Tensor completion method and system based on low-rank matrix of Tucker decomposition

By introducing prior knowledge of low rank and smoothness in tensor factor space, using Tucker to decompose low rank constraints and Frobenius norms of factor matrix, combined with ADMM algorithm, the problem of rank selection of factor matrix in tensor data recovery is solved, and the problem of rank selection of factor matrix in tensor data recovery is improved. Recovery accuracy and efficiency are reduced, and calculation costs are reduced.

CN118606617BActive Publication Date: 2024-11-12JIANGXI UNIVERSITY OF FINANCE AND ECONOMICS
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411087929.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-09
Publication Date
2024-11-12
Estimated Expiration
2044-08-09

AI Technical Summary

Technical Problem

The prior art ignores the smoothness and low-rank characteristics of the tensor decomposition factor itself in tensor data recovery, resulting in poor recovery effect and high computational cost.

Method used

Using the tensor complete method based on Tucker decomposition factor matrix, the tensor complete model is established by using the kernel norm in the horizontal and vertical gradient of the factor matrix, combined with the Frobenius norm constraint, and using the ADMM algorithm framework for Iteratively solve, introducing auxiliary variables to optimize the model.

Benefits of technology

It significantly improves the accuracy and efficiency of tensor data recovery, avoids the Tucker rank selection problem, reduces the calculation cost, and realizes more accurate reconstruction of tensor data.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118606617B_ABST
    Figure CN118606617B_ABST
Patent Text Reader

Abstract

The present invention proposes a tensor completion method and system based on the low rank of Tucker decomposition factor matrix. The method is based on Tucker decomposition, and the gradient of the factor matrix in the horizontal and vertical directions in the Tucker decomposition is constrained by the nuclear norm as a low rank constraint, and the nuclear tensor is constrained by the Frobenius norm. A tensor completion model is constructed, and auxiliary variables are introduced into the tensor completion model, and an augmented Lagrangian function is established to perform constrained optimization on the tensor completion model to obtain a constrained optimization model, and the missing data and the missing area mask tensor are tensorized and stored to obtain a tensor containing missing values ​​and a missing area mask tensor, and the missing value tensor and the missing area mask tensor are input into the constrained optimization model, and each sub-problem is iteratively solved through the ADMM algorithm framework to obtain the completed tensor. The present invention not only overcomes the difficulty of Tucker rank selection in traditional Tucker decomposition, but also achieves more accurate and efficient recovery of tensor data by characterizing the low rank and smoothness of data in the tensor factor space.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular to a tensor completion method and system based on low-rank Tucker decomposition factor matrix. Background Art

[0002] With the increasing application of multidimensional data in signal processing, physical science, machine learning and other fields, tensors (multidimensional arrays) as a natural representation of data have become increasingly important. Tensor data often encounter missing value problems during acquisition or transmission, which not only increases the cost of data processing, but also poses a challenge to multidimensional data analysis. Filling missing data based on certain assumptions and reconstructing missing elements by using preset patterns, also known as low-rank matrix completion, has attracted significant attention. However, with the emergence of high-order data, matrix completion cannot make good use of the advantages of high-order structures due to matrix or vector operations. As a natural extension of matrix completion, tensor completion has been proposed and widely used to solve the problem of missing high-order data. Tensor completion aims to estimate and complete missing or damaged elements in data, thereby improving data quality.

[0003] In practical applications, multidimensional data such as color images, videos, and hyperspectral images often have significant low-rank characteristics due to the high correlation and repeatability between pixels. Furthermore, these multidimensional data also generally show significant smoothness characteristics. Smoothness is reflected in the similar or gradual numerical characteristics between adjacent pixels or data blocks. This continuity is not only reflected in the spatial dimension, but also in the continuous and smooth transition between adjacent frames for time series data (such as videos). Many studies have shown that tensor data such as color images, videos, and hyperspectral images naturally have the dual prior properties of low rank and smoothness.

[0004] At present, tensor recovery methods mainly focus on low-rank priors, smooth priors, and comprehensive priors (low-rank + smooth) strategies. Although existing methods have achieved certain results in tensor recovery, most of them focus on analyzing and utilizing low-rank and smooth priors in the original space, ignoring the smoothness and low-rank characteristics of the tensor decomposition factors themselves. In response to this gap, exploring and utilizing low-rank and smooth priors in the decomposition factor space has become the key to improving the complete performance of tensors. Summary of the invention

[0005] In view of the above situation, the main purpose of the present invention is to propose a tensor completion method and system based on the low rank of Tucker decomposition factor matrix to solve the above technical problems.

[0006] The present invention proposes a tensor completion method based on low-rank Tucker decomposition factor matrix, the method comprising the following steps:

[0007] Step 1: Based on Tucker decomposition, the horizontal and vertical gradients of the factor matrix in Tucker decomposition are constrained by the nuclear norm as low-rank constraints, and the nuclear tensor is constrained by the Frobenius norm to construct a tensor complete model;

[0008] Step 2: Introduce auxiliary variables into the tensor complete model, establish an augmented Lagrangian function based on Lagrangian multipliers to perform constrained optimization on the tensor complete model, and obtain a constrained optimization model;

[0009] Step 3: Get the missing data image and normalize it, create a corresponding mask according to the missing data in the missing data image, store the missing data image and the mask in tensor form, and obtain a tensor containing missing values ​​and a missing area mask tensor;

[0010] Step 4: Input the tensor of missing values ​​and the mask tensor of missing regions into the constrained optimization model. Take the core tensor, factor matrix, auxiliary variables, and output target tensor as sub-problems respectively. Iterate and solve each sub-problem through the ADMM algorithm framework to obtain the complete tensor.

[0011] The present invention also proposes a tensor completion system based on a low-rank Tucker decomposition factor matrix, wherein the system applies the tensor completion method based on a low-rank Tucker decomposition factor matrix as described above, and the system comprises:

[0012] Model building modules for:

[0013] Based on Tucker decomposition, the horizontal and vertical gradients of the factor matrix in Tucker decomposition are constrained by the nuclear norm as low-rank constraints, and the nuclear tensor is constrained by the Frobenius norm to construct a tensor complete model.

[0014] Constrained Optimization Module for:

[0015] Auxiliary variables are introduced into the tensor complete model, and an augmented Lagrangian function is established based on Lagrangian multipliers to perform constrained optimization on the tensor complete model, thus obtaining a constrained optimization model.

[0016] Data acquisition module, used to:

[0017] Get the missing data image and normalize it, create a corresponding mask according to the missing data in the missing data image, store the missing data image and the mask in tensor form, and obtain a tensor containing missing values ​​and a missing area mask tensor;

[0018] Iterative solution module for:

[0019] The tensor of missing values ​​and the mask tensor of missing regions are input into the constrained optimization model. The core tensor, factor matrix, auxiliary variables and output target tensor are taken as sub-problems respectively. Each sub-problem is iteratively solved through the ADMM algorithm framework to obtain the complete tensor.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] The present invention not only solves a long-standing challenge in traditional Tucker decomposition - the problem of Tucker rank selection - by innovatively introducing prior knowledge of low rank and smoothness in the tensor factor space, but also significantly improves the accuracy and efficiency of tensor data recovery. Traditional techniques often require users to set the Tucker rank in advance, which is usually a complex task, and improper selection will seriously affect the recovery effect. In contrast, the present invention uses the inherent properties of the factor space to avoid the problem of optimal rank selection and avoids the uncertainty caused by manual adjustment.

[0022] In addition, compared with traditional technologies, the present invention shows a significant advantage of low computational cost. When processing tensor data, traditional methods are often limited by computational cost and time efficiency. They may require complex algorithms to approximate low rank and smoothness, which not only consumes a lot of resources, but also may lead to a slow recovery process and poor results. The present invention cleverly utilizes the low rank and smoothness of the tensor factor space to achieve a more accurate reconstruction of tensor data without adding additional computational burden. This innovation not only maintains the quality of data recovery, but also greatly reduces the processing time and required computing resources.

[0023] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description or learned through embodiments of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 A flowchart of the tensor completion method based on low-rank Tucker decomposition factor matrix proposed by the present invention;

[0025] Figure 2 A flowchart for iterative solution of the present invention;

[0026] Figure 3 A schematic diagram of the structure of the tensor complete system based on the low-rank Tucker decomposition factor matrix proposed in the present invention;

[0027] Figure 4 This is a comparison diagram before and after the color image is processed by the present invention;

[0028] Figure 5 This is a comparison diagram before and after the hyperspectral image is processed by the present invention;

[0029] Figure 6 This is a comparison diagram before and after the video is processed by the present invention. DETAILED DESCRIPTION

[0030] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0031] These and other aspects of the embodiments of the present invention will be apparent with reference to the following description and accompanying drawings. In these descriptions and accompanying drawings, some specific implementations of the embodiments of the present invention are specifically disclosed to represent some ways of implementing the principles of the embodiments of the present invention, but it should be understood that the scope of the embodiments of the present invention is not limited thereto.

[0032] See also Figure 1 , this embodiment provides a tensor completion method based on low-rank Tucker decomposition factor matrix, the method comprising the following steps:

[0033] Step 1: Based on Tucker decomposition, the horizontal and vertical gradients of the factor matrix in Tucker decomposition are constrained by the nuclear norm as low-rank constraints, and the nuclear tensor is constrained by the Frobenius norm to construct a tensor complete model;

[0034] In step 1, the tensor complete model has the following relationship:

[0035] ;

[0036] in, represents a tensor with missing values, , represents the field of real numbers, Respectively represent dimension information, represents the missing region mask tensor, , represents the output target tensor, , , represents different regularization parameters, represents the number of spatial dimensions, represents the number of factor matrices, represents the kernel tensor, represents the gradient multiplication operation along the kth dimension, represents the nuclear norm, represents the square of the Frobenius norm, represents the linear operation of projection, represents the factor matrix, They respectively represent that the kernel tensor is expanded along the first dimension and multiplied by the first factor matrix, the kernel tensor is expanded along the second dimension and multiplied by the second factor matrix, and the kernel tensor is expanded along the third dimension and multiplied by the third factor matrix.

[0037] Calculating the gradient of a matrix is ​​equivalent to multiplying it with a difference operator. The horizontal and vertical directions correspond to the left and right multiplication operations. In order to keep the matrix dimension unchanged before and after the difference, the rank of the n-order square matrix difference matrix under the periodic boundary condition is equal to n-1, so the rank after the gradient is calculated is at most reduced by 1 compared to the rank of the factor matrix. Therefore, the low rank of the factor matrix can be transferred to the gradient matrix of the factor matrix. According to the compatibility principle between matrix norms, the inequality In particular, the inequality and They correspond to the isotropic total variation regularization technique and the anisotropic total variation regularization technique widely used in the field of image processing. These two techniques are good at global smoothing and edge preservation. Therefore, it is easy to see that the matrix gradient is the nuclear norm, that is, It can characterize the smoothness of the matrix.

[0038] According to the definition of Tucker rank, the low rank of the original data M can be characterized by the rank of the factor matrix. In addition, the factor matrix contains the eigenvectors of each dimension, and the smoothness of the original data M can also be reflected by the linear combination of the smooth eigenvectors in the factor matrix.

[0039] The low rank of the gradient of the factor matrix can simultaneously characterize the low rank and smoothness of the factor matrix. Thus, two kinds of prior information are modeled by one term.

[0040] Step 2: Introduce auxiliary variables into the tensor complete model, establish an augmented Lagrangian function based on Lagrangian multipliers to perform constrained optimization on the tensor complete model, and obtain a constrained optimization model;

[0041] In step 2, the process of introducing auxiliary variables into the tensor complete model has the following relationship:

[0042] ;

[0043] in, represents the first auxiliary variable, represents the second auxiliary variable, represents the indicator function. The indicator function is defined as:

[0044] .

[0045] In step 2, the process of constrained optimization of the tensor complete model by establishing an augmented Lagrangian function based on Lagrangian multipliers has the following relationship:

[0046] ;

[0047] in, represents the penalty parameter, denotes different Lagrange multipliers, Represent the two sub-problems of the first auxiliary variable, They represent gradient multiplication along the first dimension and gradient multiplication along the second dimension respectively.

[0048] Step 3: Get the missing data image and normalize it, create a corresponding mask according to the missing data in the missing data image, store the missing data image and the mask in tensor form, and obtain a tensor containing missing values ​​and a missing area mask tensor;

[0049] As a generalization of the concept of matrix, tensor has stronger modeling capabilities. The missing data mentioned here can be all data with smooth low-rank characteristics caused by spatial continuity and information redundancy, such as color images, videos, hyperspectral images, etc.

[0050] Step 4: Input the tensor of missing values ​​and the mask tensor of missing regions into the constrained optimization model. Take the core tensor, factor matrix, auxiliary variables, and output target tensor as sub-problems respectively. Iterate and solve each sub-problem through the ADMM algorithm framework to obtain the complete tensor.

[0051] In step 4, the auxiliary variables include a first auxiliary variable and a second auxiliary variable, and the process of iteratively solving the sub-problem of the first auxiliary variable through the ADMM algorithm framework specifically includes the following steps:

[0052] The first auxiliary variable includes two sub-problems. The objective functions of the two sub-problems of the first auxiliary variable are constructed by fixing the core tensor, the factor matrix, the second auxiliary variable and the output target tensor. The objective functions of the two sub-problems of the first auxiliary variable are respectively expressed as follows:

[0053] ;

[0054] ;

[0055] Based on the singular value threshold algorithm, the solution relationship of the two sub-problems of the first auxiliary variable is constructed. The solutions of the two sub-problems of the first auxiliary variable have the following relationship:

[0056] ;

[0057] ;

[0058] Among them, it means that the input parameter is Singular value threshold algorithm for ;

[0059] The two subproblems of the first auxiliary variable are iteratively solved by updating the penalty parameter, regularization parameter and the corresponding Lagrange multiplier.

[0060] set up represents singular value decomposition, where U is a unit orthogonal matrix whose columns are the left singular vectors of matrix A. S is a diagonal matrix whose diagonal elements are the singular values ​​of matrix A. V is another unit orthogonal matrix whose rows are the right singular vectors of matrix A. represents the transpose operation, is the singular value threshold algorithm, which represents the compression of matrix A. Indicates that the input parameters are The singular value threshold algorithm is Indicates that the input parameters are The soft threshold operator expression is , it will be based on the parameters For input variables Perform thresholding. yes The symbol function of .

[0061] In the above scheme, the input matrix is ​​first subjected to singular value decomposition, and then the obtained singular values ​​are thresholded, retaining the singular values ​​greater than The singular value of The singular values ​​of are set to zero, which can reduce the complexity of the data while retaining as much information as possible.

[0062] In step 4, the process of iteratively solving the sub-problem of outputting the target tensor through the ADMM algorithm framework specifically includes the following steps:

[0063] The objective function of the subproblem of outputting the target tensor is constructed by fixing the core tensor, the factor matrix, the first auxiliary variable, and the second auxiliary variable. The objective function of the subproblem of outputting the target tensor has the following relationship:

[0064] ;

[0065] The solution to the subproblem of outputting the target tensor has the following relation:

[0066] ;

[0067] in, represents the projection operator, The projection operator representing the complement of the missing region mask tensor, Represents the complement of the missing region mask tensor;

[0068] The output target tensor is iteratively solved by updating the regularization parameters and the corresponding Lagrange multipliers.

[0069] In step 4, the process of iteratively solving the sub-problems of the factor matrix through the ADMM algorithm framework specifically includes the following steps:

[0070] The objective function of the subproblem of the factor matrix is ​​constructed by fixing the core tensor, the first auxiliary variable, the second auxiliary variable and the output target tensor. The objective function of the subproblem of the factor matrix has the following relationship:

[0071] ;

[0072] Known Lemma ;

[0073] in, represents the Kronecker product, , represents the Kronecker product of all other factor matrices excluding the nth factor matrix, It means to expand the tensor Y into a matrix along the nth dimension. Respectively represent N modular product operations, the Nth factor matrix, Respectively represent any number between 1 and N, the maximum number of factor matrices;

[0074] The solution to the subproblem of the factor matrix has the following relation:

[0075] ;

[0076] in, express The substitute for and Represents the core tensor , According to Dimensions are expanded into a matrix;

[0077] The factor matrix is ​​iteratively solved by updating the penalty parameters and the corresponding Lagrange multipliers.

[0078] In step 4, the process of iteratively solving the sub-problem of the second auxiliary variable through the ADMM algorithm framework specifically includes the following steps:

[0079] The objective function of the subproblem of the second auxiliary variable is constructed by fixing the core tensor, the factor matrix, the first auxiliary variable and the output target tensor. The objective function of the subproblem of the second auxiliary variable has the following relationship:

[0080] ;

[0081] After simple calculation, we can get

[0082]

[0083] Here, express The transpose operator and the right side of the formula Expressed as The closed-form solution can be derived from the following expression:

[0084] ;

[0085] in, represents the element-wise square operation, represents element-wise division, and represent Fourier transform and its inverse transform respectively;

[0086] The second auxiliary variable is iteratively solved by updating the penalty parameter and the corresponding Lagrange multiplier.

[0087] In step 4, the process of iteratively solving the sub-problem of the core tensor through the ADMM algorithm framework specifically includes the following steps:

[0088] The objective function of the subproblem of the core tensor is constructed by fixing the factor matrix, the first auxiliary variable, the second auxiliary variable and the output target tensor. The objective function of the subproblem of the core tensor has the following relationship:

[0089] ;

[0090] According to the properties of Kronecker product, tensor product multiplication is converted into vectorized form through matrix vectorization, so that the objective function of iterative solution of sub-problems of core tensor can be written as a least squares problem;

[0091] The objective function of the subproblem of the core tensor corresponds to the following relation:

[0092] ;

[0093] You can get:

[0094] ;

[0095] in, The vector form of the subproblem representing the kernel tensor, represents the identity matrix;

[0096] The solution to the subproblem for the core tensor is , yes The inverse operation of

[0097] The kernel tensor is iteratively solved by updating the regularization parameter, penalty parameter and the corresponding Lagrange multiplier.

[0098] See also Figure 2 The iteration termination condition is The relative error is less than a given positive number, and the corresponding relationship is as follows:

[0099] ;

[0100] in, Relative error, represents the X obtained in the previous iteration, represents the Frobenius norm.

[0101] Otherwise update the Lagrange multiplier , , , and penalty parameters , the specific updates are as follows:

[0102] ;

[0103] ;

[0104] ;

[0105] ;

[0106] ;

[0107] in, Represents the learning rate.

[0108] Please refer to Figure 3 This embodiment further provides a tensor completion system based on a low-rank Tucker decomposition factor matrix, wherein the system applies the tensor completion method based on a low-rank Tucker decomposition factor matrix as described above, and the system includes:

[0109] Model building modules for:

[0110] Based on Tucker decomposition, the horizontal and vertical gradients of the factor matrix in Tucker decomposition are constrained by the nuclear norm as low-rank constraints, and the nuclear tensor is constrained by the Frobenius norm to construct a tensor complete model.

[0111] Constrained Optimization Module for:

[0112] Auxiliary variables are introduced into the tensor complete model, and an augmented Lagrangian function is established based on Lagrangian multipliers to perform constrained optimization on the tensor complete model, thus obtaining a constrained optimization model.

[0113] Data acquisition module, used to:

[0114] Get the missing data image and normalize it, create a corresponding mask according to the missing data in the missing data image, store the missing data image and the mask in tensor form, and obtain a tensor containing missing values ​​and a missing area mask tensor;

[0115] Iterative solution module for:

[0116] The tensor of missing values ​​and the mask tensor of missing regions are input into the constrained optimization model. The core tensor, factor matrix, auxiliary variables and output target tensor are taken as sub-problems respectively. Each sub-problem is iteratively solved through the ADMM algorithm framework to obtain the complete tensor.

[0117] In order to verify the performance and effect of the present invention, some real data are used for experimental verification.

[0118] Since color images, hyperspectral images and videos are all typical tensor data types and exhibit low-rank and smooth characteristics, the present invention is widely applicable to these fields. In this example, three groups of experiments were designed to verify this applicability. The data used in all experiments are from public data sets, ensuring the reliability and repeatability of the test results.

[0119] In the first set of experiments, two color images "house" and "woman" were selected, and the image size was , the pixel values ​​of all experimental data are normalized to [0,1], and the constructed tensor is In the original data Randomly take 30% of the pixels and construct the observation tensor containing the missing elements The experimental setup , , , , .

[0120] The processing results of the present invention on color images are as follows: Figure 4 Specifically, Figure 4(a) shows the original color images - "house" and "woman", which are rich in color information; Figure 4 (b) shows a set of constructed image examples, in which 70% of the elements are lost and only 30% are retained, so most of the image information is significantly missing; Figure 4 (c) shows the color image restored after applying the technology of the present invention, from which the restoration effect is clearly visible. The reason why the present invention can achieve such results is that it cleverly combines the consideration of data reconstruction error and the low rank and smoothness characteristics of the characterization factor matrix. This strategy not only effectively reduces the overall error, but also ensures the structural coherence of the reconstructed image and the authenticity of the texture details.

[0121] In order to further verify the effectiveness of the present invention, this example also conducted experiments on hyperspectral images. In the second set of experiments, two common hyperspectral datasets "PaC (Pavia Center Hyperspectral Dataset)" and "Botswana (Botswana Hyperspectral Dataset)" were selected. The size of the hyperspectral image is , the pixel values ​​of all experimental data are normalized to [0,1], and the constructed tensor is Randomly select 30% of the pixels in the original data and construct an observation tensor containing the missing elements The experimental setup , , , , .

[0122] The processing results of the present invention on the hyperspectral image are as follows: Figure 5 Specifically, Figure 5 (a) depicts the 25th band of the original hyperspectral datasets “PaC” and “Botswana” that are rich in spectral information; Figure 5 (b) shows a hyperspectral image with 70% element loss, simulating the extreme data loss that may be encountered in practical applications; Figure 5 Figure (c) demonstrates the restoration capability of the present invention, which not only restores most of the lost information, but also ensures the structural authenticity and texture fineness of the restored image. Even when the original data is severely damaged, the restored image can still accurately outline the edges of the object, avoiding common blurring or breakage phenomena.

[0123] In order to verify the universality of the present invention, this example also conducts experiments on videos. In the third group of experiments, two common video data sets "Claire" and "Ajiyang" were selected. The size of the video data is , the pixel values ​​of all experimental data are normalized to [0,1], and the constructed tensor is Randomly select 30% of the pixels in the original data M and construct an observation tensor containing missing elements The experimental setup , , , , .

[0124] This experiment focuses on video data analysis, and its results are Figure 6 Intuitive display. Figure 6 (a) in the figure shows the 25th frame of two typical video sequences "Claire" and "Ajiyang" that are rich in temporal information. Figure 6 (b) shows an extreme case of up to 70% data loss in the video image. This setting simulates the complexity of real-world data loss. Figure 6 (c) in the figure demonstrates the powerful performance of the present invention in video data restoration. The present invention not only successfully restores most of the lost elements, but also ensures the structural accuracy and richness of texture details of the restored image, both of which are key indicators of image quality.

[0125] Table 1 shows the restoration results of the color images "house" and "woman", the hyperspectral images "PaC" and "Botswana", and the video sequences "Claire" and "Akiyo" at a missing rate of 70%. This example uses three image quality indicators, including PSNR (peak signal-to-noise ratio), SSIM (structural similarity index), and FSIM (feature similarity index measurement). The larger the PSNR value, the closer the SSIM and FSIM values ​​are to 1, indicating that the restored image is closer to the original image and the better the restoration effect. The experimental results show that the present invention can effectively repair observed images containing missing values, and experiments on different data sets also prove the robustness of the present invention.

[0126] Table 1: Reconstruction results of color images, hyperspectral images, and video sequences at a missing rate of 70%;

[0127]

[0128] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0129] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0130] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A tensor completion method based on low-rank Tucker decomposition factor matrix, It is characterized in that The method comprises the following steps: Step 1: Based on Tucker decomposition, the horizontal and vertical gradients of the factor matrix in Tucker decomposition are constrained by the nuclear norm as low-rank constraints, and the nuclear tensor is constrained by the Frobenius norm to construct a tensor complete model; Step 2: Introduce auxiliary variables into the tensor complete model, establish an augmented Lagrangian function based on Lagrangian multipliers to perform constrained optimization on the tensor complete model, and obtain a constrained optimization model; Step 3: Get the missing data image and normalize it, create a corresponding mask according to the missing data in the missing data image, store the missing data image and the mask in tensor form, and obtain a tensor containing missing values ​​and a missing area mask tensor; Step 4: Input the missing value tensor and the missing region mask tensor into the constrained optimization model, and use the core tensor, factor matrix, auxiliary variables, and output target tensor as sub-problems. Use the ADMM algorithm framework to iteratively solve each sub-problem to obtain the complete tensor. In step 4, the auxiliary variables include a first auxiliary variable and a second auxiliary variable, and the process of iteratively solving the sub-problem of the first auxiliary variable through the ADMM algorithm framework specifically includes the following steps: The first auxiliary variable includes two sub-problems. The objective functions of the two sub-problems of the first auxiliary variable are constructed by fixing the core tensor, the factor matrix, the second auxiliary variable and the output target tensor. The objective functions of the two sub-problems of the first auxiliary variable are respectively expressed as follows: ; ; Based on the singular value threshold algorithm, the solution relationship of the two sub-problems of the first auxiliary variable is constructed. The solutions of the two sub-problems of the first auxiliary variable have the following relationship: ; ; in, Indicates that the input parameters are Singular value threshold algorithm for ; The two subproblems of the first auxiliary variable are iteratively solved by updating the penalty parameter, the regularization parameter and the corresponding Lagrange multiplier; The process of iteratively solving the sub-problem of outputting the target tensor through the ADMM algorithm framework specifically includes the following steps: The objective function of the subproblem of outputting the target tensor is constructed by fixing the core tensor, the factor matrix, the first auxiliary variable, and the second auxiliary variable. The objective function of the subproblem of outputting the target tensor has the following relationship: ; The solution to the subproblem of outputting the target tensor has the following relation: ; in, represents the projection operator, The projection operator representing the complement of the missing region mask tensor, Represents the complement of the missing region mask tensor; The output target tensor is iteratively solved by updating the penalty parameters and the corresponding Lagrange multipliers; The process of iteratively solving the sub-problems of the factor matrix through the ADMM algorithm framework specifically includes the following steps: The objective function of the subproblem of the factor matrix is ​​constructed by fixing the core tensor, the first auxiliary variable, the second auxiliary variable and the output target tensor. The objective function of the subproblem of the factor matrix has the following relationship: ; Known Lemma ; in, represents the Kronecker product, , represents the Kronecker product of all other factor matrices excluding the nth factor matrix, It means to expand the tensor Y into a matrix along the nth dimension. Respectively represent N modular product operations, the Nth factor matrix, Respectively represent any number between 1 and N, the maximum number of factor matrices; The solution to the subproblem of the factor matrix has the following relation: ; in, express The substitute for and Represents the core tensor , According to Dimensions are expanded into a matrix; The factor matrix is ​​iteratively solved by updating the penalty parameters and the corresponding Lagrange multipliers; The process of iteratively solving the sub-problem of the second auxiliary variable through the ADMM algorithm framework specifically includes the following steps: The objective function of the subproblem of the second auxiliary variable is constructed by fixing the core tensor, the factor matrix, the first auxiliary variable and the output target tensor. The objective function of the subproblem of the second auxiliary variable has the following relationship: ; By calculation: ; in, express The transpose operator of ; Will Expressed as , the closed-form solution of the second auxiliary variable exists as follows: ; in, represents the element-wise square operation, represents element-wise division, and represent Fourier transform and its inverse transform respectively; The second auxiliary variable is iteratively solved by updating the penalty parameter and the corresponding Lagrange multiplier; The process of iteratively solving the sub-problems of the core tensor through the ADMM algorithm framework specifically includes the following steps: The objective function of the subproblem of the core tensor is constructed by fixing the factor matrix, the first auxiliary variable, the second auxiliary variable and the output target tensor. The objective function of the subproblem of the core tensor has the following relationship: ; According to the properties of the Kronecker product, the tensor product multiplication is converted into a vectorized form through matrix vectorization, so that the objective function of iteratively solving the sub-problem of the core tensor is written as a least squares problem. The corresponding relationship is as follows: ; You can get: ; in, The vector form of the subproblem representing the kernel tensor, represents the identity matrix; The solution to the subproblem for the core tensor is , represents an operation that stacks the columns of a matrix or tensor into a single column vector. yes The inverse operation of The kernel tensor is iteratively solved by updating the regularization parameter, penalty parameter and the corresponding Lagrange multiplier.

2. The tensor completion method based on low rank of Tucker decomposition factor matrix according to claim 1, It is characterized in that In step 1, the tensor complete model has the following relationship: ; in, represents a tensor with missing values, , represents the field of real numbers, Respectively represent dimension information, represents the missing region mask tensor, , represents the output target tensor, , , represents different regularization parameters, represents the number of spatial dimensions, represents the number of factor matrices, represents the kernel tensor, represents the gradient multiplication operation along the kth dimension, represents the nuclear norm, represents the square of the Frobenius norm, represents the linear operation of projection, represents the factor matrix, They respectively represent that the kernel tensor is expanded along the first dimension and multiplied by the first factor matrix, the kernel tensor is expanded along the second dimension and multiplied by the second factor matrix, and the kernel tensor is expanded along the third dimension and multiplied by the third factor matrix.

3. The tensor completion method based on low rank of Tucker decomposition factor matrix according to claim 2, It is characterized in that In step 2, the process of introducing auxiliary variables into the tensor complete model has the following relationship: ; in, represents the first auxiliary variable, represents the second auxiliary variable, represents the indicator function, which is defined as: 。 4. The tensor completion method based on low rank of Tucker decomposition factor matrix according to claim 3, It is characterized in that In step 2, the process of constraining the tensor complete model by establishing an augmented Lagrangian function based on Lagrangian multipliers has the following relationship: ; in, represents the penalty parameter, denotes different Lagrange multipliers, Represent the two sub-problems of the first auxiliary variable, They represent gradient multiplication along the first dimension and gradient multiplication along the second dimension respectively.

5. A tensor-complete system based on low-rank Tucker factorization matrix, It is characterized in that The system applies the tensor completion method based on low rank of Tucker decomposition factor matrix according to any one of claims 1 to 4, and the system comprises: Model building modules for: Based on Tucker decomposition, the horizontal and vertical gradients of the factor matrix in Tucker decomposition are constrained by the nuclear norm as low-rank constraints, and the nuclear tensor is constrained by the Frobenius norm to construct a tensor complete model. Constrained Optimization Module for: Auxiliary variables are introduced into the tensor complete model, and an augmented Lagrangian function is established based on Lagrangian multipliers to perform constrained optimization on the tensor complete model, thus obtaining a constrained optimization model. Data acquisition module, used to: Get the missing data image and normalize it, create a corresponding mask according to the missing data in the missing data image, store the missing data image and the mask in tensor form, and obtain a tensor containing missing values ​​and a missing area mask tensor; Iterative solution module for: The tensor of missing values ​​and the mask tensor of missing regions are input into the constrained optimization model. The core tensor, factor matrix, auxiliary variables and output target tensor are taken as sub-problems respectively. Each sub-problem is iteratively solved through the ADMM algorithm framework to obtain the complete tensor.