Self-correction recognition method and device for valuable document recognition device

Patent metadata

Field	Value
Publication number	US-10452941-B2
Application number	US-201615567567-A
Country	US
Kind code	B2
Filing date	Mar 30, 2016
Priority date	Apr 23, 2015
Publication date	Oct 22, 2019
Grant date	Oct 22, 2019

Abstract

A self-correction recognition method and device for a valuable document recognition device is provided. The method includes: extracting a feature of a valuable document; determining whether the feature falls in a discriminant library of any preset category, and obtaining a substitutive centroid according to a feature centroid and the feature in a case that the feature falls in a discriminant library of any preset category; and updating the discriminant library of the category by replacing the feature centroid of the category with the substitutive centroid. The discriminant library is pre-constructed according to the feature centroid of the category, and the feature centroid is calculated from a feature of the category.

First claim

The invention claimed is:

1. A self-correction recognition method for a valuable document recognition device in an intelligent processing device for valuable documents, comprising:
extracting a feature M of a valuable document entering in the valuable document recognition device;
determining whether the feature M falls in a discriminant library of a category i,
obtaining a substitutive centroid O′ i based on a feature centroid O i and the feature M and determining the valuable document as in the category i, in a case that the feature M falls in the discriminant library of the category i; and
updating the discriminant library of the category i by replacing the feature centroid O i of the category i with the substitutive centroid O′ i , wherein,
the discriminant library is pre-constructed based on the feature centroid O i of the category i; and
the feature centroid O i is calculated from a feature of the category i; and
the substitutive centroid O′ i is expressed as:
O′ i =(1−γ) O i +γM,

where an update coefficient γ is less than 1 and greater than 0.

2. The method according to claim 1 , wherein before extracting the feature M of the valuable document, the method further comprises:
extracting features of n categories of a valuable document, wherein each of the n categories comprises one or more features, and n is greater than 1;
calculating the feature centroid O i of the category i based on the features of the category i;
obtaining a relative discriminant plane group of the category i based on the feature centroid O i , wherein the relative discriminant plane group of the category i is composed of relative discriminant planes of the category i relative to the other n−1 categories, and a relative discriminant plane l iη of the category i relative to a category η satisfies that the relative discriminant plane l iη is perpendicular to a connection line between the feature centroid O i of the category i and a feature centroid O η of the category η, valuable documents of the category i and the category η are respectively divided to the two sides of the plane, all valuable documents of the category i are located on a positive side of the relative discriminant plane l iη , all valuable documents of the category η are located on a negative side of the relative discriminant plane l iη , and a minimum value of distances between valuable documents of the category i and the relative discriminant plane l iη is smaller than a minimum value of distances between valuable documents of the category η and the relative discriminant plane l iη , and i is not equal to η;
building a discriminant sphere of the category i with a radius R i , wherein R i is determined by the feature centroid O i and the relative discriminant plane group of the category i; and
determining an intersecting area between the relative discriminant plane group of the category i and the discriminant sphere of the category i as a discriminant library of the category i.

3. The method according to claim 2 , wherein the relative discriminant plane is expressed as:

w

i
⁢

⁢
η

T

*
X

+

d

i
⁢

⁢
η

=
0

,

where X is a space vector of the valuable document feature, w iη is a normal vector of the relative discriminant plane l iη , and d iη is an intercept of the relative discriminant plane l iη .

4. The method according to claim 3 , wherein the radius R i , is expressed as:

R
i

=

min
⁡

(

ɛ
i

⁢

ɛ
i

+

χ
i

2

)

,

where ε i is a minimum value of distances between the feature centroid O i of the category i and the relative discriminant plane group of the category i, and χ i is a maximum value of distances between the feature centroid O i of the category i and all the valuable documents of the category i.

5. The method according to claim 2 , wherein the calculating the feature centroid O i of the category i based on the features of the category i comprises:
obtaining the features of the category i; and
calculating an arithmetic mean of the features of the category i as the feature centroid O i .

6. The method according to claim 4 , wherein,

ɛ
i

=

min

η
≠
i

⁢

(



O
i

⁢

l

i
⁢

⁢
η



)

=

min

η
≠
i

⁢

(



w

i
⁢

Assignees

• Grg Banking Equipment Co Ltd

Inventors

• Luo Panfeng
• Wang Rongqiu
• Liu Mengtao
• Wei Dong

Classifications

• G06V30/19173

  Classification techniques · CPC title

• G06V30/414 Primary

  Extracting the geometrical structure, e.g. layout tree; Block segmentation, e.g. bounding boxes for graphics or text · CPC title

• G06F18/2411

  based on the proximity to a decision surface, e.g. support vector machines · CPC title

• G06F18/24137

  Distances to cluster centroïds · CPC title

• G06K9/00463

  Physics · mapped topic