Learning combinations of homogenous feature arrangements

• US 10,417,529 B2
• Filed: 09/15/2016
• Issued: 09/17/2019
• Est. Priority Date: 09/15/2015
• Status: Active Grant

• Current Assignee: Samsung Electronics Co. Ltd.
• Original Assignee: Samsung Electronics Co. Ltd.
• Inventors: Martineau, Justin, Currey, Jonathan J.
• Primary Examiner(s): Bitar, Nancy
• Application Number:
  US15/266,925
• Publication Number:
  US 20170076179A1
• Time in Patent Office: 1,097 Days
• CPC Class Codes: G06F 18/24 Classification techniques
• CPC Class Codes: G06F 18/2411 based on the proximity to a decision surface, e.g. support vector machines
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06N 3/08 Learning methods
• CPC Class Codes: G06V 10/82 using neural networks
• CPC Class Codes: G06V 30/19173 Classification techniques

Abstract

One embodiment provides a method comprising receiving an input, and classifying the input utilizing a learned linear combination of multi-dimensional filters. Each multi-dimensional filter identifies a multi-dimensional pattern of a homogenous feature. The method further comprises generating an output indicative of a classification of the input.

Claims

1. A method comprising:

receiving an input;

filtering, using at least one convolutional neural network (CNN), the input utilizing a learned linear combination of multi-dimensional filters, wherein each multi-dimensional filter identifies a multi-dimensional pattern of a homogenous feature; and

generating an output indicative of object recognition of the input.

2. The method of claim 1, wherein the input occupies a multi-dimensional space having a higher number of dimensions than each multi-dimensional filter.

3. The method of claim 2, wherein different semantic definitions occur along a dimension of the multi-dimensional space.

4. The method of claim 3, wherein each multi-dimensional filter has a size along a dimension of the multi-dimensional space that is less than a size of the input along the same dimension.

5. The method of claim 3, wherein each multi-dimensional filter is individually slidable across the input along each dimension of the multi-dimensional space.

6. The method of claim 4, wherein each multi-dimensional filter is individually slidable across the input along at least two dimensions of the multi-dimensional space, and the multi-dimensional filter operates at a particular value of one other dimension of the multi-dimensional space.

7. The method of claim 3, wherein each multi-dimensional filter examines the input for only one semantic definition.

8. The method of claim 3, wherein a semantic definition comprises one of a color related parameter, an audio related parameter, or a black and white related parameter.

9. The method of claim 1, wherein the input comprises at least one of an image, a video, or audio.

10. A system comprising:

at least one processor; and

a non-transitory processor-readable memory device storing instructions that when executed by the at least one processor causes the at least one processor to perform operations including;

receiving an input;

filtering, using at least one convolutional neural network (CNN), the input utilizing a learned linear combination of multi-dimensional filters, wherein each multi-dimensional filter identifies a multi-dimensional pattern of a homogenous feature; and

generating an output indicative of object recognition of the input.

11. The system of claim 10, wherein the input occupies a multi-dimensional space having a higher number of dimensions than each multi-dimensional filter, and different semantic definitions occur along a dimension of the multi-dimensional space.

12. The system of claim 10, wherein each multi-dimensional filter has a size along a dimension of the multi-dimensional space that is less than a size of the input along the same dimension.

13. The system of claim 12, wherein each multi-dimensional filter is individually slidable across the input along each dimension of the multi-dimensional space.

14. The system of claim 12, wherein each multi-dimensional filter is individually slidable across the input along at least two dimensions of the multi-dimensional space, and the multi-dimensional filter operates at a particular value of one other dimension of the multi-dimensional space.

15. The system of claim 11, wherein each multi-dimensional filter examines the input for only one semantic definition.

16. A method comprising:

receiving a plurality of inputs;

training, using at least one convolutional neural network (CNN), a first set of multi-dimensional filters based on the plurality of inputs, wherein each multi-dimensional filter of the first set identifies a first pattern type comprising a multi-dimensional pattern of a homogenous feature in the plurality of inputs;

combining the first set of multi-dimensional filters using a second set of multi-dimensional filters, wherein each multi-dimensional filter of the second set has a higher number of dimensions than each multi-dimensional filter of the first set;

learning a second pattern type comprising a learned linear combination of first pattern types identified by the combined first set of multi-dimensional filters, wherein the second pattern type has a higher number of dimensions than each of the first pattern types identified by the combined first set of multi-dimensional filters; and

generating one or more outputs indicative of object recognition for the plurality of inputs.

17. The method of claim 16, wherein each input of the plurality of inputs occupies a multi-dimensional space having a higher number of dimensions than each multi-dimensional filter of the first set, and different semantic definitions occur along a dimension of the multi-dimensional space.

18. The method of claim 16, wherein each multi-dimensional filter of the first set has a size along a dimension of the multi-dimensional space that is less than a size of each input along the same dimension, and the multi-dimensional filter examines the input for only one semantic definition.

19. The method of claim 17, wherein each multi-dimensional filter of the second set has a first size along at least two dimensions of the multi-dimensional space, the first size equal to one, each multi-dimensional filter has a second size along one other dimension of the multi-dimensional space, and the second size is equal to a size of each input of the plurality of inputs along the one other dimension.

20. The method of claim 16, wherein the plurality of inputs comprise a plurality of images.