Method of extracting feature of image to recognize object

Patent metadata

Field	Value
Publication number	US-2017236000-A1
Application number	US-201615268988-A
Country	US
Kind code	A1
Filing date	Sep 19, 2016
Priority date	Feb 16, 2016
Publication date	Aug 17, 2017
Grant date	—

Abstract

A method of converting a vector corresponding to an input image includes receiving input vector data associated with an input image including an object; and converting the received input vector data into feature data based on a projection matrix having a fixed rank, wherein a first dimension of the input vector data is higher than a second dimension of the feature data.

First claim

What is claimed is:

1 . A method of converting a vector corresponding to an input image, the method comprising:
receiving input vector data associated with an input image including an object; and converting the received input vector data into feature data based on a projection matrix having a fixed rank, wherein a first dimension of the input vector data is higher than a second dimension of the feature data.

2 . The method of claim 1 , wherein the converting comprises:
determining average vector data-subtracted input vector data by subtracting average vector data from the input vector data, and applying the projection matrix having the fixed rank to the average vector data-subtracted input vector data.

3 . The method of claim 1 , wherein the projection matrix is generated based on a covariance matrix corresponding to training vector data and a dictionary which is a set of elements constituting an object in an image.

4 . The method of claim 1 , wherein the projection matrix is generated such that feature data obtained by reflecting the projection matrix in the input vector data is to be the same as vector data obtained by adding a sparse error matrix to a product of a coefficient matrix and a dictionary which is a set of elements constituting an object in an image.

5 . The method of claim 4 , wherein the projection matrix is generated based on metric calculation using a covariance matrix and a rank of the coefficient matrix.

6 . The method of claim 1 wherein,
the receiving includes receiving first image vector data associated with a first image and second image vector data associated with at least one second image,
the converting includes converting the first image vector data and the second image vector data into second dimensional first image vector data and second dimensional second image vector data based on the projection matrix, and
the method further comprises:
comparing the first image and each of the at least one second image based on the second dimensional first image vector data and the second dimensional second image vector data.

7 . The method of claim 1 , wherein the receiving comprises:
receiving the input image; detecting an object area including the object from the input image; and extracting the input vector data from the input image.

8 . The method of claim 1 , further comprising:
determining a similarity by performing a comparison based on the feature data and feature data obtained by converting input vector data extracted from an image other than the input image; and determining that the object included in the input image is the same as an object included in the other image when the similarity is greater than a threshold similarity.

9 . The method of claim 1 , further comprising:
determining a similarity by performing a comparison based on the feature data and stored feature data corresponding to a plurality of objects; and determining that the object included in the input image is the same as an object corresponding to feature data having a similarity greater than a threshold similarity, among the stored feature data.

10 . A method of learning a projection matrix to convert a dimension of vector data associated with an input image, the method comprising:
receiving training data sets corresponding to a plurality of training images, respectively; and jointly learning a first projection matrix having a fixed rank and a dictionary which is a set of elements constituting an object in an image, based on the training data sets.

11 . The method of claim 10 , wherein the learning comprises maintaining a rank of the first projection matrix to be the fixed rank.

12 . The method of claim 10 , wherein the jointly learning comprises:
generating the first projection matrix based on a covariance matrix corresponding to training vector data.

13 . The method of claim 10 , wherein the learning comprises:
determining the first projection matrix such that a difference between elements corresponding to a same object is to be reduced and a difference between elements corresponding to different objects is to be increased in training vector data.

14 . The method of claim 10 , wherein the learning comprises:
determining the first projection matrix such that vector data obtained by reflecting the first projection matrix in input vector data corresponding to an input image is to be the same as vector data obtained by adding a sparse error matrix to a product of a coefficient matrix and the dictionary.

15 . The method of claim 14 , wherein the learning comprises:
determining the first projection matrix based on metric calculation using a covariance matrix and a rank of the coefficient matrix.

16 . The method of claim 14 , wherein the learning comprises:
generating a convergent projection matrix and a convergent dictionary by iteratively determining the first projection matrix and the dictionary.

17 . A method of recognizing an image, the method comprising:
extracting input vector data from an input image; converting the input vector data into input feature data based on a first projection matrix having a fixed rank; generating a training data set including the input vector data based on the input feature data; learning a second projection matrix and a dictionary based on the generated training data set; and correcting the first projection matrix based on the second projection matrix.

18 . The method of claim 17 , wherein the learning comprises:
determining a coefficient matrix and the second projection matrix based on the training data set; determining the dictionary based on the second projection matrix and the coefficient matrix; and iteratively determining the second projection matrix and the dictionary until the second projection matrix and the dictionary converge.

19 . The method of claim 17 , wherein the generating comprises:
mapping a label corresponding to registered feature data stored in a database to the input vector data when a similarity between the input feature data and the registered feature data is greater than or equal to a threshold similarity; and generating a training data set including the input vector data to which the label is mapped.

20 . A non-transitory computer-readable medium comprising computer-readable instructions that, when executed by one or more processors, cause the one or more processors to perform the method of claim 1 .

Assignees

• Samsung Electronics Co Ltd
• Univ Northeastern

Inventors

• Hwang Wonjun
• Suh Sungjoo
• Han Jaejoon
• Choi Chang Kyu
• Fu Yun
• Ding Zhengming
• Shao Ming

Classifications

• G06F18/2132 Primary

  based on discrimination criteria, e.g. discriminant analysis · CPC title

• G06V40/168

  Feature extraction; Face representation · CPC title

• G06F18/22

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

• G06F18/2411

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

• G06K9/6269

  Physics · mapped topic