Small Vein Image Recognition and Authorization Using Constrained Geometrical Matching and Weighted Voting Under Generic Tree Model

• US 20140016830A1
• Filed: 10/31/2012
• Published: 01/16/2014
• Est. Priority Date: 07/13/2012
• Status: Active Grant

• Current Assignee: Seiko Epson Corporation (Seiko Group)
• Original Assignee: Seiko Epson Corporation (Seiko Group)
• Inventors: Wang, Jinjun, Xiao, Jing
• Granted Patent: US 8,768,049 B2
• Time in Patent Office: Days
• US Class Current: 382/115
• CPC Class Codes: G06V 10/426 Graphical representations
• CPC Class Codes: G06V 30/1988 Graph matching graphical image representation G06V30/18181
• CPC Class Codes: G06V 30/422 Technical drawings; Geographical maps
• CPC Class Codes: G06V 40/10 Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
• CPC Class Codes: G06V 40/14 Vascular patterns

Abstract

An automated registration and authentication system combines a generative and discriminative approach to improve the matching of a query object to a database of registered objects. The discriminative approach uses a voting mechanism to identify a most likely match, and the generative approach uses ASIFT transforms to determine a best geometric match. The two results are combined using a technique base on Bayesian inference theory.

Claims

1. A method of searching for a query object within an object class, said method comprising:

(a) accessing a collection of unique training samples of multiple training objects within said object class;

(b) defining a separate training set of training item descriptors from each of said training samples;

(c) creating a composite collection of training item descriptors from the separate training sets of sample item descriptors;

(d) creating a hierarchical tree from said composite collection of training item descriptors according to relations in the training item descriptors, said hierarchical tree having a plurality of leaf nodes;

(e) accessing registration sets of registration item descriptors defined from respective registration samples obtained from registration objects of said object class, distributing said registration sets of registration item descriptors into said hierarchical tree according to said relations defined in the creation of said hierarchical tree, indexing the registration item descriptors clustered within each leaf node to their corresponding registration samples, said indexing including defining reverse index (RI) information at each leaf node specifying for each registration item descriptor within the leaf node, an ID label identifying its corresponding registration sample from which it was defined and geometric information obtained as part of its definition;

(f) accessing a query sample from said query object, defining a query set of query item descriptors from said query sample, distributing said query set of query item descriptors into said hierarchical tree according to said relations defined in the creation of said hierarchical tree, each query item descriptor that reaches a leaf node defining a separate potential descriptor-match pair with each individual registration item descriptor that is within the same reached leaf node;

(g) submitting the RI information of each leaf node reached by a query item descriptor to a first generative-and-descriminative identification process, wherein;

(i) said generative-and-descriminative identification process applies a descriminative matching model to the potential descriptor-match pairs using the ID label information provided by the RI information, the descriminative matching model identifying a first discriminatively-matched registration object with a first descirmiantive confidence;

(ii) said generative-and-descriminative identification process applies a generative matching model to the potential descriptor-match pairs using the geometric information within the RI information, said generative matching model identifying a transform that best matches the query item descriptors to a their paired registration item descriptors, and identifying as a first generative-matched registration object with a first generative confidence the registration object best represented by the registration item descriptors matched to the query item descriptors by the identified transform; and

(iii) combining the first descirmiantive confidence and the first generative confidence to determine a registration object that matches the query object.

2. The method of claim 1, wherein in (i), the applying of said descriminative matching model to the potential descriptor-match pairs omits use of any geometric information within the RI information.

3. The method of claim 1, wherein in (ii), the identified transform is a SIFT transform.

4. The method of claim 1, wherein in (ii), the identified transform is an Affine SIFT transform.

5. The method of claim 1, wherein in (g) the geometric information obtained as part of its definition include the relative position and orientation of the registration item descriptor within its respective registration sample.

6. The method of claim 1, wherein in (ii), the generative matching model is defined as:

7. The method of claim 6, wherein the alignment error is a Gaussian defined as:

8. The method of claim 1, wherein in (i), the descriminative matching model uses a voting scheme based on the number of ID labels represented at each leaf node reached by a query item descriptor.

9. The method of claim 1, wherein:

in (e), the RI information of each leaf node includes a registration path vector of each registration item descriptor through the hierarchical tree on its way to reaching a leaf node;

in (f), a query path vector is defined for each query item descriptor that reaches a leaf node, the query path vector being a path vector of each query item descriptor through the hierarchical tree on its way to reaching a leaf node; and

in (i), descriminative matching model compares the query path vectors and registration path vectors of the potential descriptor-match pairs in its identifying of the first discriminatively-matched registration object with a first descirmiantive confidence.

10. The method of claim 1, wherein the number of leaf nodes is N, and X defines the set of query item descriptors extracted from query sample, and the descriminative matching model uses a voting process for registered object l that factorizes a posterior P_O(l|X) into a per-leaf node estimation defined as:

11. The method of claim 10, wherein the descriminative matching model uses a Term Frequency—

Inverse Document Frequency (TF-IDF) technique where each tree node is given an ID-independent weight w_j defined as

12. The method of claim 11, wherein each registration sample with ID label l defines a “

path vector”

d_li at leaf node n_i, the dimension of each path vector d_li equals to the depth of leaf node n_i in the hierarchical tree, each dimension d_j of path vector d_li is equal to w_jN_j the path vector is stored in the RI information of leaf each leaf node n_i, the query sample defines a path vector v, and the descriminative matching model defines said first descirmiantive confidence as;

13. The method of claim 1, wherein (iii), the combined first descirmiantive confidence and the first generative confidence to determine a registration object that matches the query object is defined as

14. The method of claim 1, wherein the query items descriptors that that are not matched to the first discriminatively-matched registration object or to the first generative-matched registration object are re-submitted to a second generative-and-descriminative identification process to identify a second descriminatively-matched registration object and second generative-matched registration object, and the results of the first and second generative-and-descriminative identification processes are compared to determined if a registration object may be matched to the query object.

15. The method of claim 14, wherein the query object is authenticated if the first and second generative-and-descriminative identification processes agree on the matched registration object.

16. The method of claim 14, wherein the query items descriptors that that are not matched to the second descriminatively-matched registration object or to the second generative-matched registration object are re-submitted to a third generative-and-descriminative identification process to identify a third descriminatively-matched registration object and a third generative-matched registration object, and the results of the first, second, and third generative-and-descriminative identification processes are compared to determined if a registered object is matched to the query object.

17. The method of claim 1, wherein in (iii), the first descirmiantive confidence and the first generative confidence are combined using a technique based on Bayesian inference theory.

18. The method of claim 1, wherein said object class is a finger vein class.

19. A non-transient computer readable medium having computer-executable instruction for implementing the method of claim 1.