Three-dimensional reconstruction and deformation detection method and system for industrial parts

Abstract

The invention discloses a three-dimensional reconstruction and deformation detection method and system for industrial parts. According to the method, a multi-view image is collected through a monocular camera, and sparse point clouds are obtained through background removal and motion recovery structure processing; a three-dimensional Gaussian sputtering method is adopted for training reconstruction, and the reconstruction quality and efficiency are improved through a progressive strategy, edge tangential densification and a composite loss function; performing point cloud registration on the reconstructed model and a standard CAD model, and sequentially performing geometric center alignment, coarse registration based on multi-neighborhood feature descriptors and fine registration of introducing a robust loss function and dynamic weight optimization to realize high-precision alignment and generate an error cloud picture; and automatically outputting a deformation qualification judgment result based on comparison between the maximum error of the error cloud picture and a preset threshold value. According to the invention, the submillimeter-level requirement of industrial detection is pursued and met in precision, and the efficiency, the cost and the automation level are remarkably improved in practicability.

Description

Technical Field

[0001] This invention relates to the field of parts inspection technology, specifically to a method and system for three-dimensional reconstruction and deformation detection of industrial parts. Background Technology

[0002] The manufacturing industry is accelerating its transformation towards flexibility and intelligence, placing higher demands on industrial-grade deformation detection technologies involving error measurement and virtual reality interaction. On the one hand, the flexible production model of multiple varieties and small batches requires detection systems to have the ability to quickly adapt to the 3D reconstruction and dynamic switching of parts of different specifications. On the other hand, intelligent production lines require manufacturing enterprises to be able to perceive, control, and interact with the manufacturing process based on the Internet, realizing closed-loop control of "inspection-process-repair". Today, rapidly developing digital modeling technology is increasingly being integrated into intelligent manufacturing systems to generate digital virtual models of physical objects, thereby enabling the monitoring and dynamic prediction of the status of materials or semi-finished parts in the production flow. Among them, neural radiation field technology can effectively decouple the geometric parameters of objects from multi-view images, and has a natural advantage in the 3D reconstruction of highly reflective objects. At the same time, it can be deployed online / offline with only a monocular camera, breaking through the limitations of traditional 3D reconstruction systems, such as complex selection, high dependence on hardware accuracy, and low flexibility.

[0003] Three-dimensional Gaussian sputtering (3DGS) is an explicit neural radiation field technique that enables sputtering from multiple views of an object. Figure 2 A 3D editable digital model is reconstructed from a 3D image, containing the object's appearance, color information, and geometric coordinates. However, this method is difficult to apply directly to high-precision industrial 3D inspection tasks, mainly due to the inability to automatically remove irrelevant backgrounds, low reconstruction accuracy, and large digital model memory requirements. Although existing technologies focus on 3DGS methods for virtual reality visualization of aero-engines, applying this method relies on additional sensors to provide depth prior information, reducing deployability and maintainability.

[0004] The goal of industrial point cloud registration is to align reconstructed models with geometric defects with standard models, thereby quantifying their geometric differences and distribution. Scale-ICP is used to address the problem of scale-dependent point cloud registration. Inherent geometric differences are considered poor initial values, and deformed locations, as outliers, cannot find corresponding points, leading to the registration process failing to converge. Therefore, robustness improvements to Scale-ICP are needed. The Geman-McClure loss function, as a classic robust loss function, treats the robustness parameter as a free variable in the optimization process, allowing direct optimization using an optimizer without manual parameter tuning. This adaptive advantage is particularly effective in tasks with multivariate outputs. However, it still suffers from technical problems when dealing with scale-dependent models and models with inherent geometric differences, being sensitive to mismatches and prone to registration failure or accuracy degradation.

[0005] However, directly applying existing 3DGS methods to the 3D reconstruction and deformation detection of industrial parts still faces the following challenges: For industrial parts with high reflectivity, weak texture, and thin-walled structure, the edges of the reconstructed model are prone to blurring and artifacts, resulting in severe geometric distortion. During training, Gaussian points tend to diffuse into the background region, leading to increased model memory and low reconstruction efficiency. In the point cloud registration stage, if there are scale differences and local deformations between the reconstructed model and the CAD model, traditional registration methods are easily affected by mismatched points and cannot converge, resulting in registration failure or decreased accuracy.

[0006] Therefore, providing a method and system for three-dimensional reconstruction and deformation detection of industrial parts to solve the above problems is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a method and system for 3D reconstruction and deformation detection of industrial parts, solving the issues of poor reconstruction quality and insufficient registration robustness of existing technologies for highly reflective and weakly textured parts. This invention achieves millimeter-level accuracy in 3D reconstruction by optimizing the 3D Gaussian sputtering training strategy and loss function; it achieves high-precision point cloud alignment with scale and deformation by proposing a multi-neighborhood feature descriptor and an improved robust Scale-ICP registration algorithm; and finally, it achieves efficient deformation detection through a fully automated error analysis and judgment process. This method requires only a monocular camera and significantly improves accuracy, efficiency, cost, and automation.

[0008] The technical solution provided by this invention is as follows: A method for 3D reconstruction and deformation detection of industrial parts includes the following steps: After acquiring and processing multi-view RGB images of the target industrial part, a sparse point cloud is generated using the motion structure restoration method. The sparse point cloud is trained using a three-dimensional Gaussian sputtering method to construct a three-dimensional reconstruction model of the target industrial part. The point cloud of the 3D reconstructed model is registered with the standard CAD model to obtain an error cloud map. Based on the error cloud map, the detection results are output to determine whether the deformation of the target industrial part is qualified. The three-dimensional Gaussian sputtering method includes the following steps: During the iterative training process, a progressive training strategy is adopted, using downsampled images for training in the early stage of training, and gradually using higher resolution images for training as the number of iterations increases; An edge tangential densification strategy is adopted to determine the semantic region and the background region. For Gaussian points located between the semantic region and the background region, densification operation is performed along the tangent direction of the surface where they are located, so as to control the distribution of the Gaussian points in the semantic region or the background region. The model is optimized using a loss function that includes image reconstruction loss, Gaussian flattening regularization, local opacity regularization, and geometric consistency regularization.

[0009] Preferably, after acquiring and processing the multi-view RGB image of the target industrial part, a sparse point cloud is generated using the motion structure reconstruction method, including the following steps: Based on capturing multi-view RGB images of the target industrial part using a monocular camera, wherein the multi-view includes at least the view of the area to be reconstructed; The multi-view RGB image is segmented based on a preset image segmentation model, and the background region is removed to obtain the corresponding mask image. Based on the masked image and motion structure reconstruction method, a sparse point cloud of the target industrial part and the corresponding camera pose are generated.

[0010] Preferably, the step of training the sparse point cloud using a three-dimensional Gaussian sputtering method to construct a three-dimensional reconstruction model of the target industrial part includes the following steps: Each three-dimensional point in the sparse point cloud is transformed into a three-dimensional Gaussian function; The optimized 3D Gaussian sputtering method iteratively executes the following sub-steps within a preset number of iterations: Based on the camera pose, the three-dimensional Gaussian function is rendered into a predicted image at the current viewpoint through parallel rasterization. Calculate the image reconstruction loss between the predicted image and the corresponding real image, and construct the total loss function by combining at least one of the Gaussian flattening regularization term, the local opacity regularization term, and the geometric consistency regularization term. The geometric parameters of the three-dimensional Gaussian function are optimized through backpropagation based on the total loss function; Until the iteration is completed or the model convergence condition is met, the optimized three-dimensional Gaussian function set is output and used as the three-dimensional reconstruction model of the target industrial part.

[0011] Preferably, The formula for the image reconstruction loss is: ; The formula for the Gaussian flattening regularization term is: ; The formula for the local opacity regularization term is: ; The formula for the geometric consistency regularization term is: ; In the formula, This indicates the total number of pixels involved in the calculation in the current view. and This represents the color intensity value at pixel P. The flattening regularization coefficient is... It is the scale factor along the three axes of the ellipsoid. The opacity of the ellipsoid. The opacity regularization coefficient for the object region. The coordinates of the center of the Gaussian function are... For the object region, The opacity regularization coefficient for the background region. Let Gaussian total number, To in the background area Gaussian number within, and These are the rasterization-based rendering depth of pixel P projected onto the corresponding 3D point in the camera coordinate system and the backtracking depth calculated from the clustering normal vector, respectively.

[0012] Preferably, the step of registering the three-dimensional reconstructed model with a standard CAD model to obtain an error cloud map includes the following steps: Calculate the geometric centers of the 3D reconstructed model and the standard CAD model, and perform a translation transformation to align the geometric centers of the 3D reconstructed model and the standard CAD model. Based on key point matching and feature descriptors, the 3D reconstruction model and the standard CAD model are downsampled to obtain a set of key point matching pairs, and the spatial variation matrix is ​​solved by singular value decomposition to achieve coarse registration. Using the set of keypoint matching pairs as the initial value, the variable-scale iterative nearest-point algorithm based on the robust loss function of the Geman-McClure kernel function is used for iterative optimization until the objective function converges, so as to achieve fine registration. After completing the fine registration, the shortest distance from each point in the 3D reconstructed model to the surface of the standard CAD model is calculated as the error value of that point, and the error cloud map is generated based on the average error of all points.

[0013] Preferably, the step of downsampling the 3D reconstructed model and the standard CAD model based on keypoint matching and feature descriptors to obtain a set of keypoint matching pairs, and using singular value decomposition to solve the spatial variation matrix to achieve coarse registration, includes the following steps: Key points were extracted from the 3D reconstructed model and the standard CAD model based on the ISS algorithm. Build an enhanced feature descriptor for each keypoint; Calculate the Euclidean distance between the keypoint augmentation feature descriptor of the 3D reconstruction model and the keypoint augmentation feature descriptor of the standard CAD model to filter out matching point pairs and form the keypoint matching pair set; Based on the set of keypoint matching pairs, the spatial transformation matrix containing rotation, translation, and scaling is solved using the singular value decomposition method.

[0014] Preferably, the objective function of the variable-scale iterative nearest-point algorithm improved by the robust loss function based on the Geman-McClure kernel function is specifically: ; in, Let be the objective function. For the set of keypoint matching pairs, For scale parameters, For rotation matrix, It is a translation vector. Dynamic robust weights are used for each pair of matched points. Assign an independent, optimizable weight. is the weight regularization coefficient.

[0015] Preferably, the step of using the set of keypoint matching pairs as the initial value and employing a variable-scale iterative nearest-point algorithm improved with a robust loss function based on the Geman-McClure kernel function for iterative optimization until the objective function converges, in order to achieve precise registration, specifically: With the objective function as the goal, a variable-scale iterative nearest-point algorithm based on a robust loss function with the Geman-McClure kernel function is used to alternately optimize the scale parameter, the rotation matrix, the translation vector, and the set of weights to be optimized during the iteration process until the objective function converges, so as to achieve precise registration.

[0016] Preferably, the step of detecting whether the deformation of the target industrial part is qualified based on the error cloud map and outputting the detection result includes the following steps: Calculate the error value at each point in the error cloud map and generate an error histogram; The maximum error value in the error histogram is compared with a preset tolerance threshold. If the maximum error value is less than the preset tolerance threshold, the deformation of the target industrial part is defined as qualified and output as the test result; If the maximum error value is greater than the preset tolerance threshold, the deformation of the target industrial part is defined as unqualified and output as the test result.

[0017] A 3D reconstruction and deformation detection system for industrial parts includes: The sparse point cloud module is used to acquire and process multi-view RGB images of target industrial parts, and then generate sparse point clouds using the motion structure restoration method. A 3D reconstruction model building module is used to train the sparse point cloud according to the 3D Gaussian sputtering method to build a 3D reconstruction model of the target industrial part. The error cloud map acquisition module is used to perform point cloud registration between the three-dimensional reconstruction model and the standard CAD model to obtain the error cloud map. The deformation detection module is used to detect whether the deformation of the target industrial part is qualified based on the error cloud map, and output the detection result.

[0018] This invention discloses a method for 3D reconstruction and deformation detection of industrial parts. The method involves acquiring multi-view RGB images of the target industrial part using an image acquisition tool, processing them, and then generating a sparse point cloud using a motion structure recovery method. The sparse point cloud is then trained using an optimized 3D Gaussian sputtering method to construct a 3D reconstruction model of the target industrial part. The CAD model of the target industrial part is then read, and the 3D reconstruction model is registered with the standard CAD model to obtain an error cloud map. Based on this error cloud map, the deformation of the target industrial part is detected to determine if it is within acceptable limits, and the detection result is output.

[0019] This invention achieves high-precision and high-efficiency 3D reconstruction of industrial parts by optimizing the 3DGS training strategy and loss function; it achieves high-precision point cloud alignment under scale and geometric differences by proposing a multi-neighborhood feature descriptor and an improved robust Scale-ICP registration algorithm; and finally, it achieves efficient deformation detection through a fully automated error analysis and judgment process. This method requires only a monocular camera, is low-cost, and flexible in deployment. It not only pursues and achieves sub-millimeter accuracy requirements for industrial inspection but also significantly improves efficiency, cost, and automation in terms of practicality.

[0020] This invention also provides a three-dimensional reconstruction and deformation detection system for industrial parts. Since it shares the same technical concept and solves the same technical problem as this method, it should have the same beneficial effects, and will not be described in detail here. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart of a three-dimensional reconstruction and deformation detection method for industrial parts provided in an embodiment of the present invention; Figure 2 This is a flowchart of step S1 provided in an embodiment of the present invention. Figure 3 This is a flowchart of step S2 provided in an embodiment of the present invention; Figure 4 This is a flowchart of step S3 provided in an embodiment of the present invention; Figure 5 This is a flowchart of step C2 provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of a three-dimensional reconstruction and deformation detection system for industrial parts provided in an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] The embodiments of this invention are written in a progressive manner.

[0025] This invention provides a method and system for 3D reconstruction and deformation detection of industrial parts. It primarily addresses the technical problems in existing technologies, such as poor integrity and severe edge geometric distortion in the 3D reconstruction of highly reflective, weakly textured industrial parts, as well as insufficient point cloud registration accuracy and robustness under noise, initial misregistration, and scale variations.

[0026] like Figure 1 As shown, a method for 3D reconstruction and deformation detection of industrial parts includes the following steps: S1. After acquiring and processing multi-view RGB images of the target industrial part, a sparse point cloud is generated using the motion structure restoration method; S2. The sparse point cloud is trained using the three-dimensional Gaussian sputtering method to construct a three-dimensional reconstruction model of the target industrial part; S3. Register the 3D reconstructed model with the standard CAD model using point clouds to obtain an error cloud map; S4. Detect whether the deformation of the target industrial part is qualified based on the error cloud map, and output the detection results; The three-dimensional Gaussian sputtering method includes the following steps: During the iterative training process, a progressive training strategy is adopted, using downsampled images for training in the early stage of training, and gradually using higher resolution images for training as the number of iterations increases; An edge tangential densification strategy is adopted to determine the semantic region and the background region. For Gaussian points located between the semantic region and the background region, densification operation is performed along the tangent direction of the surface where they are located, so as to control the distribution of the Gaussian points in the semantic region or the background region. The model is optimized using a loss function that includes image reconstruction loss, Gaussian flattening regularization, local opacity regularization, and geometric consistency regularization.

[0027] Steps S1 to S4 detail the specific implementation of the 3D reconstruction and deformation detection method for industrial parts. This involves acquiring multi-view RGB images of the target industrial part using an image acquisition tool, processing them, and then generating a sparse point cloud using a motion structure recovery method. The sparse point cloud is then trained using an optimized 3D Gaussian sputtering method to construct a 3D reconstruction model of the target industrial part. The CAD model of the target industrial part is then read, and the 3D reconstruction model is registered with the standard CAD model to obtain an error cloud map. Based on this error cloud map, the deformation of the target industrial part is detected to determine if it is within acceptable limits, and the detection result is output.

[0028] Preferably, the optimized three-dimensional Gaussian sputtering method includes: During iterative training, a progressive training strategy is adopted. In the early stage of training, downsampled images are used for training, and higher resolution images are gradually used for training as the number of iterations increases. An edge tangential densification strategy is adopted. The semantic region and background region are determined based on the mask image. For Gaussian points located between the semantic region and the background region, densification operation is performed along the tangent direction of the surface where they are located to control the distribution of Gaussian points in the semantic region or the background region. The model is optimized using a loss function that includes image reconstruction loss, Gaussian flattening regularization, local opacity regularization, and geometric consistency regularization.

[0029] In practical applications, based on a progressive training strategy, the "coarse-to-fine" phased training solves the problems of high computational load and slow convergence speed caused by direct training with high-resolution images. Specifically, the technical effects are as follows: Accelerated initial convergence: Using downsampled images in the early stages of training allows the model to quickly capture the overall structure and coarse geometry of the scene, avoiding ineffective exploration of detailed textures; Optimized computational resource allocation: Introducing high-resolution images for fine-tuning in the later stages concentrates computational resources on improving surface details and edge sharpness, avoiding the huge overhead of training at high resolution throughout; Improved overall training efficiency: Without sacrificing final reconstruction accuracy, the total training time is significantly shortened, achieving the efficiency goal of "minute-level" reconstruction. Based on an edge tangential densification strategy, this approach directly addresses the core flaws of 3DGS, such as blurring and artifacts (density leakage) at object edges. Through geometrically guided densification operations, it achieves precise spatial control of Gaussian point distribution. Specifically, it offers the following technical benefits: Improved edge geometry quality: Densification along the surface tangent direction allows Gaussian points to more closely conform to the object's boundary contour, resulting in sharp, jagged edge reconstruction; Elimination of background interference: Forcing Gaussian points to explicitly belong to either the "object" or "background" region effectively prevents them from spreading beyond the mask into the background area, resolving the "floating artifact" problem and making the reconstructed model "cleaner"; Enhanced prior consistency with the mask: Transforming the strong semantic prior provided by image segmentation (mask) into a hard constraint on the 3D geometric distribution improves the consistency between the reconstruction result and the input prior, which is crucial for objects with well-defined structures, such as industrial parts. Based on a composite loss function model optimization strategy, four loss terms with different supervision objectives are organically combined to construct a multi-task, multi-constraint optimization objective. This objective drives the model to generate 3D models that are both visually realistic and geometrically accurate from different dimensions. Specifically: Image reconstruction loss ensures that the rendered image matches the real photograph in color and texture, providing a fundamental guarantee for a realistic appearance; Gaussian flattening regularization forces the Gaussian function to flatten to fit the surface, improving surface smoothness and geometric rationality, and avoiding "bloated" or "hollow" geometry; Local opacity regularization removes invalid high-density points inside the object and in the background space, compressing the model volume and improving geometric purity, directly contributing to the "small memory" objective; Geometric consistency regularization introduces multi-view geometric constraints, providing reliable geometric supervision in areas where color information is lost, such as weak textures, high reflectivity, and thin walls, significantly improving the accuracy of depth estimation and 3D structure, and is one of the most critical regularization terms for achieving sub-millimeter-level industrial inspection accuracy.

[0030] The above-described scheme is an improvement on 3DGS. Compared to traditional 3DGS methods, it reconstructs 3D models with sharp edges, smooth surfaces, geometric accuracy, and compact memory from multi-view images within a shorter training time. This is particularly suitable for the stringent requirements of industrial parts (characterized by high reflectivity, weak texture, thin walls, and sharp edges) for reconstruction quality, laying an irreplaceable and high-quality digital model foundation for subsequent high-precision deformation detection.

[0031] like Figure 2 As shown, preferably, after acquiring and processing multi-view RGB images of the target industrial part, a sparse point cloud is generated using the motion structure reconstruction method, including the following steps: A1. Based on a monocular camera, multi-view RGB images are captured around the target industrial part, and the multi-view includes at least the view of the area to be reconstructed; A2. Segment the multi-view RGB image based on the preset image segmentation model, remove the background region, and obtain the corresponding mask image; A3. Based on masked images and motion structure reconstruction, generate sparse point clouds of target industrial parts and their corresponding camera poses.

[0032] Steps A1 to A3 are the specific implementation details of step S1. A monocular camera is selected as the image acquisition tool. The monocular camera captures multi-view RGB images around the target industrial part, where at least one view includes the area to be reconstructed. The multi-view RGB images are segmented using a preset image segmentation model to remove the background area in each RGB image, resulting in a corresponding mask image. Based on the mask image, the motion structure restoration method is used to generate a sparse point cloud of the target industrial part and the corresponding camera pose.

[0033] In one embodiment, a titanium alloy turbine blade of a certain type of aero-engine is used as the detection object to illustrate the three-dimensional reconstruction and deformation detection method for industrial parts of the present invention. The specific implementation process of steps A1 to A3 is as follows.

[0034] Sixty-four multi-view RGB images with a resolution of 1600×1200 were captured around the turbine blades using a monocular camera. Unsupervised segmentation of the images was performed using the SAM2 model to generate masks to remove irrelevant backgrounds. Then, Colmap software was used to perform motion reconstruction on the masked image sequence to generate sparse point clouds and estimate camera pose. SAM2 is a zero-shot general segmentation model based on the Transformer architecture that supports multimodal (image and video) cues. It can automatically identify and segment any target object in an image or video based on simple user-provided interactive cues (such as points, boxes, coarse masks, or text descriptions) without requiring specialized training for specific tasks or object categories. Colmap is a software system that automatically reconstructs the sparse 3D point cloud structure and camera pose (position and orientation) parameters of a scene from an unordered or ordered multi-view 2D image sequence, and can further generate dense 3D point clouds (mesh). Essentially, it is a complete toolbox for converting a series of 2D photographs into 3D digital models. Structure of motion reconstruction (SRM) is a core technique in computer vision and photogrammetry, designed to automatically and simultaneously reconstruct the geometry (sparse 3D point cloud) of a 3D scene from a set of visually overlapping 2D images and estimate the motion trajectory (position and pose) of the camera that captured each image in 3D space.

[0035] like Figure 3 As shown, preferably, training a sparse point cloud using a three-dimensional Gaussian sputtering method to construct a three-dimensional reconstruction model of the target industrial part includes the following steps: B1. Transform each 3D point in the sparse point cloud into a 3D Gaussian function; B2. Based on the optimized 3D Gaussian sputtering method, the following sub-steps are executed repeatedly within a preset number of iterations: B21. Based on the camera pose, the three-dimensional Gaussian function is rendered into a predicted image at the current viewpoint through parallel rasterization; B22. Calculate the image reconstruction loss between the predicted image and the corresponding real image, and construct the total loss function by combining at least one of the regularization terms among Gaussian flattening regularization, local opacity regularization and geometric consistency regularization. B23. Optimize the geometric parameters of a three-dimensional Gaussian function through backpropagation based on the total loss function; B3. Until the iteration is completed or the model convergence condition is met, output the optimized three-dimensional Gaussian function set as the three-dimensional reconstruction model of the target industrial part.

[0036] Steps B1 to B3 are the specific implementation details of step S2. They involve converting each 3D point in the sparse point cloud into a 3D Gaussian function, and then iteratively executing the optimized 3D Gaussian sputtering method within a preset number of iterations. Based on the camera pose, the 3D Gaussian function is rendered into a predicted image at the current viewpoint through parallel rasterization. The image reconstruction loss between the predicted image and the corresponding real image is calculated. At least one regularization term among Gaussian flattening regularization, local opacity regularization, and geometric consistency regularization is combined to construct a total loss function. Based on this total loss function, the geometric function of the 3D Gaussian function is optimized through backpropagation. The updated 3D Gaussian function is then used as the 3D Gaussian function in step B21, and steps B21 to B23 are executed again until the iteration is completed or the model convergence condition is met. The optimized set of 3D Gaussian functions is then output and used as the 3D reconstruction model of the target industrial part.

[0037] In one embodiment, the sparse point cloud is initialized with a three-dimensional Gaussian function. The maximum number of iterations is set to 30,000, and a progressive training strategy is adopted: the first 5,000 iterations use a 1 / 4 resolution image, the middle 15,000 iterations use a 1 / 2 resolution image, and the last 10,000 iterations use a full resolution image.

[0038] During training, an edge tangential densification strategy is implemented. For Gaussian points located at the edge of the mask, they are split along their local tangential direction to ensure that the Gaussian points are strictly distributed in the blade region or background region, thus avoiding edge blurring.

[0039] The edge tangential densification strategy strictly controls the distribution of Gaussian points within two regions: those with semantic meaning (object) and those without semantic meaning (background). The specific formula is as follows: ; ; ; in, It is twice the number of tangent directions contained within a Gaussian point. For the index of the split Gaussian copy, The tangent direction vector, scale and central position Inherited from his father Gauss and Edge densification optimization binarizes the unstable Gaussian point distribution after initialization, ensuring that the Gaussian points are either located within the object region. It is either in the background area. .

[0040] The total loss function combines image reconstruction loss with at least one regularization term, as shown in the following formula: Image reconstruction loss: ; Gaussian flattening regularization term: ; Local opacity regularization term: ; Geometric consistency regularization term: ; In the formula, This indicates the total number of pixels involved in the calculation in the current view. and This represents the color intensity value at pixel P. The flattening regularization coefficient is... It is the scale factor along the three axes of the ellipsoid, in the flattening regularization coefficient. Under constraints, the Gaussian ellipsoid is compressed along the direction of the minimum scale factor, thereby flattening it to a plane that most closely approximates its original shape. Opacity The visibility of the Gaussian ellipsoid is determined by dividing image pixels into target pixels based on a mask. and Two regions, with varying opacity regularization coefficients. and Under the constraints of the background area Apply a strong penalty term to reduce the redundancy of the Gaussian distribution. Let Gaussian total number, To in the background area Gaussian number within, Let the center coordinates of the Gaussian function be the coordinates of the target region. Apply a weak penalty term to ensure a smooth transition and reduce overall model memory consumption. and These are the rasterization-based rendering depth of pixel P projected onto the corresponding 3D point in the camera coordinate system and the backtracking depth calculated from the clustering normal vector, respectively.

[0041] By optimizing the Gaussian parameters through backpropagation, a high-fidelity 3D reconstruction model of the turbine blade was finally obtained.

[0042] Among them, the optimized 3D Gaussian sputtering method refers to a technical solution that has made a series of key improvements to the original 3D Gaussian Splatting (3DGS) method for the specific task of high-precision 3D reconstruction of industrial parts; parallelizable rasterization rendering refers to the process of representing a scene as thousands of 3D Gaussian functions with spatial properties within the 3D Gaussian sputtering framework, and efficiently projecting, sorting, and blending these 3D primitives onto 2D image pixels through a highly optimized algorithm that can be fully parallelized on a GPU. It is a key technology for 3DGS to achieve real-time or near-real-time high-fidelity new perspective synthesis; image reconstruction loss refers to a scalar loss function used to measure the difference between the synthesized image generated by the model and the corresponding real-world image in learning-based rendering or reconstruction tasks. It is the core supervisory signal driving model parameter optimization and making its output approximate real observation data; the Gaussian flattening regularization term is a geometric constraint term introduced when optimizing the 3D Gaussian sputtering model. It encourages the Gaussian function to flatten within the local surface tangent plane by penalizing excessive stretching of the 3D Gaussian function in its normal direction, thereby guiding the Gaussian function to better fit the object surface and suppressing its diffusion into or out of space, thus improving the accuracy of the reconstructed geometry and the smoothness of the surface. The local opacity regularization term is a spatial constraint introduced to address a specific defect in the 3D Gaussian sputtering method. It suppresses "floating artifacts" or "density leakage" by penalizing the high opacity of the 3D Gaussian function located outside the real surface of the object (such as internal voids or background space), thereby purifying the reconstructed space and forcing high-density, high-opaque Gaussian functions to cluster on the real object surface, improving the geometric purity and spatial compactness of the reconstructed model. The geometric consistency regularization term is a multi-view geometric constraint introduced when optimizing the 3D Gaussian sputtering model. It compensates for the shortcomings of image color supervision by forcing the disparity / depth information rendered by the model to be consistent with the geometric cues calculated from multi-view images using classical geometric methods (such as stereo matching). This significantly improves the geometric accuracy and 3D structural rationality of the reconstruction model in textureless regions, thin-walled structures, and depth discontinuities. Backpropagation optimization refers to the core process in neural network-based model training where the gradient of the loss function with respect to all model parameters is calculated using the "backpropagation algorithm," and these parameters are iteratively updated using optimization algorithms (such as SGD and Adam) to gradually minimize the loss function. It is the mathematical engine that enables modern deep learning models to "learn" from data.

[0043] like Figure 4 As shown, preferably, the point cloud is registered between the 3D reconstructed model and the standard CAD model to obtain an error cloud map, including the following steps: C1. Calculate the geometric center of the 3D reconstructed model and the standard CAD model, and perform a translation transformation to align the geometric center of the 3D reconstructed model with the geometric center of the standard CAD model; C2. Based on key point matching and feature descriptors, downsampling is performed on the 3D reconstruction model and the standard CAD model to obtain a set of key point matching pairs. The singular value decomposition method is used to solve the spatial variation matrix to achieve coarse registration. C3. Using the set of keypoint matching pairs as the initial value, the variable-scale iterative nearest-point algorithm based on the robust loss function of the Geman-McClure kernel function is used for iterative optimization until the objective function converges, so as to achieve fine registration; C4. After completing the fine registration, calculate the shortest distance from each point in the 3D reconstructed model to the surface of the standard CAD model, and use it as the error value of that point. Generate an error cloud map based on the average error of all points.

[0044] Steps C1 to C4 are the specific implementation details of step S3. They involve calculating the geometric centers of the 3D reconstructed model and the standard CAD model, aligning their geometric centers through translation transformation, and then downsampling based on keypoint matching and feature descriptors to obtain a set of keypoint matching pairs. The spatial transformation matrix is ​​solved using singular value decomposition to achieve coarse registration. Using the set of keypoint matching pairs as initial values, a robust loss function of the Geman-McClure kernel function is introduced, and an improved variable-scale iterative nearest-point algorithm is used for iterative optimization until the objective function converges to achieve fine registration. After fine registration, the error value of each point in the 3D reconstructed model is calculated, and an error cloud map is generated based on the average error of all points.

[0045] In practical application, based on step C1, a simple translation transformation is used to make the centroids of the two models coincide, eliminating any large positional offset that may exist between the reconstructed model and the CAD model. This provides a physically close initial state for the subsequent feature matching algorithm, preventing the subsequent optimization algorithm from getting stuck in an incorrect local optimum due to an excessively poor initial position. Based on step C2, on the downsampled point cloud, a reliable sparse point pair correspondence is established using key points and feature descriptors, and an initial transformation matrix containing rotation, translation, and scaling is solved. This solves the problem of any arbitrary rotation and overall scale difference that may exist between the two models. The directly estimated scale parameters and the obtained near-true alignment state transformation matrix provide an excellent optimization starting point for fine registration (C3), ensuring that it quickly converges to the global optimum. Based on the high-quality initial values ​​provided by coarse registration in step C3, a variable-scale ICP algorithm with a robust loss function incorporating the Geman-McClure kernel function is used for iterative optimization until convergence. Due to the actual deformation of the part or the local error of the reconstructed model, a large number of "mismatched" point pairs may occur. The robust loss function automatically reduces the weight of these outlier point pairs, preventing them from "skewing" the registration results. After coarse registration has resolved large-scale biases, fine registration focuses on fine-tuning and alignment details. The improved algorithm ensures stable convergence even with local deformations, without oscillations or divergence, achieving sub-millimeter level or even higher alignment accuracy. It is insensitive to reconstruction noise, local missing parts, and real deformations, ensuring the reliability of the detection system. Based on the registration in step C4, an error cloud map is generated by calculating the directed distance from each point in the reconstructed model to the surface of the CAD model, mapping the error magnitude with color. Global statistics, including maximum error, average error, median error, and failure rate, can be calculated (where the failure rate is a key indicator for measuring the overall accuracy and pass rate of the reconstructed model, and its formula is:). ,in, This refers to the set of fault points where the error exceeds a preset threshold (e.g., 0.2 mm). The total set of points; the failure rate is the ratio between the number of fault points and the total set of points. The error cloud map can clearly and accurately locate the location, range and severity of deformation (e.g., red indicates a large error). The failure rate and the maximum error value together constitute the core quantitative basis for automated deformation qualification judgment, providing objective and comprehensive data support for quality assessment, process improvement and tolerance analysis.

[0046] like Figure 5 As shown, preferably, based on keypoint matching and feature descriptors, downsampling is performed on the 3D reconstructed model and the standard CAD model to obtain a set of keypoint matching pairs, and the spatial variation matrix is ​​solved using the singular value decomposition method to achieve coarse registration, including the following steps: D1. Extract key points from the 3D reconstructed model and the standard CAD model based on the ISS algorithm; D2. Construct an enhanced feature descriptor for each keypoint; D3. Calculate the Euclidean distance between the keypoint augmentation feature descriptors of the 3D reconstruction model and the keypoint augmentation feature descriptors of the standard CAD model to filter out matching point pairs and form a set of keypoint matching pairs; D4. Based on the set of keypoint matching pairs, use the singular value decomposition method to solve the spatial transformation matrix containing rotation, translation and scaling.

[0047] Steps D1 to D4 are the specific implementation details of step C2. They involve extracting key points from the 3D reconstructed model and the standard CAD model using the ISS algorithm, constructing an enhanced feature descriptor for each key point, calculating the Euclidean distance between the enhanced key point descriptors of the 3D reconstructed model and the enhanced key point descriptors of the standard CAD model to filter out matching point pairs, forming a set of key point matching pairs, and using the singular value decomposition method to solve the spatial transformation matrix that includes rotation, translation, and scaling.

[0048] In practical applications, based on step D1, the ISS algorithm automatically selects feature points with significant geometric changes (high curvature, corners, etc.) at multiple scales, avoiding the huge computational overhead of dense matching on the entire point cloud, achieving efficient downsampling. The extracted key points have strong invariance to viewpoint changes, noise, and local deformation, improving the repeatability of subsequent matching. Flat and featureless regions are filtered out, concentrating computational resources on geometrically significant regions that substantially contribute to registration. Based on step D2, an enhanced feature descriptor with greater discriminative power and robustness than traditional descriptors is constructed for each key point. By fusing geometric information under multiple neighborhood radii, the descriptor simultaneously possesses local detail resolution and regional context awareness, avoiding sensitive dependence on neighborhood radii. Based on traditional covariance features (describing the shape of point distribution), principal curvature information is fused, enhancing the descriptor's ability to characterize surface curvature direction and subtle geometric differences. In step D3, the Euclidean distance of the enhanced feature descriptors in high-dimensional space is used to implement a strict binary decision to filter matching point pairs. Only point pairs with highly similar descriptors (distance less than a threshold) are accepted, which greatly improves the accuracy of matching pairs and reduces erroneous correspondences (outliers) from the source. Based on the high-quality matching point pair set obtained in step D4, the optimal rigid body transformation parameters are analytically solved using the SVD method. Given the corresponding point set, SVD provides the optimal rotation, translation, and scaling solutions in the least squares sense. It is mathematically rigorous and reliable. The solved transformation matrix can align the two models to a very close state in terms of rotation, translation, and scaling, providing near-ideal initial values ​​for the fine registration algorithm that relies on iterative optimization, ensuring its fast and stable convergence.

[0049] Preferably, the objective function of the variable-scale iterative closest point algorithm improved by the robust loss function based on the Geman-McClure kernel function is as follows: ; in, Let be the objective function. For the set of keypoint matching pairs, For scale parameters, For rotation matrix, It is a translation vector. Dynamic robust weights are used for each pair of matched points. Assign an independent, optimizable weight. is the weight regularization coefficient.

[0050] In practical applications, the objective function aims to simultaneously solve for the optimal spatial transformation (scaling). Rotation Translation ) and the credibility weight of the matching point pair This method aims to robustly and accurately align 3D reconstructed models with CAD models even in the presence of numerous mismatches (outliers) and scale variations. The overall effect is to transform the registration problem into a joint optimization problem, using automated weight adjustment to mitigate interference. This is the core error term, used to calculate the current transform parameters. Distance error of the next matching point pair; by fixing , , ,right about By taking the derivative and setting it to zero, the optimal weights can be derived. Closed solution: ; During the optimization process, the optimizer tends to assign a preference to point pairs with small errors and high confidence. To ensure that its contribution is fully preserved, the optimizer tends to assign more weight to point pairs with large errors that may be outliers. This allows for the automatic and dynamic removal of its contribution from the objective function; This is a weight regularization term used to penalize excessively small values. ,when When the value is close to 1, the regularization term is close to 0. When the value is much less than 1 or greater than 1, the value of the regularization term increases, thus increasing the total error. This means that the optimization process cannot simply cheat by setting all weights to zero, but must find a balance between reducing the error term and keeping the weights reasonable. The regularization coefficient controls the constraint strength, determining the level of error at which point pairs begin to be weighted less. As the algorithm iterates, the coefficient gradually decreases. This means that the optimization process will become increasingly "picky," trusting only those perfectly aligned point pairs, thereby gradually filtering out incorrect matches.

[0051] Preferably, using the keypoint matching pair set as the initial value, a variable-scale iterative nearest-point algorithm based on a robust loss function improved by the Geman-McClure kernel function is used for iterative optimization until the objective function converges, in order to achieve precise registration. Specifically: With the objective function as the goal, a variable-scale iterative nearest-point algorithm based on the robust loss function of the Geman-McClure kernel function is adopted to alternately optimize the scale parameter, rotation matrix, translation vector and the set of weights to be optimized during the iteration process until the objective function converges, so as to achieve precise registration.

[0052] The above scheme is the specific implementation detail of step C3, which will jointly optimize the transformation parameters. and weight parameters This highly coupled nonlinear problem is decomposed into two relatively simple subproblems, which are solved alternately in the iterations. Since directly performing global joint optimization on all weights and transformation parameters easily leads to local optima or difficulty in convergence, alternating optimization is chosen to transform it into a sequential convex approximation process. This allows for fixing one set of variables and optimizing the other set in each iteration, reducing the difficulty and computational complexity of the problem. Each sub-optimization is performed under more controllable conditions, making the entire iterative process clear in direction, controllable in step size, and more stable in convergence behavior.

[0053] The fine registration strategy provided by the above scheme ensures that the fine registration algorithm can automatically and robustly find the globally optimal scale and spatial transformation parameters, starting from the initial values ​​provided by the coarse registration, even in the presence of numerous uncertainties and local disturbances (mismatches, reconstruction noise, and actual deformation). This dynamic interaction between weight adjustment and transformation optimization improves the system's accuracy, robustness, and automation level.

[0054] In one embodiment, the reconstructed model is registered with a standard CAD model of the turbine blade: Geometric center alignment: Calculate the geometric centers of the two models and translate them.

[0055] Coarse registration: Key points are extracted using the ISS algorithm, a multi-neighborhood feature descriptor (FDN) is constructed, matching point pairs are filtered by Euclidean distance threshold, and the initial transformation matrix is ​​solved using SVD.

[0056] Among them, the ISS (Intrinsic Shape Signatures) algorithm is a key point (feature point) detection algorithm based on the local geometric features of 3D point clouds. It automatically identifies points with significant and stable geometric structures at multiple spatial scales, such as corner points and edge points, by analyzing the distribution characteristics of each point in the point cloud within its neighborhood, and uses these points as "key points" representing the features of the entire model. The multi-neighborhood feature descriptor is an enhancement method for describing the local features of 3D point clouds. For a given 3D point, instead of using a single neighborhood radius, it selects multiple neighborhood radii of different sizes, calculates the local geometric features at each radius, and fuses these feature vectors at multiple scales (e.g., concatenation, weighting, or pooling) to form a more comprehensive, robust, and discriminative high-dimensional feature descriptor. Singular Value Decomposition (SVD) is a mathematical method that uses the linear algebraic tool SVD to analytically solve for the optimal rigid body transformations (rotation, translation) and scale parameters between 3D point clouds. It is used to solve the following problem: given two sets of 3D points with a one-to-one correspondence (i.e., matching point pairs), find a transformation that includes scaling, rotation and translation, such that the sum of squared distances between the two sets of points is minimized after the transformation.

[0057] The specific process for constructing a multi-domain feature descriptor (FDN) is as follows: A comprehensive multi-neighborhood local feature descriptor (FDN) was constructed. For single-neighborhood local features (n=1), the descriptor uses the following FD formula to extract the three eigenvalues ​​of the point cloud covariance matrix constructed by the ISS. , and The inherent linear, planar, and spherical shape information is fused with the principal curvature fusion characteristics of that point. These features are jointly encoded into a more discriminative feature vector. This fusion strategy significantly improves the descriptor's sensitivity to subtle local geometric differences, as shown in the FD formula below: ; During the matching phase, a multi-neighborhood local feature descriptor (FDN) is employed to further enhance the feature description capability of each key point by adjusting the neighborhood radius. The FDN formula is obtained by integrating FDs of different scales and normalizing them: ; Computational source point cloud a certain point in the middle With target point cloud a certain point in the middle Enhanced feature descriptors and Euclidean distance between As shown in the Euclidean distance formula, only when this distance... Less than the preset distance threshold Only then is it considered that the point is correct. A reliable candidate correspondence is constructed and preserved. This screening mechanism, based on multiple local features and feature space consistency, effectively filters out erroneous matches caused by excessively high descriptor similarity, thus providing a higher-quality and more reliable initial set of correspondence points for subsequent accurate registration. The Euclidean distance formula is as follows:

[0058] Fine registration: The objective function of the improved Scale-ICP algorithm is iteratively optimized, and the transformation parameters (scaling) are updated alternately. Rotation Translation ) and the credibility weight of the matching point pair Until it converges.

[0059] The obtained feature matching set Assign a set of weights to be optimized { The objective function can be rewritten as follows: ; pass right Taking the partial derivative, we get: ; As the optimization iterations proceed, the distance coefficient continuously decreases. The sensitivity to outliers can be dynamically adjusted, making the robust kernel function increasingly sensitive to them, gradually eliminating incorrect matches and guiding the optimization process toward the correct solution. Introducing linear process variables transforms the original non-convex problem into a grouping of alternative optimization variables. In this form, each step solves for the optimal solution of the current group variable, ensuring that the residuals in the Gauss-Newton method are perfectly linear with respect to a single variable, thus guaranteeing that the objective function value decreases at each step. Therefore, we have... That is, the objective function is monotonically decreasing and has a lower bound, thus ensuring convergence.

[0060] Generate error cloud map: Calculate the shortest distance from each point in the reconstructed model to the surface of the CAD model, and generate an error cloud map after coloring.

[0061] Preferably, the detection of whether the deformation of the target industrial part is qualified based on the error cloud map and the output of the detection result include the following steps: Calculate the error value at each point in the error cloud map and generate an error histogram; The maximum error value in the error histogram is compared with a preset tolerance threshold. If the maximum error value is less than the preset tolerance threshold, the deformation of the target industrial part is defined as qualified and output as the test result; If the maximum error value is greater than the preset tolerance threshold, the deformation of the target industrial part is defined as unqualified and output as the test result.

[0062] The above scheme is a specific implementation detail of step S4. It calculates the error value of each point in the error cloud map, generates an error histogram, selects the maximum error value (instead of the average error or median) as the judgment criterion, and compares it with a preset and clear tolerance threshold. Based on the logical comparison in the previous step, the system automatically gives a clear "qualified" or "unqualified" conclusion as the final output.

[0063] The above solution translates advanced, high-precision 3D deformation analysis into a stable, reliable, and directly applicable automated tool for production quality control. It hides the complexity of the underlying algorithm, presenting users with a simple logic similar to a "go / no-go gauge," but behind it lies sub-millimeter-level full-field deformation measurement capability.

[0064] In one embodiment, the maximum error value (0.118 mm in this example) is extracted from the error cloud map and compared with a preset tolerance threshold (e.g., 0.2 mm). Since the maximum error is less than the threshold, the system automatically determines that the turbine blade deformation is acceptable.

[0065] To verify the effectiveness of this invention, comparative experiments were conducted on DTU, Stanford, and self-made industrial datasets. In the table, 2D Gaussian Sputtering (2DGS) is a method for reconstructing geometrically accurate radiation fields by representing a scene as a series of two-dimensional oriented Gaussian disks, achieving improved reconstruction quality compared to 3DGS. Planar Gaussian Sputtering Reconstruction (PGSR) is a technique that focuses on achieving high-fidelity surface reconstruction from multi-view images by introducing unbiased depth rendering and multi-view geometric regularization. FPFH (Fast Point Feature Histogram) is a fast algorithm for describing the local geometric features of 3D point clouds, generating feature histograms by analyzing the spatial distribution of the neighborhood around a point. 3DSC (3D Shape Context) is a method for forming feature descriptors in 3D point clouds by statistically analyzing the spatial distribution of points within a spherical neighborhood, exhibiting good robustness to noise. Intrinsic Shape Signature (ISS) establishes local geometric feature descriptors by analyzing the covariance matrix and eigenvalues ​​of the spherical neighborhood around a point. Scale-Invariant Feature Transform (SIFT) is a classic algorithm for matching key points. The feature descriptors it generates are highly invariant to changes in image scale, rotation, brightness, etc.

[0066] 3D reconstruction results: As shown in Table 1, the method of this invention achieved the lowest average chamfer distance (Mean CD=0.52mm), the shortest training time (8.2 minutes), and the smallest model memory usage (56MB) on the DTU dataset, and its overall performance is better than methods such as 2DGS and PGSR.

[0067] Table 1. Comparison of evaluation metrics between the method of this invention and other 3DGS methods

[0068] Point cloud registration results: As shown in Tables 2 and 3, the registration method of this invention has the lowest RMSE error and the shortest time on Stanford Bunny, Dragon and other models and DTU scanning scenarios.

[0069] Table 2. Comparison of registration quality between the method of this invention and other point cloud registration methods.

[0070] Table 3. Comparison of registration efficiency between the method of this invention and other point cloud registration methods.

[0071] Industrial Parts Inspection Results: When processing highly reflective surfaces, structured light suffers from several challenges. The camera can only receive extremely bright, overexposed spots along one surface light reflection path from multiple angles, while receiving almost no light at other locations. This leads to a loss of pattern information, making it impossible for the system to establish a reliable correspondence between the projector's projection point and the camera's imaging point. Furthermore, due to the weak texture of metal surfaces, structured light reconstruction relying on feature stitching struggles to find decodeable patterns. Additionally, the self-occluding structure of the blades blocks some of the structured light projection, causing reflection. These challenges result in less than ideal reconstruction of titanium alloy turbine blade models using structured light, achieving only the lowest possible completeness.

[0072] The turbine blade model reconstructed using raw 3DGS demonstrates the advantages of the Gaussian sputtering method compared to structured light reconstruction. Specifically, it effectively overcomes the challenges of reflective objects and self-occluding structures through multi-angle optical photography and motion decoupling. Furthermore, its model architecture incorporates various methods to enhance geometric information, enabling the successful decoding of smooth surfaces and faithful geometric information from weakly textured surfaces. Nevertheless, this case study still reveals issues with geometric regularization and depth estimation based on the local plane assumption. Specifically, for thin-walled structures or areas with drastic changes in edge normals, the inconsistency between geometric and depth estimations leads to more pronounced reconstruction distortion, making it difficult to obtain a complete geometric structure.

[0073] To comprehensively and quantitatively evaluate the performance of different methods in turbine blade reconstruction, this invention not only calculates traditional statistics such as mean error and median error, but also introduces failure rate as a key indicator for evaluating whether the reconstruction model meets industrial-grade testing standards. Failure rate reflects the proportion of points exceeding the allowable error range and is a direct basis for judging the reliability of batch testing. Experiments were conducted by accurately registering the CAD model and the reconstruction model to obtain a point cloud error distribution map. Points with an absolute distance exceeding 0.2 mm between corresponding points were defined as the failure point set, and the failure rate of each method was then calculated.

[0074] Based on experimental data from 3D reconstruction of turbine blades, the optimized Gaussian sputtering framework proposed in this invention demonstrates significant advantages over baseline methods and structured light. As shown in Table 4, for highly reflective, thin-walled turbine blades, the structured light reconstruction failure rate is as high as 46.73%, the original 3DGS (baseline method) failure rate is 27.91%, while the method of this invention significantly reduces the failure rate to 16.42%. This means that the model reconstructed using the method of this invention has 83.58% of its points with errors controlled within a tolerance threshold of 0.2 mm, fully meeting the accuracy and reliability requirements of sub-millimeter-level industrial inspection. This result verifies the effectiveness of the key improvements in regularization loss function and edge densification in this invention.

[0075] Table 4 Quantitative Table of Turbine Blade Reconstruction

[0076] Experimental results show that the method described in this invention achieves an average chamfer distance of 0.52 mm on the DTU dataset, requires only 8.2 minutes of training time, and occupies 56 MB of model memory. In turbine blade reconstruction, the failure rate is reduced to 16.42%, and 83.58% of the point errors are below 0.2 mm, reaching sub-millimeter level industrial inspection standards. Compared with traditional structured light and original 3DGS methods, it significantly improves reconstruction accuracy, efficiency, and robustness.

[0077] In summary, this invention not only provides a complete detection method and process, but also establishes a closed-loop quality assessment system from 3D reconstruction and high-precision registration to automated pass / fail determination by introducing the failure rate as a quantitative indicator. This method requires only a monocular camera to achieve sub-millimeter-level accuracy in 3D reconstruction and deformation detection, and outperforms existing technologies in terms of reconstruction accuracy, computational efficiency, memory usage, and failure rate control. It is particularly suitable for online inspection scenarios of industrial parts with high reflectivity, weak texture, and complex structures.

[0078] like Figure 6 As shown, a 3D reconstruction and deformation detection system for industrial parts includes: The sparse point cloud module is used to acquire and process multi-view RGB images of target industrial parts, and then generate sparse point clouds using the motion structure restoration method. The 3D reconstruction model building module is used to train sparse point clouds based on the 3D Gaussian sputtering method to build a 3D reconstruction model of the target industrial part. The error cloud map acquisition module is used to register the point cloud of the 3D reconstruction model with the standard CAD model to obtain the error cloud map. The deformation detection module is used to detect whether the deformation of the target industrial part is qualified based on the error cloud map and output the detection results.

[0079] This invention also provides a 3D reconstruction and deformation detection system for industrial parts. The system uses a sparse point cloud module to acquire and process multi-view RGB images of the target industrial part, and then generates a sparse point cloud using the motion structure recovery method. A 3D reconstruction model construction module trains the sparse point cloud using a 3D Gaussian sputtering method to construct a 3D reconstruction model of the target industrial part. An error cloud map acquisition module registers the 3D reconstruction model with a standard CAD model to obtain an error cloud map. Finally, a deformation detection module detects whether the deformation of the target industrial part is acceptable based on the error cloud map and outputs the detection results.

[0080] One or more embodiments in this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments in this application should be included within the protection scope of this application.

[0081] If a flowchart is used in this application, it is used to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, the steps can be processed in reverse order or simultaneously. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0082] The foregoing has provided a detailed description of a method and system for three-dimensional reconstruction and deformation detection of industrial parts provided in this application. The above description of the disclosed embodiments enables those skilled in the art to implement or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for three-dimensional reconstruction and deformation detection of industrial parts, characterized in that, Includes the following steps: After acquiring and processing multi-view RGB images of the target industrial part, a sparse point cloud is generated using the motion structure restoration method. The sparse point cloud is trained using a three-dimensional Gaussian sputtering method to construct a three-dimensional reconstruction model of the target industrial part. The point cloud of the 3D reconstructed model is registered with the standard CAD model to obtain an error cloud map. Based on the error cloud map, the detection results are output to determine whether the deformation of the target industrial part is qualified. The three-dimensional Gaussian sputtering method includes the following steps: During the iterative training process, a progressive training strategy is adopted, using downsampled images for training in the early stage of training, and gradually using higher resolution images for training as the number of iterations increases; An edge tangential densification strategy is adopted to determine the semantic region and the background region. For Gaussian points located between the semantic region and the background region, densification operation is performed along the tangent direction of the surface where they are located, so as to control the distribution of the Gaussian points in the semantic region or the background region. The model is optimized using a loss function that includes image reconstruction loss, Gaussian flattening regularization, local opacity regularization, and geometric consistency regularization.

2. The method for three-dimensional reconstruction and deformation detection of industrial parts as described in claim 1, characterized in that, After acquiring and processing multi-view RGB images of the target industrial part, a sparse point cloud is generated using the motion structure reconstruction method, including the following steps: Based on capturing multi-view RGB images of the target industrial part using a monocular camera, wherein the multi-view includes at least the view of the area to be reconstructed; The multi-view RGB image is segmented based on a preset image segmentation model, and the background region is removed to obtain the corresponding mask image. Based on the masked image and motion structure reconstruction method, a sparse point cloud of the target industrial part and the corresponding camera pose are generated.

3. The method for three-dimensional reconstruction and deformation detection of industrial parts as described in claim 2, characterized in that, The step of training the sparse point cloud using a three-dimensional Gaussian sputtering method to construct a three-dimensional reconstruction model of the target industrial part includes the following steps: Each three-dimensional point in the sparse point cloud is transformed into a three-dimensional Gaussian function; The optimized 3D Gaussian sputtering method iteratively executes the following sub-steps within a preset number of iterations: Based on the camera pose, the three-dimensional Gaussian function is rendered into a predicted image at the current viewpoint through parallel rasterization. Calculate the image reconstruction loss between the predicted image and the corresponding real image, and construct the total loss function by combining at least one of the Gaussian flattening regularization term, the local opacity regularization term, and the geometric consistency regularization term. The geometric parameters of the three-dimensional Gaussian function are optimized through backpropagation based on the total loss function; Until the iteration is completed or the model convergence condition is met, the optimized three-dimensional Gaussian function set is output and used as the three-dimensional reconstruction model of the target industrial part.

4. The method for three-dimensional reconstruction and deformation detection of industrial parts as described in claim 3, characterized in that, The formula for the image reconstruction loss is: ; The formula for the Gaussian flattening regularization term is: ; The formula for the local opacity regularization term is: ; The formula for the geometric consistency regularization term is: ; In the formula, This indicates the total number of pixels involved in the calculation in the current view. and Indicated in pixels The color intensity value at that location, The flattening regularization coefficient is... It is the scale factor along the three axes of the ellipsoid. The opacity of the ellipsoid. The opacity regularization coefficient for the object region. The coordinates of the center of the Gaussian function are... For the object region, This is the opacity regularization coefficient for the background region. Let Gaussian total number, To in the background area Gaussian number within, and These are the rasterization-based rendering depth of pixel P projected onto the corresponding 3D point in the camera coordinate system and the backtracking depth calculated from the clustering normal vector, respectively.

5. The method for three-dimensional reconstruction and deformation detection of industrial parts as described in claim 1, characterized in that, The step of registering the 3D reconstructed model with a standard CAD model to obtain an error cloud map includes the following steps: Calculate the geometric centers of the 3D reconstructed model and the standard CAD model, and perform a translation transformation to align the geometric centers of the 3D reconstructed model and the standard CAD model. Based on key point matching and feature descriptors, the 3D reconstruction model and the standard CAD model are downsampled to obtain a set of key point matching pairs, and the spatial variation matrix is ​​solved by singular value decomposition to achieve coarse registration. Using the set of keypoint matching pairs as the initial value, the variable-scale iterative nearest-point algorithm based on the robust loss function of the Geman-McClure kernel function is used for iterative optimization until the objective function converges, so as to achieve fine registration. After completing the fine registration, the shortest distance from each point in the 3D reconstructed model to the surface of the standard CAD model is calculated as the error value of that point, and the error cloud map is generated based on the average error of all points.

6. The method for three-dimensional reconstruction and deformation detection of industrial parts as described in claim 5, characterized in that, The process of downsampling the 3D reconstructed model and the standard CAD model based on keypoint matching and feature descriptors to obtain a set of keypoint matching pairs, and then using singular value decomposition to solve the spatial variation matrix to achieve coarse registration, includes the following steps: Key points were extracted from the 3D reconstructed model and the standard CAD model based on the ISS algorithm. Build an enhanced feature descriptor for each keypoint; Calculate the Euclidean distance between the keypoint augmentation feature descriptor of the 3D reconstruction model and the keypoint augmentation feature descriptor of the standard CAD model to filter out matching point pairs and form the keypoint matching pair set; Based on the set of keypoint matching pairs, the spatial transformation matrix containing rotation, translation, and scaling is solved using the singular value decomposition method.

7. The method for three-dimensional reconstruction and deformation detection of industrial parts as described in claim 5, characterized in that, The objective function of the variable-scale iterative nearest-point algorithm improved by the robust loss function based on the Geman-McClure kernel function is specifically: ； in, Let be the objective function. For the set of keypoint matching pairs, For scale parameters, For rotation matrix, It is a translation vector. Dynamic robust weights are used for each pair of matched points. Assign an independent, optimizable weight. is the weight regularization coefficient.

8. The method for three-dimensional reconstruction and deformation detection of industrial parts as described in claim 7, characterized in that, The process involves using the set of keypoint matching pairs as initial values, and then iteratively optimizing the algorithm using a variable-scale iterative nearest-point algorithm improved with a robust loss function based on the Geman-McClure kernel function until the objective function converges, thereby achieving precise registration. Specifically: With the objective function as the goal, a variable-scale iterative nearest-point algorithm based on a robust loss function with the Geman-McClure kernel function is used to alternately optimize the scale parameter, the rotation matrix, the translation vector, and the set of weights to be optimized during the iteration process until the objective function converges, so as to achieve precise registration.

9. The method for three-dimensional reconstruction and deformation detection of industrial parts as described in claim 1, characterized in that, The process of detecting whether the deformation of the target industrial part is qualified based on the error cloud map and outputting the detection result includes the following steps: Calculate the error value at each point in the error cloud map and generate an error histogram; The maximum error value in the error histogram is compared with a preset tolerance threshold. If the maximum error value is less than the preset tolerance threshold, the deformation of the target industrial part is defined as qualified and output as the test result; If the maximum error value is greater than the preset tolerance threshold, the deformation of the target industrial part is defined as unqualified and output as the test result.

10. A three-dimensional reconstruction and deformation detection system for industrial parts, characterized in that, include: The sparse point cloud module is used to acquire and process multi-view RGB images of target industrial parts, and then generate sparse point clouds using the motion structure restoration method. A 3D reconstruction model building module is used to train the sparse point cloud according to the 3D Gaussian sputtering method to build a 3D reconstruction model of the target industrial part. The error cloud map acquisition module is used to perform point cloud registration between the three-dimensional reconstruction model and the standard CAD model to obtain the error cloud map. The deformation detection module is used to detect whether the deformation of the target industrial part is qualified based on the error cloud map, and output the detection result.

Images