Multi-view vector processing method and multi-view vector processing device

Patent metadata

Field	Value
Publication number	US-10796205-B2
Application number	US-201815971549-A
Country	US
Kind code	B2
Filing date	May 4, 2018
Priority date	May 16, 2017
Publication date	Oct 6, 2020
Grant date	Oct 6, 2020

Abstract

A multi-view vector processing method and a multi-view vector processing device are provided. A multi-view vector x represents an object containing information on at least two non-discrete views. A model of the multi-view vector, where the model includes at least components of: a population mean μ of the multi-view vector, view component of each view of the multi-view vector and noise is established. The population mean μ, parameters of each view component and parameters of the noise , are obtained by using training data of the multi-view vector x. The device includes a processor and a storage medium storing program codes, and the program codes implements the aforementioned method when being executed by the processor.

First claim

The invention claimed is:

1. A method of multi-view vector processing by a processor, where a multi-view vector x represents an object containing information on at least two non-discrete views, the method comprising:
establishing a model of the multi-view vector x, where the model includes at least components of: a population mean μ of the multi-view vector x, a view component of a view among the at least two non-discrete views of the multi-view vector x and noise ; and
using training data of the multi-view vector x to obtain the population mean μ, parameters of the view component and parameters of the noise ,
where,
the multi-view vector is obtained by processing a feature vector with a classifier, and the feature vector is obtained by directly vectorizing the object, and
the classifier is configured to relatively separate the multi-view vector from the feature vector obtained by directly vectorizing the object to be represented, and a discreteness between an excluded view and the two non-discrete views of the multi-view vector x is higher than a discreteness between the two non-discrete views of the multi-view vector x.

2. The method according to claim 1 , where the population mean μ, is set as zero.

3. The method according to claim 1 , where the view component of the view is based on a product of a space basis S i corresponding to the view and a coefficient u i —selected for the view, where i is a sequence number of the view.

4. The method according to claim 3 , where the noise is set to meet a Gauss distribution taking a diagonal matrix Σ as a covariance.

5. The method according to claim 4 , where to use the training data includes: obtaining the population mean μ, space base S n of the view and the Σ, based on the training data by using an expectation-maximization algorithm.

6. The method according to claim 5 , where in the expectation-maximization algorithm, mean expected values of a plurality of samples for the multi-view vector x with respect to the selected coefficient u i for the view component of the view, and expected values related to covariance with respect to the selected coefficient u i for the view component of the view, are calculatable based on μ, S n and Σ, and μ, S n and Σ are recalculatable until the mean expected values of the samples for the multi-view vector x and the expected values related to covariance converge.

7. The method according to claim 4 , where space bases of the at least two non-discrete views are respectively recorded as S and T, and the multi-view vector x is represented as
x ijk =μ+ Su i +Tv j +ϵ ijk

where μ represents the population mean, u i represents a coefficient corresponding to an i-th selection for the view corresponding to the space basis S, v j represents a coefficient corresponding to a j-th selection for the view corresponding to the space basis T, ϵ ijk represents the noise , and k represents a k-th sample under the i-th selection and the j-th selection.

8. The method according to claim 7 , where if θ={μ, S, T, Σ} and B=[S T], then the following distribution is met:
P ( x ijk |u i ,v j ,θ)= N ( x ijk |μ+Su i +Tv j ,Σ),

P ( u i )= N ( u i |0, I ),
P ( v j )= N ( v j |0, I ),
where N(x|μ, Σ) is a normal distribution with a mean of μ, and a variance of Σ, and | is a unit matrix.

9. The method according to claim 7 , where the multi-view vector x ijk represents a voiceprint for a k-th sample of a j-th type of text by an i-th speaker, u i is a coefficient of the i-th speaker and v j is a coefficient of the j-th type of text.

10. The method according to claim 1 , further including;
calculating a first likelihood representing that at least one view component is same among view components of at least two non-discrete views in two multi-view vectors and a second likelihood representing that the at least one view component is different among the view components of the at least two non-discrete views in the two multi-view vectors, by using population mean μ, parameters of a view component and parameters of noise of respective two multi-view vectors; and
determining whether the at least one view component is same in the two multi-view vectors based on the first and second likelihoods.

11. The method according to claim 10 , further including calculating a first probability representing that at least one view component is same among view components of at least two non-discrete views in the two multi-view vectors and a second probability representing that the at least one view component is different among the view components of the at least two non-discrete views in the two multi-view vectors based on the calculated first and second likelihoods, and determining whether the at least one view component is same in the two multi-view vectors based on the first and second probabilities.

12. The method according to claim 10 , further including: determining whether at least two of the view components are same among view components of at least two non-discrete views in the two multi-view vectors.

13. The method according to claim 8 , further including:
calculating a first likelihood representing that two view components both are same among view components of at least two non-discrete views in two multi-view vectors and a second likelihood representing that the two view components are different among the view components of the at least two non-discrete views in the two multi-view vectors based on the determined parameters of the multi-view vector model, and determining whether the two view components are both the same in the two multi-view vectors based on the first and second likelihoods, where
the first likelihood representing that a plurality of the view components are same,

A
=

𝒩
⁡

(

[

x
t

x
s

]

|

[

μ

μ

]

,

[

SS
T

+

Assignees

• Fujitsu Ltd

Inventors

• Shi Ziqiang
• Liu Liu
• Liu Rujie

Classifications

• G06V10/806 Primary

  of extracted features · CPC title

• G06F18/2415

  based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate · CPC title

• G06F18/22

  Matching criteria, e.g. proximity measures · CPC title

• G06F2218/08 Primary

  Feature extraction · CPC title

• G06F18/217

  Validation; Performance evaluation; Active pattern learning techniques · CPC title