Compact and robust signature for large scale visual search, retrieval and classification

Patent metadata

Field	Value
Publication number	US-9864928-B2
Application number	US-201414903590-A
Country	US
Kind code	B2
Filing date	Jul 7, 2014
Priority date	Jul 8, 2013
Publication date	Jan 9, 2018
Grant date	Jan 9, 2018

Abstract

A method for deriving a representation of an image, by processing signals corresponding to the image, comprises identifying a plurality of feature points in the image, deriving feature descriptors of feature points, and assigning feature descriptors to pre-defined center points, wherein each feature descriptor is assigned to a plurality of center points, the method further comprising, for each center point, calculating the difference between each feature descriptor assigned to said center point, deriving a value descriptor for each center point from said calculated differences, and deriving the representation from said value descriptors for said center points.

First claim

The invention claimed is:

1. A method of deriving a representation of an image, by processing signals corresponding to the image, comprising:
identifying a plurality of feature points in the image, deriving feature descriptors of the feature points, and assigning the feature descriptors to pre-defined centre points,
wherein each feature descriptor is assigned to a plurality of centre points,
the method further comprising, for each centre point, calculating a vector difference between each feature descriptor assigned to said centre point and said centre point, deriving a value descriptor for each centre point from said calculated vector differences, and deriving the representation of the image from said value descriptors for said centre points.

2. The method of claim 1 , wherein each feature descriptor is assigned to a plurality of centre points with a same weight.

3. The method of claim 1 , wherein deriving a value descriptor for each centre point from said calculated vector differences comprises transforming each calculated vector difference by a robust function.

4. The method of claim 3 wherein said robust function is such that the influence of each feature descriptor assigned to a centre point on the corresponding value descriptor is substantially equalized.

5. The method of claim 3 , wherein said robust function is a normalisation function.

6. The method of claim 1 wherein deriving a value descriptor for each centre point from said calculated differences comprises summing said calculated vector differences or said transformed calculated vector differences.

7. The method of claim 1 wherein each feature descriptor is assigned to a predetermined number of centre points.

8. The method of claim 1 wherein each feature descriptor is assigned to the nearest centre points.

9. The method of claim 1 further comprising, for each feature descriptor assigned to a centre point, assigning an associated rank, indicating the order of proximity of the feature descriptor from said respective centre point.

10. The method of claim 9 , wherein deriving a value descriptor for each centre point from said calculated differences comprises summing said distances or said transformed distances for each rank independently.

11. The method of claim 9 , wherein deriving a value descriptor for each centre point from said calculated vector differences comprises transforming each calculated vector difference by a robust function and summing transformed calculated vector differences for each rank independently.

12. The method of claim 10 , further comprising summing over ranks.

13. The method of claim 12 , wherein weights are assigned to ranks, and used in the sum over ranks.

14. The method of claim 13 , wherein higher ranks have higher weights.

15. The method of claim 12 , comprising assigning a corresponding weight to each of the ranks and using said assigned weights to determine reliability.

16. The method of claim 1 comprising, for the value descriptor for each centre point, determining a factor indicating reliability of the value descriptor.

17. The method of claim 1 , comprising selecting centre points, and/or associated value descriptors, based on the number of feature points assigned to said centre points and/or reliability factor.

18. The method of claim 1 comprising concatenating selected value descriptors to form a descriptor.

19. The method of claim 18 further comprising indicating the respective centre points, using binary flags.

20. The method of claim 1 further comprising performing coarse quantization on the elements of the value descriptors, preferably binary or ternary quantization.

21. The method of claim 1 comprising selecting a subset of elements for each value descriptor, optionally different subsets for different centre points.

22. A method of matching images, comprising deriving a value descriptor of each image using the method of claim 1 and comparing each of the value descriptors for corresponding centre points.

23. The method of claim 22 wherein deriving a value descriptor of each image comprises identifying a plurality of feature points in the image, deriving feature descriptors of feature points, and assigning feature descriptors to pre-defined centre points, and wherein each feature descriptor is assigned to a plurality of centre points.

24. Apparatus comprising an image processor for determining feature points in an image, and a processor for deriving an image representation by executing the steps of—
identifying a plurality of feature points in the image, deriving feature descriptors of the feature points, and assigning the feature descriptors to pre-defined centre points,
wherein each feature descriptor is assigned to a plurality of centre points;
and for each centre point, calculating a vector difference between each feature descriptor assigned to said centre point and said centre point, deriving a value descriptor for each centre point from said calculated vector differences, and deriving the image representation from said value descriptors for said centre points.

25. A non-transitory computer-readable storage medium storing computer-executable instructions, arranged for causing a processor to perform the steps of—
identifying a plurality of feature points in the image, deriving feature descriptors of the feature points, and assigning the feature descriptors to pre-defined centre points,
wherein each feature descriptor is assigned to a plurality of centre points,
the method further comprising, for each centre point, calculating a vector difference between each feature descriptor assigned to said centre point and said centre point, deriving a value descriptor for each centre point from said calculated vector differences, and deriving the representation from said value descriptors for said centre points.

Assignees

• Univ Surrey
• Visual Atoms Ltd

Inventors

• Bober Miroslaw
• Husain Syed

Classifications

• G06V10/764

  using classification, e.g. of video objects · CPC title

• G06V10/462 Primary

  Salient features, e.g. scale invariant feature transforms [SIFT] · CPC title

• G06F18/23

  Clustering techniques · CPC title

• G06F18/24137

  Distances to cluster centroïds · CPC title

• G06K9/6272

  Physics · mapped topic