Method and apparatus for classifying data, and method and apparatus for segmenting region of interest (roi)

Patent metadata

Field	Value
Publication number	US-2016162805-A1
Application number	US-201514944717-A
Country	US
Kind code	A1
Filing date	Nov 18, 2015
Priority date	Dec 3, 2014
Publication date	Jun 9, 2016
Grant date	—

Abstract

A method and an apparatus are described to classify data. The method and apparatus includes selecting a hypothesis class among entire classes. The method and corresponding apparatus generate output data with regard to the entire classes by applying a classification algorithm to input data, and modify the input data to increase a value of the hypothesis class among the output data in response to a re-classification condition being met. The modified input data is set to be new input data.

First claim

What is claimed is:

1 . A method of classifying data, comprising:
selecting a hypothesis class among entire classes; generating output data with regard to the entire classes by applying a classification algorithm to input data; modifying the input data to increase a value of the hypothesis class among the output data in response to a re-classification condition being met; and setting the modified input data to be new input data.

2 . The method of claim 1 , wherein the re-classification condition comprises at least one of a value of the hypothesis class being lower than a preset threshold and a number of a re-classification iteration being lower than a preset number of the re-classifications.

3 . The method of claim 1 , further comprising:
outputting the input data and the output data in response to the determination that the re-classification condition is not met.

4 . The method of claim 1 , wherein the modifying of the input data comprises:
defining a loss function of the classification algorithm using the hypothesis class, calculating a gradient vector of the defined loss function, and modifying the input data on a basis of the gradient vector.

5 . The method of claim 4 , wherein the modifying of the input data comprises reducing each value of the input data by a preset positive value in a direction a gradient.

6 . The method of claim 4 , wherein the modifying of the input data on a basis of the gradient vector comprises:
for each value of the input data, modifying to 0 a value of the input data in which a result from multiplying a sign and a gradient of each value is greater than or equal to a reference value, or a value of the input data, in which an absolute value of the gradient is greater than or equal to the reference value.

7 . The method of claim 4 , wherein the modifying of the input data on a basis of the gradient vector comprises:
reducing, in the direction in which a gradient descends, each value of the input data by a positive value.

8 . The method of claim 1 , further comprising:
generating initial output data of the entire classes by applying the classification algorithm to the received input data, and the selecting of the hypothesis class is performed on a basis of a size of each value of the initial output data.

9 . The method of claim 1 , wherein the classification algorithm is one of a neutral network, a convolutional neutral network (CNN), and a recurrent neural network (RNN).

10 . An apparatus to classify data, comprising:
a hypothesis class selector configured to select one hypothesis class among entire classes; a data classifier configured to generate output data with regard to the entire classes by applying a classification algorithm to input data; and a data setter configured to modify input data to increase a value of the hypothesis class among the output data and set the modified input data to new input data in response to a determination that a re-classification condition is met.

11 . The apparatus of claim 10 , wherein the re-classification condition comprises at least one of a value of the hypothesis class being lower than a preset threshold and a value of the hypothesis class being lower than a preset number of the re-classifications.

12 . The apparatus of claim 10 , in response to a determination that the re-classification condition is not met, further comprising:
a result output configured to output the input data and the output data.

13 . The apparatus of claim 10 , wherein the data setter comprises:
a loss function definer configured to define a loss function of the classification algorithm by using the hypothesis class, a gradient calculator configured to calculate a gradient vector with respect to the defined loss function, and a data modifier configured to modify the input data on a basis of the gradient vector.

14 . The apparatus of claim 13 , wherein the data modifier is configured to reduce each value of the input data by a preset positive in a direction of a gradient.

15 . The apparatus of claim 13 , wherein the data modifier is configured to, for each value of the input data, modify to 0 a value of the input data in which a result from multiplying a sign and a gradient of each value is greater than or equal to a reference value, or a value of the input data in which an absolute value of the gradient is greater than or equal to the reference value.

16 . The apparatus of claim 13 , wherein the data modifier modifies the input data on a basis of the gradient vector by reducing, in the direction in which a gradient descends, each value of the input data by a positive value.

17 . The apparatus of claim 10 , wherein the hypothesis class selector is configured to generate initial output data with respect to the entire classes by applying the classification algorithm to the received input data and select the hypothesis class on a basis of a size of each value of the initial output data.

18 . The apparatus of claim 10 , wherein the classification algorithm is one of a neutral network, a convolutional neutral network (CNN), and a recurrent neural network (RNN).

19 . A method of segmenting a region of interest (ROI), comprising:
selecting one hypothesis class among entire classes; generating output data with regard to the entire classes by applying a classification algorithm to input data; modifying the input data to increase a value of the hypothesis class among the output data; and segmenting, as ROIs, an area from the modified input data based on the modifying.

20 . The method of claim 19 , further comprising:
in response to a determination that the ROI is to be re-segmented, generating new input data that comprises the segmented area that is continuous; and repeatedly performing operations subsequent to the generating of the output data.

21 . The method of claim 19 , wherein the segmenting of the area comprises segmenting a continuous area, of which a value is increased as ROIs from the modified input data by using a segmentation algorithm.

22 . The method of claim 21 , wherein the segmentation algorithm comprises at least one of a graph cut algorithm and a conditional random field (CRF) algorithm.

23 . The method of claim 19 , wherein the modifying of the input data comprises:
defining a loss function of the classification algorithm using the hypothesis class; calculating a gradient vector of the defined loss function; and modifying the input data based on the gradient vector.

24 . An apparatus to segment a region of interest (ROI), comprising:
a hypothesis class selector configured to select a hypothesis class among entire classes; a data classifier configured to generate output data about the entire classes by applying a classification algorithm to input data; a data setter configured to modify the input data to increase a value of the hypothesis class among the output data and outputting a modification result indicative thereof; and an ROI segmentor configured to segment, as ROIs, an area from the modified input data based on the modification result.

25 . The apparatus of claim 24 , wherein, in response to a determination that the ROI is to be re-segmented, the data setter is configured to generate new input data that comprises the one or more segmented areas.

26 . The apparatus of claim 24 , wherein the ROI segmentor is configured to segment a continuous

Assignees

• Samsung Electronics Co Ltd

Inventors

• Kwon Keun Joo

Classifications

• G06N3/044

  Recurrent networks, e.g. Hopfield networks · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06Q10/00

  Administration; Management · CPC title

• G06N3/08 Primary

  Learning methods · CPC title

• G06F17/00 Primary

  Digital computing or data processing equipment or methods, specially adapted for specific functions (information retrieval, database structures or file system structures therefor G06F16/00) · CPC title