are... Add to point cloud middle.

[0014] As a further improvement to one embodiment of the present invention, the point cloud... Information as constraints on point clouds Optimize and obtain point cloud. Specifically, this includes: constructing a two-dimensional continuous depth correction field, and generating the point cloud based on the two-dimensional continuous depth correction field. As a priori constraint, a smoothness constraint is generated for the two-dimensional continuous depth correction field based on the point cloud. The prior constraints and the smoothness constraints generate a scalar correction field, and the point cloud is generated based on the scalar correction field. The optimized coordinates of each point in the graph are used to obtain the optimized point cloud. .

[0015] As a further improvement to one embodiment of the present invention, the method further includes the following step: generating the point cloud. The corresponding confidence level.

[0016] This invention also provides a point cloud completion device for a workpiece in a grinding system, comprising the following modules: a data acquisition module, used to acquire a point cloud D of the workpiece to be processed with defects and an RGB image corresponding to the point cloud D. The point cloud D and its corresponding RGB image All data were obtained from the same viewpoint using a preset camera; all hollow pixels in the point cloud D were acquired, and a set of missing regions was formed from all the acquired hollow pixels. Using the optical axis of the camera as the z-axis, and with a preset origin, the point cloud D is transformed to obtain a point cloud in xyz format under a normalized viewpoint. Hollow regions after transformation of missing region set ; for the RGB image Normalization is performed to obtain the RGB image. The processing module is used to process the point cloud. The input is fed into a pre-trained ResNet neural inference model to obtain a point cloud. Field function; generating point cloud SDF field, generating point cloud The three-dimensional Gaussian representation of the SDF field is obtained, and the SDF field is optimized based on the three-dimensional Gaussian representation. Extract the SDF field Corresponding point cloud , with point clouds Information as constraints on point clouds Optimize and obtain point cloud. .

[0017] Compared with the prior art, the technical advantages of the present invention are as follows: The embodiments of the present invention provide a method and apparatus for completing the point cloud of a workpiece in a grinding system. The completion method includes the following steps: obtaining the point cloud D of the workpiece to be processed with defects and the corresponding RGB image; inputting the point cloud into a pre-trained ResNet neural inference model to obtain a field function; generating the corresponding SDF field and a three-dimensional Gaussian representation; and optimizing the SDF field based on the three-dimensional Gaussian representation to obtain the SDF field. Extract SDF field Point clouds , with point clouds Information as constraints on point clouds Optimize and obtain point cloud. This completion method can complete the point cloud of a workpiece that has defects. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the point cloud completion method for a workpiece in an embodiment of the present invention. Figure 2 This is a schematic diagram of a field function in an embodiment of the present invention. Detailed Implementation

[0019] The following description and accompanying drawings fully illustrate specific embodiments described herein to enable those skilled in the art to practice them. Some embodiments may include or substitute parts and features of other embodiments. The scope of the embodiments herein encompasses the entire scope of the claims and all available equivalents thereof. Throughout this document, the terms “first,” “second,” etc., are used only to distinguish one element from another without requiring or implying any actual relationship or order between the elements. Indeed, a first element can also be referred to as a second element, and vice versa. Furthermore, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-parallel inclusion, such that a structure, apparatus, or device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a structure, apparatus, or device. Without further limitation, an element defined by the phrase “comprising one…” does not preclude the presence of other identical elements in the structure, apparatus, or device that includes said element. The various embodiments described herein are presented in a progressive manner, with each embodiment focusing on its differences from other embodiments; similar or identical parts between embodiments can be referred to interchangeably.

[0020] The terms "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer" used in this document to indicate orientation or positional relationships are based on the orientation or positional relationships shown in the accompanying drawings and are used only for the convenience of describing this document and simplifying the description. They do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as limiting the invention. In the description herein, unless otherwise specified and limited, the terms "installed," "connected," and "linked" should be interpreted broadly. For example, they can refer to mechanical or electrical connections, or internal connections between two elements, or direct connections or indirect connections through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.

[0021] Embodiment 1 of the present invention provides a method for completing the point cloud of a workpiece in a grinding system, such as... Figure 1 As shown, it includes the following steps: Step 101: Obtain the point cloud D of the workpiece to be processed with defects and the corresponding RGB image of the point cloud D. The point cloud D and its corresponding RGB image All data were obtained from the same viewpoint using a preset camera; all hollow pixels in the point cloud D were acquired, and a set of missing regions was formed from all the acquired hollow pixels. Using the optical axis of the camera as the z-axis, and with a preset origin, the point cloud D is transformed to obtain a point cloud in xyz format under a normalized viewpoint. Hollow regions after transformation of missing region set ; for the RGB image Normalization is performed to obtain the RGB image. The data entry preprocessing stage of the entire workpiece reconstruction algorithm takes as input a TIFF depth map with holes and an RGB original image from the same viewpoint acquired by a 3D camera, and outputs standardized data with a unified coordinate system, unified pixel value range, and marked hole positions. This provides regular input for all downstream steps such as SDF generation and 3D Gaussian optimization, and eliminates computational interference caused by shooting pose, image value range, and hole positions.

[0022] Original input data description, depth map TIFF format 3D coordinate plot, dimensions Each pixel stores three-dimensional spatial coordinates. Pixels in areas of camera reflection or occlusion have a value of NaN, representing missing data (holes). RGB image. A color image of the same resolution, with pixel values ​​ranging from 0 to 255, is used for subsequent neural network extraction of polishing texture, scratches, and highlight features. Processing workflow breakdown: Step 1: Mark the original hole pixel positions, traverse all pixels in the depth map, record the coordinates of all 2D pixels with a value of NaN, and summarize them to form the initial hole coordinate set. Function: To pre-identify all missing data collection areas, which are then used for subsequent coordinate transformations, hole weighting, and confidence calculations.

[0023] Step 2: Normalize the point cloud view coordinates. Perform a rigid coordinate transformation on the 3D coordinates of the depth map. Two core constraints are: align the Z-axis of the coordinate system with the camera's optical axis; and translate the centroid of the entire workpiece point cloud to the origin. The transformed point cloud is then obtained. A unified, center-aligned set of XYZ measurement points; void areas after transformation. Synchronously follow the coordinate transformation to obtain the range of the void space in the new coordinate system.

[0024] Design objectives: 1. Eliminate differences in robot and camera positioning poses, unify point clouds from different workpieces and shooting positions to the same standard coordinate system, and ensure that network training and inference are not affected by the positioning pose; 2. Center the point cloud and use the optical axis as the depth axis to match the imaging geometry of a monocular 3D camera, reducing the calculation deviation of SDF field and 3D Gaussian modeling.

[0025] Step 3: RGB image pixel value normalization. Divide each pixel value in the original image by 255, mapping pixels from 0 to 255 to the interval [0,1], generating a normalized image. Advantages: Neural networks are more stable when trained with floating-point values ​​in the range of 0 to 1, avoiding gradient explosion and convergence difficulties caused by large values ​​in the range of 0 to 255, and facilitating subsequent fusion with point cloud features.

[0026] Step 102: Transfer the point cloud The input is fed into a pre-trained ResNet neural inference model to obtain a point cloud. Field function; generating point cloud SDF field, generating point cloud The three-dimensional Gaussian representation of the SDF field is obtained, and the SDF field is optimized based on the three-dimensional Gaussian representation. Extract the SDF field Corresponding point cloud , with point clouds Information as constraints on point clouds Optimize and obtain point cloud. .

[0027] The completion method includes the following steps: 1. Data standardization preprocessing: Collect point clouds and RGB images with holes from the same viewpoint, and mark the hole pixels; perform coordinate normalization with the camera optical axis as the Z-axis and the centroid of the point cloud as the origin, and simultaneously transform the hole region; normalize the RGB pixels to the 0~1 interval to obtain the standardized point cloud, image, and hole region label. 2. Generate an initial neural SDF field: Input the normalized point cloud into the pre-trained ResNet model, generate an initial SDF field covering the workpiece bounding box based on implicit functions, and initially fill the collected holes by relying on the continuity of the neural field. 3. Construct a weighted 3D Gaussian constraint and optimize the SDF: Use K-means++ to cluster the point cloud to generate a 3D Gaussian, set exponentially decaying hole weights according to the distance from the measurement point to the hole, apply Z-axis regularization constraint to the Gaussian covariance to obtain a 3D Gaussian set with confidence; combine Gaussian local geometric loss and SDF intrinsic Eikonal loss to jointly optimize the initial SDF to obtain an optimized SDF field that takes into account both the global shape and the micro-texture of the polishing. 4. Extract dense, complete point cloud to match the original depth map resolution. Locate the SDF zero isosurface along the depth axis to extract a hole-free dense point cloud. 6. Point cloud consistency correction: Construct a continuous depth correction field. Use the original measured points as high-weight alignment constraints, combined with smoothing constraints, to solve for the optimal correction field, correcting the dense point cloud and ensuring unchanged measured coordinates and smooth transitions between holes. 7. Point-by-point confidence labeling: Set the confidence level of the measured points to 1. For hole-filling points, integrate SDF deviation, Gaussian support strength, and correction field curvature to calculate the confidence level. Finally, output a complete workpiece point cloud with point-by-point confidence for grinding defect detection.

[0028] It boasts the following advantages: 1. Unified coordinates, complete retention of void information, and standardized image value range eliminate computational interference caused by shooting pose and missing highlights, resulting in strong production line generalization. 2. Neural SDF completes large-area voids, and 3D Gaussian constraints preserve grinding micro-textures, solving the problems of surface fracture in traditional reconstruction and excessive smoothing of pure SDF. 3. Prioritizes locking the original measurement point coordinates without modification, smoothing only the AI-completed area, meeting industrial metrological traceability standards. 4. Distinguishes between real measurement points and algorithm-completed points, quantifies the distortion risk of the completed area, filters reconstruction artifacts, and improves defect detection accuracy. 5. End-to-end data interoperability with no coordinate transformation errors, reuses intermediate calculation parameters, has high computational efficiency, and is suitable for online real-time quality inspection on grinding production lines.

[0029] In this embodiment, the RGB image is... Normalization is performed to obtain the RGB image. Specifically, this includes: RGB images Pixel normalization is performed by dividing each pixel value by 255 to obtain a normalized RGB image. .

[0030] This section processes two types of industrial data collected in conjunction with the data, each corresponding to the same camera and shooting angle: a depth map D with holes (TIFF format). Where H is the image height in pixels and W is the image width in pixels; 3 represents that each pixel stores a set of three-dimensional coordinates. , pixel coordinates The corresponding 3D spatial coordinates of the workpiece have a key defect: some pixels have no effective depth and the values ​​are filled with NaN, which are the empty / missing areas in the reconstruction, and are the parts that need to be filled in by AI later.

[0031] Synchronized RGB color images , The size and depth maps are fully aligned, and each pixel stores the R / G / B three-channel grayscale value, which is an integer from 0 to 255. Its function is to provide the surface texture, grinding scratches, and reflective features of the workpiece, and to provide color detail constraints for the subsequent neural network.

[0032] Breakdown of the complete three-step processing flow Step 1: Mark all holes in the pixels to generate a set of missing regions. Traverse the entire depth map D and convert the image plane coordinates of all pixels with NaN values. Record everything and package it into a collection. Objective: To accurately identify locations lacking 3D measurement data, as subsequent Gaussian weighting, confidence calculation, and region completion determination all rely on this missing region marker; Distinguishing features: Pixels with coordinates = actual camera measurement data, NaN pixels = holes that need to be predicted and filled by the algorithm.

[0033] Step 2, 3D coordinate view normalization transformation (core geometry preprocessing), perform rigid body translation + rotation transformation on the 3D coordinates D of the depth map, and unify the coordinate system standard: rotation constraint, align the camera's shooting optical axis with the Z-axis of the new coordinate system; translation constraint, calculate the 3D centroid of all valid points, and move the centroid to the coordinate origin (0, 0, 0); output two types of results: (1) normalized complete point cloud , The total number of all valid 3D points. , , (1) Three-dimensional coordinates under the unified coordinate system after transformation; (2) Missing region after transformation The set of pixel holes recorded in step 1 Synchronous mapping following coordinate transformation yields a hollow region adapted to the new coordinate system. .

[0034] This step addresses the engineering problem of inconsistent shooting poses in industrial settings. Each time a workpiece is placed or the camera is moved, the origin and orientation of the original point cloud coordinates change completely, causing neural network training / inference to be affected by pose variations. Normalization unifies the coordinate system for all workpiece data, eliminating biases caused by shooting position and angle, making it easier for the network to learn the general geometric features of the workpiece.

[0035] Step 3: Normalize the pixel value range of the RGB image. The original RGB pixels are integers from 0 to 255. Perform linear scaling: divide the pixel value of each channel by 255 to obtain the normalized image. The value range is compressed from integers 0-255 to floating-point numbers 0-1; Engineering significance: Neural network activation functions (Sigmoid, ReLU, etc.) are adapted to the input range of 0-1 by default. A unified value range avoids training instability and gradient explosion caused by differences in numerical scale.

[0036] In summary, hole marking can distinguish between measured reliable regions and regions to be completed by AI, providing a basic marker for subsequent weighting, confidence, and completion logic; geometric normalization can unify the 3D point cloud coordinate system, eliminate camera pose interference, and ensure the model's generalization ability; image normalization can unify the RGB value range and adapt to neural network input specifications; and output normalization can normalize the point cloud with supporting data. Normalized texture map Hollow areas This serves as the standard input for subsequent neural SDF and 3D Gaussian reconstruction modules.

[0037] In this embodiment, the training method of the ResNet neural inference model includes: Acquire multiple training data sets, each including: point clouds in xyz format. The point cloud Corresponding RGB image and the point cloud The corresponding ground truth SDF field is used to remove some points from the point cloud in the training data; a ResNet neural inference model is created, in which the gradient descent optimizer is the Adam optimizer and the loss function is the RMSE function; the ResNet neural inference model is then solidified into a function. ,in, For the normalized view of xyz format point cloud, It is the corresponding RGB image. It is the sampling space. It is the field function used to describe the surface of the point cloud. For any point in space, For point The shortest distance to the point cloud surface; the ResNet neural inference model is trained based on all training data.

[0038] Specifically, 1) Construct each sample in the training set including a point cloud in xyz format. RGB image It consists of the true SDF field. , Real SDF scenarios, such as Figure 2 As shown, it is defined as: ,in It is the true and complete surface of the workpiece. It is an arbitrary spatial sampling point. For symbolic functions, specifically: Point inside the workpiece → SDF value is negative; The point falls exactly on the workpiece surface → SDF=0 (zero isosurface, representing the solid surface of the workpiece). Point outside the workpiece → SDF value is positive. The shortest Euclidean distance from this point to the workpiece surface.

[0039] 2) The data comes from complete point cloud data and corresponding RGB images of no less than 100 workpieces. The SDF field is automatically generated using the Poisson reconstruction algorithm.

[0040] 3) Randomly select no less than 5% of the point cloud and remove no more than 1% of the random-shaped point cloud as data augmentation samples.

[0041] 4) The gradient descent optimizer is Adam, the network architecture is ResNet, and the loss function is RMSE.

[0042] 5) The model is solidified into a function. ,in, For the normalized view of xyz format point cloud, It is the corresponding RGB image. It is the sampling space, which defaults to three-dimensional space unless otherwise defined. , It is the field function used to describe the surface of the point cloud. For any point in space, For point The shortest distance to the point cloud surface. Figure 2 A field function is shown.

[0043] The core roles of this module in the overall completion scheme are: 1. Providing a global continuous geometric prior discrete point cloud. Discrete point clouds only have sparse observation data, resulting in gaps and breaks; the neural SDF is a continuous field, so even with local missing data, a complete surface can be inferred from the global shape, resolving geometric discontinuities caused by occlusion and reflection. 2. Providing a foundational field for subsequent 3D Gaussian co-optimization. This module outputs the initial global SDF field. The subsequent 3D Gaussian module will overlay local detail constraints on this basis to correct the defects of excessive smoothing and loss of micro-texture in pure SDF. 3. The underlying basis for dense point cloud extraction is to find the 3D coordinates of SDF=0 to extract the complete dense point cloud. Without the SDF field, it is impossible to generate a complete workpiece surface after filling the voids. 4. Adapting to single-view industrial acquisition scenarios. Traditional multi-view reconstruction equipment is costly. This network only needs a single depth map + RGB to infer the complete geometry, which is suitable for real-time detection in workshop production lines.

[0044] In this embodiment, the point cloud generation The "SDF field" specifically includes: generating point clouds minimum bounding cube The minimum enclosing cube frame The internal space is defined as a three-dimensional sampling space. ; to point clouds RGB images and sampling space Input the ResNet neural inference model to generate an SDF field.

[0045] Here, will As a sampling space, calculate SDF field function: .

[0046] The input is a preprocessed normalized point cloud. Normalized RGB image Output the initial global neural symbol distance field This provides a complete and continuous prior knowledge of the workpiece shape for subsequent 3D Gaussian optimization and SDF co-optimization.

[0047] II. Two-step process breakdown Step 1: Define the three-dimensional sampling space Input normalized point cloud Calculate the minimum bounding cube B of all valid 3D measurement points; delineate the entire 3D space inside cube B as the sampling space. This limits the computational scope of the neural network, calculating the SDF only in the space surrounding the workpiece entity, avoiding invalid calculations in blank spaces without workpieces, and reducing computational overhead; it also unifies the three-dimensional computational boundary of the global field, ensuring that the spatial range of subsequent 3D Gaussian and point cloud extraction steps is completely matched, and the coordinate system is unified and without deviation.

[0048] Step 2: Call the neural inference function to generate the initial SDF field. The inference function formula is as follows: ,in, The workpiece point cloud after viewpoint normalization and coordinate system alignment; A matching RGB texture map with a value range scaled to [0,1]. This represents the sampling space of the workpiece bounding box obtained in the previous step. Output meaning: For the sampling space... any three-dimensional point The function outputs the signed SDF distance d corresponding to the point: d>0: the point is outside the workpiece; d=0: the point lies on the workpiece surface; d<0: the point is inside the workpiece entity. Underlying logic: DLsdf is the ResNet neural inference model trained previously, fusing point cloud geometric features with RGB polishing texture features to output a continuous, closed global distance field; even if the original point cloud has holes or missing parts, the network can infer the complete geometry based on the global shape, solving the problem of discrete point cloud breaks and holes.

[0049] Advantages: 1. Generates globally continuous geometric priors. Unlike discrete point clouds and voxel meshes, neural SDF is a continuous function, naturally possessing the ability to fill in gaps. The complete workpiece shape can be inferred from a single-view residual point cloud. 2. Limited bounding box sampling space, computationally efficient. It only operates within the smallest bounding cube of the workpiece, significantly reducing the number of 3D space samples, and is suitable for real-time inspection in industrial production lines. 3. Dual input of geometry and texture, resulting in higher reconstruction accuracy. Simultaneous input of point cloud coordinates and RGB polishing texture allows the network to learn surface features such as scratches and highlights, avoiding excessive smoothing distortion of pure geometric SDF. 4. Full-link connection between upstream and downstream modules, outputting the initial SDF field. It serves directly as the basic input for collaborative optimization, and the entire process space boundary and coordinate system are completely unified, with no transformation error.

[0050] In this embodiment, the point cloud generation The "3D Gaussian representation" specifically includes: using the K-means++ algorithm to represent point clouds Clustering is performed to obtain an initial 3D Gaussian cloud; then, based on the point cloud... From the measuring point to the cavity area The distance generates the hole weight coefficient; the Gaussian center is fixed, and the covariance is corrected by weighting with the hole weight coefficient and regularizing along the Z-axis of the camera; the Gaussian geometric confidence is calculated by combining the hole weight coefficient, the point cluster size and the Gaussian volume; based on the Gaussian center, the corrected covariance and geometric confidence, an optimized three-dimensional Gaussian set is output.

[0051] Initialize point cloud The three-dimensional Gaussian representation is as follows: 1) Use K-means++ clustering algorithm to analyze point clouds The process is performed, and K cluster centers are generated to obtain the initial three-dimensional Gaussian representation: ,in, For the first One cluster, It is the location of the center of Gauss. It is the covariance matrix of the Gaussian center. The specific process includes: Input: Normalized effective point cloud Algorithm: K-means++ clustering, dividing the workpiece point cloud into K local point clusters. Output the initial Gaussian parameters for each cluster: : The coordinates of the three-dimensional center of the k-th Gaussian cluster; The initial covariance matrix is ​​used to characterize the spatial distribution of the point cloud cluster (major and minor axes of the ellipsoid, stretching direction). Its function is to replace massive discrete point clouds with a small number of 3D Gaussian ellipsoidal patches fitting the workpiece surface, balancing computational efficiency with the ability to represent local details.

[0052] 2) Calculate point cloud The void weight coefficient for each measuring point in the data : ,in, , It is the distance from the i-th point to the cavity region. The minimum Euclidean distance, The distance attenuation constant is preferred. Exponential function property: the closer the point is to the hole... The weight is extremely large; the point is far from the hole ( (Get bigger) → The exponential decay approaches 0, resulting in extremely small weights. The hole edge is the region most prone to distortion, jagged edges, and artifacts during completion. This weight amplifies the constraint of the Gaussian distribution on the effective measurement points around the hole, ensuring that the Gaussian distribution near the hole perfectly matches the actual measurement points and suppressing geometric distortion at the hole edge. The weight is reduced in regions far from the hole to decrease computational redundancy.

[0053] 3) Calculate the optimized and , , ,in, Gauss Center: The cluster center positions remain unchanged; only the shape (covariance) of the Gaussian ellipsoid is corrected to ensure that the overall workpiece segment center does not shift. Two layers of correction logic: First layer: using hollow weights. The weighted calculation of covariance gives higher weight to measurement points near the cavity, which dominates the shape of the Gaussian ellipsoid and prevents deformation of the cavity edge surface; the second layer: Z-axis normalization suppresses excessive stretching by subtracting the covariance component along the camera Z-axis, which limits the abnormal elongation of the Gaussian ellipsoid along the depth direction and avoids the common 3DGS problems of "depth direction ghosting and surface expansion", making the Gaussian more consistent with the two-dimensional surface of the workpiece and accurately restore the micro-texture of the polished plane / curved surface.

[0054] 4) Calculate the geometric confidence level of the 3D Gaussian representation. Item meanings: Mean distance from void The smaller the value, the farther away from the void. The larger the value, the higher the value for clusters that are far from the cavity and have complete measurement points; the lower the value for clusters that are at the edge of the cavity. Point cluster density term The number of valid points within a cluster. It represents the volume enclosed by a Gaussian ellipsoid; the denser the points, the higher the confidence level; overall physical meaning: Characterizing the reliable support strength of this 3D Gaussian on the actual surface of the workpiece: Gaussian far from voids and with dense measuring points → Large, high geometric reliability; Gaussian at the cavity edge and with sparse measuring points → The point cloud is small and has low confidence; this confidence level will be passed to the subsequent SDF collaborative optimization and final point cloud quality inspection stages to distinguish between high-confidence measured regions and low-confidence AI-completed regions. This yields the 3D Gaussian representation of the point cloud. Each Gaussian simultaneously contains: center location, weighted regular covariance (shape), and geometric confidence; serving as the input local geometric constraints for the SDF co-optimization in the next section.

[0055] The core innovative advantages of this solution compared to traditional 3D Gaussian are: (1) The void perception weighting mechanism is customized for void scenes of industrial depth cameras, strengthens the constraints of measurement points around the missing area, and solves the industry pain points of void edge surface distortion and jagged artifacts in ordinary 3DGS. (2) Z-direction normal covariance regularization suppresses excessive stretching of Gaussian along the depth axis, adapts to single-view 3D camera acquisition scenes, preserves micro-details such as grinding scratches and small bumps, and does not cause surface expansion distortion. (3) Gaussian hierarchical confidence quantification assigns a confidence index to each local Gaussian patch, realizing the hierarchical classification of "high confidence in dense areas and low confidence in sparse void areas", providing a quantitative basis for subsequent quality inspection defect identification. (4) Taking into account both global stability and local details, only the ellipsoid shape is corrected without changing the cluster center, ensuring that the overall shape of the workpiece does not shift; at the same time, the piecewise Gaussian finely depicts the local grinding texture, making up for the defects of excessive smoothing and loss of small features in pure neural SDF fields. (5) Adapt to the input and output of the entire pipeline and directly reuse normalized point cloud and hole marker set, seamlessly connect the preprocessing in the preceding text and the SDF global field joint optimization in the following text to form a complete single-view workpiece reconstruction link.

[0056] In this embodiment, the SDF field is obtained by optimizing the SDF field based on the three-dimensional Gaussian representation. Specifically, this includes: calculating the local geometric constraints of the three-dimensional Gaussian representation. and the constraints of the inherent properties of the SDF field. Calculate the comprehensive constraints Based on comprehensive constraints The SDF field is obtained by optimizing the SDF field. .

[0057] The optimization process for the SDF field using the 3D Gaussian representation is as follows: (1) Calculate the local geometric constraints of the three-dimensional Gaussian representation: Where K is the total number of 3D Gaussians; The three-dimensional coordinates of the k-th Gaussian center; The Gaussian geometric confidence level calculated above; This represents the SDF value corresponding to the Gaussian center position. Physical-logical SDF=0 represents the actual surface of the workpiece. The loss objective is to ensure that all Gaussian centers fall as close as possible to the workpiece surface; if the Gaussian centers deviate from the surface, As the value increases, the loss also increases; confidence level A higher Gaussian density (dense measurement points, far from holes) results in a greater weight on the loss function, prioritizing the fit of high-confidence regions to the real surface. Function: Injects the micro-texture information carried by the 3D Gaussian density into the SDF field, correcting the over-smoothing problem of pure SDF.

[0058] (2) Calculate the constraints on the intrinsic properties of the SDF field: ,in, for The number of midpoints. Fundamental mathematical principle (inherent property of SDF): The standard symbolic distance function satisfies the Eikonal equation: the magnitude of the gradient at any point in space. This indicates that the rate of change of the distance field is uniform; if the gradient magnitude deviates from 1, it indicates that the SDF field is distorted and the distance calculation is inaccurate. Parameter explanation: P is the normalized point cloud. Total number of valid measuring points within the area; For measuring points The SDF gradient is calculated; the loss is calculated by taking the squared error between the gradient magnitude and 1, and averaging the errors across all measurement points. Purpose: To ensure the optimized SDF field conforms to the mathematical definition and to prevent global field distortion and distance scale distortion after introducing Gaussian constraints.

[0059] (3) Calculate the comprehensive constraints: Among them, the weighting ratio is as follows: the two losses are equally weighted (both with a coefficient of 0.5) and synchronously constrained; the balancing logic is as follows: 0.5 times Gaussian loss: forces SDF to fit the local micro-polishing texture; 0.5 times Eikonal loss: maintains the mathematical integrity of the global field of SDF and avoids local texture constraints from destroying the overall shape; to achieve a two-way balance of "no global distortion and no loss of details".

[0060] (4) Calculation results Recorded as The optimized SDF field representation is as follows: The input is the initial global SDF field. The solution logic is based on the initial field. Starting from this point, iteratively minimize the overall loss. Find the SDF field that minimizes the loss, and denote it as the optimized field. The output is It is a continuous symbolic distance field that takes into account both the global outline and the local polished texture, and serves as the input for subsequent extraction of dense and complete point clouds.

[0061] Advantages of this solution compared to traditional single-constraint SDF optimization The traditional SDF only uses Eikonal single constraint and has no local texture guidance; this solution superimposes 3D Gaussian surface loss to specifically restore the tiny scratches and micro-features of metal grinding parts, resulting in higher accuracy in quality inspection defect identification.

[0062] Confidence-weighted constraints and built-in Gaussian loss to resist void distortion The confidence level is low. The low confidence Gaussian at the edge of the hole has a weak influence on the SDF correction and will not cause surface distortion due to missing measurement points. It is suitable for depth camera hole and reflection acquisition scenarios.

[0063] Equal weight balance constraint, no bias distortion. Equal weight balances geometric details and field mathematical rationality, and will not cause the two extreme problems of sacrificing the overall shape to fit the texture or losing micro-features to ensure global smoothness.

[0064] The entire process seamlessly integrates and directly reuses the preprocessing, 3D Gaussian optimization, and neural SDF network output results to form a complete single-view workpiece reconstruction pipeline of "point cloud preprocessing → Gaussian weighted modeling → global SDF collaborative optimization", which is suitable for industrial online visual inspection.

[0065] In this embodiment, the extraction of the SDF field Corresponding point cloud Specifically, this includes: generating point clouds. The point cloud Empty; preserve depth map At the same sampling density, Extract the two-dimensional coordinate set of the sampling points, and for each sampling point... The following operations are performed: Generate sampling points Sampling points satisfy: ,and ,in, , The coordinates are on the X-axis. The coordinates on the Y-axis. , and All coordinates are Z-axis coordinates, and the sampling points are... Add to point cloud middle.

[0066] By utilizing the SDF zero isosurface to represent the characteristics of the workpiece surface, the NaN void regions of the depth camera are automatically filled, generating a new point cloud that covers the entire workpiece surface. This is to prepare for subsequent alignment and correction of the measuring points.

[0067] This may include the following steps: Step 1: Search for SDF zero-value 3D surface points along the Z-axis. The sampling rule is: the sampling resolution must be completely consistent with the original depth map D to ensure a one-to-one correspondence between pixels, facilitating subsequent matching and alignment with the original measured points; traverse each set of 2D pixel coordinates Define the intervals along the Z-axis of the camera. Find within the interval that satisfy both conditions. , forming three-dimensional sampling points Condition 1: The point lies on the physical surface of the workpiece (SDF zero isosurface); Condition 2: Mathematical meaning: The opposite signs of the SDF values ​​at the upper and lower ends of the Z-axis interval indicate that the interval must pass through the workpiece surface, and there exists a condition that satisfies... The intersection of these points is used to avoid invalid searches.

[0068] Principle: In the SDF field, positive and negative values ​​distinguish between the inside and outside of the workpiece. The SDF value is negative (inside the workpiece). When SDF is positive (outside the workpiece), the product ≤ 0, and the interval passes through the surface; solve by bisection / interpolation along the Z-axis to find the positive-negative boundary position, which is the current position. Corresponding workpiece surface depth The locations of pixels with NaN holes in the original depth map can also be calculated using the SDF field. This enables automatic filling of voids.

[0069] Step 2: Generate a complete dense point cloud, which involves summing up the 3D sampling points obtained by matching all pixel coordinates to obtain a dense point cloud without missing pixels and aligned with the resolution of the original image. , It equals the total number of pixels in the original depth map. Regardless of whether the original image has holes, each pixel corresponds to a three-dimensional surface point.

[0070] Advantages: 1. Pixel alignment is compatible with the original acquisition equipment, and the sampling density is completely consistent with the 3D camera depth map. The original effective measurement points and AI-completed hole points correspond one-to-one with pixels, which facilitates the subsequent differentiation between the measured area and the completed area and the calculation of confidence level point by point; 2. Automatically fills all holes without the need for a separate interpolation algorithm. It relies on the global continuous SDF field to naturally infer the surface depth of the hole position, without jagged edges or discontinuities; 3. The search logic is efficient and stable. It uses the positive and negative signs of the SDF in the Z-axis interval to determine whether there are surface intersections, skipping empty intervals without workpieces and reducing invalid calculations; 4. It connects the complete pipeline before and after. The input is an optimized SDF field that takes into account both the global contour and the micro-polishing texture. The output is a complete and dense point cloud, which is directly sent to the next stage for original measurement point alignment correction, forming a closed-loop reconstruction process.

[0071] In this embodiment, the point cloud Information as constraints on point clouds Optimize and obtain point cloud. Specifically, this includes: constructing a two-dimensional continuous depth correction field, and generating the point cloud based on the two-dimensional continuous depth correction field. As a priori constraint, a smoothness constraint is generated for the two-dimensional continuous depth correction field based on the point cloud. The prior constraints and the smoothness constraints generate a scalar correction field, and the point cloud is generated based on the scalar correction field. The optimized coordinates of each point in the graph are used to obtain the optimized point cloud. .

[0072] Extracted dense point cloud It is generated using pure SDF field inference, and will be compared with the original camera's measured point cloud. There are slight depth offsets; industrial quality inspection scenarios require that the actual measurement data cannot be altered, and only correction of the void areas filled by AI is allowed. This module constructs a continuous depth correction field to force the original measurement points to strictly conform to the actual measurement values, while ensuring that the surface of the void-filled area is smooth and free of jagged edges, outputting a final point cloud that balances measurement accuracy and surface continuity. .

[0073] Step-by-step process breakdown: Step 1, Define the continuous depth correction field , It is a two-dimensional continuous function that corresponds one-to-one with image pixels, representing the depth offset that needs to be compensated for at each pixel position. This is for the original measured points. The correction field satisfies the following constraints: Meaning: After depth compensation, the depth value of the original measured point is strictly equal to the actual measurement value of the camera, ensuring that the actual measured point will not be tampered with by the algorithm and meeting the requirements of industrial metrology traceability.

[0074] Step 2, Construct the measurement point alignment loss , The squared error between the corrected depth and the measured depth is calculated for all original valid measuring points. Minimizing this loss ensures that all corrected real measuring points completely coincide with the original measurements, thus locking in the coordinates of the reliable measured area.

[0075] Step 3, Construct the correction field smoothing loss , The integration range covers all valid measurement point areas plus areas with missing cavities; Represents the gradient of the correction field on the image plane. The integral of the squared gradient is used to constrain the gradual change of the correction field. Its function is to prevent abrupt changes in depth and jagged steps in the hole-filled area, and to make the surface transition from the measured area to the AI-filled area natural and smooth.

[0076] Step 4: Solve for the optimal correction field , Weighting: Alignment loss weight 0.9 (higher priority), smoothing loss weight 0.1; Optimization logic: Prioritize ensuring the original measurement points align with the actual measured values, while also considering surface smoothing, without altering the measured data for the sake of smoothing; Output the globally optimal continuous depth correction field. .

[0077] Step 5: Correct depth coordinates pixel by pixel. The optimal correction offset is superimposed on the depth corresponding to each pixel of the dense point cloud to obtain the corrected depth value: the original measured pixel has a corrected depth equal to the original measurement value of the camera, with no change; the hole-filled pixel is only smoothed to eliminate the small offset caused by SDF inference.

[0078] Step 6: Generate the final point cloud By integrating all corrected 3D coordinates, the final workpiece point cloud is output as a complete, hole-free, and accurate measurement result with a smooth surface, which can be directly used for grinding defect detection.

[0079] Advantages: 1. Rigid and authentic measured data, conforming to industrial metrology standards. The original measurement point coordinates are locked through high-weight alignment loss, unlike traditional smoothing algorithms that tamper with the real measurement data, and the quality inspection results are traceable; 2. Layered weight balance. "Authenticity preservation" and "smoothing" are distinguished by a weight of 0.9 / 0.1 to differentiate the primary and secondary constraints, prioritizing measurement accuracy and then optimizing the visual geometry of the void area, without over-smoothing and loss of workpiece contour; 3. Continuous two-dimensional correction field, without pixel discrete tortuosity. The correction field is a continuous scalar field, rather than point-by-point independent offset, resulting in a smooth transition between the measured area and the void completion area, eliminating jagged edges at the completion boundary; 4. Perfectly adaptable to the input and output of the entire reconstruction pipeline. The input consists of dense point cloud extracted by SDF, pre-processed marked void areas, and the original normalized measured point cloud. The output can be directly used for subsequent confidence calculation and automated grinding inspection.

[0080] In this embodiment, the following step is also included: generating the point cloud. The corresponding confidence level.

[0081] Industrial grinding quality inspection requires distinguishing between actual camera measurement points and AI algorithm hole completion points. This module calculates the quantified confidence level of each 3D point in the final product point cloud: the original camera-collected measured points are directly marked with the highest confidence value; the virtual points generated by hole completion are combined with three geometric indicators to calculate the confidence level, quantifying the reliability of the completed area; finally, the complete point cloud + point-by-point confidence level are output synchronously, which can filter out low-confidence artifacts during quality inspection and avoid misjudging AI completion distortion as grinding defects.

[0082] Step-by-step process breakdown Step 1: Calculate the strength of the 3D Gaussian support. The essence of the formula is to perform a weighted summation of Gaussian kernels on all optimized 3D Gaussian surfaces to represent how many highly reliable 3D Gaussian surfaces support the current spatial point. Where M is the total number of 3D Gaussians. For a single Gaussian geometric confidence level; and These are the Gaussian center and covariance matrix, respectively; physical meaning: the denser the area around a point, the more high-confidence Gaussians there are. The larger the value, the more sufficient the surface reconstruction constraints are and the higher the geometric reliability at that location.

[0083] Step 2: Assign confidence scores to the original measured points. If the two-dimensional pixel coordinates of the points are... Existing in the original normalized measured point cloud This indicates that the data at this point is real data directly collected by the camera, and the confidence level is set to the maximum value. It represents 100% reliability, with no algorithmic prediction errors, and serves as a reliable benchmark for industrial testing.

[0084] Step 3: Calculate the confidence score of the hole-filling point using multi-factor fusion. If the pixel coordinates are not within the original measured point cloud, it means that the point was generated by SDF inference to fill the hole. The confidence score is calculated by multiplying three exponential decay factors together. Each of the three factors controls a type of distortion risk, and all use exponential decay. A deterioration in any single indicator will rapidly lower the confidence level: The first factor: Initial SDF distance decay term. This represents the deviation of the point from the theoretical workpiece surface. The farther it is from the standard surface, the smaller the exponential term and the lower the confidence level; it also represents the error caused by the constraint SDF inference deviating from the actual shape. The second term is the Gaussian support normalization term. Globally normalize the Gaussian support strength. To prevent the denominator from being zero; the more abundant the surrounding Gaussian coverage, the higher this term's value; sparse measuring points and weak Gaussian support at the cavity edge significantly reduce the reliability of this term. Third term: Correction field curvature attenuation term. After consistency optimization, the curvature of the depth correction field is increased. The larger the curvature, the more severe the surface abrupt change and jagged distortion in the completed area. The exponential decay decreases rapidly, suppressing the confidence level and inhibiting abnormal wrinkles on the surface after smoothing correction.

[0085] Step 4: Output a complete point cloud with confidence level, and simultaneously output two sets of data: a complete 3D point cloud without holes. The confidence score value corresponding to each three-dimensional point. The entire dataset can be directly input into the grinding defect detection algorithm. By setting a confidence threshold, unreliable AI-completed areas can be eliminated, thereby reducing the false detection rate.

[0086] Advantages: 1. A binary hierarchical confidence system distinguishes between measured and supplementary data. Measured points have the highest confidence, while supplementary points quantify risks across multiple dimensions, fully aligning with the industrial metrology standard of "measured data first, predictions requiring confidence level labeling"; 2. Comprehensive coverage of three-dimensional risk factors quantifies multi-source reconstruction errors, simultaneously covering three types of reconstruction distortion: SDF global shape deviation, insufficient 3D Gaussian local support, and depth-corrected surface distortion. A single defect automatically lowers the confidence level, resulting in a more comprehensive assessment; 3. An exponential decay mechanism sensitively identifies low-reliability areas, using an exponential function rather than a linear function. 4. Scaling: If any geometric metric deteriorates, the confidence level drops rapidly, accurately marking high-risk areas such as hole edges and surface distortion; 5. End-to-end reuse of pre-module results: No additional computational overhead. Direct reuse of the SDF field, 3D Gaussian set, and depth correction field curvature from the previous section, without additional reconstruction calculations, resulting in high pipeline computational efficiency; 6. Adaptable to automated quality inspection: The output point-by-point confidence level can be used as a basis for defect filtering thresholds, solving the industry pain point that traditional reconstruction solutions cannot distinguish between real data and AI virtual completion points, and are prone to misjudging workpiece defects.

[0087] Because this invention provides a global continuous surface prior through a pre-trained neural implicit SDF model, and further combines a 3D Gaussian representation containing only geometric information to constrain and optimize local regions, it can simultaneously take into account the integrity of the overall shape and the accuracy of local details, and achieve high-precision 3D completion of the point cloud cavity region of a high-reflectivity polished workpiece from a single viewpoint. It is particularly suitable for surface restoration of complex curved surfaces, edge missing regions, and local reflective distortion regions.

[0088] Because this invention explicitly introduces multiple constraints during the SDF optimization process, it can still generate continuous, stable, and physically reasonable patch even when there are local holes in the depth map, unstable boundary measurements, or noise disturbances. This effectively reduces the geometric breaks, surface drift, and local artifacts that are common in traditional interpolation methods or pure data-driven methods.

[0089] Since this invention does not introduce local geometric constraints during the SDF optimization process, but instead uses a three-dimensional Gaussian distribution as the optimization condition, it can still generate physically reasonable geometric completions when facing challenging scenarios such as sparse image information, noise interference, or high reflection.

[0090] Because this invention further constructs a continuous Z-axis correction field after point cloud extraction and requires the final completed point cloud to strictly coincide with the original point cloud in the observed area, it can complete and correct only the missing area while retaining the true measurement results, thereby ensuring that the completed result is both complete and consistent with the original detection data. It is especially suitable for the requirements of measurement traceability and result verifiability in industrial testing.

[0091] Since this invention fuses and outputs multi-scale confidence scores for each point in the final completed point cloud, it can provide a quantitative reliability assessment of the completion results, providing an objective basis for industrial quality inspection, polishing effect analysis, and subsequent automatic decision-making, and reducing the risk of misjudgment caused by low-confidence completion results directly participating in measurement and analysis.

[0092] Embodiment 2 of the present invention provides a point cloud completion device for a workpiece in a grinding system, comprising the following modules: The data acquisition module is used to obtain the defective point cloud D of the workpiece to be processed and the corresponding RGB image of the point cloud D. The point cloud D and its corresponding RGB image All data were obtained from the same viewpoint using a preset camera; all hollow pixels in the point cloud D were acquired, and a set of missing regions was formed from all the acquired hollow pixels. Using the optical axis of the camera as the z-axis, and with a preset origin, the point cloud D is transformed to obtain a point cloud in xyz format under a normalized viewpoint. Hollow regions after transformation of missing region set ; for the RGB image Normalization is performed to obtain the RGB image. ; The processing module is used to process the point cloud. The input is fed into a pre-trained ResNet neural inference model to obtain a point cloud. Field function; generating point cloud SDF field, generating point cloud The three-dimensional Gaussian representation of the SDF field is obtained, and the SDF field is optimized based on the three-dimensional Gaussian representation. Extract the SDF field Corresponding point cloud , with point clouds Information as constraints on point clouds Optimize and obtain point cloud. .

[0093] It should be noted that although the steps are described in a specific order above, it does not mean that the steps must be executed in the above specific order. In fact, some of these steps can be executed concurrently, or even in a different order, as long as the required function can be achieved.

[0094] This invention can be a system, method, and / or computer program product. A computer program product may include a readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of the invention.

[0095] A readable storage medium can be a tangible device that holds and stores instructions for use by an instruction execution device. Readable storage media can include, for example, but not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination thereof.

[0096] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for completing the point cloud of a workpiece in a grinding system, characterized in that, Includes the following steps: Obtain the point cloud D of the workpiece to be processed, which contains defects, and the corresponding RGB image of the point cloud D. The point cloud D and its corresponding RGB image All data were obtained from the same viewpoint using a preset camera; all hollow pixels in the point cloud D were acquired, and a set of missing regions was formed from all the acquired hollow pixels. Using the optical axis of the camera as the z-axis, and with a preset origin, the point cloud D is transformed to obtain a point cloud in xyz format under a normalized viewpoint. Hollow regions after transformation of missing region set ; for the RGB image Normalization is performed to obtain the RGB image. ; The point cloud The input is fed into a pre-trained ResNet neural inference model to obtain a point cloud. Field function; generating point cloud SDF field, generating point cloud The three-dimensional Gaussian representation of the SDF field is obtained, and the SDF field is optimized based on the three-dimensional Gaussian representation. Extract the SDF field Corresponding point cloud , with point clouds Information as constraints on point clouds Optimize and obtain point cloud. .

2. The completion method according to claim 1, characterized in that, The RGB image Normalization is performed to obtain the RGB image. Specifically, it includes: For RGB images Pixel normalization is performed by dividing each pixel value by 255 to obtain a normalized RGB image. .

3. The completion method according to claim 1, characterized in that, The training method for the ResNet neural inference model includes: Acquire multiple training data sets, each including: point clouds in xyz format. The point cloud Corresponding RGB image and the point cloud The corresponding ground truth SDF field removes some points from the point cloud in some training data. A ResNet neural inference model is created, in which the gradient descent optimizer is the Adam optimizer and the loss function is the RMSE function; the ResNet neural inference model is then fixed as a function. ,in, For the normalized view of xyz format point cloud, It is the corresponding RGB image. It is the sampling space. It is the field function used to describe the surface of the point cloud. For any point in space, For point The shortest distance to the point cloud surface; The ResNet neural inference model is trained based on all the training data.

4. The completion method according to claim 1, characterized in that, The generated point cloud The "SDF field" specifically includes: Generating point clouds minimum bounding cube , frame the smallest cube The internal space is defined as a three-dimensional sampling space. ; to point clouds RGB images and sampling space Input the ResNet neural inference model to generate an SDF field.

5. The completion method according to claim 1, characterized in that, The generated point cloud The "three-dimensional Gaussian representation" specifically includes: The K-means++ algorithm was used to analyze the point cloud. Clustering is performed to obtain an initial 3D Gaussian cloud; then, based on the point cloud... From the measuring point to the cavity area The distance generates the hole weight coefficient; the Gaussian center is fixed, and the covariance is corrected by weighting with the hole weight coefficient and regularizing along the Z-axis of the camera; the Gaussian geometric confidence is calculated by combining the hole weight coefficient, the point cluster size and the Gaussian volume; based on the Gaussian center, the corrected covariance and geometric confidence, an optimized three-dimensional Gaussian set is output.

6. The completion method according to claim 1, characterized in that, The SDF field is obtained by optimizing the SDF field based on the three-dimensional Gaussian expression. Specifically, it includes: Calculate the local geometric constraints of the three-dimensional Gaussian representation. and the constraints of the inherent properties of the SDF field. Calculate the comprehensive constraints Based on comprehensive constraints The SDF field is obtained by optimizing the SDF field. .

7. The completion method according to claim 1, characterized in that, The extraction of the SDF field Corresponding point cloud Specifically, it includes: Generating point clouds The point cloud Empty; Preservation and depth map At the same sampling density, Extract the two-dimensional coordinate set of the sampling points, and for each sampling point... The following operations are performed: Generate sampling points Sampling points satisfy: ,and ,in, , The coordinates are on the X-axis. The coordinates on the Y-axis. , and All coordinates are Z-axis coordinates, and the sampling points are... Add to point cloud middle.

8. The completion method according to claim 1, characterized in that, The point cloud Information as constraints on point clouds Optimize and obtain point cloud. Specifically, it includes: Construct a two-dimensional continuous depth correction field, and generate the point cloud based on the two-dimensional continuous depth correction field. As a priori constraint, a smoothness constraint is generated for the two-dimensional continuous depth correction field based on the point cloud. The prior constraints and the smoothness constraints generate a scalar correction field, and the point cloud is generated based on the scalar correction field. The optimized coordinates of each point in the graph are used to obtain the optimized point cloud. .

9. The completion method according to claim 1, characterized in that, It also includes the following steps: Generate the point cloud The corresponding confidence level.

10. A point cloud completion device for a workpiece in a grinding system, characterized in that, Includes the following modules: The data acquisition module is used to obtain the defective point cloud D of the workpiece to be processed and the corresponding RGB image of the point cloud D. The point cloud D and its corresponding RGB image All data were obtained from the same viewpoint using a preset camera; all hollow pixels in the point cloud D were acquired, and a set of missing regions was formed from all the acquired hollow pixels. ; Using the optical axis of the camera as the z-axis, and performing a coordinate transformation on the point cloud D with a preset origin, a point cloud in xyz format under normalized viewpoint is obtained. Hollow regions after transformation of missing region set ; for the RGB image Normalization is performed to obtain the RGB image. ; The processing module is used to process the point cloud. The input is fed into a pre-trained ResNet neural inference model to obtain a point cloud. Field function; generating point cloud SDF field, generating point cloud The three-dimensional Gaussian representation of the SDF field is obtained, and the SDF field is optimized based on the three-dimensional Gaussian representation. Extract the SDF field Corresponding point cloud , with point clouds Information as constraints on point clouds Optimize and obtain point cloud. .

Images