Robustness image restoration method for automobile part surface quality detection

Abstract

The invention provides a robust image restoration method for automobile part surface quality detection, and the method achieves the precise separation of target data and noise through the collection of the point cloud data of the automobile part surface, the construction of high-dimensional structured representation, and the combination of a noise separation data model and a multi-dimensional data decomposition technology. According to the method, the optimization model including the fidelity term, the regular term and the noise reduction term is constructed, the error between the recovered data and the collected data is optimized, data initialization and feature enhancement are performed through the feature extraction module, and the precision and geometric consistency of the recovered data are improved. And solving the model by adopting an iterative optimization method, and finally obtaining noise-free restored point cloud data. According to the method, noise can be effectively suppressed in a complex industrial environment, the recovery precision and robustness are improved, and efficient and reliable technical support is provided for surface quality detection of automobile parts.

Description

Technical Field

[0001] The present invention relates to the field of industrial vision, and in particular to a robust image restoration method for surface quality inspection of automobile parts, aiming to improve the accuracy and applicability of surface inspection of automobile parts in complex industrial environments through point cloud data modeling and restoration technology. Background Art

[0002] With the acceleration of the global industrialization process, manufacturing technology is constantly innovating and upgrading, especially in the field of automobile manufacturing. Product quality, production efficiency and automation level have become one of the important indicators to measure its core competitiveness. The surface quality of automobile parts, as an important measure of the quality and performance of the whole vehicle, plays a key role in improving the appearance consistency and product life of the whole vehicle. Therefore, higher requirements are placed on the quality inspection of the surface of automobile parts. Traditional manual inspection methods are difficult to meet the needs of modern industrialized production in terms of efficiency and accuracy. Especially in large-scale, high-speed production lines, manual inspection is often limited by the subjectivity of manual judgment and the complexity of the inspection environment, and there are problems such as missed inspections, false inspections and low efficiency. In recent years, the emergence of industrial visual inspection technology has provided a new solution for the surface quality inspection of automobile parts, and has gradually become an important means to replace manual inspection.

[0003] Industrial visual inspection systems usually use 3D point cloud data to obtain the geometric shape and feature information of the surface of parts. Point cloud data is a data form that can describe the geometric structure of objects with high precision. It is widely used in fields such as autonomous driving, reverse engineering, virtual reality and industrial inspection. By analyzing and processing 3D point cloud data on the surface of automotive parts, efficient identification of surface defects such as surface flaws, cracks, and dents can be achieved. However, point cloud data is usually interfered by a variety of complex environmental factors during the acquisition process, such as noise, uneven lighting, occlusion, reflection, etc., which leads to a decrease in the quality of the collected data, large noise or distortion, and directly affects the subsequent detection accuracy and effect. Especially in actual industrial environments, factors such as changes in ambient lighting, mechanical vibration, and sensor accuracy will have a negative impact on the quality of point cloud data acquisition, making it impossible for the detection system to effectively cope with the diverse noise interference on the surface of complex parts, reducing the detection accuracy and consistency.

[0004] In the prior art, point cloud data restoration methods have been gradually applied to the surface quality inspection of automobile parts. Traditional point cloud restoration methods usually use global or local smoothing techniques to eliminate noise to improve data quality, but these methods often find it difficult to strike a good balance between noise suppression and detail retention in complex geometric structures and texture scenes. Although the existing global smoothing technology can effectively eliminate large-area noise, it will cause edge details to be blurred or even misjudged as noise and removed, resulting in the loss of key geometric features. For local smoothing methods, although they can retain details to a certain extent, they often have poor effects when dealing with large-area unstructured noise, and it is difficult to meet the requirements of high precision and high stability for the surface quality inspection of automobile parts. In addition, the complex surfaces of automobile parts often have high reflectivity, curved surface features, and rich subtle structures and textures. These factors further increase the difficulty of point cloud data restoration and put forward higher applicability and robustness requirements for data restoration technology.

[0005] In order to balance noise suppression and detail preservation, some point cloud restoration algorithms have been improved based on different mathematical models and data features. Common methods include using sparse representation, low-rank matrix decomposition and deep learning technology to restore point cloud data. Among them, sparse representation and low-rank decomposition methods are based on the sparsity and low-rank characteristics of noise and signals. By constructing specific mathematical models, the noise components in the data are separated from the real signal, thereby removing noise and improving the quality of point cloud data. However, these methods often need to make specific assumptions about the type of noise, data characteristics, etc., such as assuming that the noise has a sparse distribution or presents a Gaussian distribution, which limits its applicability when dealing with diverse noises in complex industrial scenarios and cannot flexibly deal with different types of noise. When processing large-scale data, the low-rank matrix decomposition method has a high computational complexity and is difficult to meet the requirements of real-time detection.

[0006] In recent years, deep learning has achieved remarkable results in image restoration and noise suppression, and some image restoration methods based on deep neural networks have been applied to the field of point cloud data restoration. Deep learning methods usually build deep models such as convolutional neural networks or generative adversarial networks and use a large amount of labeled data for training, so that the model can achieve a good balance between denoising and detail preservation. However, deep learning methods also face many challenges when applied to actual industrial scenarios. First, deep learning models usually rely on large-scale labeled data, and it is very difficult to obtain large-scale and high-quality labeled data in point cloud data. Secondly, when dealing with complex surfaces of automotive parts, deep learning models often lack the ability to express details and are prone to losing key geometric features. In addition, the training and reasoning processes of deep learning models are computationally intensive and difficult to meet the requirements of real-time detection.

[0007] The existing point cloud restoration methods still have many deficiencies in terms of applicability and robustness in complex scenarios. Especially in the scene of surface inspection of automotive parts, the collected point cloud data has complex surface geometry and high reflectivity, which puts higher requirements on the applicability of the restoration method. When faced with such complex situations, traditional point cloud restoration methods often find it difficult to achieve effective noise suppression and detail retention at the same time, which easily leads to poor restoration results. Specifically, the deficiencies of existing methods are mainly reflected in the following aspects:

[0008] 1. Poor applicability: Existing methods usually rely on specific noise assumptions or geometric models, assuming that the noise is of a specific distribution or only processing data of regular geometric shapes. However, for automotive parts with complex shapes, surface features and multi-level details, this assumption is very limited and difficult to adapt to diverse industrial environments.

[0009] 2. Loss of detail information: Most existing point cloud restoration methods use global or local smoothing techniques to remove noise. Although this method can improve data quality to a certain extent, it will cause edge details to be blurred or misjudged as noise and removed, especially on the surface of automotive parts with complex geometric shapes. This loss of detail will seriously affect the accuracy of subsequent detection.

[0010] 3. High computational complexity: As the amount of point cloud data increases, traditional restoration algorithms often perform poorly in terms of computational efficiency. Especially in industrial production that requires large-scale real-time detection, restoration algorithms with high computational complexity are difficult to meet real-time requirements. Existing methods often fail to achieve high-quality restoration while taking into account efficiency.

[0011] In summary, in the field of surface quality inspection of automotive parts, existing point cloud data restoration methods have certain limitations in terms of applicability, detail retention ability and computational efficiency. Especially in the face of complex industrial environments and diverse noise interference, traditional methods are difficult to meet the needs of industrial production for high-precision and high-consistency inspection. To this end, a more robust, adaptable and efficient point cloud data restoration technology is urgently needed to improve the accuracy and applicability of surface quality inspection of automotive parts and provide reliable data support for subsequent inspection and quality control. This restoration technology not only needs to have the ability to adaptively suppress noise in complex noise environments, but also needs to be able to effectively retain the geometric features and detail information of the surface of automotive parts, so as to ensure that the quality of the restored data meets the requirements of high-precision inspection. Summary of the invention

[0012] In order to solve the above problems in the prior art, the present invention proposes a robust image restoration method for surface quality inspection of automobile parts, which is characterized by comprising the following steps:

[0013] Collecting automobile parts surface point cloud data containing three-dimensional geometric information, and constructing the automobile parts surface point cloud data into a high-dimensional structured representation;

[0014] The point cloud data includes target data and different types of noise components, and a data model for noise separation is established based on the point cloud data;

[0015] The point cloud data is processed by multidimensional data decomposition method, and the signal and noise are separated by the noise separation data model;

[0016] Construct an optimization model including fidelity terms, regularization terms, and noise reduction terms to optimize the error between the restored data and the collected data;

[0017] The data is initialized and feature enhanced through a feature extraction module; the optimization model is solved by an iterative optimization method, and finally noise-free restored point cloud data is obtained.

[0018] Using a third-order tensor Model the point cloud data of the surface of automobile parts and define the collected noise pollution point cloud data as tensor Y, satisfying Y=X+N g +S, where N g represents Gaussian noise, S represents sparse noise, and X represents the target tensor of noise-free point cloud data in the ideal case.

[0019] The point cloud data is represented as a high-dimensional target tensor X = Φ(G), where G = {G (1) ,G (2) ,…,G (n)} is a third-order core tensor sequence, and each core tensor The weights of different dimensions are adaptively adjusted through the third-order tensor ring decomposition structure to achieve robust processing of different noise types and data missing situations, where r n-1 and r n represents the rank of the tensor ring; I n is the number of features in the nth dimension.

[0020] The point cloud image is restored by optimizing the model, and the optimization model is:

[0021]

[0022] in, is the fidelity term, g(X) is the non-local denoising function; the parameter Y represents the collected point cloud data, which is the original data containing noise and is used as the input of the model; the parameter X is the restored point cloud data; the sparse noise S represents the non-structural noise in the collected data; the tensor decomposition result Φ(G) is the structured expression of the point cloud data; the parameter β1 controls the matching degree between the restored data X and the tensor decomposition result Φ(G), the parameter β2 is used to adjust the regularization strength of the sparse noise S, and the parameter μ affects the balance between the non-local denoising function g(X) in denoising and detail preservation.

[0023] Using deep convolutional neural network f θ (X0) performs feature extraction and data initialization, where the input is a randomly initialized tensor X0, and the image edge and texture detail information is extracted through convolution operations, and the regional threshold is calculated using grayscale average. The image reconstruction is completed; where u1 and u2 represent the grayscale averages of adjacent areas respectively.

[0024] The optimization model is solved step by step using the alternating direction method of multipliers (ADMM), including the following steps:

[0025] Solving for G (1) Sub-problem: The optimization objective is in and is the modal expansion of the core tensor;

[0026] Solving for G (2) Sub-problem: Update

[0027] Solving for G (3) Sub-problem: Update

[0028] Solve the subproblem of S: The optimization goal is And update S (t+1) =shr ink(Y-Φ(G) t ,μ);

[0029] Solve the subproblem of X: Update

[0030] Subproblem of solving θ: Update

[0031] Where: G (1) ,G (2) ,G (3) It is a third-order core tensor; it expresses the local features and geometric structure of point cloud data through tensor ring decomposition;

[0032] Tensor G(1) The second mode expansion and G (2) , G (3) The multimodal expansion results of

[0033] fo ld2: modal contraction operation;

[0034] The expanded form of the tensor X in different modes; the matrices after the first mode, second mode, and third mode of X are respectively represented;

[0035] t+1 represents the iterative update;

[0036] G (t+1) : tensor decomposition result updated in the t+1th iteration;

[0037] S (t+1) : The sparse noise after optimization in the t+1th iteration;

[0038] X (t+1) : The restored tensor calculated in the t+1th iteration;

[0039] Y: collected point cloud data;

[0040] Φ(G) t : tensor decomposition result of the current iteration;

[0041] The square of the Froben i us norm;

[0042] S: sparse noise;

[0043] shr ink(x,μ): soft threshold function;

[0044] μ: threshold parameter of soft threshold function;

[0045] β1,β2: penalty parameters of the optimization model;

[0046] g(X): non-local denoising function;

[0047] f θ (X0): feature tensor extracted by the deep network;

[0048] θ: weight parameter of the deep network;

[0049] X0: Initial input tensor.

[0050] Beneficial effects:

[0051] The present invention provides a robust image restoration method for surface quality inspection of automobile parts. By constructing point cloud data as a high-dimensional structured representation, combined with a noise separation data model and multidimensional data decomposition technology, the target data and different types of noise are accurately separated. The present invention significantly improves the accuracy and geometric consistency of the restored data by introducing an optimization model of fidelity terms, regularization terms and noise reduction terms, while effectively suppressing noise interference. The feature extraction module is used for data initialization and feature enhancement to make the detailed expression of the restored data richer. Through the iterative optimization method, the convergence and efficiency of the restoration process are further ensured. The present invention can perform high-quality restoration of point cloud data on the surface of automobile parts in complex industrial environments, significantly improve detection accuracy and robustness, and provide strong technical support for industrial quality inspection. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present application, but do not constitute an improper limitation of the present invention. In the drawings:

[0053] Figure 1 A flow chart of the method of the present invention is shown. DETAILED DESCRIPTION

[0054] The present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments, wherein the illustrative embodiments and descriptions are only used to explain the present invention but are not intended to limit the present invention.

[0055] This embodiment provides a method for restoring noise-contaminated surface point cloud data of automobile parts through tensor modeling and decomposition. The method uses a third-order tensor to finely model and restore point cloud data, solving the technical problem of the difficulty in balancing noise suppression and detail retention in complex industrial environments in the prior art. The implementation process of the method will be described in detail below in conjunction with the accompanying drawings.

[0056] Step 1: Tensor modeling of point cloud images of automobile parts surface

[0057] The present invention adopts a third-order tensor To represent the point cloud data of the surface of automobile parts, M, N, and P represent the spatial length, width, and number of channels of the tensor, respectively, so as to construct a high-dimensional data structure with spatial and channel dimensions. Through the tensor modeling method, the geometric information and texture features of the surface of automobile parts can be effectively expressed in high-dimensional space, thereby providing an accurate structural description for the subsequent data restoration process.

[0058] Specifically, the collected point cloud data is defined as a tensor Y. In the actual collection process, the tensor is usually contaminated by various noises in the industrial environment, resulting in data distortion. In order to accurately characterize such noise, the present invention assumes that the point cloud image data Y can be decomposed into three parts: target tensor X, Gaussian noise N g and sparse noise S, that is, satisfying the relationship Y = X + N g +S. Among them, the target tensor X represents the noise-free point cloud data in an ideal situation, which is used to truly reflect the structural information of the surface of automobile parts; Gaussian noise N g It is the noise component introduced by random factors such as ambient light and sensor jitter during the acquisition process, which manifests as overall interference; sparse noise S represents some random strong interference, such as reflection, occlusion and other strong interference signals unique to industrial environments. These noise components are mixed together to reduce the quality of the collected data Y, thereby affecting the subsequent surface quality inspection results.

[0059] Therefore, the goal of this embodiment is to process the noise-affected tensor Y into a clean target tensor X through tensor restoration technology, that is, to extract clean point cloud information from the noise, so as to more accurately describe the surface morphology and detail features of automotive parts and provide high-quality data support for quality inspection.

[0060] The present invention constructs a representation method of high-dimensional space through third-order tensor modeling, which effectively improves the expression ability of noise and details in point cloud data, so that the noise components can be removed more accurately in the subsequent restoration processing.

[0061] Step 2: Non-local Deep Tensor Ring Decomposition Model

[0062] In order to achieve efficient and accurate restoration of point cloud data, this embodiment adopts a non-local deep tensor ring decomposition model. This model decomposes the high-dimensional representation of the tensor to extract the core structural features of the data, while effectively suppressing complex noise and retaining high details.

[0063] First, the point cloud data of the surface of automobile parts is represented as a high-dimensional tensor X, which constructs the structural relationship of the point cloud data through the cyclic multilinear product of a series of third-order tensor ring factors. In this structural relationship, the core features of the point cloud data are efficiently represented through multi-level decomposition. Specifically, through tensor ring decomposition, the tensor X is represented as X = Φ (G), where G = {G (1) ,G (1) ,…,G (N)} is a sequence of third-order core tensors, representing local features and multi-dimensional geometric structure relationships in the data. Each core tensor Defines the characteristic components of point cloud data in a specific dimension, where r n-1 and rn The rank of the tensor ring is a parameter determined by the specific structural characteristics of the data.

[0064] The core purpose of tensor ring decomposition is to compactly represent the original point cloud data as the product of multiple low-rank tensors. This compact representation can adaptively adjust the weights between different dimensions to a certain extent, so that the model can adapt to a variety of noise environments and data loss situations. Especially in complex industrial environments, this method can significantly improve the robustness to noise. Even if there are large areas of noise or missing areas in the point cloud data, the decomposition model can still maintain an accurate expression of the real signal through the low-rank structure. In addition, through this ring decomposition method, the model can discriminate different types of noise, thereby achieving a good balance between detail preservation and noise suppression.

[0065] Furthermore, the formula X = Φ(G) describes a tensor decomposition method for high-dimensional point cloud data, where the point cloud data X is represented as a structured expression generated by a third-order tensor ring decomposition. This tensor ring decomposition method is an efficient and compact data representation technology that can decompose high-dimensional data into a combination of multiple low-rank tensors, facilitating the extraction of data features and the processing of noise. Through this decomposition structure, the weights of different dimensions of the data can be adaptively adjusted, thereby effectively suppressing noise interference and having the robustness to handle data missing.

[0066] In this formula, X represents the restored point cloud data tensor. Ideally, it is a high-dimensional structured representation of noise-free point cloud data, which can accurately preserve the geometric shape and texture characteristics of the surface of automotive parts. Φ(G) is an expression generated by tensor ring decomposition, which represents the high-dimensional structure of point cloud data and consists of multiple third-order tensor factors G (1) ,G (2) ,…,G (1) The core tensor sequence G is the key to describing the local features and multi-dimensional geometric structure of point cloud data. Each core tensor G (n) The dimension is r n-1 ×I n × n Here, r n-1 and r n is the rank of the tensor ring, which determines the compactness of the decomposition, and I n It is the feature number of the nth dimension, which is used to reflect the information complexity of this dimension.

[0067] Tensor ring decomposition is an efficient multidimensional data decomposition method, and its specific expression is: In this expression, Trace is the matrix trace operation, which is used to simplify the multilinear product results between core tensors into scalars, and G(n) (i n ) is the core tensor G (n) In the i n Through this structure, the overall features of the point cloud data can be compactly represented as a combination of low-rank tensors, making data storage and calculation more efficient while maintaining high fidelity of the geometric structure and texture features of the original data.

[0068] This decomposition structure has multiple functions. First, it can adaptively adjust the weights between different dimensions so that the decomposition results can better adapt to the characteristic distribution of point cloud data. Second, through the decomposition model, different types of noise such as sparse noise and Gaussian noise can be effectively separated and processed to improve the quality of point cloud data. In addition, even if part of the point cloud data is missing, this decomposition method can still maintain an accurate expression of the overall geometric features through its compact structure, thus having strong robustness. Finally, by reducing data redundancy through low-rank decomposition, the efficiency of data processing and the economy of storage are further improved.

[0069] The specific meanings of the parameters can be further clarified: X is the restored point cloud data tensor, which represents the high-dimensional data result generated during the restoration process; G is the core parameter sequence of tensor decomposition, which is used to compactly represent the features of point cloud data; G (n) It is the basic unit of decomposition, and each core tensor captures the features of a specific dimension.

[0070] Through the synergy of these parameters and the tensor ring decomposition formula, efficient restoration of the surface point cloud data of automotive parts can be achieved. Whether it is noise suppression or geometric feature retention, this decomposition method shows strong adaptability and provides high-quality data support for subsequent quality inspection tasks.

[0071] Step 3: Optimize the model definition

[0072] In order to achieve efficient restoration of point cloud data, this embodiment constructs an optimization model that uses multiple fidelity terms and balance parameters to ensure the best balance between noise suppression and detail preservation. Specifically, the goal of the optimization model is to minimize the error between the restored data and the acquired data, while suppressing noise and preserving the geometric details and texture features in the data. The model is defined as follows:

[0073]

[0074] This optimization model is used to restore point cloud image data through mathematical optimization methods. It comprehensively considers data fidelity, regularization constraints and noise reduction effects in order to solve the noise problem in the collected data. The goal of the model is to efficiently process the collected data and generate high-quality restored images, thereby providing reliable support for subsequent surface quality inspection.

[0075] This optimization goal finds the best restoration result of the point cloud image by comprehensively optimizing the restored data X, sparse noise S and tensor decomposition core parameter G. is the fidelity term, which measures the error between the restored data X and the collected data Y, while taking into account the impact of sparse noise S on the collected data. This fidelity constraint ensures that the restored point cloud data X can be as close as possible to the original collected data Y, thereby retaining the core information of the collected data.

[0076] Item 2 is a regularization term used to constrain the difference between the restored data X and the tensor ring decomposition result Φ(G). Through this regularization constraint, the restored data X can conform to the structural characteristics of the tensor decomposition, thereby improving the performance of the restored data in terms of geometric consistency and texture detail preservation. The introduction of regularization effectively improves the applicability of the restored data in complex environments and makes the restoration results more accurate.

[0077] Item 3 It is a regularization constraint on the sparse noise S, which aims to ensure the sparsity of S and thus suppress the influence of unstructured noise. This constraint avoids overfitting of the model to random interference while retaining the key features of the data.

[0078] The fourth term g(X) is a non-local denoising function, which is used to extract non-local features in the restored data X. Through this function, the model can further optimize the smoothness of the restored data and enhance geometric consistency while suppressing local noise, thereby improving the overall quality of the point cloud image.

[0079] Parameter Y represents the collected point cloud data, which is the original data containing noise and serves as the input of the model. Parameter X is the restored point cloud data. The optimization goal is to generate a noise-free, high-quality point cloud image through model processing. Sparse noise S represents the non-structural noise in the collected data, such as random interference such as occlusion and reflection. The tensor decomposition core parameter G is an important part of the model. The core tensor sequence generated by tensor ring decomposition is used to describe the geometric and texture features of the point cloud data. The tensor decomposition result Φ(G) is a structured expression of the point cloud data, ensuring that the restored data conforms to the essential properties of the data.

[0080] The weight parameters β1, β2, and μ in the model determine the relative importance of each constraint. Parameter β1 controls the matching degree between the restored data X and the tensor decomposition result Φ(G), parameter β2 is used to adjust the regularization strength of the sparse noise S, and parameter μ affects the balance between the non-local denoising function g(X) in denoising and detail preservation.

[0081] The square of the Frobenius norm It is used to calculate the Euclidean distance between matrices or tensors and is a commonly used error metric. By using the norm, the model can intuitively quantify the differences between data and provide a clear goal for optimization.

[0082] In summary, the optimization model comprehensively balances data fidelity, geometric consistency and noise reduction effects through the synergy of fidelity terms, regularization terms, sparse noise terms and non-local noise reduction terms. Under the regulation of penalty parameters, the model can effectively generate high-quality, noise-free point cloud data, providing reliable technical support for subsequent surface quality inspection. Through this optimization model, the restoration accuracy of point cloud images has been significantly improved, and the robustness in complex industrial environments has also been enhanced.

[0083] In order to further improve the restoration accuracy, this embodiment also uses a deep convolutional neural network f θ (X0) is used for feature extraction and data initialization. The deep convolutional neural network uses its powerful feature extraction ability to effectively identify and retain the edge and texture details of the image, which helps to build a clearer restored image. In the initialization stage, the model input is a randomly initialized tensor X0, and the gradient features of the tensor are extracted after multiple layers of convolution operations. The convolution process enhances the detail performance of the image by identifying the edge changes of the image.

[0084] In the feature extraction process, the model uses a convolutional network to generate a gradient image Z and calculates the threshold by averaging the regional grayscale. Among them, u1 and u2 represent the grayscale average values ​​of adjacent areas. Through the threshold calculation and feature reconstruction process, the model can accurately capture the local features and edge information in the image, ensuring that the restored point cloud data has high resolution and detail expression, thus providing a more reliable data basis for subsequent quality inspection.

[0085] Step 4: ADMM solves the model

[0086] In order to effectively solve the above optimization model, this embodiment uses the alternating direction multiplier method (ADMM) to optimize each variable in the model step by step. The ADMM method decomposes the complex optimization problem into a series of relatively simple sub-problems, updates each variable in a step-by-step iteration, and finally converges to the global optimal solution. The specific solution process is as follows:

[0087] Solving for G (1) Sub-problem

[0088] To optimize the tensor G (1) To express , we first need to minimize the following objective function:

[0089]

[0090] in, G (1) The second modal expansion of Then the tensor G (2) and G (3) The second mode expansion after multiple linear products. This step can optimize G (1) The core tensor expression of can better capture and adapt to different noise structures in point cloud data, thereby more effectively improving the quality of restored data.

[0091] Solving for G (2) Sub-problem

[0092] In solving G (2) When updating, the calculation formula is:

[0093]

[0094] This step obtains G by expanding and solving the multi-modal tensor. (2) The updated value of G is used to further reduce the impact of noise on the data. Through this process, G (2) A more optimized expression is obtained, which can more accurately represent the feature information in the point cloud data.

[0095] Solving for G (3) Sub-problem

[0096] For the tensor G (3) The update is calculated as follows:

[0097]

[0098] This step is in G (3) The optimization is performed under the multi-modal expansion to further extract and preserve the geometric structure information of the point cloud data. This optimization process helps the model to more accurately restore the point cloud details in a complex data environment.

[0099] Solve the subproblem of S

[0100] The solution goal of sparse noise S is as follows:

[0101]

[0102] After obtaining the optimal solution, the value of S is updated to:

[0103] S (t+1) =shr ink(Y-Φ(G) t ,μ)

[0104] This step effectively removes the noise in S through sparse constraints, making it more consistent with the real data distribution. In the sparse noise optimization process, the soft threshold function shr ink is introduced to suppress S, so that the model can effectively reduce the impact of noise while retaining more real image information.

[0105] Solve the subproblem of X

[0106] For the solution of tensor X, iterate through the following update formula:

[0107]

[0108] In this step, by iteratively updating X, the restored image can be gradually approached. The g(X) in the formula is an improved non-local denoising function that can further suppress noise and ensure the clarity and detail integrity of X. This iterative update process helps the model extract a restored image with high fidelity in a complex noisy environment.

[0109] The parameters are described as follows:

[0110] X (t+1) : The restored point cloud data obtained by the current iteration. The point cloud data obtained after the update through this iteration calculation gradually approaches the ideal restoration result without noise.

[0111] Y: The original point cloud data collected. The input data containing noise is the basis of the restoration process.

[0112] S (t+1) : The sparse noise of the current iteration. The latest S value obtained by optimizing the sparse noise subproblem is used to remove the non-structural noise components in the point cloud data.

[0113] β1Φ(G (t+1) ): Tensor ring decomposition constraint. Φ(G (t+1) ): The latest iteration of tensor ring decomposition result, representing the structured representation of point cloud data.

[0114] β1: Weight parameter of the tensor ring decomposition constraint, used to adjust the impact of ring decomposition on the restoration result.

[0115] The geometric features of the restored data X are kept consistent with the tensor decomposition structure to ensure that the restored result conforms to the high-dimensional structure of the data.

[0116] g(X): Non-local denoising function. The non-local denoising constraint imposed on X captures the non-local characteristics of the data, smooths the noise and retains important detail information. The restored image is further optimized to ensure clarity and detail integrity.

[0117] 1+β1+μ: normalization factor.

[0118] 1: The default weight of the collected data Y.

[0119] β1: Weight parameter of the tensor ring decomposition constraint.

[0120] μ: Weight of the non-local denoising function.

[0121] Balance the weights of various constraints to ensure the rationality of the restoration results.

[0122] Solve the subproblem of θ

[0123] In order to further improve the restoration accuracy of the model, the solution goal for the deep network weight θ is:

[0124]

[0125] By optimizing the weights θ of the deep network, this step further improves the model's ability to identify and suppress noise, making it more robust in complex data environments. The optimization of the deep network can enhance the model's ability to capture edge features and detail information, thereby improving the accuracy and clarity of the restored image.

[0126] The further explanation of the above solution process is:

[0127] In order to effectively solve the above optimization model, the present invention uses the alternating direction multiplier method (ADMM) to optimize each variable in the model step by step. The ADMM method is an iterative optimization algorithm that decomposes a complex optimization problem into a series of simpler sub-problems, gradually updates each variable, and finally achieves an approximation of the global optimal solution. This method can effectively balance the fidelity, noise suppression and geometric consistency of the data, ensuring that the restored point cloud data has higher accuracy and robustness.

[0128] First, optimize the tensor G (1) The objective function is optimized by minimizing the error Update tensor G (1) Here, Represents the tensor G (1) The second modal expansion of is the tensor G (2) and G (3) The second mode expansion result of the multilinear product of . This step optimizes G(1) The core expression ability of point cloud data can more accurately capture the geometric features in point cloud data while reducing noise interference.

[0129] Next, optimize the tensor G (2) The update formula is in, is the second modal expansion of the tensor X, and It's G (3) and G (1) The first mode expansion result of the multilinear performance of . This step optimizes G through multi-mode expansion and solution. (2) The expression method makes it more adaptable to the interference of noise in point cloud data and further improves the quality of restored data.

[0130] Then, for the tensor G (3) Optimize. The update formula is In this process, through the G (1) and G (2) The multimodal expansion results of Utilization, optimization of G (3) The value of is used to better preserve the geometric features in the point cloud data. This step further enhances the expressiveness and applicability of the model in complex data environments.

[0131] Then, the expression of sparse noise S is optimized. The objective function is By using the sparse constraint, the update formula is S (t+1) =shr ink(Y-Φ(G) t ,μ), where shr ink is a soft threshold function used to control the sparsity of noise. This step effectively suppresses non-structural noise, making the model more consistent with the actual data distribution while avoiding overfitting of noise.

[0132] When optimizing the tensor X, the update formula is used Among them, g(X) is a non-local denoising function, which further enhances the smoothness and consistency of the restored data. Through this iterative update, the tensor X is gradually optimized to make it closer to the accurately restored high-quality image data.

[0133] Finally, the deep network weight θ is optimized. The objective function is By optimizing and adjusting the weights of the deep network, the model can better identify and suppress noise, while enhancing the ability to capture edge features and detail information in point cloud data, thereby further improving the accuracy and clarity of the restored image.

[0134] Through the above step-by-step optimization, the ADMM method decomposes the complex global optimization problem into multiple independent but closely related sub-problems. The solution goal of each sub-problem is to optimize the requirements of specific variables, thereby ensuring that the fidelity, noise suppression ability and geometric consistency of the restored data are at a high level. Ultimately, this method significantly improves the accuracy and robustness of point cloud data restoration, especially in complex industrial environments, showing excellent adaptability, providing strong technical support for surface quality inspection of automotive parts.

[0135] Through the above-mentioned step-by-step solution, the ADMM method effectively decomposes the original complex optimization problem into multiple manageable sub-problems and gradually converges to the global optimal solution. Ultimately, this embodiment can achieve effective noise suppression and detail retention while ensuring data accuracy, providing high-quality restoration data for surface quality inspection of automotive parts.

[0136] Through the above steps, this embodiment achieves high-precision restoration of the surface quality of automobile parts, and shows significant advantages in noise suppression and detail retention in complex industrial environments.

[0137] The above description is only a preferred embodiment of the present invention, so all equivalent changes or modifications made according to the structure, characteristics and principles described in the scope of the patent application of the present invention are included in the scope of the patent application of the present invention.

Claims

1. A robust image restoration method for surface quality inspection of automobile parts, characterized by: The following steps are involved: Collecting automobile parts surface point cloud data containing three-dimensional geometric information, and constructing the automobile parts surface point cloud data into a high-dimensional structured representation; The point cloud data includes target data and different types of noise components, and a data model for noise separation is established based on the point cloud data; The point cloud data is processed by multidimensional data decomposition method, and the signal and noise are separated by the noise separation data model; Construct an optimization model including fidelity terms, regularization terms, and noise reduction terms to optimize the error between the restored data and the collected data: The data is initialized and feature enhanced through a feature extraction module; the optimization model is solved by an iterative optimization method, and finally noise-free restored point cloud data is obtained.

2. The robust image restoration method for automobile parts surface quality inspection according to claim 1, characterized in that: Using a third-order tensor Model the point cloud data of the surface of automobile parts and define the collected noise pollution point cloud data as tensor Y, satisfying Y=X+N g +S, where N g represents Gaussian noise, S represents sparse noise, and X represents the target tensor of noise-free point cloud data in the ideal case.

3. The robust image restoration method for automobile parts surface quality inspection according to claim 1, characterized in that: The point cloud data is represented as a high-dimensional target tensor X = Φ(G), where G = {G (1) , G (2) , ..., G (n) } is a third-order core tensor sequence, and each core tensor The weights of different dimensions are adaptively adjusted through the third-order tensor ring decomposition structure to achieve robust processing of different noise types and data missing situations, where r n-1 and r n represents the rank of the tensor ring; I n is the number of features in the nth dimension.

4. The robust image restoration method for automobile parts surface quality inspection according to claim 1, characterized in that: The point cloud image is restored by optimizing the model, and the optimization model is: in, is the fidelity term, g(X) is the non-local denoising function; the parameter Y represents the collected point cloud data, which is the original data containing noise and is used as the input of the model; the parameter X is the restored point cloud data; the sparse noise S represents the non-structural noise in the collected data; the tensor decomposition result Φ(G) is the structured expression of the point cloud data; the parameter β1 controls the matching degree between the restored data X and the tensor decomposition result Φ(G), the parameter β2 is used to adjust the regularization strength of the sparse noise S, and the parameter μ affects the balance between the non-local denoising function g(X) in denoising and detail preservation.

5. The robust image restoration method for automobile parts surface quality inspection according to claim 1, characterized in that: Using deep convolutional neural network f θ (X0) performs feature extraction and data initialization, where the input is a randomly initialized tensor X0, and the image edge and texture detail information is extracted through convolution operations, and the regional threshold is calculated using grayscale average. The image reconstruction is completed; where u1 and u2 represent the grayscale averages of adjacent areas respectively.

6. The robust image restoration method for automobile parts surface quality inspection according to claim 1, characterized in that: The optimization model is solved step by step using the alternating direction method of multipliers (ADMM), including the following steps: Solving for G (1) Sub-problem: The optimization objective is in and is the modal expansion of the core tensor; Solving for G (2) Sub-problem: Update Solving for G (3) Sub-problem: Update Solve the subproblem of S: The optimization goal is And update S (t+1) =shrink(Y-Φ(G) t , μ); Solve the subproblem of X: Update Subproblem of solving θ: Update Where: G (1) , G (2) , G (3) It is a third-order core tensor; it expresses the local features and geometric structure of point cloud data through tensor ring decomposition; Tensor G (1) The second mode expansion and G (2) , G (3) The multimodal expansion results of fo|d2: modal contraction operation; The expanded form of the tensor X in different modes; the matrices after the first mode, second mode, and third mode of X are respectively represented; t+1 represents the iterative update; G (t+1) : tensor decomposition result updated in the t+1th iteration; S (t+1) : The sparse noise after optimization in the t+1th iteration; X (t+1) : The restored tensor calculated in the t+1th iteration; Y: collected point cloud data; Φ(G) t : tensor decomposition result of the current iteration; The square of the Frobenius norm; S: sparse noise; shrink(x, μ): soft threshold function; μ: threshold parameter of soft threshold function; β1, β2: penalty parameters of the optimization model; g(X): non-local denoising function; f θ (X0): feature tensor extracted by the deep network; θ: weight parameter of the deep network; X0: Initial input tensor.

Images