A Tensor Recovery Method Based on Exponentialized Kernel Norm and Mixed Singular Value Truncation

A recovery method and nuclear norm technology, applied in image data processing, instrumentation, computing, etc., can solve problems such as difficulty in obtaining local optimum, difficulty in directly optimizing CP rank, difficulty in selecting weights, etc., to achieve tensor recovery, fast Efficient Tensor Recovery, Effect of Reduced Complexity

Active Publication Date: 2016-09-07
WENZHOU UNIV
View PDF4 Cites 0 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

However, the geometric meaning of the Tucker rank definition is not clear, and it is difficult to choose the weights of the tensor in different modes. If the rank of the expanded matrix in a certain mode is large, but the corresponding weight is small, then the above definition will not be correct. Describe the rank structure of the tensor, which leads to the unsatisfactory low-rank recovery effect of the tensor
The CP rank of the tensor is an extension of the matrix rank, and its geometric meaning is clearer than the Tucker rank. However, directly optimizing the CP rank is a very difficult problem, and it is even difficult to obtain a local optimum.

Method used

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View more

Image

Smart Image Click on the blue labels to locate them in the text.
Viewing Examples
Smart Image
  • A Tensor Recovery Method Based on Exponentialized Kernel Norm and Mixed Singular Value Truncation
  • A Tensor Recovery Method Based on Exponentialized Kernel Norm and Mixed Singular Value Truncation
  • A Tensor Recovery Method Based on Exponentialized Kernel Norm and Mixed Singular Value Truncation

Examples

Experimental program
Comparison scheme
Effect test

Embodiment Construction

[0025] see figure 1 ,本发明提供一种基于指数化核范数与混合奇异值截断的张量恢复方法,包括以下三个步骤:

[0026] (1)提出一种新的张量秩定义:张量不同模态下展开矩阵秩的最大值;该定义是张量CP秩的下界,能够有效的逼近CP秩,并消除了权重参数的影响,采用核范数指数和的对数来逼近该张量秩定义,将其转化为凸函数;

[0027] (2)为了消除张量不同模态下展开的矩阵的相关性,引入一系列辅助变量来代替展开矩阵,并将约束条件利用拉格朗日乘子法转化为增广拉格朗日函数;

[0028] (3)采用交替方向法对增广拉格朗日函数中各类变量进行迭代优化,直到收敛;其中,对于核范数的指数和中的优化变量,采用混合奇异值截断算法来获得解析解。

[0029] 作为优选的,本实施例所述的步骤(1)具体包括以下子步骤:

[0030] 首先,根据张量CP秩和Tucker秩的优缺点,提出一种新的张量秩定义:张量展开矩阵秩的最大值;

[0031] 其次,将展开矩阵的秩松弛为展开矩阵的核范数,并且利用核范数的指数和的对数来逼近最大值函数,从而将上述张量的秩定义转化为凸函数。

[0032] 作为优选的,本实施例所述的步骤(2)具体包括以下子步骤:

[0033] 首先,由于张量在不同模态下的展开矩阵具有相关性,引入一系列辅助矩阵变量来替换不同模态下的展开矩阵,并增加对应的约束条件;

[0034] 其次,采用拉格朗日乘子法将所有约束条件加入到目标函数中,获得增广拉格朗日函数。

[0035] 作为优选的,本实施例所述的步骤(3)具体包括以下子步骤:

[0036] 首先,为了对增广拉格朗日函数中的不同变量进行分别优化,采用交替方向法对增广拉格朗日函数中的各类变量进行迭代优化;

[0037] 其次,对于核范数指数中的优化变量,采用混合奇异值截断算法来获得解析解。

[0038] 本发明的方法具体运行的硬件和编程语言并不限制,用任何语言编写都可以完成,为此其它工作模式不再赘述。

[0039] 本发明的实施例采用一台具有Intel Core-i3中央处理器和4G字节内存的计算机并用Matlab语言编制了基于指数化核范数与混合奇异值截断的张量恢复的工作程序,实现了本发明的方法。

[0040] 本发明的...

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to view more

PUM

No PUM Login to view more

Abstract

The invention discloses a tensor recovery method based on exponential nuclear norm and mixed singular value truncation, which mainly includes the following steps: firstly, a new definition of tensor rank is proposed, and the maximum rank of the expanded matrix under different tensor modes is Value, the logarithm of the nuclear norm exponent sum is used to approximate the tensor rank definition, and it is transformed into a convex function; secondly, in order to eliminate the correlation of the matrix expanded in different modes of the tensor, a series of auxiliary variables are introduced to replace the expanded matrix, and transform the constraints into augmented Lagrangian functions using the Lagrangian multiplier method; finally, the alternating direction method is used to iteratively optimize various variables in the augmented Lagrangian function until convergence. Among them, the present invention is a general method for optimizing variables in the exponent sum of the nuclear norm. Compared with other classical tensor recovery methods, this method can better describe the internal structure of high-dimensional data, thereby obtaining more Good recovery results.

Description

technical field [0001] 本发明涉及计算机模式识别技术领域,具体涉及一种基于指数化核范数与混合奇异值截断的张量恢复方法。 Background technique [0002] 张量恢复(tensor completion),即高维矩阵的恢复问题,对于一个部分元素缺失的高维矩阵,通过观察其已有位置的元素,从而恢复出缺失部分元素的一般性问题,是计算机模式识别领域中的研究热点之一,被广泛应用于图像去噪、图像恢复、信息推荐系统等众多领域。总的来说,大多数现有的张量恢复方法是基于低秩结构假设,即要求恢复的缺失元素使得整个张量的秩尽可能的小。目前有两种传统的定义张量秩的方法:基于张量的CP(CANDECOMP / PARAFAC)分解方法(CP秩)和Tucker分解方法(Tucker秩)。具体来说,CP秩可以定义为:用秩一张量(rank-one tensor)之和来表示给定张量需要的秩一张量的最小个数。Tucker秩可以定义为:不同模态下展开矩阵的秩的线性加权。无论是CP秩还是Tucker秩,最小化该张量秩的优化问题被证明是一个NP难问题。 [0003] 为了解决上述问题,Gandy等人采用不同模态下展开矩阵秩的和来表示张量的秩,在计算过程中,将展开矩阵的秩用矩阵的核范数来近似代替。Signoretto等人提出一种Shatten-p范数来代替展开矩阵的核范数,并由此定义了张量的秩,最后讨论了该方法与核范数之间的关系。随后,Liu等人采用不同模态下展开矩阵核范数的线性加权来近似代替Tucker秩,并将该方法应用于图像恢复和医学图像去噪。最后,Tomioka等人对张量恢复方法进行了总结,认为有两种方式可以实现张量恢复:(1)将张量按某一个模态展开成二阶矩阵,可以将张量恢复问题转化为了二阶矩阵的恢复问题;(2)采用不同模态下展开矩阵核范数的线性加权来近似代替Tucker秩。 [0004] 可以看出,上述方法的目的都是寻找张量Tucker秩的近似。然而Tucker秩定义的几何意义不清晰,并且张量不同模态下的权重难以选择,如果某个模态下展开矩阵的秩很大,而其对应的权重很小,那么上述定义将无法正确地描述张量的秩结构,从而导致张量低秩恢复效果不够理想。张量的CP秩是对矩阵秩的一个推广,它的几何意义比Tucker秩更为明确,然而直接优化CP秩...

Claims

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to view more

Application Information

Patent Timeline
no application Login to view more
Patent Type & Authority Patents(China)
IPC IPC(8): G06T5/00
Inventor 张笑钦王迪
Owner WENZHOU UNIV
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Try Eureka
PatSnap group products