Intelligent defect detection and control point automatic repair method for industrial product curved surface design

Abstract

This invention discloses an intelligent defect detection and automatic control point repair method for curved surface design of industrial products, relating to the field of industrial design. The method includes the following steps: acquiring a sequence of physical light field images and corresponding original design model data, wherein the original design model data includes local control point tensor data of a non-uniform rational B-spline surface; and extracting features from the physical light field image sequence data to obtain the spatial coordinate range data of the initial defect features. This invention constructs a joint loss function containing a geometric fidelity loss term and a surface smoothing gain term during the iterative optimization calculation process. This ensures that while the system adjusts the control points to repair light and shadow defects, it is also strongly constrained by the original mechanical outline of the design, avoiding assembly interference or abrupt changes in structural wall thickness caused by excessive smoothing. This guarantees that the generated solution conforms to both industrial design aesthetics and meets the real physical constraints of mechanics and manufacturing.

Description

Technical Field

[0001] This invention relates to the field of industrial design, specifically to an intelligent defect detection and automatic repair method for control points in the curved surface design of industrial products. Background Technology

[0002] With the development of high-end manufacturing and computer vision technology, 3D visual inspection has gradually become an important means of evaluating the appearance quality of industrial products and avoiding the risk of duplication. Existing industrial design defect detection systems are mostly based on structured light scanning or binocular vision acquisition. By extracting macroscopic geometric indicators such as the absolute deviation of spatial coordinates between the point cloud of the object and the original 3D model, they identify and evaluate abnormal design defects such as surface fractures and assembly interference.

[0003] However, because high-gloss materials are extremely sensitive to the continuity of micro-surfaces, their slight geometric deformations are often amplified into severe light and shadow distortions under ambient light fields. Traditional systems often rely on absolute distance deviation as a judgment threshold or static comparison rule, which makes it difficult to fully reflect and quantify the changes in the normal partial derivatives and high-order smoothness of complex surfaces under actual light fields.

[0004] In this context, existing methods lack a mapping mechanism at the underlying parameter level for the connection between defect characterization and model repair, which can easily lead to a serious disconnect between detection results and the repair logic of industrial design software. For example, when the system detects a microscopic optical fracture on a surface on a prototype, because the system can only output discrete mesh deviation maps and lacks an automatic convergence mechanism, it fails to directly map and intervene in the non-uniform rational B-spline (NURBS) control points at the underlying level of the 3D CAD software. This forces designers to blindly modify feature tree nodes based on manual experience. Due to the lack of quantitative constraints in the repair process, the generated repair solutions often fail to meet the real physical limitations of industrial manufacturing. Ultimately, this results in a lengthy trial-and-error cycle for the product from design and mold making to prototyping verification, causing significant mold making losses and wasting R&D resources. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent defect detection and automatic repair method for control points in the curved surface design of industrial products.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] A method for intelligent defect detection and automatic repair of control points in the curved surface design of industrial products includes the following steps:

[0008] Acquire physical light field image sequence data and corresponding original design model data, wherein the original design model data includes local control point tensor data of non-uniform rational B-spline surface;

[0009] Feature extraction is performed on the sequence data of physical light field images to obtain the spatial coordinate range data of the initial defect features;

[0010] Based on spatial coordinate range data, the corresponding physical object normal partial derivative distribution field data is calculated from the physical object light field image sequence data;

[0011] The physical normal partial derivative distribution field data is three-dimensionally registered with the original design model data to calculate and generate smoothness residual map data.

[0012] Extract the normal gradient variance data from the smoothness residual map data, and determine whether the normal gradient variance data is greater than the preset continuity threshold data;

[0013] Based on the smoothness residual map data, iterative optimization calculations are performed on the local control point tensor data to generate control point correction parameter data that makes the normal gradient variance data converge.

[0014] Based on the control point correction parameter data, smooth reconstruction instruction data is generated to instruct the modification of local control point tensor data in the original design model data.

[0015] An intelligent defect detection and control point automatic repair system for curved surface design of industrial products includes:

[0016] Data acquisition module: used to acquire physical light field image sequence data and corresponding original design model data, wherein the original design model data includes local control point tensor data of non-uniform rational B-spline surface;

[0017] Feature extraction module: used to extract features from the sequence data of physical light field images to obtain the spatial coordinate range data of the initial defect features;

[0018] Data processing module: used to calculate the corresponding physical normal partial derivative distribution field data from the physical light field image sequence data based on spatial coordinate range data;

[0019] Data registration module: used to perform three-dimensional registration of the physical normal partial derivative distribution field data with the original design model data, and calculate and generate smoothness residual map data;

[0020] Data Judgment and Optimization Module: Used to extract the normal gradient variance data from the smoothness residual map data and determine whether the normal gradient variance data is greater than the preset continuity threshold data; if so, it performs iterative optimization calculation on the local control point tensor data based on the smoothness residual map data to generate control point correction parameter data that makes the normal gradient variance data converge.

[0021] Data output module: Used to generate smooth reconstruction instruction data based on control point correction parameter data, instructing modification of local control point tensor data in the original design model data.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0023] This invention constructs a joint loss function during the iterative optimization process, comprising a "geometric fidelity loss term" (first spatial distance deviation data) and a "surface smoothing gain term" (reconstructed normal partial derivative variance data). This ensures that while the system adjusts control points to correct lighting defects, it is also strongly constrained by the original mechanical shape and contour, avoiding assembly interference or abrupt changes in structural wall thickness caused by excessive smoothing. This guarantees that the generated solution conforms to both industrial design aesthetics and meets the real physical limitations of mechanics and manufacturing. Furthermore, it provides parametric modeling for complex robotic arm structures or precision equipment shells, making the system no longer... Instead of outputting uneditable discrete patches, it directly uses the smoothness residual map as input to the neural network convolution process. Through iterative optimization calculation, it directly outputs the control point correction parameter data for non-uniform rational B-splines (NURBS), namely the modification of spatial coordinates and rational weight values. This parameter can be directly converted into an API script to drive 3D mechanical design software to rewrite feature tree nodes, completely realizing an automated closed loop from visual perception to source file modification. This effectively shortens the lengthy trial and error cycle that products need to go through from design and mold making to prototyping verification, effectively reduces mold making losses, and saves R&D resource costs. Attached Figure Description

[0024] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0025] Figure 1 This is a diagram illustrating the method of the present invention;

[0026] Figure 2 This is a flowchart of the present invention. Detailed Implementation

[0027] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0028] Application Overview:

[0029] Traditional industrial design defect detection and repair systems often rely on macroscopic spatial distance deviations (such as absolute point cloud distances) as fixed judgment thresholds, which cannot meet the stringent requirements of microscopic curvature continuity for high-gloss material shells. When a physical surface undergoes minute nonlinear geometric distortions under actual light fields, the system cannot establish a dynamic correlation between optical distortion features and the underlying parametric geometric model, leading to a mismatch between defect saliency criteria and actual industrial design aesthetic standards (such as G2 / G3 continuity). This static comparison mechanism based on discrete grids reduces the accuracy of surface defect detection, causing the features extracted by the vision system to contain redundant noise that is difficult to guide CAD repair, ultimately affecting the timeliness and specificity of model repair instructions.

[0030] For example, in the prototyping scenario of high-gloss casings for 3C products, a 0.01mm microscopic surface fracture caused by injection molding will result in obvious light and shadow breakage under strong light. In this case, traditional systems can only screen for macroscopic contour differences through the nearest point iteration algorithm (ICP), leading to critical high-frequency optical distortion features being misjudged as normal manufacturing tolerances or measurement noise and rejected. The discrete mesh deviation report output by the system loses the key dimension characterizing the rate of change of the surface normal vector, making it impossible for designers to directly locate the non-uniform rational B-spline (NURBS) control points at the bottom layer of the 3D software for precise repair. Relying solely on blind manual adjustment is prone to causing local deformation, resulting in repair solutions that cannot meet industrial-grade smoothness requirements.

[0031] If the above problems are not addressed, the missed detection and misidentification of microscopic optical defects will lead to substandard designs flowing into the next production stage, increasing the risk of product scrap and mold opening losses. Data silos between the detection and design ends will hinder the system from capturing the critical path of visual appearance deviations and modifying the layer feature tree, delaying the optimal time for prototype iteration. In addition, conventional AI 3D repair without physical constraints will also result in the loss of geometric fidelity and mechanical strength, reducing the reliability of the model in downstream assembly verification, ultimately forming a negative feedback loop of repeated prototyping and mold modification, seriously affecting the iterative optimization of personalized industrial design solutions.

[0032] In this regard, such as Figure 1 and Figure 2 As shown, an intelligent defect detection and automatic repair method for control points in the curved surface design of industrial products includes the following steps:

[0033] Real object light field image sequence data: This contains a sequence of images of the appearance of a real object taken from multiple different angles and under different lighting phases. Unlike ordinary photographs, the light field not only records the intensity of the light but also the direction of the light. It is the core raw data that reflects the subtle undulations (light and shadow distortion) of the surface of a high-gloss material.

[0034] Original design model data: This refers to the theoretical digital source file data of the initial design by the engineer in 3D CAD software (such as SolidWorks, Rhino, etc.), representing a perfect geometric state without any manufacturing defects.

[0035] Non-uniform rational B-spline surface: The international standard model used in industrial 3D design software for the precise mathematical representation of free and complex surfaces. All smooth surfaces in advanced industrial design are calculated by NURBS formulas at the computer level.

[0036] Local control point tensor data: skeleton nodes that control the shape of NURBS surfaces. Tensors can be understood as multidimensional matrices. Here, it specifically refers to the set of coordinates and weights of control points that determine the surface undulation shape and tension of a local defect area.

[0037] Initial defect feature spatial coordinate range data: Through early image recognition networks, the specific three-dimensional bounding box location information of suspected defects (such as dents, chamfer errors) is roughly delineated in three-dimensional space.

[0038] Real-world normal partial derivative distribution field data: A normal is a line perpendicular to a surface, and its partial derivative represents the rate at which this line tilts in space, i.e., the degree of curvature or bending of the surface. This data field is a mapping map covering the entire surface of the object, precisely quantifying the intensity of the object's actual distortion at the microscopic level.

[0039] 3D registration: Performing mathematical operations such as translation, rotation, and scaling to perfectly align the actual data captured by the camera with the theoretical source data in the computer within the same 3D coordinate system, eliminating misalignment caused by shooting pose.

[0040] Smoothness residual map data: After registration, the actual surface change rate of the physical object is subtracted from the perfect surface change rate of the original design to obtain the difference matrix map. This accurately eliminates the curvature of the normal design and purely exposes the distortion caused by defects.

[0041] Normal gradient variance data: A statistical term applied to geometry. Variance measures the degree of fluctuation in data. The larger the variance, the more drastic and irregular the curvature change in that local area, which means that there is a serious break in the light and shadow.

[0042] Preset continuity threshold data: In the field of industrial design, the quantitative passing line for judging whether a surface is qualified is usually the G2 (curvature continuity) or G3 (curvature change rate continuity) standard in advanced surface design. It is the basis for determining whether to start repair.

[0043] Iterative optimization calculation: The machine simulates the process of continuous trial and error to approach the perfect result. Based on a set goal (to reduce the distortion variance), the system continuously fine-tunes the control point parameters in the loop, and calculates the effect after each fine-tuning until the most perfect repair combination is found.

[0044] Feature fusion: The discovered residual distortions and the original underlying control point structure are interwoven and spliced ​​together at the mathematical matrix level.

[0045] Initial state input matrix data: a high-dimensional data package formed after feature fusion, which is also the starting point for subsequent deep analysis by the neural network.

[0046] Neural network convolution processing: a mathematical operation that uses convolution kernels in deep learning algorithms to slide across the input matrix to extract features. It can effectively capture the local spatial correlation of data.

[0047] Spatial feature data characterizing spatial distortion relationships and surface topological relationships: high-dimensional condensed information extracted after convolution processing, which includes not only distortion relationships but also topological relationships.

[0048] Network layer mapping computation: The fully connected layer at the tail of the neural network processes abstract high-dimensional spatial features and converts them into specific numerical outputs that can be used directly through a series of complex function mappings.

[0049] Predicted spatial displacement vector data: The vector (direction and distance) that each control point should move in three-dimensional space, which is ultimately predicted.

[0050] Control point correction parameter data: After iterative optimization and convergence, the specific values ​​of the displacement and weight adjustment of the NURBS control points adopted by the system and finally determined.

[0051] Smoothly reconstruct instruction data: Correction parameters are encapsulated and translated into low-level API (Application Programming Interface) scripts or macro code that CAD software can directly recognize and run. This directly drives external software to automatically modify the feature tree, enabling machine-based drawing modification.

[0052] Specifically, step one: acquire the sequence data of the physical light field image and the corresponding original design model data. The original design model data includes the local control point tensor data of the non-uniform rational B-spline surface.

[0053] By constructing a multi-angle programmable light source array (such as an LED dome light source) and a binocular industrial camera, the surface of the object is photographed synchronously under different lighting phases.

[0054] Let the acquired sequence of physical light field images be... ,in Index number for the angle of the light source ( ), To determine the total number of images taken, the original design model data (such as STEP format CAD files) in the digital twin system is simultaneously read, and the local control point tensor data of the non-uniform rational B-spline (NURBS) that determines the model surface is extracted. This tensor data contains the spatial coordinates of the control points. and the corresponding rational weights ,in and The control points are in the surface parameter space. and Topological index in the direction.

[0055] In the aforementioned technologies, existing visual inspection systems typically only acquire discrete point cloud data or two-dimensional images, and are disconnected from the underlying parametric data (such as control points) of industrial design software, making it impossible to establish a connection between physical entities and underlying mathematical models. By acquiring a sequence of multi-angle light field images of a physical object containing rich optical reflection information on the hardware side, and simultaneously extracting the original CAD source file containing tensor data features of local control points from non-uniform rational B-splines (NURBS) on the software side, the data barrier between the physically measured appearance and the underlying parameters of the digital model is broken down. This lays a solid data foundation for subsequent closed-loop modifications from optical appearance directly to the underlying mathematical parameters.

[0056] Step 2: Extract features from the sequence of physical light field images to obtain the spatial coordinate range data of the initial defect features;

[0057] To avoid computational overload caused by high-precision full-image processing, a deep residual network (such as Mask R-CNN) with instance segmentation capabilities is used as a feature extractor to process the image sequence. The input network extracts abnormal high-frequency reflective or shadow truncation regions using edge operators, outputting a two-dimensional defect mask. Then, utilizing the parallax principle of a binocular camera, the two-dimensional mask is reprojected onto a three-dimensional world coordinate system. The final output is initial spatial coordinate range data containing suspected defects (such as surface fractures or tensile scratches). This data is represented as a three-dimensional bounding box, defined by its minimum boundary point. and the maximum boundary point It is a component used to define the physical space for subsequent precise calculations.

[0058] In the aforementioned technologies, due to the extremely large amount of data in high-resolution light field images, directly performing high-precision microscopic surface calculations and three-dimensional analysis on the entire system would result in excessive computational power consumption and extremely low detection efficiency. By using a feature extraction network to perform preliminary screening and classification of high-frequency local features in the image matrix, the three-dimensional bounding box of suspected defects can be quickly located, achieving coarse localization in the macroscopic dimension. This process eliminates a massive amount of invalid background data, effectively reducing the overall computational consumption of the system and improving detection efficiency and real-time performance.

[0059] Step 3: Based on the spatial coordinate range data, calculate the corresponding physical object normal partial derivative distribution field data from the physical object light field image sequence data;

[0060] Within the spatial coordinate range to Internally, using the Photometric Stereo measurement algorithm, based on the Lambertian or micro-facet reflection model, the three-dimensional true normal vector corresponding to each pixel on the real surface is calculated. .

[0061] To quantify the degree of curvature of the surface, the normal vector is... Orthogonal coordinate axes along the spatial tangent plane and Calculate the partial derivatives to generate the distribution field data of the partial derivatives of the physical normal, i.e., the Jacobian matrix. :

[0062] ;

[0063] This matrix accurately maps the microscopic undulation rate of the actual surface.

[0064] In the aforementioned technologies, traditional spatial distance measurements (such as ICP point cloud distance comparison) cannot capture the light and shadow distortion caused by extremely small deformations (such as 0.01mm level) on highly glossy curved surfaces, and lack the ability to quantify microscopic optical distortion. By utilizing photometric stereoscopic solution algorithms to process light field sequence data, extract the three-dimensional normal vector of the object surface and calculate its spatial gradient, a partial derivative matrix reflecting the true curvature state of the surface is generated. This breaks through the accuracy limitations of traditional physical size measurements, successfully transforming the light and shadow changes perceived by the human eye into a geometric curvature gradient matrix, and achieving precise quantification of microscopic surface smoothness (optical-grade defects).

[0065] Step 4: Perform three-dimensional registration between the physical normal partial derivative distribution field data and the original design model data, and calculate and generate the smoothness residual map data;

[0066] Since there is an inevitable error between the pose of the actual object and the digital model, the Iterative Closest Point (ICP) algorithm is first used to calculate the spatial transformation parameters, i.e., the rotation matrix. (3x3 matrix) and translation vector (3 rows and 1 column vector) Align the partial derivative field of the physical normal to the coordinate system of the original design model.

[0067] Meanwhile, by taking the second derivative of the original NURBS surface equation, the partial derivative matrix of the theoretical normal in this region of the digital model was calculated. .

[0068] The registered physical matrix is ​​subtracted point by point from the theoretical matrix to remove distortions caused by manufacturing or design defects, generating smoothness residual map data. :

[0069] .

[0070] In the aforementioned techniques, the placement and pose of the actual object during photography inevitably deviates from the macroscopic coordinates of the digital model, and the inherent design curvature (normal bending) of the product itself can interfere with the system's judgment of abnormal defect distortion. By calculating the transformation matrix of the measured partial derivative field of the normal and the normal field of the theoretical digital model in three-dimensional space and performing rigid registration, the data of the corresponding spatial nodes of the two are subtracted point by point to generate a difference matrix (residual map). This eliminates the interference of spatial pose error and the original design curvature of the product, accurately identifying the abnormal distortion caused purely by manufacturing or design defects, and providing a pure source of analysis for subsequent judgment.

[0071] Step 5: Extract the normal gradient variance data from the smoothness residual map data, and determine whether the normal gradient variance data is greater than the preset continuity threshold data;

[0072] Residual plot This only represents the absolute value of the error. To evaluate the smoothness and continuity of light and shadow, it is necessary to calculate the degree of dispersion of the residual within this local area.

[0073] Assume that there are a total of [number] defects in the area. There are 10 valid sampling points, and their indices are: And the average residual is Then calculate the normal gradient variance data. :

[0074] ;

[0075] Extract the calculated value of this region and compared with the preset continuity threshold data Compare. If If the surface smoothness is found to be severely substandard, the subsequent iterative repair mechanism will be triggered.

[0076] In the aforementioned technologies, traditional systems lack a unified industrial aesthetic evaluation standard when faced with detected geometric distortions. They are unable to distinguish between minor normal production tolerances and severe light and shadow breakage defects, which can easily lead to misjudgment or over-repair. By extracting the normal gradient variance (representing the severity of distortion) from the residual map and comparing it rigidly with a preset threshold representing stringent industrial design standards (such as characterizing curvature continuity and curvature change rate continuity), the machine vision system is endowed with the aesthetic judgment criteria of an expert-level industrial designer, realizing the scientific classification, quantitative diagnosis, and precise interception of defect severity.

[0077] Step 6: If so, perform iterative optimization calculations on the local control point tensor data based on the smoothness residual map data to generate control point correction parameter data that makes the normal gradient variance data converge.

[0078] The gradient descent optimization algorithm is used to fine-tune the local control point tensors of the original CAD file. To ensure that the repaired surface is both smooth and undistorted, a joint penalty objective function is constructed. :

[0079] ;

[0080] in:

[0081] and These are preset weighting coefficients;

[0082] As a geometric fidelity loss term, calculate the coordinates of the points on the repaired surface. Coordinates of the original design point Sum of squared Euclidean distances ( For grid sampling point index, );

[0083] This is the surface smoothing benefit term, whose value is dynamically equal to the variance of the normal gradient of the updated surface. ;

[0084] By taking the partial derivative, the loss function with respect to the spatial coordinates of the control points can be calculated. The gradient is calculated, and a preset learning rate is used. Perform iterative updates:

[0085] ;

[0086] when The iteration stops when the value decreases and falls below the convergence condition. At this point, the final displacement vector of each control point after convergence is extracted. And the weight changes, and solidify them into control point correction parameter data.

[0087] In the aforementioned techniques, traditional automatic repair (or manual retouching) lacks dual mathematical constraints on the geometric fidelity and surface smoothness of the model, easily leading to distortions such as out-of-tolerance errors and assembly interference after repair. By initiating a parameter optimization mechanism that includes a target loss function for the defective region, the machine automatically searches for the optimal solution that minimizes the variance of the normal gradient in multiple iterations. Finally, it outputs the spatial displacement and adjustment weights of the control points for the NURBS underlying surface, ensuring that the generated repair scheme meets both extremely high optical continuity requirements and is subject to strict physical assembly boundary constraints, avoiding random generation and thus achieving the optimal solution for model reconstruction.

[0088] Step 7: Generate smooth reconstruction instruction data based on the control point correction parameter data to indicate the modification of local control point tensor data in the original design model data.

[0089] The pure mathematical matrix output from the previous step is converted using a built-in script converter. It translates into secondary development interface commands (APIs) that can be directly executed by 3D mechanical design software (such as SolidWorks or NX). For example, it generates a smooth reconstruction command data containing SelectByID2 (select a specific NURBS surface) and SetControlPoints (rewrite control point array parameters). This script runs silently in the background and directly modifies the underlying definition of control points in the feature tree of the original design model. It eliminates the need for manual dragging of surfaces and achieves a fully automated closed loop from visual detection to reverse repair of parametric source files.

[0090] In the aforementioned technologies, due to the data gap between current visual inspection systems and industrial design software, after the system detects defects, it can typically only output discrete grid deviation maps or reports. Designers still need to manually compare and search for feature tree nodes in the CAD software for tedious blind adjustments, making closed-loop automation impossible. By converting the previously generated control point correction parameter data into standard application programming interface (API) scripts or macro instructions that can directly drive the underlying operations of external CAD software, and directly overwriting the source file data, the transition from the visual inspection end to the CAD design end is bridged. This achieves true machine-automated discovery and automatic drawing modification, effectively reducing the trial-and-error costs of manual drawing modification.

[0091] Specific Implementation Example: Scenario Setting:

[0092] Test object: The transition curved surface near the crater of the camera lens on the back cover of a mobile phone.

[0093] Design standard: It must achieve G3 continuity (continuous rate of curvature change) to ensure that the reflected lines under the highlights are absolutely smooth.

[0094] Step 1: Obtaining Data

[0095] Real-world light field data: Binocular camera in Shot under a programmed light source angle to obtain pixel coordinates grayscale sequence at .

[0096] Original model data: Read the CAD file and extract a NURBS control point at the center of the defect area. .

[0097] Initial coordinates: .

[0098] Initial weighting coefficients: .

[0099] Step 2: Feature Extraction

[0100] CNN output: A break in the highlight stripes in this area has been detected.

[0101] Spatial coordinate range: Generate a 3D bounding box ROI.

[0102] , , .

[0103] Step 3: Calculate the partial derivative of the physical normal:

[0104] Calculation results: Through photometric stereoscopic calculation, the surface of the actual object was obtained. Jacobian matrix of position :

[0105] The numerical value represents the rate of change of the normal vector with respect to its spatial position.

[0106] Step 4: 3D registration and residual calculation:

[0107] Theoretical model value: The theoretical partial derivative matrix of the CAD model at this point. :

[0108] ;

[0109] Smoothness residual :

[0110] .

[0111] Step 5: Variance assessment:

[0112] Variance calculation: The variance of the normal gradient is statistically obtained within the ROI region. .

[0113] Threshold comparison: Preset G3 continuity threshold .

[0114] result: The defect was identified as a curvature discontinuity defect, triggering repair.

[0115] Step 6: Iterative optimization calculation:

[0116] Set constraints: weights (Geometric Fidelity) (Smoothing returns).

[0117] Iterative process: The neural network calculates the control points. Need to Fine-tune in the negative axis direction and increase weight to tighten the surface.

[0118] Convergence results (control point correction parameters):

[0119] Coordinate correction: .

[0120] Weighting adjustment: .

[0121] Final recommended value: , .

[0122] Step 7: Generate instructions:

[0123] Generate instructions: encapsulated as a low-level API script.

[0124] The CAD model was automatically updated, and the repaired surfaces were verified using virtual ray tracing. Down to It meets the design requirements.

[0125] The core innovation of this application lies in constructing a joint loss function during the iterative optimization calculation process. This function includes a "geometric fidelity loss term" (first spatial distance deviation data) and a "surface smoothing benefit term" (reconstructed normal partial derivative variance data). This ensures that while the system adjusts control points to correct lighting defects, it is also strongly constrained by the original mechanical shape contour. This avoids assembly interference or abrupt changes in structural wall thickness caused by excessive smoothing, guaranteeing that the generated solution conforms to both industrial design aesthetics and meets the real physical limitations of mechanics and manufacturing. Furthermore, it provides parametric modeling for complex robotic arm structures or precision equipment shells, enabling the system to... Instead of outputting uneditable discrete patches, the system directly uses the smoothness residual map as input to the neural network convolution process. Through iterative optimization calculation, it directly outputs the control point correction parameter data for non-uniform rational B-splines (NURBS), namely the modification of spatial coordinates and rational weight values. This parameter can be directly converted into API scripts to drive 3D mechanical design software to rewrite feature tree nodes, thus completely realizing the automated closed loop from visual perception to source file modification. This effectively shortens the lengthy trial and error cycle that products need to go through from design and mold making to prototyping verification, effectively reduces mold making losses, and saves R&D resource costs.

[0126] Perform iterative optimization calculations on the local control point tensor data, specifically including:

[0127] The smoothness residual map data and the local control point tensor data are fused to generate the initial state input matrix data.

[0128] The initial state input matrix data is processed by neural network convolution to extract spatial feature data representing spatial distortion relationships and surface topological relationships.

[0129] Network layer mapping calculations are performed based on spatial feature data, and the predicted spatial displacement vector data for local control point tensor data is output.

[0130] in:

[0131] Initial state input matrix data: refers to the composite data matrix after spatial alignment and tensor quantization of the microscopic optical error (residual map) of the area to be repaired with the underlying geometric structure (control points).

[0132] Spatial distortion relationship: refers to the mapping law between the physical deformation of the surface of an object and the deviation of the reflected light field from the theoretical state.

[0133] Topological relationships of a surface: refers to the logical connection between control points in a non-uniform rational B-spline (NURBS) surface and their weighted influence on the surface shape.

[0134] Spatial feature data refers to intermediate feature vectors with high-order semantic information extracted by deep learning models, which can quantitatively represent the unevenness and structural continuity of a surface.

[0135] Network layer mapping computation: refers to the process of converting abstract feature space data into specific physical coordinate offsets through the linear or nonlinear layers of a neural network.

[0136] Predicted spatial displacement vector data: refers to the direction and distance that each control point needs to move in three-dimensional space, calculated by the system. .

[0137] Specifically, feature fusion generates the initial state input matrix:

[0138] First, extract the smoothness residual map data. and local control point tensor data The residual map data is discretized to match the spatial resolution of the control point grid. Then, a feature stitching operation is used to merge the initial coordinate data of each control point with the corresponding normal residual value to generate the initial state input matrix. :

[0139] ,in For the coordinates of the control points, This is the residual scalar at that location.

[0140] Neural network convolution processing for feature extraction:

[0141] The generated input matrix The data is input into a pre-defined convolutional processing layer, where a sliding window convolutional kernel performs localized perception of the data. The convolutional kernel slides across the matrix, passing through weight parameters. By performing a weighted summation of control points and residuals within the neighborhood, this process captures the rate of change of the surface within a local range, thereby extracting the spatial distortion relationship reflecting the degree of surface distortion, as well as the surface topological relationship reflecting the mutual attraction between control points. The final result is spatial feature data containing multi-granularity information. .

[0142] Network layer mapping calculates the output displacement vector:

[0143] High-dimensional spatial feature data This is mapped to specific physical correction commands. It utilizes a weight matrix. The fully connected layer, for Perform a linear transformation to calculate the predicted displacement vector of each control point in three-dimensional space. :

[0144] Mathematical formula: ;

[0145] in, Predict spatial displacement vector data;

[0146] : is a preset activation function used to limit the displacement amplitude within a reasonable physical range;

[0147] : This is the weight matrix mapped to the network layer;

[0148] : For calculating bias term data.

[0149] Feature fusion: Generating initial state input matrix data:

[0150] Assuming in coordinates residual of the normal partial derivative at the point .

[0151] The coordinates of the NURBS control point at this location .

[0152] The two features are fused to generate an element in the initial state input matrix:

[0153] .

[0154] Convolutional processing: Extracting spatial feature data: The system scans the neighborhood using convolutional kernels to identify "distortion trends".

[0155] Set a 3×3 convolution kernel The control point grid is scanned, assuming residuals around the center point are 0.031, 0.035, etc. Spatial feature data is extracted through weighted convolution operations. This value quantifies the combined degree of "twisting" and "topological stretching" of the surface in this local area.

[0156] Mapping calculation: Outputs predicted spatial displacement vector data, mapping abstract features to physical-level modification instructions.

[0157] Utilizing fully connected layer weights and bias Perform mapping calculations, referencing linear update logic, to calculate the predicted displacement. :

[0158] ;

[0159] The output is the predicted spatial displacement vector for this control point. .

[0160] Performing iterative optimization calculations on local control point tensor data also includes:

[0161] Reconstructed surface geometry data is generated based on predicted spatial displacement vector data and local control point tensor data.

[0162] The first spatial distance deviation between the reconstructed surface geometric data and the original design model data is calculated and used as the geometric fidelity loss term.

[0163] Calculate the variance of the partial derivative of the reconstructed normal corresponding to the geometric data of the reconstructed surface, and use it as the surface smoothing benefit term data.

[0164] The geometric fidelity loss term data and the surface smoothing gain term data are weighted and summed to calculate the objective function output value data for the current iteration round.

[0165] in:

[0166] Reconstructing surface geometry data: refers to the process of generating a temporary, draft 3D test surface using mathematical formulas after superimposing the predicted displacements onto the original control points in the current iteration.

[0167] First spatial distance deviation data: refers to the difference in physical distance between the temporarily generated draft surface and the original CAD surface in three-dimensional space. It is used to measure whether the machine has deformed the part during the retouching process.

[0168] Geometric fidelity loss term data: In the algorithm's scoring mechanism, this represents the penalty score for the degree of shape deformation. The greater the deformation, the higher the score (the heavier the penalty).

[0169] Reconstructed normal partial derivative variance data: refers to the degree of dispersion of the rate of change of the normal vector recalculated on the draft surface.

[0170] Surface smoothing benefit data: In the algorithm scoring mechanism, the score represents the smoothness of light and shadow. The smaller the variance, the more continuous the light and shadow, and the better the performance of this item.

[0171] The objective function output value data, which is the final total score, is a weighted total evaluation value that combines two indicators: no distortion (fidelity) and must be smoothed (profit).

[0172] Specifically, because existing generative AI or traditional automatic repair algorithms often only pursue a single objective (such as simply checking the smoothness of the surface), if CAD control points are stretched without constraints to eliminate optical defects, it is very easy to cause excessive deformation of the model's appearance contour, i.e., over-smoothing. The repaired part, although perfectly lit and shadowed, cannot be physically assembled with the motherboard or other casing. Therefore, the following steps are used to solve this problem:

[0173] Generating reconstructed surface geometry data: Extract the predicted spatial displacement vector data output from the previous step, and perform vector addition with the local control point tensor data in the original digital model to obtain the updated temporary control point coordinates. Subsequently, call the underlying surface equations of Non-Uniform Rational B-Splines (NURBS) to generate the discretized 3D surface mesh point cloud for the current iteration using these temporary control points. Let the temporary control points of the current iteration be... The original control point is The predicted spatial displacement vector is ,but ,based on Generate reconstructed surface geometry data, i.e., point sets. .

[0174] Calculating geometrical fidelity loss terms: It is necessary to ensure that the repaired shell will not undergo severe deformation, otherwise it will lead to assembly interference. Chamfer distance or nearest-point matching algorithms are used to reconstruct the surface point set. and the original design surface point set Find corresponding points between them and calculate the average of the sum of squared Euclidean distances between them. The specific formula is as follows:

[0175] ;

[0176] in:

[0177] : That is, the calculated first spatial distance deviation data (geometric fidelity loss term data).

[0178] : The total number of surface nodes in the discrete sampling.

[0179] : The traversal index for the surface sampling node.

[0180] : Reconstructing the first on the surface The spatial three-dimensional coordinates of each node.

[0181] : On the original design surface and the first The spatial three-dimensional coordinates of the reference point that matches each node.

[0182] Calculate the surface smoothing benefit term data: on the newly generated reconstructed surface The normal vector of each mesh node is recalculated, and its partial derivative with respect to the spatial coordinate system is calculated to obtain a new matrix of partial derivatives of the reconstructed normal vectors. Then, the variance of all partial derivative values ​​within this region is calculated to quantify the continuity of light and shadow on the current draft surface. The specific formula is as follows:

[0183] ;

[0184] This refers to reconstructing the variance data of the partial derivatives of the normal (data of the surface smoothing benefit term).

[0185] : The total number of normal feature sampling points.

[0186] : Traversal index of the sampling point.

[0187] : Reconstructing the first on the surface The partial derivative of the normal at each point.

[0188] : The mathematical average of the partial derivatives of all normals within the reconstructed region.

[0189] Calculate the objective function output value data for the current iteration: [The data represents the physical deformation.] With representing optical smoothness A linearly weighted sum is then performed; this sum represents the overall loss cost of the current machine prediction scheme. The specific formula is as follows:

[0190] ;

[0191] : That is, the output value of the objective function in the current iteration.

[0192] The preset first weight parameter determines the degree of importance the system places on "preventing model deformation".

[0193] The preset second weight parameter determines the system's emphasis on "pursuing perfect light and shadow continuity".

[0194] Specific Implementation: Continuing with the scenario of repairing the micro-curved surface of a smartphone back cover, we assume the system is currently in the iterative optimization phase... Rounds.

[0195] Generating reconstructed surface geometry data based on predicted spatial displacement vector data and local control point tensor data:

[0196] Original control point Located at the center of the defect, with coordinates: .

[0197] This round of predicted spatial displacement vector The neural network has just output a predicted value; it is suggested that the price be pushed down at this point. .

[0198] Generate reconfiguration control points:

[0199] .

[0200] Generate a temporary draft surface:

[0201] New control points Substituting the NURBS curve equation, a graph containing 1000 sampling points is rendered and discretized. Reconstructed surface geometry data .

[0202] Calculate the first spatial distance deviation data as the geometric fidelity loss term data. ):

[0203] Check how much these 1000 sampling points have deviated from the original design surface (to prevent parts from deforming or failing to assemble during retouching). Calculate the reconstructed surface. Each point on the original surface The Euclidean distance between the corresponding points. Assuming that the control point is pushed down by 0.020 mm, the sampling points on the surface shift down by an average of approximately 0.010 mm. Calculate the sum of squares and average of the distances (mean square error): Geometric Fidelity Loss Item Data If this value is greater than 0, it means that the model has indeed undergone physical deformation, and the system records this deformation penalty score.

[0204] Calculate the variance of the partial derivatives of the reconstructed normals as data for the surface smoothing benefit term. On this compressed draft surface, the rate of change of the algorithmic lines was remeasured to see if the lighting and shadows became smoother. The partial derivative matrix of the normal vector at these 1000 sampling points was recalculated, and the spatial variance of these partial derivative values ​​was calculated. Before the repair, the variance of the lighting and shadow breakage in this area was as high as 0.0012. Due to the compression of the control points, the surface became taut, and the smoothness was greatly improved, with the new variance measured to be reduced to 0.0005. Surface smoothing benefit data. This value has dropped significantly compared to the initial 0.0012, and is getting closer and closer to the G3 level industrial perfection standard of 0.0003. The system records this smoothing bonus score.

[0205] The two data points are weighted and summed to calculate the objective function output value for the current iteration. ):

[0206] Pre-set weighting: Engineers set the first weight. (Requires extremely strict dimensional fidelity to prevent assembly interference), engineers set a second weight. (Smooth lighting and shadows are required).

[0207] Weighted summation:

[0208] .

[0209] The objective function output value data of the current iteration round .

[0210] In the aforementioned technology, by constructing a loss function model with dual-objective joint constraints, the first spatial distance deviation data (geometric deformation penalty) and the variance data of the partial derivative of the reconstructed normal (optical smoothing reward) are simultaneously introduced when the machine attempts to repair in each round. The direction of the repair process is strictly monitored by weighted summation, which points out an extremely strict convergence direction that conforms to the laws of physical manufacturing for the optimization of the neural network. This ensures that while the system thoroughly repairs the G2 / G3 high-order optical fracture defects, the physical deformation of the model is strictly locked within the industrially permissible assembly tolerance range. This avoids assembly interference or abrupt changes in structural wall thickness caused by excessive smoothing, and ensures that the generated scheme conforms to both industrial design aesthetics and meets the real physical constraints of mechanics and manufacturing.

[0211] The iterative optimization calculation process also includes:

[0212] Determine whether the output value of the objective function meets the preset convergence conditions.

[0213] If the conditions are not met, the gradient data is calculated using the backpropagation algorithm, and the weight parameters of the network nodes performing neural network convolution processing are updated based on the gradient data, triggering the next round of iteration calculation;

[0214] If the conditions are met, the iteration stops, and the predicted spatial displacement vector data that currently meets the convergence condition is extracted and solidified into control point correction parameter data.

[0215] in:

[0216] Preset convergence condition data: usually refers to the difference between the output values ​​of the objective function in two consecutive iterations being less than a very small value, or the value itself falling to the industrially permissible lower limit of error.

[0217] Gradient data: The multidimensional partial derivative matrix calculated through backpropagation indicates which direction each weight parameter in the neural network should be changed in, and by how much, to make the surface smoother.

[0218] Network node weight parameter data: Learnable parameters in the convolutional and fully connected layers inside the neural network. These parameters are like empirical values ​​in the process of optimizing control points. By continuously modifying them, the network's ability to predict displacement will become more and more accurate.

[0219] Specifically, traditional surface trimming often relies on manual trial and error based on experience. If the control point is misaligned even once, the process is undone and restarted. This is not only extremely inefficient, but the human brain cannot simultaneously handle the coupled tension relationships of dozens of control points on a complex surface. If only a static model (non-iterative model) based on a single calculation is used, the entire trimming process will fail completely if the initial prediction is flawed. Therefore, the following steps address this problem:

[0220] Determine whether the output value of the objective function meets the preset convergence criteria:

[0221] After completing the first After predicting and scoring each round, the objective function output value data for the current round is extracted and combined with the data from the previous round to determine the convergence of calculus.

[0222] Let the current number be... The objective function output value of each round is The previous round The value of the round Calculate the absolute value of the difference between the two iterations. and with the preset minimum convergence threshold Compare:

[0223] ;

[0224] if ,or Directly less than the preset lower limit of absolute error If the condition is met, the convergence condition is satisfied. If the condition is not met, it means that the current control point displacement has not been corrected, the model still has deformation or the lighting is still broken. In this case, the backpropagation algorithm is called, and the objective function is calculated based on the chain rule of calculus. The partial derivative matrix (i.e., gradient data) of all internal node weights of the current neural network is used. Then, a gradient descent optimizer (such as Adam or SGD) is used to update the network node weights in the opposite direction of the gradient, and the input matrix is ​​read again to trigger the forward prediction in the (t+1)th round.

[0225] Calculate gradient data :

[0226] ;

[0227] Update network node weight parameter data :

[0228] ;

[0229] in:

[0230] : Current number Network node weight matrix for each round.

[0231] : The updated network node weight matrix used in the next round.

[0232] The calculated gradient matrix data indicates the direction of the fastest error descent.

[0233] The preset optimization learning rate controls the step size of each parameter modification.

[0234] If the convergence condition is met (i.e.) This indicates that the current predicted displacement has achieved the optimal solution for both geometric fidelity and lighting smoothness. Therefore, the iteration loop should be terminated immediately, the network weights frozen, and the [presumably referring to a specific step or setting] should be [further details needed]. The most perfect set of predicted spatial displacement vector data output in each round is extracted from temporary memory, converted into a structured external instruction format (solidification operation), and formally output as control point correction parameter data to the downstream CAD software API interface. That is: (when hour).

[0235] : Predicted spatial displacement vector data generated during the convergence rounds.

[0236] : Final control point correction parameter data after curing.

[0237] Specific implementation: Following the scenario of repairing the micro-curved surface of the smartphone back cover mentioned above, assuming that the current preset convergence condition data (minimum convergence threshold) is set to 0.00001, it means that when the score difference between two iterations is less than one ten-thousandth, the system believes that it has found the extreme optimal solution, and there is no point in continuing to repair.

[0238] If the convergence condition is not met (continue optimization and updating):

[0239] Preceding data (round t):

[0240] Previous round (number) Total score for round: .

[0241] Total score for the current round (round t) (using the data above): .

[0242] Determine if the convergence condition is met:

[0243] Calculate the absolute value of the difference between the two scores: .

[0244] Comparison threshold: .

[0245] Result: Convergence condition not met. This indicates that the current control point displacement has not yet reached a perfect state, and there is still room for optimization. Perform backpropagation to calculate gradient data:

[0246] The system uses the chain rule to calculate the weight node of a specific convolutional kernel in the network that caused the 0.00026 error. .

[0247] Calculate the partial derivative (gradient data) of this weight node: This indicates that reducing the weight will further decrease the total error. Update the network node weight parameters and trigger the next round:

[0248] Assuming a pre-defined optimization learning rate Current weight .

[0249] Update weight formula: .

[0250] Update calculation: .

[0251] The system will update the weights Load the network, extract the residual image, and automatically start the first... Forward prediction of the wheel.

[0252] If the convergence condition is met (stop iteration and solidify the output):

[0253] Prerequisite data (entering the first) (After round)

[0254] Based on the updated weights, the neural network provides new predictions: suggested control points. Shaft pressing .

[0255] Regenerate the draft and score it, then calculate the total score for this round: .

[0256] Re-evaluate the convergence condition:

[0257] Calculate the absolute value of the difference between the two scores: .

[0258] Comparison threshold: .

[0259] Judgment result: The convergence condition is met, indicating that although the control point was pressed down by 0.001mm, the combined benefits of optical smoothness and geometric fidelity have reached Pareto optimality. Further modifications would only increase computational load or lead to overfitting. Stop the iteration and extract the solidified data, immediately trigger the stop command, exit the iteration loop, and set the... The perfect spatial displacement vector predicted by the wheel ( Extract the data from dynamic memory, assign a conclusion to this set of data, convert it into static control point correction parameter data, and encapsulate it into a Python / API script to send to the 3D CAD software to perform the final model reconstruction.

[0260] The aforementioned technology incorporates closed-loop backpropagation and gradient descent optimization mechanisms from deep learning. Through rigorous objective function scoring, convergence determination, chain-like differentiation for gradient calculation, and weight parameter updates, the system gains the ability to autonomously test and correct itself in the mathematical space. It only solidifies the output after finding the optimal parameters that absolutely meet the industrial threshold, effectively eliminating the blindness of manual trial and error. This ensures that the control point correction parameters output to the 3D design software have absolute mathematical reliability and physical manufacturability, greatly improving the yield rate and design output speed.

[0261] Feature extraction is performed on the sequence of physical light field images to obtain the spatial coordinate range data of the initial defect features, specifically including:

[0262] Extracting local high-frequency geometric feature data from a two-dimensional pixel matrix from a sequence of physical light field image data;

[0263] Defect classification and boundary regression calculation are performed based on local high-frequency geometric feature data, and the output is initial defect feature data containing suspected curved surface fracture and chamfer abnormality areas.

[0264] The initial defect feature data is mapped to a preset three-dimensional world coordinate system to generate spatial coordinate range data.

[0265] in:

[0266] Two-dimensional pixel matrix: The underlying digital image format acquired by industrial camera sensors. Each pixel in the image corresponds to an element in the matrix, and its value represents the light intensity (grayscale value) or color information at that location.

[0267] Local high-frequency geometric feature data: In the field of image processing, high frequency refers to the areas where pixel values ​​change drastically. On the surface of industrial products, smooth areas are low-frequency; while defects such as scratches, surface fractures, and chamfering errors will produce sharp edges, abrupt highlights or shadows under specific lighting conditions. These are high-frequency geometric features.

[0268] The three-dimensional world coordinate system is a global and absolute three-dimensional physical space reference system. Physical parts, industrial cameras, and digital CAD models must all be uniformly converted to this coordinate system to ensure that the "position seen in the photo" and the "position of the part in reality" correspond perfectly.

[0269] Specifically, local high-frequency geometric feature data is extracted from the two-dimensional pixel matrix: a single image from the real-object light field image sequence is converted into a two-dimensional pixel matrix. Then, multiple convolutional layers in a convolutional neural network (such as the ResNet feature extraction backbone network with residual structures) are used to perform a sliding scan on this matrix. The lower-level convolutional kernels (equivalent to various edge extraction filters) generate high-response activation values ​​in regions of abrupt grayscale changes in the image, thereby filtering out large areas of smooth background and extracting feature maps containing information about structural abrupt changes. Let the input two-dimensional pixel matrix be... ,in and These represent the horizontal and vertical pixel coordinates of the image, respectively. After convolution, the output local high-frequency geometric feature data is denoted as the feature tensor. .

[0270] Perform defect classification and boundary regression calculations to output initial defect feature data: feed the extracted feature tensors into two parallel computing branches of the detection network (similar to the detection head structure of Faster R-CNN).

[0271] Classification branch: Using a fully connected layer and a softmax activation function, calculate the probability that the feature region belongs to a preset industrial defect category (such as suspected surface fracture, abnormal chamfer, etc.). .

[0272] Regression branch: The bounding box boundary parameters of the defect region on the two-dimensional image are calculated and fine-tuned using a linear regression algorithm.

[0273] The output contains initial defect feature data including category labels and two-dimensional bounding boxes. Let the output two-dimensional bounding box data be... .in and The pixel coordinates of the top-left corner of the bounding box. and The pixel coordinates of the bottom right corner of the bounding box constitute the initial defect feature data.

[0274] Mapping to a 3D world coordinate system to generate spatial coordinate range data: Since subsequent control point modifications are performed in 3D CAD space, the 2D pixel coordinates must be converted to 3D spatial coordinates. Based on the calibration parameters (intrinsic and extrinsic parameters) of the binocular industrial camera, combined with the depth information calculated from binocular parallax, the 2D bounding box is transformed using the inverse perspective projection. The four corner rays are projected back into the three-dimensional physical space, extracting a three-dimensional bounding box containing the defect. The core calculation of spatial mapping follows the physical model of pinhole camera perspective projection:

[0275] ;

[0276] and : These are the two-dimensional pixel coordinates of the defect boundary obtained in the previous step.

[0277] The physical depth distance of the point measured by binocular parallax or structured light.

[0278] : Camera intrinsic parameter matrix (containing focal length and optical center data, which are known constants).

[0279] : Camera extrinsic matrix (containing the camera's rotation and translation vectors in the 3D world, which are known constants).

[0280] , , : That is, the actual physical space coordinates that the system ultimately solves and maps to in the three-dimensional world coordinate system.

[0281] By solving the above equations, the three-dimensional boundary set of the defect is obtained, that is, the spatial coordinate range data.

[0282] Based on spatial coordinate range data, the corresponding partial derivative distribution field data of the object's normal is calculated from the object's light field image sequence data, specifically including:

[0283] Extract high-frequency photometric stereo information data from the physical light field image sequence data within the defined spatial coordinate range;

[0284] Photometric stereoscopic calculations are performed on high-frequency photometric stereoscopic information data to obtain three-dimensional surface normal vector data of the object surface;

[0285] Calculate the gradient of the change of the three-dimensional surface normal vector data with respect to the spatial coordinates, and generate the distribution field data of the partial derivative of the physical normal.

[0286] in:

[0287] High-frequency photometric 3D information data refers to the sequence of pixel grayscale values ​​located within a suspected defect area, captured under different lighting angles. These grayscale values ​​contain information about the rapid changes in light and shadow (high frequency) caused by minute surface undulations.

[0288] Photometric stereo algorithm: a classic computer vision 3D reconstruction algorithm. It continuously switches the direction of the light source from a fixed camera viewpoint and uses the laws of light reflection on the object's surface (such as Lambert reflection) to deduce the normal vector of each point on the surface.

[0289] 3D surface normal vector data: represents the absolute orientation of every extremely small surface element (pixel) on the surface of a physical object in 3D space. The normal vector is always perpendicular to the tiny tangent plane in which it is located.

[0290] Gradient variation: Here, it refers to spatial differentiation. That is, how much the orientation of the normal vector changes between two adjacent pixels. Mathematically, it is directly equivalent to the curvature of a surface.

[0291] Specifically, the high-frequency photometric stereo information data is extracted from the physical light field image sequence data within a defined range: The three-dimensional spatial coordinate range data (ROI bounding box) generated in the previous step is received. Using the inverse mapping of the camera's perspective projection matrix, this three-dimensional bounding box is reprojected back onto the two-dimensional camera imaging plane. The corresponding local pixel region is cropped, and all pixels within that region are extracted within a preset range. A set of grayscale values ​​under illumination from light sources at different angles. Let the coordinates of the extracted local pixel region be... The system constructs an observation brightness vector. This vector contains the current pixel. Gray values ​​under each light source ( ),Right now: This vector is the high-frequency photometric three-dimensional information data, which records the light and shadow patterns of this point as the illumination changes.

[0292] A three-dimensional photometric solution is performed to obtain the three-dimensional surface normal vector data of the actual object's surface: a lighting equation is constructed based on the Lambertian reflection model. According to the principles of physical optics, the observed brightness of a pixel is equal to the dot product of the light source direction vector, the surface normal vector, and the surface reflectivity (albedo). The system substitutes multiple light source directions and their corresponding observed brightness into this equation, solves the overdetermined system of equations using the least squares method, and calculates the unit normal vector for each pixel. The matrix form of the lighting equation is: ;

[0293] : is the known light source direction matrix (size is Each row represents a unit direction vector of a light source in three-dimensional space.

[0294] : An unknown normal vector with albedo (size is...) ).

[0295] Solve using the least squares method : .

[0296] Finally, for Normalization is performed to extract the pure geometric unit normal vectors that are unaffected by color and material. By traversing all pixels within the ROI, dense 3D surface normal vector data can be generated.

[0297] Calculating the gradient of the normal vector relative to spatial coordinates generates data on the distribution of partial derivatives of the physical normal vectors. Normal vectors alone can only represent the orientation of the surface and cannot reflect the "degree of distortion" of the surface. Using the finite difference method, such as applying the central difference operator, the generated dense normal vector field is analyzed. Along the lateral direction and longitudinal Partial derivatives are taken in both directions to generate a Jacobian matrix field. At the pixel point... At this point, the gradient of each component of the normal relative to the spatial coordinates is calculated, generating a data matrix of the partial derivative distribution field of the physical normal. :

[0298] ;

[0299] in: The matrix The extremely precise quantification of the curvature change (i.e., bending rate) at this point in various directions is the core mathematical basis for subsequent judgment of whether the smoothness meets the standard.

[0300] The physical partial derivative distribution field data of the normal is three-dimensionally registered with the original design model data to calculate and generate smoothness residual map data, specifically including:

[0301] The theoretical surface topology data is extracted from the original design model data, and theoretical normal partial derivative distribution field data is generated based on the theoretical surface topology data;

[0302] Calculate the spatial transformation matrix between the physical partial derivative distribution field data and the theoretical partial derivative distribution field data;

[0303] Align the physical partial derivative distribution field data with the theoretical partial derivative distribution field data based on the spatial transformation matrix data, and calculate the difference of the normal vector derivative of the corresponding spatial nodes to generate smoothness residual map data.

[0304] in:

[0305] Theoretical surface topology data refers to the underlying mathematical parameters describing the shape of a surface in the source file of 3D CAD software (such as NX and SolidWorks). For NURBS surfaces, this includes node vectors, surface order, and control point mesh. This data defines an "absolutely perfect surface" without any manufacturing defects.

[0306] Theoretical normal partial derivative distribution field data: The theoretical rate of change of the normal vector at each spatial location is calculated by performing pure mathematical differentiation on an absolutely perfect surface. This is used as the absolute standard answer to judge whether a physical object is qualified.

[0307] Spatial transformation matrix data: a mathematical matrix containing rotation and translation parameters (usually 10 ... The homogeneous transformation matrix is ​​used to "transfer" the coordinate system of the physical object on the inspection table to perfectly match the coordinate system of the CAD model in the computer, since the coordinate system of the physical object is different from that of the CAD model in the computer.

[0308] Normal vector derivative difference: After complete alignment in three-dimensional space, at the same physical location, the difference matrix is ​​obtained by subtracting the theoretical normal vector change rate from the measured normal vector change rate.

[0309] Smoothness residual map data: A high-dimensional matrix map formed by the difference of the derivatives of the normal vectors of all spatial nodes. It no longer records the shape of the surface itself, but the pure amount of defect deformation.

[0310] Specifically, theoretical surface topological data is extracted to generate theoretical normal partial derivative distribution field data: The original design model data is analyzed to extract the NURBS surface parametric equations for the target region. High-density discretization sampling is performed within the surface's parameter domain. Based on calculus principles, the first-order partial derivative of the surface equation is first obtained to yield the theoretical normal vector. Then, the partial derivative of the theoretical normal vector is calculated to obtain the accurate theoretical curvature variation matrix field. Let the extracted theoretical NURBS surface equation be... ,in , Let be the two-dimensional parametric coordinates of the surface. Calculate the strictly orthogonal unit theoretical normal vector at this point. :

[0311] ;

[0312] Taking the derivative of the theoretical normal vector with respect to the spatial coordinate system again generates the theoretical normal partial derivative distribution field data matrix. :

[0313] ;

[0314] The matrix This represents the standard curvature flow trend that the designer gave to the surface.

[0315] Calculating the spatial transformation matrix: Since the measured data extracted in the previous steps is located in the camera coordinate system, while the theoretical data is located in the CAD world coordinate system, the system needs to perform 3D rigid registration to extract the 3D contours of the measured point cloud and the CAD model. Using an improved iterative nearest-point algorithm, the optimal spatial transformation matrix is ​​calculated by minimizing the mean square error between the two point sets. Let the extracted measured spatial point set be... The corresponding theoretical CAD surface matching point set is ,in For matching point pairs, the traversal index ( The system solves for a problem involving a rotation matrix. (3×3 matrix) and translation vector (3×1 vector) spatial transformation matrix Its goal is to minimize the following objective function:

[0316] ;

[0317] Calculated This refers to spatial transformation matrix data, which can perfectly overlay physical data onto digital models.

[0318] Align the data and calculate the difference in normal vector derivatives to generate smoothness residual map data: Using the calculated spatial transformation matrix, perform coordinate system transformation and attitude correction on the distribution field of the partial derivatives of the physical normal. After alignment, at each corresponding node in three-dimensional space, subtract the theoretical partial derivative matrix corresponding to that node from the corrected measured partial derivative matrix. Let the corresponding node... The original measured partial derivative of the normal at point is After rotation matrix correction (the normal derivative belongs to the second-order tensor property and requires attitude alignment), the aligned partial derivative of the physical normal is obtained. :

[0319] ;

[0320] Subsequently, the theoretical partial derivative corresponding to that point is calculated. Difference between the normal vector derivatives :

[0321] ;

[0322] All spatial nodes difference matrix Combined, this generates the final high-dimensional tensor data, namely the smoothness residual map data. .

[0323] The preset continuity threshold data includes a first threshold data representing the continuity of curvature and a second threshold data representing the continuity of the rate of change of curvature, and the magnitude of the first threshold data is greater than that of the second threshold data.

[0324] Determining whether the normal gradient variance data is greater than a preset continuity threshold data specifically includes:

[0325] Determine whether the normal gradient variance data is greater than the first threshold data;

[0326] If the data exceeds the first threshold, the current defect is determined to be a surface fracture defect, and a global position reconstruction instruction for the local control point tensor data is triggered.

[0327] If the data is not greater than the first threshold, then it is further determined whether the normal gradient variance data is greater than the second threshold.

[0328] If the data exceeds the second threshold, the current defect is determined to be a curvature discontinuity defect, and a local tension fine-tuning instruction for the local control point tensor data is triggered.

[0329] If the value is not greater than the second threshold, the current region is determined to meet the industrial design continuity requirements, and the iterative optimization calculation is terminated.

[0330] in:

[0331] The first threshold data characterizing curvature continuity corresponds to the G2 continuity standard in industrial design. G2 continuity means that two adjacent surfaces not only have no gaps and identical tangents, but also have exactly the same degree of curvature at their junction. The first threshold data is the upper limit of mathematical tolerance for determining whether G2 continuity is acceptable.

[0332] The second threshold value characterizing the continuity of the rate of change of curvature corresponds to the more advanced G3 continuity standard in industrial design. G3 continuity requires not only the same degree of curvature but also a smooth transition in the rate of change of curvature. This is the ultimate standard to ensure that the zebra stripes reflected on extremely glossy surfaces (such as piano lacquer and ceramics) do not twist or jitter. Because the standard is more stringent, the tolerance is smaller, so the magnitude of the first threshold must be numerically greater than that of the second threshold.

[0333] Surface fracture defects: These refer to severe geometric defects that fail to even meet the most basic G2 continuity. Visually, they appear as obvious creases, pits, or hard edges.

[0334] Global position reconstruction instruction: Allows the neural network to significantly modify the three-dimensional spatial coordinates of control points in order to rebuild the basic framework of the surface.

[0335] Curvature discontinuity defect: refers to a microscopic imperfection that reaches G2 but not G3. The surface shape may appear fine, but under strong light, the reflected highlight lines will show slight distortion or unevenness.

[0336] Local tension fine-tuning command: A micro-shaping command issued for microscopic imperfections. It locks or greatly restricts the movement of coordinates, and instead allows the neural network to fine-tune the rational weight factors in the NURBS control point formula, thereby smoothing out the light and shadow by changing the local surface tension.

[0337] Specifically, different manufacturing defects cause vastly different degrees of damage to the underlying model. Using a single threshold and a one-size-fits-all repair strategy will lead to severe repair mismatch. Therefore, this problem can be addressed through the following specific steps:

[0338] Threshold setting and first-level judgment (G2 diagnosis and reconstruction): During each iteration of scoring, the calculated normal gradient variance data of the current region is first extracted. A first threshold representing the G2 tolerance and a second threshold representing the G3 tolerance are pre-set in memory. First-level logic comparison is performed to determine whether the current variance has fallen below the relatively lenient G2 tolerance threshold. Let the currently calculated normal gradient variance data be... Let the preset first threshold data be... The second threshold data is And satisfy .

[0339] First-level judgment: If A surface fracture defect is detected. At this point, the global position reconstruction instruction is activated, and the learning rate for coordinate displacement is increased in the neural network update formula, allowing control point coordinates to... Generates a large spatial correction vector ,Right now:

[0340] The focus at this stage is on repairing severe physical structural collapse.

[0341] Second-level judgment (G3 diagnosis and fine-tuning): If the reconstruction in the previous stage is completed, or if the initial defect itself is not serious, making the first-level judgment invalid (i.e., the variance has been reduced to the G2 tolerance range), then the second-level, more stringent optical inspection stage is entered to determine whether the current variance is still greater than the G3 tolerance bottom line.

[0342] Second-level judgment: If and The underlying geometric surface is determined to be unbroken, but contains high-order curvature discontinuities (i.e., uneven lighting). At this point, the local tension fine-tuning command is activated, and a displacement penalty lock is introduced into the iterative algorithm, causing the coordinate correction vector to... The optimization objective is forcibly switched to the rational weight variables of NURBS control points. By generating subtle tension correction parameters Perform iterations:

[0343] ;

[0344] Change weight It's like stretching a flexible curtain; without changing the hook coordinates, by finely adjusting the tension, the gloss reflection on the curtain surface achieves an absolutely smooth transition.

[0345] Termination of iteration (standard determination): After multiple rounds of reconstruction or fine-tuning, when the system detects that the current variance no longer triggers even the most stringent G3 tolerance bottom line, it indicates that the mathematical surface of the region has reached the extremely high standard of industrial aesthetic requirements, with light and shadow flowing smoothly like water.

[0346] Termination determination: If The system determines that the current region meets the highest industrial design continuity requirements, immediately sends an interrupt signal to the control unit, forcibly terminates the gradient descent optimization calculation, and freezes the current coordinates. and weight The status is then pushed into the output queue to prepare for generating CAD rewrite instructions.

[0347] The aforementioned technology incorporates a dual-threshold hierarchical diagnostic mechanism aligned with industrial G2 / G3 standards. Through tiered conditional judgments, two distinct optimization paths are planned for the neural network. Once the basic structure (G2) is deemed satisfactory, the spatial coordinates are immediately locked, and lighting is corrected solely by modifying NURBS weights. This effectively eliminates the risk of overall part deformation and scrap due to the pursuit of perfect highlights, and avoids the algorithm blindly searching in a vast high-dimensional solution space of coordinates and weights. The hierarchical instructions enable the network to rapidly reduce dimensionality in the later stages of repair (searching only for the optimal weight solution), resulting in a several-fold increase in iterative convergence speed and ensuring the high efficiency of the industrial design self-healing system.

[0348] Based on the control point correction parameter data, smooth reconstruction instruction data is generated, specifically including:

[0349] Convert control point correction parameter data into standard application programming interface script data;

[0350] The standard application programming interface script data is encapsulated into smooth reconstruction instruction data to trigger the rewriting of the coordinates and weight values ​​of the non-uniform rational B-spline control points in the original design model data, thereby obtaining candidate reconstruction model data.

[0351] Virtual ray tracing rendering is performed on the candidate reconstruction model data to generate simulated light field verification data;

[0352] Extract the virtual normal gradient variance data from the simulated light field verification data;

[0353] If the variance of the virtual normal gradient is not greater than the continuity threshold, the feature tree difference log data between the candidate reconstruction model data and the original design model data is extracted and packaged into a repair scheme feedback report data and sent to the terminal interface.

[0354] in:

[0355] Standard application programming interface script data: executable code (such as code segments based on C#, Python or VBA macro instructions) written for the secondary development interface protocols that are open at the underlying level of mainstream 3D mechanical design software (such as SolidWorks, NX, etc.).

[0356] Candidate reconstruction model data: Temporary 3D digital model files whose underlying parameters have been modified by the algorithm, but which have not yet undergone final optical simulation for inspection.

[0357] Virtual ray tracing rendering: A virtual optical laboratory is built inside the computer, exactly the same as the real testing platform. By emitting and tracing the reflection and refraction paths of a massive number of virtual photons, the reflective images of high-gloss materials under realistic light fields are simulated.

[0358] Simulated light field verification data: A virtual two-dimensional image matrix output by the rendering engine, used to predict "what the actual object will look like when photographed under strong light if it is manufactured according to this scheme".

[0359] Virtual normal gradient variance data: Based on the virtual rendered image, the photometric stereo solution process described above is repeated to calculate the secondary verification variance value.

[0360] Feature tree difference log data: records detailed information about exactly what changes were made to the model within the 3D design software.

[0361] Repair solution feedback report data: The final diagnosis and treatment closure report presented to the engineer includes both visual comparisons of the effects and a precise list of modified parameters.

[0362] Specifically, the control point correction parameters are converted into API scripts and the model is overwritten to obtain candidate reconstructed models. The process receives the pure mathematical correction matrix locked in the previous step, and includes an embedded instruction translator. For the target mechanical design software (such as SolidWorks), the control point coordinate indices and specific offsets in the matrix are converted line by line into standard API call code. Then, the 3D design software is silently awakened via inter-process communication technology (such as COM component calls), and the script is injected as a macro instruction to directly overwrite the parameters in the model's underlying feature tree. The software automatically performs feature reconstruction updates and outputs temporarily modified candidate reconstructed model data. Let the extracted control point correction parameter matrix be... Includes the target node identifier and correction vector The translator converts it into standard application interface script data in string format. .For example:

[0363] ;

[0364] Execute this After that, the original model The memory parameters are overwritten to generate candidate reconstruction model data. .

[0365] Virtual ray tracing rendering is performed on the candidate reconstruction model to generate simulated light field verification data: To prevent unpredictable rendering folds or topological breaks when the AI's mathematical calculations are converted into actual curved surfaces, a virtual optical twin environment with a 1:1 scale to the physical detection environment from the first step is constructed. Import the rendering engine (such as Physically Based Renderer PBR), assign it the corresponding specular ceramic or metallic material properties, and perform multi-view virtual ray tracing calculations under the illumination of a preset virtual light source array to generate high-fidelity simulated lighting and shadow images. Let the virtual camera pose matrix be... The virtual light source array is The bidirectional reflection distribution function of the material is .

[0366] The rendering equation calculates and outputs simulated light field verification data. :

[0367] ;

[0368] in The normal vector of the candidate reconstructed model surface. Let be the incident ray vector. This represents the area of ​​the hemisphere.

[0369] Extract virtual variance and perform final validation, generating a feedback report: for the generated virtual image The normal calculation and calculus differentiation are performed again to extract the virtual normal gradient variance data. This virtual variance is then compared with a preset continuity pass threshold. Once confirmed to be acceptable, the operation history of the 3D software is retrieved, and the parameter differences before and after the modification are extracted to generate a feature tree difference log. Finally, the digital model file, virtual rendering effect map, and difference log are packaged and pushed to the engineer's terminal interface. Let the extracted virtual normal gradient variance data be... The final continuous error prevention threshold data is :like The reconstructed model was determined to have perfect smoothness under real lighting conditions, and feature tree difference log data was extracted. Generate and send feedback report data on the repair solution. .like If this occurs, it indicates that geometric distortion has occurred during the copying process. The system will then revert to the previous update and re-trigger the optimization calculation or issue a manual alarm.

[0370] An intelligent defect detection and control point automatic repair system for curved surface design of industrial products includes:

[0371] Data acquisition module: used to acquire physical light field image sequence data and corresponding original design model data. The original design model data includes local control point tensor data of non-uniform rational B-spline surface.

[0372] Feature extraction module: used to extract features from the sequence data of physical light field images to obtain the spatial coordinate range data of the initial defect features;

[0373] Data processing module: used to calculate the corresponding physical normal partial derivative distribution field data from the physical light field image sequence data based on spatial coordinate range data;

[0374] Data registration module: used to perform three-dimensional registration of the physical normal partial derivative distribution field data with the original design model data, and calculate and generate smoothness residual map data;

[0375] Data Judgment and Optimization Module: Used to extract the normal gradient variance data from the smoothness residual map data and determine whether the normal gradient variance data is greater than the preset continuity threshold data; if so, it performs iterative optimization calculation on the local control point tensor data based on the smoothness residual map data to generate control point correction parameter data that makes the normal gradient variance data converge.

[0376] Data output module: Used to generate smooth reconstruction instruction data based on control point correction parameter data, instructing modification of local control point tensor data in the original design model data.

[0377] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A method for intelligent defect detection and automatic repair of control points in the curved surface design of industrial products, characterized in that, Includes the following steps: Acquire physical light field image sequence data and corresponding original design model data, wherein the original design model data includes local control point tensor data of non-uniform rational B-spline surface; Feature extraction is performed on the sequence data of physical light field images to obtain the spatial coordinate range data of the initial defect features; Based on spatial coordinate range data, the corresponding physical object normal partial derivative distribution field data is calculated from the physical object light field image sequence data; The physical normal partial derivative distribution field data is three-dimensionally registered with the original design model data to calculate and generate smoothness residual map data. Extract the normal gradient variance data from the smoothness residual map data, and determine whether the normal gradient variance data is greater than the preset continuity threshold data; Based on the smoothness residual map data, iterative optimization calculations are performed on the local control point tensor data to generate control point correction parameter data that makes the normal gradient variance data converge. Based on the control point correction parameter data, smooth reconstruction instruction data is generated to instruct the modification of local control point tensor data in the original design model data.

2. The intelligent defect detection and automatic repair method for control points in the curved surface design of industrial products according to claim 1, characterized in that: Perform iterative optimization calculations on the local control point tensor data, specifically including: The smoothness residual map data and the local control point tensor data are fused to generate the initial state input matrix data. The initial state input matrix data is processed by neural network convolution to extract spatial feature data representing spatial distortion relationships and surface topological relationships. Network layer mapping calculations are performed based on spatial feature data, and the predicted spatial displacement vector data for local control point tensor data is output.

3. The intelligent defect detection and automatic repair method for control points in the curved surface design of industrial products according to claim 2, characterized in that: Performing iterative optimization calculations on local control point tensor data also includes: Reconstructed surface geometry data is generated based on predicted spatial displacement vector data and local control point tensor data. The first spatial distance deviation between the reconstructed surface geometric data and the original design model data is calculated and used as the geometric fidelity loss term. Calculate the variance of the partial derivative of the reconstructed normal corresponding to the geometric data of the reconstructed surface, and use it as the surface smoothing benefit term data. The geometric fidelity loss term data and the surface smoothing gain term data are weighted and summed to calculate the objective function output value data for the current iteration round.

4. The intelligent defect detection and automatic repair method for control points in the curved surface design of industrial products according to claim 3, characterized in that: The iterative optimization calculation process also includes: Determine whether the output value of the objective function meets the preset convergence conditions. If the conditions are not met, the gradient data is calculated using the backpropagation algorithm, and the weight parameters of the network nodes performing neural network convolution processing are updated based on the gradient data, triggering the next round of iteration calculation; If the conditions are met, the iteration stops, and the predicted spatial displacement vector data that currently meets the convergence condition is extracted and solidified into control point correction parameter data.

5. The intelligent defect detection and automatic repair method for control points in the curved surface design of industrial products according to claim 1, characterized in that: Feature extraction is performed on the sequence of physical light field images to obtain the spatial coordinate range data of the initial defect features, specifically including: Extracting local high-frequency geometric feature data from a two-dimensional pixel matrix from a sequence of physical light field image data; Defect classification and boundary regression calculation are performed based on local high-frequency geometric feature data, and the output is initial defect feature data containing suspected curved surface fracture and chamfer abnormality areas. The initial defect feature data is mapped to a preset three-dimensional world coordinate system to generate spatial coordinate range data.

6. The intelligent defect detection and automatic repair method for control points in the curved surface design of industrial products according to claim 1, characterized in that: Based on spatial coordinate range data, the corresponding partial derivative distribution field data of the object's normal is calculated from the object's light field image sequence data, specifically including: Extract high-frequency photometric stereo information data from the physical light field image sequence data within the defined spatial coordinate range; Photometric stereoscopic calculations are performed on high-frequency photometric stereoscopic information data to obtain three-dimensional surface normal vector data of the object surface; Calculate the gradient of the change of the three-dimensional surface normal vector data with respect to the spatial coordinates, and generate the distribution field data of the partial derivative of the physical normal.

7. The intelligent defect detection and automatic repair method for control points in the curved surface design of industrial products according to claim 1, characterized in that: The physical partial derivative distribution field data of the normal is three-dimensionally registered with the original design model data to calculate and generate smoothness residual map data, specifically including: The theoretical surface topology data is extracted from the original design model data, and theoretical normal partial derivative distribution field data is generated based on the theoretical surface topology data; Calculate the spatial transformation matrix between the physical partial derivative distribution field data and the theoretical partial derivative distribution field data; Align the physical partial derivative distribution field data with the theoretical partial derivative distribution field data based on the spatial transformation matrix data, and calculate the difference of the normal vector derivative of the corresponding spatial nodes to generate smoothness residual map data.

8. The intelligent defect detection and automatic repair method for control points in the curved surface design of industrial products according to claim 1, characterized in that: The preset continuity threshold data includes a first threshold data representing the continuity of curvature and a second threshold data representing the continuity of the rate of change of curvature, and the magnitude of the first threshold data is greater than that of the second threshold data. Determining whether the normal gradient variance data is greater than a preset continuity threshold data specifically includes: Determine whether the normal gradient variance data is greater than the first threshold data; If the data exceeds the first threshold, the current defect is determined to be a surface fracture defect, and a global position reconstruction instruction for the local control point tensor data is triggered. If the data is not greater than the first threshold, then it is further determined whether the normal gradient variance data is greater than the second threshold. If the data exceeds the second threshold, the current defect is determined to be a curvature discontinuity defect, and a local tension fine-tuning instruction for the local control point tensor data is triggered. If the value is not greater than the second threshold, the current region is determined to meet the industrial design continuity requirements, and the iterative optimization calculation is terminated.

9. The intelligent defect detection and automatic repair method for control points in the curved surface design of industrial products according to claim 1, characterized in that: Based on the control point correction parameter data, smooth reconstruction instruction data is generated, specifically including: Convert control point correction parameter data into standard application programming interface script data; The standard application programming interface script data is encapsulated into smooth reconstruction instruction data to trigger the rewriting of the coordinates and weight values ​​of the non-uniform rational B-spline control points in the original design model data, thereby obtaining candidate reconstruction model data. Virtual ray tracing rendering is performed on the candidate reconstruction model data to generate simulated light field verification data; Extract the virtual normal gradient variance data from the simulated light field verification data; If the virtual normal gradient variance data is not greater than the continuity threshold data, then the feature tree difference log data between the candidate reconstruction model data and the original design model data is extracted and packaged into a repair scheme feedback report data and sent to the terminal interface.

10. An intelligent defect detection and control point automatic repair system for curved surface design of industrial products, characterized in that, include: Data acquisition module: used to acquire physical light field image sequence data and corresponding original design model data, wherein the original design model data includes local control point tensor data of non-uniform rational B-spline surface; Feature extraction module: used to extract features from the sequence data of physical light field images to obtain the spatial coordinate range data of the initial defect features; Data processing module: used to calculate the corresponding physical normal partial derivative distribution field data from the physical light field image sequence data based on spatial coordinate range data; Data registration module: used to perform three-dimensional registration of the physical normal partial derivative distribution field data with the original design model data, and calculate and generate smoothness residual map data; Data Judgment and Optimization Module: Used to extract the normal gradient variance data from the smoothness residual map data and determine whether the normal gradient variance data is greater than the preset continuity threshold data; if so, it performs iterative optimization calculation on the local control point tensor data based on the smoothness residual map data to generate control point correction parameter data that makes the normal gradient variance data converge. Data output module: Used to generate smooth reconstruction instruction data based on control point correction parameter data, instructing modification of local control point tensor data in the original design model data.

Images