Multi-view aluminum product identification method, system, equipment and medium

Abstract

The invention discloses a multi-view aluminum product identification method, system, equipment and medium, and relates to the technical field of object identification, and the method comprises the steps: superposing subspace representation matrixes of different views of an aluminum product into a tensor, applying low-rank constraint to the tensor, introducing Frobenius norm constraint to process the subspace representation matrixes, and obtaining a subspace representation matrix; carrying out sparse constraint on the transpose of the consistent subspace representation matrix in the processed class and the product of the matrix, and obtaining a target function for identifying the multi-view aluminum product; and by optimizing various results in the target function and applying spectral clustering to the optimized subspace representation matrix when a convergence condition is met, a final clustering result is obtained, and the clustering result is used as an aluminum product identification result. The block diagonal structure of the subspace representation matrix is enhanced by performing sparse constraint on transpose of the subspace representation matrix and the product of the subspace representation matrix, so that the recognition performance of the construction method on the aluminum product can be improved to a certain extent.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of object recognition, in particular to a multi-view aluminum product recognition method, system, device and medium. BACKGROUND

[0002] Aluminum and its alloys have unique material properties and are widely used in important industries such as aviation, construction and automobiles. Due to the characteristics and rich application prospects of this new type of metal aluminum, its production and use have been greatly promoted. Aluminum has a very wide range of applications and provides many important materials and products for various industries.

[0003] When recycling or otherwise processing aluminum products, the aluminum products need to be identified to achieve fast recycling and processing. In the existing field, multi-view learning is often used to identify aluminum products. At the same time, multi-view subspace clustering, as an important branch of multi-view learning clustering, has certain advantages in clustering tasks, but it does not usually consider the high-order correlation between different views in multi-view data.

[0004] In the prior art, tensors are high-order generalizations of vectors and matrices, which can explore the high-order relationships in multi-view data and mine the consistency information between different views. In the field, the diversity graph and consistency graph of aluminum products are combined into a tensor, and tensor kernel norm constraints are applied to it. This not only obtains complete information between multiple views, but also enables mutual learning and mutual enhancement of the diversity graph and the consistency graph, thereby achieving accurate identification of aluminum products.

[0005] Existing research shows that although existing multi-view subspace clustering methods based on tensor constraints have shown considerable advantages in aluminum product identification tasks, they still have some problems: existing subspace clustering methods generate subspace representation matrices through tensor constraints when identifying aluminum products. However, experiments have shown that the generated subspace representation matrices have poor intra-class consistency and insufficient inter-class sparsity, and when fusing multi-view data information to construct subspace representations, it is difficult to guarantee the integrity and consistency of the structure, resulting in poor accuracy of the final aluminum product identification results. SUMMARY

[0006] The present application aims to overcome the shortcomings of the prior art and provide a multi-view aluminum product recognition method, system, device and medium to solve the problems of the prior art.

[0007] The present application specifically provides the following technical solutions: a multi-view aluminum product recognition method, comprising the following steps: Obtaining a subspace representation matrix of different views of an aluminum product image; The subspace representation matrices of different views of the aluminum product are superimposed into a tensor, a low rank constraint is applied to the tensor, and a Frobenius norm constraint is introduced to process the subspace representation matrices, to obtain the subspace representation matrices with intra-class consistency, and the product of the transpose of the subspace representation matrices with itself is subjected to a sparsity constraint; The result nuclear norms of the tensor, the low rank constraint, the Frobenius norm constraint and the sparsity constraint are balancedly weighted to obtain a target function of multi-view aluminum product recognition; The result nuclear norms in the target function of multi-view aluminum product recognition are optimized through an iterative updating rule, and when the optimized target function satisfies a convergence condition, the optimized sparsity constraint is subjected to spectral clustering to obtain a final clustering result, which is taken as the aluminum product recognition result.

[0008] Preferably, the result nuclear norms of the tensor, the low rank constraint, the Frobenius norm constraint and the sparsity constraint are balancedly weighted to obtain a target function of multi-view aluminum product recognition, and the specific expression is: ; Among them, is the result nuclear norm of the tensor, F (v) is the subspace representation matrix of the vth view, S (v) is the reconstruction error matrix of the vth view, is a low rank constraint term, 、 and is a balance parameter, is a sparsity constraint term, T is a transpose, is a Frobenius norm constraint, V is the total number of views, A (v) is the data matrix corresponding to the vth view, and is a 3-order tensor constructed by merging the subspace representation matrices of different views (v) , S is a matrix formed by splicing the reconstruction error matrices of each view.

[0009] Preferably, the result nuclear norms in the target function of multi-view aluminum product recognition are optimized through an iterative updating rule, and the specific process is as follows: An auxiliary variable G n is introduced to replace the matrix F (n) expanded from the tensor, and an ALM method is used to obtain an intermediate target function; The subspace representation matrix F (v) , the reconstruction error matrix S, the auxiliary variable G n and the balance parameter μ, Lagrange multiplier B v and α n , until the convergence condition is met.

[0010] Preferably, the introduced auxiliary variable G n is substituted into the matrix F (n) after tensor expansion, and the ALM method is used to obtain an intermediate target function, specifically: The augmented Lagrange multiplier method target function is used for optimization, and the specific expression is: ; wherein f and g n are the vectorization of the optimized tensor F and auxiliary variable G n , P n is a permutation matrix, t n , σ and τ are different balancing parameters, N is the number of auxiliary variables, is the result kernel norm of the auxiliary variable; The ALM method is used to update the optimized result to obtain an intermediate target function, and the specific expression is: ; wherein, is the intermediate target function value.

[0011] Preferably, the intermediate target function is solved by iteratively updating the subspace representation matrix F (v) , the reconstruction error matrix S, the auxiliary variable G n , the balancing parameter μ , the Lagrange multiplier B v and α n , specifically: The update rule of the subspace representation matrix in the intermediate target function is obtained, and the specific expression is: ; wherein, , indicates that the element corresponding to the first v view is selected, I is an identity matrix, I1 indicates a matrix full of 1, σ is a balancing parameter; When F (v) and the auxiliary variable G n are fixed, the matrix A (v) -A (v) F (v) +B vA matrix Z is constructed by vertically connecting the columns of A, and the i-th column of Z is denoted as Zi. i The reconstructed error matrix is obtained by optimizing the relationship between the columns of Z , and the specific expression is: ; wherein, Zi represents the i-th column of the matrix Z; The auxiliary variable is updated by iteration , and the specific expression is: ; wherein, the parameter , is obtained according to the n-th mode of the expansion P n f+ a n is reshaped into the corresponding matrix, prox is the proximal operator, tr represents the constraint condition of the truncated nuclear norm; The update rule of the balance parameter μ is obtained by minimizing the relationship between the balance parameter and the maximum balance parameter, and the specific expression is: μ ; wherein, is the correlation coefficient, is the maximum balance parameter; The update rule of the Lagrange multiplier is obtained by summing the matrix A (v) -A (v) F (v) -S (v) or the matrix P n f-g n and the response initial Lagrange multiplier, and the specific expression is: ; ; wherein, is the updated Lagrange multiplier based on B v , and is the updated Lagrange multiplier based on α n .

[0012] Preferably, the optimized sparse constraint is subjected to spectral clustering to obtain the final clustering result, and the specific expression is: The clustering matrix is calculated based on the updated subspace representation matrix F (v) , and the specific expression is: ; ​Apply the spectral clustering algorithm to the clustering matrix E to output the final clustering results.

[0013] Preferably, when obtaining the subspace representation matrix of different views of the aluminum product image, the method further includes: extracting low-frequency components in the aluminum product image as views using Coiflets, Daubechies, and Symlets techniques, and using principal component analysis (PCA) to reduce the dimension of each view to construct the subspace representation matrix.

[0014] This invention provides a multi-view aluminum product identification system, comprising: The data acquisition module is used to acquire the subspace representation matrix of different views of the aluminum product image; The constraint module is used to superimpose the subspace representation matrices of different views of aluminum products into a tensor, apply low-rank constraints to the tensor, and introduce Frobenius norm constraints to process the subspace representation matrices to obtain subspace representation matrices with intra-class consistency. Sparse constraints are applied to the product of the transpose of the intra-class consistent subspace representation matrix and itself. The objective construction module is used to balance and weight the nuclear norms of the results of tensors, low-rank constraints, Frobenius norm constraints and sparse constraints to obtain the objective function for multi-view aluminum product recognition. The identification module is used to optimize the nuclear norm of various results in the objective function of multi-view aluminum product identification by iteratively updating the rules. When the optimized objective function meets the convergence condition, spectral clustering is performed on the optimized sparse constraints to obtain the final clustering result, which is then used as the aluminum product identification result.

[0015] The present invention provides a computer device, including a memory and a processor. The memory stores a program, and when the program is executed by the processor, the processor performs the steps of the above-described multi-view aluminum product identification method.

[0016] The present invention provides a storage medium storing a computer program thereon, which, when executed by a processor, performs the steps of the above-described multi-view aluminum product identification method.

[0017] Compared with the prior art, the present invention has the following significant advantages: This invention proposes a multi-view aluminum product recognition method. This method superimposes the subspace representation matrices of different views into a tensor to explore higher-order correlations between multi-view data. Simultaneously, a low-rank constraint is applied to this tensor to reduce redundant information in the subspace representation. Frobenius norm constraints are used to process the subspace representation matrix, obtaining a subspace representation matrix with intra-class consistency. Furthermore, sparse constraints are applied to the product of the transpose and itself of the intra-class consistent subspace representation matrix to enhance the connection between similar data points. Sparse constraints on the product of the transpose and itself of the subspace representation matrix can enhance the block diagonal structure of the subspace representation matrix. Based on this, an objective function and related optimizations are constructed, and finally, clustering is used to obtain the aluminum product recognition result. Under the premise of optimizing the accuracy of the subspace representation matrix, the recognition performance of aluminum products after clustering can be improved to a certain extent. Attached Figure Description

[0018] Figure 1 This is a framework diagram proposed in this invention; Figure 2 The present invention provides an overall flowchart of a multi-view aluminum product identification method. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0020] Capturing effective representations of data and acquiring effective representational information is crucial in pattern recognition and data mining. Multi-view learning obtains discriminative information from data from different perspectives, resulting in more accurate and complete representations than single-view learning. Therefore, multi-view learning has attracted considerable attention from researchers both domestically and internationally. In recent years, multi-view learning has been widely and successfully applied in numerous fields, such as object tracking, text analysis, image clustering, biomedicine, and sentiment analysis.

[0021] Multi-view subspace clustering is an important branch of multi-view clustering, and many researchers have conducted extensive research on it. Kan et al. used a self-representation method to obtain similarity maps for each view and then fused these maps into a consensus map. Furthermore, they applied a rank constraint to the Laplacian matrix of the consensus map to learn the optimal clustering structure. Yao et al. used a subspace representation method to reconstruct the data matrix and obtained the similarity matrix by minimizing the reconstruction error. They further utilized rank constraints to improve the model's clustering performance. Gan et al. proposed a novel multi-stage fusion method that cleverly integrates late and early fusion into a unified framework to capture more complete representational information in multi-view data and effectively reduce the impact of low-quality views. Although the above multi-view subspace learning models have certain advantages in clustering tasks, they do not consider the higher-order correlations between different views in multi-view data.

[0022] Tensors, as higher-order generalizations of vectors and matrices, can explore higher-order relationships in multi-view data, thereby mining consistency information between different views. Zhang et al. proposed a low-rank tensor-constrained multi-view subspace clustering (LT-MSC) method. This method constructs a tensor based on the subspace representation matrices of different views, capturing higher-order correlations between multi-view data. Furthermore, it applies low-rank constraints to this tensor, reducing redundant information in learning subspace representations and improving the accuracy of the clustering algorithm. Chen et al. jointly studied multiple low-rank similarity matrices and a consensus graph reflecting the final performance within a unified framework. They also stacked multiple affinity matrices in a tensor and applied low-rank constraints to explore higher-order correlations in multi-view data. Mu et al. focused on exploring the shared consensus latent representation and the subspace representation of each view, combining subspace representation learning and tensor low-rank constraints to enhance consensus graph learning. Wu et al. combined the diversity graph and the consistency graph into a tensor and imposed tensor nuclear norm constraints on it. This not only allows for the acquisition of complete information across multiple views but also enables mutual learning and reinforcement between the diversity graph and the consistency graph. Although the multi-view subspace clustering models discussed above utilize tensor concepts to mine higher-order relationships in multi-view data to improve model performance, they cannot simultaneously enhance the block diagonal structure of the subspace representation matrix and the connections between similar data points, or weaken the correlation between dissimilar data points.

[0023] Existing research indicates that the clarity of the block diagonal structure of a subspace representation matrix plays a crucial role in clustering tasks. The block diagonal structure of a subspace representation matrix can reveal the inherent clustering structure of the data; the clearer the block diagonal structure of the subspace representation matrix, the better its clustering effect.

[0024] Among them, the Subspace Segmentation via Quadratic Programming (SSQP) method is an effective subspace segmentation method. The SSQP method differs from the sparse subspace clustering (SSC) method in its sparse constraint term; the SSC method performs [further processing] on the coefficient matrix. Norm constraints, while the SSQR method treats the transpose of the coefficient matrix multiplied by itself as a whole, and then performs a process on this whole. Norm constraints. This approach offers several advantages. Firstly, it enhances the block diagonal structure of the coefficient matrix, resulting in better segmentation. Secondly, it allows for the acquisition of a sparse representation of the data. Its objective function is:

[0025] (1); in, Representing a matrix Norm, and They are equivalent, where e is a vector with all elements equal to 1.

[0026] Least squares regression subspace clustering method: Least Squares Regression (LSR) subspace clustering is a method that leverages the correlation between data points. It encourages a grouping effect that tends to group highly correlated data together. The objective function of the LSR model is expressed as follows:

[0027] (2); in, λ It is a balance parameter. This represents the Frobenius norm of the matrix.

[0028] like Figure 1 and Figure 2 As shown, this invention proposes a multi-view aluminum product identification method, which specifically includes the following steps: Step S1: Obtain the subspace representation matrix of different views of the aluminum product image.

[0029] Step S2: Superimpose the subspace representation matrices of different views of the aluminum product into a tensor, obtain the high-order correlation between the multi-view data, apply low-rank constraints to the tensor, and introduce Frobenius norm constraints to process the subspace representation matrix to obtain a subspace representation matrix with intra-class consistency. Apply sparsity constraints to the product of the transpose of the intra-class consistent subspace representation matrix and itself to enhance the block diagonal structure of the subspace representation matrix.

[0030] Method of the present invention: The LT-MSC method introduces the concept of tensors into the framework of low-rank representation (LRR). The definition of the tensor nuclear norm is as follows: (3); in, The nuclear norm of a tensor. ζ n It is a constant greater than 0, and satisfies .

[0031] in, It is an N-order tensor, F (n) Let F be the matrix expanded over the nth pattern, and its expression is as follows: (4); The LT-MSC method effectively captures high-order correlations between multi-view data by using a tensor to integrate the subspace representations corresponding to each single-view data point. Furthermore, it applies a low-rank constraint to this tensor to reduce redundant information in the learned subspace representations. Its objective function is as follows:

[0032] (5); Among them, A (v) F (v) and S (v) They represent the first v The data matrix, subspace representation matrix, and reconstruction error matrix corresponding to each view. Ψ(.) is the subspace representation matrix F of different views. (v) A third-order tensor constructed after merging . It is a matrix formed by stitching together the reconstruction error matrices of each view.

[0033] This invention introduces the Frobenius norm into the LT-MSC framework and applies sparse constraints to the product of the transpose of the subspace representation matrix and itself, thereby constructing a unified multi-view learning framework. The advantages of this approach are: first, it obtains subspace representation matrices with intra-class consistency, thus enhancing the correlation between similar data points; second, it enhances the block diagonal structure of the subspace representation matrix while obtaining sparse data representation, thereby improving the recognition performance of the method for aluminum products.

[0034] Step S3: Balance and weight the resulting nuclear norms of tensor, low-rank constraint, Frobenius norm constraint and sparse constraint to obtain the objective function for multi-view aluminum product recognition.

[0035] Figure 1 This is a framework diagram of the method of the present invention, and its objective function is as follows: (6); in, F is the result norm of the tensor. (v) Let S be the subspace representation matrix of the v-th view. (v) Let v be the reconstruction error matrix of the v-th view. For low-rank constraint terms, , and For balancing parameters, Here, T represents the sparse constraint term, and T is the transpose. For Frobenius norm constraints, V is the total number of views, A (v) Let v be the data matrix corresponding to the v-th view. The new sparsity constraint is used to enhance the block diagonal structure of the subspace representation matrix, and the Frobenius norm constraint is introduced to enhance the correlation between similar data points.

[0036] According to formula (3), formula (6) can be reformulated as follows: (7); in, , , .

[0037] Step S4: Optimize the nuclear norm of various results in the objective function of multi-view aluminum product recognition by iteratively updating the rules. When the optimized objective function meets the convergence condition, perform spectral clustering on the optimized sparse constraints to obtain the final clustering result. Use the clustering result as the aluminum product recognition result.

[0038] The augmented Lagrange multiplier method (ALM) was used to solve problem (7), and an auxiliary variable G was introduced into formula (7).n To replace F (n) Then we have: (8); Among them, f and g n These are the optimized tensor F and the auxiliary variable G. n Vectorization, P n It is a permutation matrix used to calibrate F (n) and G n The corresponding elements between them t n , σ and τ For different equilibrium parameters, N For the number of auxiliary variables, It is the nuclear norm of the auxiliary variable.

[0039] Using the ALM method, the optimized result formula (8) is updated to obtain the intermediate objective function, the specific expression of which is: (9); in, It is the intermediate objective function value.

[0040] By iteratively updating the rules, the subspace representation matrix F in the intermediate objective function is solved sequentially. (v) Reconstruction error matrix S, auxiliary variable G n Balance parameters μ Lagrange multiplier B v and α n Continue until the convergence condition is met.

[0041] By iteratively updating the rules, the subspace representation matrix F in the intermediate objective function is solved sequentially. (v) Reconstruction error matrix S, auxiliary variable G n Balance parameters μ Lagrange multiplier B v and α n Specifically: When S(v) and G n When fixed, F (v) This can be obtained by solving the following subproblem: ; According to formula (10), the update rule for the subspace representation matrix in the intermediate objective function is obtained, and the specific expression is as follows: (11); in, , Indicates selection and the first v The elements corresponding to each view are then used to reconstruct a matrix, where I is the identity matrix and I1 represents a matrix of all ones. σ These are the balancing parameters.

[0042] In F (v) and auxiliary variable G n When fixed, S is obtained. * as follows: (12); Through matrix A (v) -A (v) F (v) +B v The matrix Z formed by vertically connecting the columns, and the matrix Z's first column... i Obtain the optimized reconstruction error matrix from the column relations The specific expression is: (13); in, Let B represent the i-th column of matrix Z. v It represents the Lagrange multiplier.

[0043] By iterating through the smallest auxiliary variable, we obtain the update rule for the auxiliary variable, specifically expressed as: (14); Among them, parameters , It is based on the nth pattern of the expansion that P is... n f+ a n Reshape into the corresponding matrix, where prox is the proximal operator. tr This represents the constraint condition for truncating the nuclear norm.

[0044] By minimizing the balance parameter μ The relationship with the maximum equilibrium parameter is used to obtain the equilibrium parameters. μ The update rule, specifically the expression, is: ; in, The correlation coefficient, This is the maximum balance parameter.

[0045] Through matrix A (v) -A(v) F (v) -S (v) or matrix P n fg n The update rule for obtaining the Lagrange multipliers by responding to the sum of the initial Lagrange multipliers is specifically expressed as follows: (15); (18); in, For B-based v The updated Lagrange multipliers, For based on α n The updated Lagrange multipliers.

[0046] For the updated F (v) You can directly replace F (v) Update f using the corresponding elements in the table.

[0047] f*←F (v) (16); g n The update of is similar to the update of f, and the following can be obtained: ←G n (17); Based on the optimized sparsity constraints, the updated subspace representation matrix F is used. (v) Calculate the clustering matrix. Apply the spectral clustering algorithm to the clustering matrix E to output the final clustering results.

[0048] The specific process of this invention is as follows: Input: Multi-view data matrix: A (1) A (1) , ... A (V) ,parameter t n , σ , τ Cluster number K n .

[0049] Output: Clustering results C K .

[0050] initialization: .

[0051] .

[0052] While not converged do: Update F according to formula (11).

[0053] Update S according to formula (12).

[0054] Update G according to formula (14) n .

[0055] Update B according to formula (15) v .

[0056] Update f according to formula (16).

[0057] Update g according to formula (17) n .

[0058] Update according to formula (18) αn .

[0059] renew μ .

[0060] Check convergence conditions: .

[0061] calculate .

[0062] Apply spectral clustering to matrix E.

[0063] Output: Clustering results C K .

[0064] The key contributions of the method of this invention are as follows: (1) The method of the present invention explores the higher-order correlation in multi-view data of aluminum products using the concept of tensors.

[0065] (2) The method of the present invention constructs a new sparse constraint, which can not only obtain the sparse representation of data, but also enhance the block diagonal structure of the subspace representation matrix.

[0066] (3) The method of the present invention applies Frobenius norm constraints to obtain a class-consistent subspace representation matrix.

[0067] Experimental results and analysis: This section validates the effectiveness of the proposed method. All experiments were conducted on a desktop computer running Windows 11, with a 13th Gen Intel(R) Core(TM) i9-13900K 3.00 GHz processor, 128GB RAM (64-bit), and MATLAB R2018b. The proposed method is compared with several classic and state-of-the-art methods, including the single-view learning method NMF and five multi-view learning methods: LT-MSC, LMSC, GMC, JSMC, and MCDHSIC. Furthermore, specific comparison methods are considered as follows:

[0068] NMF: It is a local feature extraction method that attempts to approximate the original data matrix as the product of two non-negative matrices.

[0069] LT-MSC: This method uses the concept of tensors to explore higher-order relationships in multi-view data.

[0070] LMSC: This method mines potential complementary information in multi-view data.

[0071] GMC: This method merges the data graph matrices of all views to generate a unified graph matrix, so that the learning of each view's graph matrix and the learning of the unified graph matrix mutually reinforce each other.

[0072] JSMC: This model incorporates commonalities and inconsistencies across views into subspace representation learning.

[0073] MCDHSIC: This model uses diversity constraints to capture nonlinear relationships between different views and also imposes... Norm constraints and Frobenius norm constraints are used to obtain subspace representation matrices with inter-class sparsity and intra-class consistency.

[0074] Dataset used in the experiment: The simulation experiments of this invention were conducted on a multi-view dataset of aluminum tableware and a multi-view dataset of aluminum packaging containers. The aluminum tableware multi-view dataset contains 14 classes, with 8 images per class, for a total of 112 images of aluminum tableware. The aluminum packaging container multi-view dataset contains 8 classes, with 16 images per class, for a total of 128 images of aluminum packaging containers. For the above two aluminum multi-view datasets, drawing on the view data processing methods in the literature, three techniques—Coiflets, Daubechies, and Symlets—were used to obtain the three low-frequency components (i.e., three views) of each image. To increase the stability of the method, Principal Component Analysis (PCA) was used to reduce the dimension of each view to approximately 100 dimensions.

[0075] Experimental parameter settings.

[0076] Based on previous research experience, the value ranges of different parameters were set for the proposed method. The specific value ranges of the experimental parameters for the method of this invention are shown in Table 1. Experimental parameters for other methods are referenced from relevant literature. The number of repeated runs for all methods in this invention is set to 20.

[0077] Table 1. Parameter settings for a multi-view aluminum product identification method Recognition performance: This section presents the mean and standard deviation of 20 recognition results for all methods, which are typically equipped with K-means. The recognition results are shown in Table 2, with bolded values ​​representing the optimal experimental results.

[0078] Table 2. Recognition results (%) of each method on the aluminum tableware multi-view dataset. Table 3. Recognition results (%) of each method on the multi-view dataset of aluminum packaging containers Based on the experimental results in Table 2, the following observations can be obtained: 1) Of all the methods, the method of the present invention exhibits the best recognition performance.

[0079] 2) The recognition performance of the multi-view method is better than that of the single-view method.

[0080] 3) Generally, methods that consider tensors have better recognition performance than methods that do not consider tensors.

[0081] This invention constructs a multi-view aluminum product recognition method. This method superimposes the subspace representation matrices of different views into a tensor to explore higher-order correlations between multi-view data. Simultaneously, a low-rank constraint is applied to this tensor, effectively reducing redundant information in the subspace representation. Furthermore, the method uses Frobenius norm constraints to enhance the connections between similar data points, and applies sparsity constraints to the product of the transpose of the subspace representation matrix and itself to enhance the block diagonal structure of the subspace representation matrix. This, to a certain extent, improves the clustering performance of the constructed model. Experimental results on multiple multi-view datasets of aluminum products verify the effectiveness of the proposed method.

[0082] Based on the above methods and statements, the present invention provides a multi-view aluminum product recognition system, comprising: a data acquisition module, a constraint module, a target construction module, and a recognition module.

[0083] The system comprises the following modules: a data acquisition module for obtaining subspace representation matrices of different views of aluminum product images; a constraint module for superimposing these subspace representation matrices into a tensor, applying low-rank constraints to the tensor, and introducing Frobenius norm constraints to process the subspace representation matrices to obtain intra-class consistent subspace representation matrices, and applying sparsity constraints to the transpose of the intra-class consistent subspace representation matrix multiplied by itself; a target construction module for balancing and weighting the kernel norms of the tensor, low-rank constraints, Frobenius norm constraints, and sparsity constraints to obtain the target function for multi-view aluminum product recognition; and a recognition module for optimizing the kernel norms of various results in the target function for multi-view aluminum product recognition through iterative rule updates, and performing spectral clustering on the optimized sparse constraints when the optimized target function meets the convergence condition to obtain the final clustering result, which is then used as the aluminum product recognition result.

[0084] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a program, and when the program is executed by the processor, the processor performs the steps of a multi-view aluminum product identification method.

[0085] According to the disclosed embodiments, the computer device can communicate with one or more external devices (e.g., keyboard, pointing device, Bluetooth communication, etc.) or with any device that enables the computing device to communicate with one or more other computing devices (e.g., router, demodulator, etc.).

[0086] The present invention also provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of a multi-view aluminum product identification method.

[0087] According to the disclosed embodiments, the storage medium can be a non-volatile computer-readable storage medium, such as, but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, the storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0088] The above description, in conjunction with specific preferred embodiments, provides a more detailed explanation of the present invention. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such deductions or substitutions should be considered to fall within the scope of protection of the present invention.

Claims

1. A multi-view aluminum article identification method characterized by, The method comprises the following steps: obtaining a subspace representation matrix of different views of an aluminum product image; stacking the subspace representation matrices of different views of the aluminum product into a tensor, applying a low-rank constraint to the tensor, introducing a Frobenius norm constraint to the subspace representation matrix, obtaining a subspace representation matrix with intra-class consistency, and performing sparse constraint on the product of the transpose of the intra-class consistent subspace representation matrix and the intra-class consistent subspace representation matrix; balancing and weighting the kernel norms of the results of the tensor, the low-rank constraint, the Frobenius norm constraint and the sparse constraint to obtain a target function for multi-view aluminum product recognition; optimizing the kernel norms of the results of each class in the target function for multi-view aluminum product recognition through an iterative updating rule, and performing spectral clustering on the optimized sparse constraint when the optimized target function meets a convergence condition to obtain a final clustering result, which is taken as the aluminum product recognition result.

2. A method of identifying a multi-view aluminum article as defined in claim 1, wherein, The balancing and weighting of the kernel norms of the results of the tensor, the low-rank constraint, the Frobenius norm constraint and the sparse constraint to obtain the target function for multi-view aluminum product recognition has the following specific expression: ； where, is the result kernel norm of a tensor, F (v) is the subspace representation matrix of the vth view, S (v) is the reconstruction error matrix of the vth view, is the low-rank constraint term, , and are the balancing parameters, is the sparsity constraint term, T is the transpose, is the Frobenius norm constraint, V is the total number of views, A (v) is the data matrix corresponding to the vth view, Ψ(.) is a 3rd order tensor constructed by merging the subspace representation matrices F (v) of different views , S is a matrix formed by concatenating the reconstruction error matrices of each view.

3. A method of identifying a multi-view aluminum article as defined in claim 2, wherein, The optimization of the kernel norms of the results of each class in the target function for multi-view aluminum product recognition through the iterative updating rule has the following specific expression: Introducing auxiliary variable G n Instead of the matrix F after tensor expansion (n) And using the ALM method, the intermediate objective function is obtained; By iteratively updating the rules, the subspace representation matrix F in the intermediate target function is solved in turn (v) , reconstruction error matrix S, auxiliary variable G n , balance parameters μ , Lagrange multiplier B v and α n , until the convergence condition is met.

4. The method of claim 3, wherein, The introduction of the auxiliary variable G n Instead of the matrix F after tensor expansion (n) And using the ALM method, the intermediate objective function is obtained, specifically: The optimization of the target function through the augmented Lagrange multiplier method has the following specific expression: ； where f and g n are the optimized tensor F and auxiliary variable G n respectively, P n is a permutation matrix, t n , σ and τ are different equalization parameters, N is the number of auxiliary variables, is the resulting kernel norm of auxiliary variables; The updating of the results after optimization through the ALM method to obtain an intermediate target function has the following specific expression: ； wherein is the intermediate objective function value.

5. A method of identifying a multi-view aluminum product as recited in claim 4, wherein, Solving the intermediate target function in subspaces represented by matrix F (v) , reconstruction error matrix S, auxiliary variable G n , balance parameters μ , Lagrange multiplier B v and α n , specifically: The updating rule of the subspace representation matrix in the intermediate target function has the following specific expression: ； wherein, , represents selecting elements corresponding to the first v view, I is an identity matrix, I1represents a matrix of all ones, σ is an equalization parameter; In F (v) and auxiliary variable G n is fixed, the matrix A (v) -A (v) F (v) +B v The matrix Z is composed of the vertical connection along the column, and the first i column relationship of the matrix Z obtains the optimized reconstruction error matrix , the specific expression is: ； wherein denotes the i-th column of the matrix Z; By iterating the least auxiliary variable The update rule of auxiliary variable is obtained, and the specific expression is as follows: ； where the parameters , are reshaped into respective matrices according to the nth mode of unfolding, n f+ a n prox is a proximity operator, tr denotes the constraint of truncated nuclear norm; By minimizing the balance parameter μ The relationship with the maximum balance parameter, the balance parameter is obtained μ The update rule of the balance parameter is as follows: ； wherein is a correlation coefficient, is a maximum balance parameter; By matrix A (v) -A (v) F (v) -S (v) Or matrix P n f-g n And the response initial Lagrange multiplier sum to obtain the update rule of Lagrange multiplier, the specific expression is: ； ； wherein is based on B v updated Lagrange multiplier, is based on α n updated Lagrange multiplier.

6. A method of identifying a multi-view aluminum product as recited in claim 5, wherein, The spectral clustering of the optimized sparse constraint to obtain the final clustering result has the following specific expression: Based on the optimized sparse constraint, through the updated subspace representation matrix F (v) The clustering matrix is calculated, and the specific expression is as follows: ； The spectral clustering algorithm is applied to the clustering matrix E to output the final clustering result.

7. The method of claim 1, wherein, When the subspace representation matrix of different views of the aluminum product image is obtained, the low-frequency components in the aluminum product image are extracted as the views through Coiflets, Daubechies and Symlets technologies, and the dimension of each view is reduced through principal component analysis (PCA) for constructing the subspace representation matrix.

8. A multi-view aluminum article identification system characterized by, The method comprises the following steps: a data acquisition module for obtaining a subspace representation matrix of different views of an aluminum product image; a constraint module for stacking the subspace representation matrices of different views of the aluminum product into a tensor, applying a low-rank constraint to the tensor, introducing a Frobenius norm constraint to the subspace representation matrix, obtaining a subspace representation matrix with intra-class consistency, and performing sparse constraint on the product of the transpose of the intra-class consistent subspace representation matrix and the intra-class consistent subspace representation matrix; a target construction module for balancing and weighting the kernel norms of the results of the tensor, the low-rank constraint, the Frobenius norm constraint and the sparse constraint to obtain a target function for multi-view aluminum product recognition; an identification module for optimizing the kernel norms of the results of each class in the target function for multi-view aluminum product recognition through an iterative updating rule, and performing spectral clustering on the optimized sparse constraint when the optimized target function meets a convergence condition to obtain a final clustering result, which is taken as the aluminum product recognition result.

9. A computer device, comprising: A computer program product comprising a memory having stored therein a program, the program being executable by a processor to cause the processor to perform the steps of the method of any one of claims 1 to 7.

10. A storage medium having stored thereon a computer program, characterized in that The computer program product is executable by a processor to cause the processor to perform the steps of the method of any one of claims 1 to 7.

Images