A mobile phone support quality detection method based on image recognition

Abstract

The application relates to a mobile phone support quality detection method based on image recognition and belongs to the technical field of vision-based quality detection. The method comprises the following steps: extracting assembly constraints and mechanical benchmark parameters of a qualified mobile phone support standard three-dimensional model to construct a standard support model; inputting the standard support model and a multi-view image sequence of a to-be-detected support into a topography deviation solving network to generate a constraint feature map and a multi-view feature map for cross-view registration, outputting a topography deviation field and generating a component integrity detection result; identifying surface defects and structural deformation from the topography deviation field and constructing a topography-mechanical response field to map and quantize mechanical performance influence and generate an appearance defect detection result; and performing gradient field tensor analysis on key structural feature points in a low-risk area to obtain sub-pixel level deviation, executing deviation transmission solving, and generating a structural precision detection result. The application realizes high-precision and layer-by-layer progressive mobile phone support multi-level comprehensive quality detection.

Description

Technical Field

[0001] This invention belongs to the field of vision-based quality inspection technology, specifically relating to a mobile phone holder quality inspection method based on image recognition. Background Technology

[0002] With the increasing precision of the smartphone and microelectronics industries, phone stands, as key structural components that support and secure the core internal components of a phone (such as the motherboard, camera, clips, and various sensors), directly determine the structural stability and reliability of the entire device through their manufacturing and assembly quality. Phone stands typically have complex three-dimensional topological shapes, containing various micro-components and mating surfaces. Quality inspection of these stands requires not only verifying the assembly integrity of the components but also accurately identifying appearance defects and precisely evaluating the micron-level geometric tolerances after assembly.

[0003] Currently, the quality inspection of mobile phone holders mainly relies on manual visual inspection or traditional two-dimensional machine vision methods. Manual visual inspection suffers from low efficiency, strong subjectivity, and inconsistent inspection standards, making it difficult to meet the needs of large-scale automated production. Although traditional machine vision inspection solutions incorporate image processing algorithms, they often use single-viewpoint or fixed-angle shooting, which can easily limit the field of view for mobile phone holders with complex three-dimensional structures due to viewpoint obstruction. In addition, most existing vision inspection solutions are limited to qualitative two-dimensional plane pixel-level edge detection, with each detection module operating independently and lacking a progressive linkage mechanism.

[0004] More importantly, traditional technologies cannot fully establish a quantitative correlation between manufacturing defects and service mechanical properties. When there are minor scratches, deformations, or burrs on the surface of the support, existing solutions can only make qualitative judgments and cannot quantitatively assess the stress concentration and material damage risks caused by these geometric deviations under actual assembly conditions, making it impossible to scientifically classify the risk level of defects. At the same time, in terms of structural accuracy inspection, traditional solutions are unable to achieve micron-level precision measurements and cannot effectively simulate the dynamic tolerance transfer process of geometric deviations between multi-level assembly mating surfaces in space.

[0005] In summary, existing mobile phone bracket quality inspection solutions cannot simultaneously address multi-dimensional unified testing of component integrity, mechanical assessment of appearance defects, and multi-level tolerance transfer, making it difficult to meet the multi-level and quantitative quality inspection needs of modern high-precision mobile phone brackets under actual service conditions. Summary of the Invention

[0006] To address the aforementioned problems in the existing technology, this invention provides a mobile phone holder quality inspection method based on image recognition. The objective of this invention can be achieved through the following technical solution: S1: Obtain a standard 3D model of a qualified mobile phone holder, extract the assembly constraint parameters and mechanical reference parameters of each assembly mating surface from the standard 3D model, construct a standard holder model with assembly constraint information; collect a multi-view detection image sequence of the mobile phone holder to be tested; S2: Input the standard support model and the multi-view detection image sequence into the topography deviation calculation network, perform joint feature encoding on the assembly constraint parameters and the multi-view detection image sequence to generate constraint feature maps and multi-view feature maps; perform constraint-guided cross-view registration and deviation calculation on the multi-view feature maps through the constraint feature maps, output the topography deviation field of each component region, identify components with abnormal assembly integrity in the topography deviation field, and generate component integrity detection results; S3: Using the morphology deviation field and the mechanical reference parameters as input, identify surface defects and structural deformations from the morphology deviation field to generate preliminary appearance defect detection results; construct a morphology-mechanical response field, perform mechanical property mapping and quantitative evaluation on each defect area in the preliminary appearance defect detection results to obtain the quantitative value of the mechanical performance influence of each defect area, and generate appearance defect detection results. S4: For component regions where the quantification value of the impact on mechanical properties is lower than the preset level, sub-pixel-level deviation data of key structural feature points of assembly are obtained from the morphological deviation field through gradient field tensor analysis. The sub-pixel-level deviation data is then used to perform assembly deviation transfer calculation along the assembly tolerance chain of the assembly mating surface in the standard bracket model to generate structural accuracy detection results.

[0007] Specifically, the standard scaffold model construction process includes: Locate each assembly mating surface from the standard three-dimensional model, and calculate the spatial position parameters and mating tolerance parameters of each assembly mating surface as the assembly constraint parameters; Apply load boundary conditions simulating assembly conditions to each component area, solve the stress field distribution of each component area using finite element method, and extract the stress distribution reference value and strength safety factor from the solution results as the mechanical reference parameters; The assembly constraint parameters are associated with and stored in relation to the corresponding mating surfaces, and the mechanical reference parameters are associated with and stored in relation to the corresponding component regions, thereby constructing a standard bracket model with attached assembly constraint information.

[0008] Specifically, the step of solving the stress field distribution of each component region using the finite element method includes: Each component region is divided into tetrahedral meshes, and the surface height data of each component region is extracted from the standard 3D model. The surface height data is then converted into displacement constraints for the mesh nodes after mesh division. Based on the assembly constraint parameters in the standard support model, the load distribution of each assembly mating surface in the assembly state is obtained, and the load distribution is used as the load boundary condition of the mesh model. The displacement constraint condition and the load boundary condition are input into the static finite element solution process to obtain the stress tensor of each mesh node, forming a stress field distribution.

[0009] Specifically, the construction process of the topography deviation calculation network includes: An initial network is constructed, comprising a parametric coding layer, an image feature extraction layer, a constraint-guided attention layer, and a bias decoder. The initial network includes a three-dimensional geometric coding branch and a two-dimensional image feature extraction branch. The initial network is trained using multi-view training image samples with known topographic deviation annotations and the assembly constraint parameters of the corresponding standard scaffold model as inputs to form a topographic deviation solution network. The parameter encoding layer maps the assembly constraint parameters into a high-dimensional topological space vector, and the image feature extraction layer extracts multi-scale features from the training images. The constraint-guided attention layer calculates the association weights between the high-dimensional topological space vector and the multi-scale features and performs fusion updates. The deviation decoder outputs a predicted topographic deviation field and adjusts the network weights based on the loss function of the predicted topographic deviation field and the ground truth annotations until the network converges.

[0010] Specifically, the generation process of the constraint feature map and the multi-view feature map includes: The assembly constraint parameters in the standard support model are input into the three-dimensional geometric coding branch of the topography deviation solution network, and nonlinear feature mapping is performed through a multilayer perceptron to obtain a constraint feature map that characterizes the geometric topological relationship and tolerance range of the mating surface. The multi-view detection image sequence is input into the two-dimensional image feature extraction branch of the shape deviation solution network. Cross-channel and cross-scale feature extraction is performed using a densely connected convolutional neural network, and alignment is performed in the spatial dimension to obtain a multi-view feature map containing the geometric contours and texture features of the appearance of the mobile phone bracket under test from each viewpoint.

[0011] Specifically, the cross-view registration and bias calculation includes: The constraint feature map is used as the query vector and cross-attention calculation is performed with the multi-view feature map to obtain the three-dimensional spatial reconstruction transformation matrix of each view image in the spatial coordinate system of the standard support model. Based on the transformation matrix, the multi-view feature map is projected into a unified three-dimensional space for point cloud-level registration to obtain a reconstructed three-dimensional model; The reconstructed 3D model is iteratively registered with the standard bracket model to the nearest point, and the spatial distance deviation of each spatial grid node in the normal direction is calculated. The spatial distance deviations are arranged in order to generate a shape deviation field covering the region of each component of the mobile phone bracket to be tested.

[0012] Specifically, the construction process of the morphology-mechanical response field includes: Extract the spatial three-dimensional discrete deviation values ​​of each grid node in the topography deviation field; use each spatial three-dimensional discrete deviation value as a boundary perturbation quantity to correct the spatial geometric position of the corresponding initial component grid in the standard support model under the standard coordinate system, and reconstruct the equivalent analysis entity with manufacturing topography defects; Based on the preset load and constraint conditions in the mechanical reference parameters, the equivalent analysis entity is subjected to structural mechanics calculation, and continuous stress response data characterizing the stress change trajectory of each defect region is output. The continuous stress response data is sequentially mapped back to the spatial coordinate nodes corresponding to the standard three-dimensional model to generate a topography-mechanical response field.

[0013] Specifically, the reconstructing of the equivalent analytical entity with manufacturing morphology defects further includes: Identify sensitive edge regions in the equivalent analysis entity where curvature abruptly changes due to geometric position correction; Obtain the baseline curvature of the sensitive edge region before the abrupt change, and calculate the curvature distortion rate of the sensitive edge region before and after the abrupt change; Using the curvature distortion rate, the stiffness response coefficient of the sensitive edge region in the structural mechanics solution process is linearly reduced, and the corrected stress response data containing the early deformation displacement evolution components caused by local morphological defects is output.

[0014] Specifically, the mechanical property mapping and quantitative evaluation includes: The stress values ​​at each node in the topography-mechanical response field are compared with the corresponding standard stress reference values ​​in the mechanical reference parameters to extract the stress concentration response increment of each region. Based on the yield strength limit of the support material in the mechanical reference parameters, calculate the safety margin attenuation corresponding to each stress concentration response increment; When the safety margin attenuation exceeds the set safety boundary threshold, the risk level of the appearance defect detection result is determined by combining the stress concentration response increment, so as to quantitatively characterize the mechanical attenuation degree of the mobile phone bracket under actual service conditions.

[0015] Specifically, the gradient field tensor analysis includes: Extract the normal deviation vectors of each spatial grid node in the neighborhood of the key structural feature point in the topography deviation field; calculate the gradient of the normal deviation vectors to construct the local topography gradient tensor matrix corresponding to the key structural feature point in the assembly. The local topography gradient tensor matrix is ​​decomposed into eigenvalues, and the eigenvectors corresponding to the eigenvalues ​​are extracted as the distortion gradient direction. The normal deviation vector is subjected to subpixel-level parabolic fitting extreme value extraction along the distortion gradient direction, and the subpixel-level deviation data of the assembly key structural feature points are output.

[0016] Specifically, the assembly deviation propagation calculation includes: Extract the spatial geometric dependency tree between each assembly mating surface from the standard bracket model, and convert the spatial geometric dependency tree into an assembly tolerance chain with a directed graph structure; The subpixel-level deviation data of the key structural feature points of the assembly is used as the initial deviation input and converted into a kinematic transformation matrix along the assembly tolerance chain of the directed graph structure. Through the multi-level cascaded product of the Jacobian matrix, the cumulative deviation of the assembly mating surface of the previous level is transferred to the tolerance range of the assembly mating surface of the next level. Calculate the comprehensive assembly deviation vector of each component region at the end of the assembly tolerance chain to complete the assembly deviation propagation solution.

[0017] Specifically, the process of generating the structural accuracy detection results includes: Using the comprehensive assembly deviation vector as node position constraints and the theoretical design dimensions between components in the standard support model as distance constraints, a support spatial geometric accuracy constraint network is constructed. The least squares global closed-loop adjustment calculation is performed on the spatial geometric accuracy constraint network of the bracket to output the form and position tolerance deviation values ​​of each assembly mating surface; The form and position tolerance deviation value is compared with the preset industry factory size tolerance limit by a threshold comparison to generate a structural accuracy test result that characterizes the overall geometric accuracy level of the bracket.

[0018] The beneficial effects of this invention are as follows: (1) This invention solves the technical problem that traditional visual inspection can only qualitatively identify appearance defects but cannot assess service risks by constructing a morphology-mechanical response field. The solution utilizes geometric position correction and linear stiffness reduction techniques to reconstruct the equivalent analytical entity with manufacturing defects and perform structural mechanics calculations. This allows the appearance inspection results to not only include geometric morphological features but also implicitly include the stress concentration response increment and safety margin attenuation caused by the defects. This progressive evaluation mechanism can scientifically and quantitatively determine the severity level and damage risk of defects under actual assembly and service conditions, providing accurate mechanical data support for the quality release of mobile phone brackets and effectively avoiding the omission of high-risk hidden defects.

[0019] (2) This invention overcomes the shortcomings of existing solutions that are limited to pixel-level edge detection and cannot meet the micrometer-level assembly accuracy requirements by combining gradient field tensor analysis and assembly tolerance chain calculation. The solution first extracts sub-pixel-level high-precision deviations of key structural feature points through eigenvalue decomposition and parabolic fitting of local topography gradient tensor; then, it uses this deviation as the initial manufacturing error input, performs dynamic deviation propagation calculation along the assembly tolerance chain using the multi-level cascaded product of the Jacobian matrix, and performs global closed-loop adjustment through a spatial geometric accuracy constraint network. This process comprehensively simulates the geometric dependence and error accumulation effect between components, and realizes a refined quantitative evaluation of the form and position tolerance deviation values ​​of multi-level assembly mating surfaces. Attached Figure Description

[0020] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0021] Figure 1 This is a flowchart illustrating a mobile phone holder quality inspection method based on image recognition according to the present invention. Figure 2 This is an architecture diagram of a mobile phone holder quality inspection method based on image recognition according to the present invention. Detailed Implementation

[0022] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following detailed description of the specific implementation methods, structures, features and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0023] Please see Figure 1-2 A method for quality inspection of mobile phone holders based on image recognition, comprising: S1: Obtain a standard 3D model of a qualified mobile phone holder, extract the assembly constraint parameters and mechanical reference parameters of each assembly mating surface from the standard 3D model, construct a standard holder model with assembly constraint information; collect a multi-view detection image sequence of the mobile phone holder to be tested; S2: Input the standard support model and the multi-view detection image sequence into the topography deviation calculation network, perform joint feature encoding on the assembly constraint parameters and the multi-view detection image sequence to generate constraint feature maps and multi-view feature maps; perform constraint-guided cross-view registration and deviation calculation on the multi-view feature maps through the constraint feature maps, output the topography deviation field of each component region, identify components with abnormal assembly integrity in the topography deviation field, and generate component integrity detection results; S3: Using the morphology deviation field and the mechanical reference parameters as input, identify surface defects and structural deformations from the morphology deviation field to generate preliminary appearance defect detection results; construct a morphology-mechanical response field, perform mechanical property mapping and quantitative evaluation on each defect area in the preliminary appearance defect detection results to obtain the quantitative value of the mechanical performance influence of each defect area, and generate appearance defect detection results. S4: For component regions where the quantification value of the impact on mechanical properties is lower than the preset level, sub-pixel-level deviation data of key structural feature points of assembly are obtained from the morphological deviation field through gradient field tensor analysis. The sub-pixel-level deviation data is then used to perform assembly deviation transfer calculation along the assembly tolerance chain of the assembly mating surface in the standard bracket model to generate structural accuracy detection results.

[0024] Specifically, the standard scaffold model construction process includes: Locate each assembly mating surface from the standard three-dimensional model, and calculate the spatial position parameters and mating tolerance parameters of each assembly mating surface as the assembly constraint parameters; Apply load boundary conditions simulating assembly conditions to each component area, solve the stress field distribution of each component area using finite element method, and extract the stress distribution reference value and strength safety factor from the solution results as the mechanical reference parameters; The assembly constraint parameters are associated with and stored in relation to the corresponding mating surfaces, and the mechanical reference parameters are associated with and stored in relation to the corresponding component regions, thereby constructing a standard bracket model with attached assembly constraint information.

[0025] Specifically, the step of solving the stress field distribution of each component region using the finite element method includes: Each component region is divided into tetrahedral meshes, and the surface height data of each component region is extracted from the standard 3D model. The surface height data is then converted into displacement constraints for the mesh nodes after mesh division. Based on the assembly constraint parameters in the standard support model, the load distribution of each assembly mating surface in the assembly state is obtained, and the load distribution is used as the load boundary condition of the mesh model. The displacement constraint condition and the load boundary condition are input into the static finite element solution process to obtain the stress tensor of each mesh node, forming a stress field distribution.

[0026] Specifically, the construction process of the topography deviation calculation network includes: An initial network is constructed, comprising a parametric coding layer, an image feature extraction layer, a constraint-guided attention layer, and a bias decoder. The initial network includes a three-dimensional geometric coding branch and a two-dimensional image feature extraction branch. The initial network is trained using multi-view training image samples with known topographic deviation annotations and the assembly constraint parameters of the corresponding standard scaffold model as inputs to form a topographic deviation solution network. The parameter encoding layer maps the assembly constraint parameters into a high-dimensional topological space vector, and the image feature extraction layer extracts multi-scale features from the training images. The constraint-guided attention layer calculates the association weights between the high-dimensional topological space vector and the multi-scale features and performs fusion updates. The deviation decoder outputs a predicted topographic deviation field and adjusts the network weights based on the loss function of the predicted topographic deviation field and the ground truth annotations until the network converges.

[0027] Specifically, the generation process of the constraint feature map and the multi-view feature map includes: The assembly constraint parameters in the standard support model are input into the three-dimensional geometric coding branch of the topography deviation solution network, and nonlinear feature mapping is performed through a multilayer perceptron to obtain a constraint feature map that characterizes the geometric topological relationship and tolerance range of the mating surface. The multi-view detection image sequence is input into the two-dimensional image feature extraction branch of the shape deviation solution network. Cross-channel and cross-scale feature extraction is performed using a densely connected convolutional neural network, and alignment is performed in the spatial dimension to obtain a multi-view feature map containing the geometric contours and texture features of the appearance of the mobile phone bracket under test from each viewpoint.

[0028] Specifically, the cross-view registration and bias calculation includes: The constraint feature map is used as the query vector and cross-attention calculation is performed with the multi-view feature map to obtain the three-dimensional spatial reconstruction transformation matrix of each view image in the spatial coordinate system of the standard support model. Based on the transformation matrix, the multi-view feature map is projected into a unified three-dimensional space for point cloud-level registration to obtain a reconstructed three-dimensional model; The reconstructed 3D model is iteratively registered with the standard bracket model to the nearest point, and the spatial distance deviation of each spatial grid node in the normal direction is calculated. The spatial distance deviations are arranged in order to generate a shape deviation field covering the region of each component of the mobile phone bracket to be tested.

[0029] Specifically, the construction process of the morphology-mechanical response field includes: Extract the spatial three-dimensional discrete deviation values ​​of each grid node in the topography deviation field; use each spatial three-dimensional discrete deviation value as a boundary perturbation quantity to correct the spatial geometric position of the corresponding initial component grid in the standard support model under the standard coordinate system, and reconstruct the equivalent analysis entity with manufacturing topography defects; Based on the preset load and constraint conditions in the mechanical reference parameters, the equivalent analysis entity is subjected to structural mechanics calculation, and continuous stress response data characterizing the stress change trajectory of each defect region is output. The continuous stress response data is sequentially mapped back to the spatial coordinate nodes corresponding to the standard three-dimensional model to generate a topography-mechanical response field.

[0030] Specifically, the reconstructing of the equivalent analytical entity with manufacturing morphology defects further includes: Identify sensitive edge regions in the equivalent analysis entity where curvature abruptly changes due to geometric position correction; Obtain the baseline curvature of the sensitive edge region before the abrupt change, and calculate the curvature distortion rate of the sensitive edge region before and after the abrupt change; Using the curvature distortion rate, the stiffness response coefficient of the sensitive edge region in the structural mechanics solution process is linearly reduced, and the corrected stress response data containing the early deformation displacement evolution components caused by local morphological defects is output.

[0031] Specifically, the mechanical property mapping and quantitative evaluation includes: The stress values ​​at each node in the topography-mechanical response field are compared with the corresponding standard stress reference values ​​in the mechanical reference parameters to extract the stress concentration response increment of each region. Based on the yield strength limit of the support material in the mechanical reference parameters, calculate the safety margin attenuation corresponding to each stress concentration response increment; When the safety margin attenuation exceeds the set safety boundary threshold, the risk level of the appearance defect detection result is determined by combining the stress concentration response increment, so as to quantitatively characterize the mechanical attenuation degree of the mobile phone bracket under actual service conditions.

[0032] Specifically, the gradient field tensor analysis includes: Extract the normal deviation vectors of each spatial grid node in the neighborhood of the key structural feature point in the topography deviation field; calculate the gradient of the normal deviation vectors to construct the local topography gradient tensor matrix corresponding to the key structural feature point in the assembly. The local topography gradient tensor matrix is ​​decomposed into eigenvalues, and the eigenvectors corresponding to the eigenvalues ​​are extracted as the distortion gradient direction. The normal deviation vector is subjected to subpixel-level parabolic fitting extreme value extraction along the distortion gradient direction, and the subpixel-level deviation data of the assembly key structural feature points are output.

[0033] Specifically, the assembly deviation propagation calculation includes: Extract the spatial geometric dependency tree between each assembly mating surface from the standard bracket model, and convert the spatial geometric dependency tree into an assembly tolerance chain with a directed graph structure; The subpixel-level deviation data of the key structural feature points of the assembly is used as the initial deviation input and converted into a kinematic transformation matrix along the assembly tolerance chain of the directed graph structure. Through the multi-level cascaded product of the Jacobian matrix, the cumulative deviation of the assembly mating surface of the previous level is transferred to the tolerance range of the assembly mating surface of the next level. Calculate the comprehensive assembly deviation vector of each component region at the end of the assembly tolerance chain to complete the assembly deviation propagation solution.

[0034] Specifically, the process of generating the structural accuracy detection results includes: Using the comprehensive assembly deviation vector as node position constraints and the theoretical design dimensions between components in the standard support model as distance constraints, a support spatial geometric accuracy constraint network is constructed. The least squares global closed-loop adjustment calculation is performed on the spatial geometric accuracy constraint network of the bracket to output the form and position tolerance deviation values ​​of each assembly mating surface; The form and position tolerance deviation value is compared with the preset industry factory size tolerance limit by a threshold comparison to generate a structural accuracy test result that characterizes the overall geometric accuracy level of the bracket.

[0035] Example 1: Construction of a Standard Scaffold Model Step 1: Extraction of assembly constraint parameters The system acquires the standard 3D model of the mobile phone holder's CAD design. This model contains the precise geometric information of each component of the holder in the design coordinate system and the assembly topology relationships between the components. The system locates each assembly mating surface from the standard 3D model. Taking the assembly mating surface between the buckle component A and the slot component B as an example, this mating surface is the contact area between the bottom surface of the buckle component A and the top surface of the slot component B. Boundary extraction is performed on this mating surface using an edge detection algorithm based on curvature changes to extract the boundary point sequence along the contour of the mating surface. Geometric feature fitting is performed on the extracted boundary point sequence, and least-squares plane fitting is used to obtain the spatial plane equation of the mating surface, which has the following form: ax + by + cz + d = 0, where (a, b, c) is the unit normal vector of the fitted plane, and d is the directed distance from the plane to the origin. The 3D coordinates of the center point of the mating surface are taken as the arithmetic mean of the coordinates of each point in the boundary point sequence. Simultaneously, the mating tolerance parameters of the mating surface are read from the design annotations of the standard 3D model, including the dimensional tolerance range and geometric tolerance requirements. The calculated spatial position parameters and mating tolerance parameters are used as the assembly constraint parameters of the mating surface. The above extraction operation is performed on all assembly mating surfaces in the standard 3D model one by one to complete the full extraction of assembly constraint parameters.

[0036] Step 2: Extraction of mechanical reference parameters Finite element analysis was performed on each component region to extract mechanical reference parameters. First, tetrahedral meshes were generated for each component region. Taking snap-fit ​​component A as an example, its geometric model under its standard morphology was discretized into tetrahedral elements of preset size.

[0037] Load boundary conditions simulating assembly conditions are applied to the meshed model. Surface height data of snap-fit ​​component A is extracted from the standard 3D model and converted into displacement constraints for the mesh nodes after mesh generation. Each sampling point of the surface height data is matched with a mesh node using nearest neighbor matching, and the initial displacement value of each node is taken as the height value of the corresponding sampling point. Based on the assembly constraint parameters in the standard bracket model, the load distribution of each assembly mating surface in the assembly state is obtained. The contact pressure distribution of the assembly mating surfaces is calculated using Hertzian contact theory based on the mating tolerance parameters and assembly preload. The calculated contact pressure distribution is used as the load boundary condition for the mesh model of the component region.

[0038] The displacement constraints and load boundary conditions are input into the static finite element method (FEM) solution process. The FEM solution is based on the fundamental equations of linear elasticity, with the core governing equations being: KU=F, Where K is the system stiffness matrix, assembled from the stiffness matrices of each element; U is the nodal displacement vector to be solved; and F is the nodal load vector, determined by the load boundary conditions. The assembly process of the system stiffness matrix K is as follows: for each tetrahedral element, calculate its element stiffness matrix K. e : Where B is the element strain-displacement matrix, D is the material elasticity matrix, and Ω e The element integration domain is defined. The element stiffness matrices of all elements are assembled into the system stiffness matrix K. After applying displacement constraints, the nodal displacement vectors U are obtained by solving the linear equations. The strain and stress fields of each element are then solved from the nodal displacement vectors. The formulas for calculating element strain ε and stress σ are as follows: ε=BUe, σ=Dε, where Ue is the displacement vector of the element node. When the residual solution satisfies the convergence condition, the stress tensor data of each mesh node are obtained, forming the stress field distribution.

[0039] The stress distribution baseline value and strength safety factor are extracted from the obtained stress field distribution. Statistical analysis is performed on the stress values ​​of each component region, and the average stress value is taken. The stress distribution benchmark value for this component region is used; the strength safety factor Sf is taken as: , where σ y σ is the yield strength of the material. max This represents the maximum stress value within the component region. The aforementioned stress distribution benchmark value and the strength safety factor are used as the mechanical benchmark parameters for this component region.

[0040] Step 3: Assembly of the standard support model The extracted assembly constraint parameters are associated with and stored along with the corresponding assembly mating surfaces, and the mechanical datum parameters are associated with and stored along with the corresponding component regions. The association storage method is as follows: in the data structure of the standard 3D model, assembly constraint parameter attribute fields are added to each mating surface node, and mechanical datum parameter attribute fields are added to each component region. After completing the association storage, a standard support model with attached assembly constraint information is constructed.

[0041] Example 2: Generation of Morphology Deviation Field and Component Integrity Detection Step 1: Construction of the topography deviation solution network An initial network is constructed, comprising a parametric encoding layer, an image feature extraction layer, a constraint-guided attention layer, and a bias decoder. This initial network includes a 3D geometric encoding branch and a 2D image feature extraction branch. Multi-view training image samples of sample mobile phone holders and corresponding standard topographic bias field annotation data are obtained. The multi-view training image samples, assembly constraint parameters of the sample holders, and standard topographic bias field annotations are combined into a training dataset.

[0042] The training dataset is input into the initial network for forward propagation: the parameter encoding layer maps the assembly constraint parameters into a high-dimensional topological space vector—a multilayer perceptron is used to perform nonlinear feature mapping on the spatial position parameters and fit tolerance parameters, outputting a fixed-dimensional high-dimensional topological space vector; the image feature extraction layer performs multi-scale feature extraction on multi-view training image samples—a densely connected convolutional neural network is used to perform cross-channel and cross-scale feature extraction, outputting a multi-scale feature map containing the appearance geometric contours and texture features of each viewpoint; the constraint-guided attention layer uses the high-dimensional topological space vector as the query vector, performs cross-attention calculation with the multi-scale feature map, generates association weights, and fuses and updates the multi-scale features; the deviation decoder receives the fused and updated features and outputs the predicted morphology deviation field.

[0043] Calculate the deviation between the predicted topographic deviation field and the standard topographic deviation field annotation. The deviation value is calculated using the mean square error function: Where N is the total number of spatial grid nodes in the topography deviation field. Let be the prediction bias vector for the i-th node. Let ||·|| be the annotation bias vector of the i-th node, and ||·|| represent the Euclidean norm. The network parameters of the parameter encoding layer, image feature extraction layer, constraint-guided attention layer, and bias decoder are updated based on the bias values. During training, when the validation set loss decreases by less than 1e for 10 consecutive rounds... -5 When the network converges, the trained topography deviation solution network is obtained. The convergence condition is determined based on the following technical basis: The number of training rounds is set based on the following: a continuous 10-round observation window is an industry standard for early stopping in deep learning regression tasks, balancing model convergence sufficiency and training efficiency. If the window is too short (<5 rounds), it is prone to premature stopping due to training fluctuations; if the window is too long (>20 rounds), it will increase unnecessary computational overhead and may cause model overfitting. This invention presets 100 training rounds, with the 10-round observation window accounting for 10% of the total training rounds, conforming to the training specifications of industrial-grade deep learning models.

[0044] Threshold setting basis: 1e -5 The loss reduction threshold is specifically optimized for the 3D topography deviation regression task of this invention. Unlike classification tasks, the topography deviation calculation network outputs continuous spatial distance deviation values, and its loss function (mean squared error) is inherently small in magnitude. If the threshold is too high (>1e...), the loss reduction threshold will be significantly reduced. -4 Insufficient model convergence can lead to morphological deviation prediction errors exceeding 0.05mm, failing to meet the micron-level detection accuracy requirements for mobile phone brackets; if the threshold is too low (<1e -6 If this happens, the training will be difficult to converge, and the model will overfit the noise in the training data.

[0045] Step 2: Generation of Constrained Feature Maps and Multi-View Feature Maps A multi-view image sequence of the mobile phone bracket to be inspected is acquired. The assembly constraint parameters from the standard bracket model are input into the 3D geometric encoding branch of the topography deviation solution network. Nonlinear feature mapping is performed using a multilayer perceptron to obtain constraint feature maps representing the geometric topological relationship and tolerance range of the mating surfaces. The multi-view image sequence is then input into the 2D image feature extraction branch of the topography deviation solution network. Densely connected convolutional neural networks are used for cross-channel and cross-scale feature extraction, and alignment is performed in the spatial dimension to obtain a multi-view feature map containing the geometric contours and texture features of the mobile phone bracket from various perspectives.

[0046] Step 3: Cross-view registration and bias calculation The constraint feature map is used as the query vector, and cross-attention is calculated between it and the multi-view feature map. The cross-attention calculation process is as follows: the constraint feature map is transformed linearly to generate the query matrix Q, and the multi-view feature map is transformed linearly to generate the key matrix K and the value matrix V. Q=W Q F c K=W K F m V=W V F m F c For constraining the feature map, F m For multi-view feature maps, W Q W K W V Let A be a learnable linear transformation weight matrix. The formula for calculating the attention weight matrix A is: , where dk is the dimension of the key vector, used to scale the dot product result. The attention output matrix O=AV, after decoding, yields the 3D spatial reconstruction transformation matrix of each viewpoint image in the standard scaffold model spatial coordinate system.

[0047] Based on the 3D spatial reconstruction transformation matrices from various perspectives, the multi-view feature maps are projected onto a unified 3D space for point cloud-level registration, resulting in a reconstructed 3D model. The reconstructed 3D model is then iteratively closest to the target point (ICP) registration with a standard scaffold model, and the spatial distance deviation of each spatial mesh node in the normal direction is calculated. For each spatial mesh node p in the reconstructed 3D model... i Search the surface of the standard support model for the point q that is closest to the node. i Node p i Spatial distance deviation δ in the normal direction i for: δ i =(p i q i )·n i , where n i For the standard support model at point q i The unit normal vector at the location. The distance deviations of the normal directions of each spatial grid node are arranged in spatial order to generate a topographic deviation field covering the regions of each component of the mobile phone bracket to be tested.

[0048] Step 4: Identification of Components with Assembly Integrity Issues Identify components with abnormal assembly integrity within the generated topographic deviation field. Statistically analyze the topographic deviation of each component region and calculate the average deviation of all spatial grid nodes within that region. Compare the average deviation with a preset integrity threshold; component regions with average deviations exceeding the threshold are identified as components with abnormal assembly integrity. For identified abnormal component regions, determine the anomaly type based on the spatial distribution characteristics of the topographic deviation field: if the topographic deviation within the abnormal component region exhibits an overall offset characteristic with a consistent offset direction, it is determined to be component misalignment; if the topographic deviation within the abnormal component region exhibits a large-area disordered distribution without a clear standard topographic correspondence, it is determined to be component missing. Summarize the identification information and anomaly type of all abnormal components to generate component integrity detection results.

[0049] Example 3: Construction of Morphology-Mechanical Response Field and Detection of Appearance Defects Step 1: Preliminary identification of appearance defects: Surface defects and structural deformations are identified from a morphological deviation field. The deviation of the normal direction of each spatial grid node in the morphological deviation field is extracted, and spatial continuity analysis is performed on the deviation: nodes with adjacent deviations and similar values ​​are connected into the same region. Morphological features of each connected region are extracted, including region area, aspect ratio, and edge curvature. Based on these morphological features, each connected region is classified as follows: when the area of ​​a connected region is within a preset range and the aspect ratio exceeds a preset threshold, it is classified as a scratch defect; when a connected region exhibits a raised morphology and its edge curvature characteristics match those of a burr, it is classified as a burr defect; when a connected region exhibits continuous bending characteristics along structural feature lines, it is classified as a structural deformation. All identified surface defects and structural deformation regions are summarized to generate preliminary appearance defect detection results.

[0050] Step 2: Construction of the morphological-mechanical response field: Extract the spatial three-dimensional discrete deviation values ​​of each grid node in the topography deviation field. Use these spatial three-dimensional discrete deviation values ​​as boundary perturbations to correct the spatial geometric position of the corresponding initial component grid in the standard support model under the standard coordinate system. The correction method is as follows: for nodes in the initial grid... Its corrected position for: = +δ i ·n i , where δ i Let n be the deviation of the normal direction at that node. i Let be the unit normal vector of the node under the standard topography. The corrected mesh node reconstruction yields an equivalent analytical entity with manufacturing topography defects.

[0051] During the reconstruction process, sensitive edge regions in the equivalent analysis entity where curvature abruptly changes due to geometric position correction are identified. The average curvature H and Gaussian curvature K of the surfaces within the neighborhood of each mesh node are calculated before and after correction. H = 1 / 2(κ1 + κ2), K = κ1·κ2, where κ1 and κ2 are the principal curvatures at that point on the surface. Nodes whose average curvature change exceeds a preset curvature change threshold are marked as sensitive edge nodes. The baseline curvature of the sensitive edge region before the abrupt change is obtained, and the curvature distortion rate η of the sensitive edge region before and after the abrupt change is calculated: H pert H is the corrected mean curvature. nom The value represents the average curvature before correction. The stiffness response coefficients of sensitive edge regions in the structural mechanics solution process are linearly reduced using the curvature distortion rate, resulting in corrected stress response data that includes early deformation displacement evolution components caused by local topographic defects.

[0052] Based on the preset loads and constraints in the mechanical reference parameters, structural mechanics calculations are performed on the corrected equivalent analysis entity, outputting continuous stress response data characterizing the stress change trajectory of each defect region. This continuous stress response data is then sequentially mapped back to the corresponding spatial coordinate nodes of the standard 3D model to generate a morphological-mechanical response field.

[0053] Step 3: Mechanical property mapping and quantitative evaluation The stress values ​​at each node in the topographic-mechanical response field are compared with the corresponding standard stress reference values ​​in the mechanical reference parameters. The stress concentration response increment Δσ at each node is calculated. i : ,in Let be the current stress value at node i. The standard stress reference value for node i.

[0054] The yield strength σ of the support material in the mechanical reference parameters y Calculate the safety margin attenuation M corresponding to each stress concentration response increment. i : M i =Δσ i / σ y When the safety margin decays by M i Exceeding the preset security boundary threshold M th At the same time, the risk level of the preliminary appearance defect detection results is determined by combining the stress concentration response increment, so as to quantitatively characterize the mechanical attenuation degree of the mobile phone bracket under actual service conditions and generate appearance defect detection results with risk level.

[0055] Example 4: Subpixel accuracy detection and assembly deviation propagation calculation Step 1: Gradient field tensor analysis and sub-pixel level bias data acquisition: For component regions where the quantified impact value on mechanical properties is below a preset level, sub-pixel-level deviation data of key structural feature points in the assembly are obtained from the topographic deviation field through gradient field tensor analysis. The normal deviation vectors of each spatial mesh node within the neighborhood of the key structural feature points in the topographic deviation field are extracted. The gradient of the normal deviation vectors of each spatial mesh node within the neighborhood is calculated. For the normal deviation field δ(x,y,z), its gradient vector in three-dimensional space is obtained. δ is: The partial derivatives are calculated using the central difference scheme. The gradient vectors of each node in the neighborhood are arranged using partial differentials to construct the local topographic gradient tensor matrix G corresponding to the key structural feature points of the assembly: , Perform eigenvalue decomposition on the local topographic gradient tensor matrix G, and solve the characteristic equation det(G) Given λI)=0, we obtain three eigenvalues ​​λ1≥λ2≥λ3 and their corresponding eigenvectors v1,v2,v3. We take the eigenvector v1 corresponding to eigenvalue λ1 as the distortion gradient direction.

[0056] Subpixel-level parabolic fitting and extreme value extraction are performed on the normal deviation vector along the distortion gradient direction. Along the distortion gradient direction, with the feature point as the center, the normal deviation of each sampling point in the neighborhood is extracted at a preset step size, forming a deviation sequence {dk}. A parabolic function f(t) = at is applied to the deviation sequence {dk}. 2 Using least-squares fitting with +bt+c, the extreme point position tpeak of the fitted parabola is: tpeak= b / 2a calculates the difference between the subpixel-level precise position and the standard design position of the feature point to obtain the subpixel-level deviation data of the key structural feature points in the assembly.

[0057] Step 2: Assembly Deviation Transfer Calculation: The spatial geometric dependency tree between each assembly mating surface is extracted from the standard bracket model and transformed into a directed graph structure assembly tolerance chain. Subpixel-level deviation data of key structural feature points are used as initial deviation inputs. The initial deviations are then transformed into a kinematic transformation matrix T along the directed graph structure assembly tolerance chain. , where R is the rotation matrix and t is the translation vector.

[0058] The cumulative deviation of the mating surfaces at the higher level is transferred to the tolerance range of the mating surfaces at the lower level through a multi-level cascaded product of the Jacobian matrix. The cumulative deviation transfer process for the k-th level mating surface is as follows: ,in For the kth The cumulative transformation matrix of the first-order mating surface. Let be the Jacobian matrix of the k-th order mating surface. Let be the cumulative transformation matrix of the k-th order mating surface. Jacobian matrix. The elements are the partial derivatives of the positional deviations of the geometric feature points of the mating surfaces with respect to the assembly posture deviations. These deviations are propagated step-by-step to the end of the assembly tolerance chain, and the comprehensive assembly deviation vector for each component region at the end of the assembly tolerance chain is calculated, thus completing the assembly deviation propagation solution.

[0059] Step 3: Generation of structural accuracy test results: Using the comprehensive assembly deviation vectors as node position constraints and the theoretical design dimensions between components in the standard support model as distance constraints, a spatial geometric accuracy constraint network for the support is constructed. Least squares global closed-loop adjustment is then performed on this constraint network. The core objective function of the global closed-loop adjustment is: , The constraints are: ,in For the optimal position estimate of node i, The measured position of node i (obtained by correction from the integrated assembly deviation vector), w i For observation weights, Let E be the theoretical design dimension between node i and node j, E be the set of all edges in the constraint network, and m be the total number of nodes. After adjustment, the geometric tolerance deviation values ​​of each assembly mating surface are output. The geometric tolerance deviation values ​​are compared with the preset industry factory size tolerance limits to generate structural accuracy test results used to characterize the overall geometric accuracy level of the support.

[0060] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for quality inspection of mobile phone holders based on image recognition, characterized in that, include: S1: Obtain a standard 3D model of a qualified mobile phone holder, extract the assembly constraint parameters and mechanical reference parameters of each assembly mating surface from the standard 3D model, construct a standard holder model with assembly constraint information; collect a multi-view detection image sequence of the mobile phone holder to be tested; S2: Input the standard support model and the multi-view detection image sequence into the topography deviation calculation network, perform joint feature encoding on the assembly constraint parameters and the multi-view detection image sequence to generate constraint feature maps and multi-view feature maps; perform constraint-guided cross-view registration and deviation calculation on the multi-view feature maps through the constraint feature maps, output the topography deviation field of each component region, identify components with abnormal assembly integrity in the topography deviation field, and generate component integrity detection results; S3: Using the topography deviation field and the mechanical reference parameters as input, identify surface defects and structural deformations from the topography deviation field to generate preliminary appearance defect detection results; construct a topography-mechanical response field, perform mechanical property mapping and quantitative evaluation on each defect area in the preliminary appearance defect detection results to obtain the mechanical performance influence quantification value corresponding to each defect area, and generate appearance defect detection results with risk level. S4: For component regions where the quantification value of the impact on mechanical properties is lower than the preset level, sub-pixel-level deviation data of key structural feature points of assembly are obtained from the morphological deviation field through gradient field tensor analysis. The sub-pixel-level deviation data is then used to perform assembly deviation transfer calculation along the assembly tolerance chain of the assembly mating surface in the standard bracket model to generate structural accuracy detection results.

2. The method according to claim 1, characterized in that, The standard scaffold model construction process specifically includes: Locate each assembly mating surface from the standard three-dimensional model, and calculate the spatial position parameters and mating tolerance parameters of each assembly mating surface as the assembly constraint parameters; Apply load boundary conditions simulating assembly conditions to each component area, solve the stress field distribution of each component area using finite element method, and extract the stress distribution reference value and strength safety factor from the solution results as the mechanical reference parameters; The assembly constraint parameters are associated with and stored in relation to the corresponding mating surfaces, and the mechanical reference parameters are associated with and stored in relation to the corresponding component regions, thereby constructing a standard bracket model with attached assembly constraint information.

3. The method according to claim 2, characterized in that, The specific steps of solving the stress field distribution of each component region using the finite element method include: Each component region is divided into tetrahedral meshes, and the surface height data of each component region is extracted from the standard 3D model. The surface height data is then converted into displacement constraints for the mesh nodes after mesh division. Based on the assembly constraint parameters in the standard support model, the load distribution of each assembly mating surface in the assembly state is obtained, and the load distribution is used as the load boundary condition of the mesh model. The displacement constraint condition and the load boundary condition are input into the static finite element solution process to obtain the stress tensor of each mesh node, forming a stress field distribution.

4. The method according to claim 1, characterized in that, The construction process of the topography deviation calculation network includes: An initial network is constructed, comprising a parametric coding layer, an image feature extraction layer, a constraint-guided attention layer, and a bias decoder. The initial network includes a three-dimensional geometric coding branch and a two-dimensional image feature extraction branch. The initial network is trained by taking multi-view training image samples with known topographic deviation annotations and the assembly constraint parameters of the corresponding standard scaffold model as inputs, thus forming a topographic deviation solution network.

5. The method according to claim 1, characterized in that, The process of generating the constraint feature map and the multi-view feature map specifically includes: The assembly constraint parameters in the standard support model are input into the three-dimensional geometric coding branch of the topography deviation solution network, and nonlinear feature mapping is performed through a multilayer perceptron to obtain a constraint feature map that characterizes the geometric topological relationship and tolerance range of the mating surface. The multi-view detection image sequence is input into the two-dimensional image feature extraction branch of the shape deviation solution network. Cross-channel and cross-scale feature extraction is performed using a densely connected convolutional neural network, and alignment is performed in the spatial dimension to obtain a multi-view feature map containing the geometric contours and texture features of the appearance of the mobile phone bracket under test from each viewpoint.

6. The method according to claim 1, characterized in that, The cross-view registration and bias calculation specifically include: The constraint feature map is used as the query vector, and cross-attention calculation is performed with the multi-view feature map to obtain the three-dimensional spatial reconstruction transformation matrix of each view image in the spatial coordinate system of the standard support model. Based on the transformation matrix, the multi-view feature map is projected into a unified three-dimensional space for point cloud-level registration to obtain a reconstructed three-dimensional model; The reconstructed 3D model is iteratively registered with the standard bracket model at the nearest point. The spatial distance deviation of each spatial grid node in the normal direction is calculated. The spatial distance deviations are arranged in order to generate a shape deviation field covering the region of each component of the mobile phone bracket to be tested.

7. The method according to claim 1, characterized in that, The construction process of the morphology-mechanical response field specifically includes: Extract the spatial three-dimensional discrete deviation values ​​of each grid node in the topography deviation field; use each spatial three-dimensional discrete deviation value as a boundary perturbation quantity to correct the spatial geometric position of the corresponding initial component grid in the standard support model under the standard coordinate system, and reconstruct the equivalent analysis entity with manufacturing topography defects; Based on the preset load and constraint conditions in the mechanical reference parameters, the equivalent analysis entity is subjected to structural mechanics calculation, and continuous stress response data characterizing the stress change trajectory of each defect region is output. The continuous stress response data is sequentially mapped back to the spatial coordinate nodes corresponding to the standard three-dimensional model to generate a topography-mechanical response field.

8. The method according to claim 7, characterized in that, The reconstructed equivalent analysis entity with manufacturing morphology defects also includes: Identify sensitive edge regions in the equivalent analysis entity where curvature abruptly changes due to geometric position correction; Obtain the baseline curvature of the sensitive edge region before the abrupt change, and calculate the curvature distortion rate of the sensitive edge region before and after the abrupt change; Using the curvature distortion rate, the stiffness response coefficient of the sensitive edge region in the structural mechanics solution process is linearly reduced, and the corrected stress response data containing the early deformation displacement evolution components caused by local morphological defects is output.

9. The method according to claim 1, characterized in that, The mechanical property mapping and quantitative evaluation specifically includes: The stress values ​​at each node in the topography-mechanical response field are compared with the corresponding standard stress reference values ​​in the mechanical reference parameters to extract the stress concentration response increment of each region. Based on the yield strength limit of the support material in the mechanical reference parameters, calculate the safety margin attenuation corresponding to each stress concentration response increment; When the safety margin attenuation exceeds the set safety boundary threshold, the risk level of the appearance defect detection result is determined by combining the stress concentration response increment, so as to quantitatively characterize the mechanical attenuation degree of the mobile phone bracket under actual service conditions.

10. The method according to claim 1, characterized in that, The gradient field tensor analysis specifically includes: Extract the normal deviation vectors of each spatial grid node in the neighborhood of the key structural feature point in the topography deviation field; calculate the gradient of the normal deviation vectors to construct the local topography gradient tensor matrix corresponding to the key structural feature point in the assembly. The local topography gradient tensor matrix is ​​decomposed into eigenvalues, and the eigenvectors corresponding to the eigenvalues ​​are extracted as the distortion gradient direction. The normal deviation vector is subjected to subpixel-level parabolic fitting extreme value extraction along the distortion gradient direction, and the subpixel-level deviation data of the assembly key structural feature points are output.

11. The method according to claim 1, characterized in that, The assembly deviation propagation calculation specifically includes: Extract the spatial geometric dependency tree between each assembly mating surface from the standard bracket model, and convert the spatial geometric dependency tree into an assembly tolerance chain with a directed graph structure; The subpixel-level deviation data of the key structural feature points of the assembly is used as the initial deviation input and converted into a kinematic transformation matrix along the assembly tolerance chain of the directed graph structure. Through the multi-level cascaded product of the Jacobian matrix, the cumulative deviation of the assembly mating surface of the previous level is transferred to the tolerance range of the assembly mating surface of the next level. Calculate the comprehensive assembly deviation vector of each component region at the end of the assembly tolerance chain to complete the assembly deviation propagation solution.

12. The method according to claim 1 or 11, characterized in that, The process of generating the structural accuracy test results specifically includes: Using the comprehensive assembly deviation vector as node position constraints and the theoretical design dimensions between components in the standard support model as distance constraints, a support spatial geometric accuracy constraint network is constructed. The least squares global closed-loop adjustment calculation is performed on the spatial geometric accuracy constraint network of the bracket to output the form and position tolerance deviation values ​​of each assembly mating surface; The form and position tolerance deviation value is compared with the preset industry factory size tolerance limit by a threshold comparison to generate a structural accuracy test result that characterizes the overall geometric accuracy level of the bracket.

Images