Feature Detection in Numeric Data

• US 20130308860A1
• Filed: 05/16/2013
• Published: 11/21/2013
• Est. Priority Date: 05/16/2012
• Status: Active Grant

• Current Assignee: IMEC International, Katholieke Universiteit Leuven
• Original Assignee: IMEC International, Katholieke Universiteit Leuven
• Inventors: Mainali, Pradip, LaFruit, Gauthier
• Granted Patent: US 9,042,645 B2
• Time in Patent Office: Days
• US Class Current: 382/170
• CPC Class Codes: G06V 10/449 Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters
• CPC Class Codes: G06V 10/462 Salient features, e.g. scale invariant feature transforms [SIFT]
• CPC Class Codes: G06V 10/757 Matching configurations of points or features

Abstract

A method for detecting features in digital numeric data comprises obtaining digital numeric data comprising values corresponding to a plurality of sampling points over a domain space having at least one dimension, computing a plurality of scale-space data comprising filtering said digital numeric data using a filter bank, determining a plurality of feature regions each corresponding to a local extremum in scale and location of the scale-space data; and determining a feature region descriptor for each of said plurality of feature regions. The filter bank is a Cosine Modulated Gaussian filter bank in which the standard deviation parameter of the Gaussian equals

$\frac{1}{\pi }\sqrt{\frac{\ln \left(2\right)}{2}}\frac{2^b+1}{2^b-1}$

multiplied by the cosine wavelength, in which b is in the range of 0.75 to 1.25, or said filter bank is an N^th-order Gaussian Derivative filter bank with N being in the range of 5 to 20.

Claims

1. A method for detecting features in digital numeric data, said method comprising:

obtaining digital numeric data comprising values corresponding to a plurality of sampling points over a domain space having at least one dimension;

computing a plurality of scale-space data, each of said plurality of scale-space data comprising data over said domain space at a corresponding scale in said domain space, wherein the computing of the plurality of scale-space data comprises filtering said digital numeric data using a filter bank;

determining a plurality of feature regions in said plurality of scale-space data, each feature region corresponding to a local extremum in scale and location of the scale-space data; and

determining a feature region descriptor for each of said plurality of feature regions;

wherein said filter bank is a Cosine Modulated Gaussian filter bank in which a standard deviation parameter of the Gaussian equals

2. The method according to claim 1, wherein said filter bank comprises a plurality of infinite impulse response filters, each infinite impulse response filter corresponding to a different scale parameter value for a common mathematical filter kernel having a controllable scale parameter, the common mathematical filter kernel being either:

3. The method according to claim 2, in which said filter bank is constructed by approximating said common mathematical filter kernel by a sum of exponentials.

4. The method according to claim 1, wherein computing the plurality of scale-space data further comprises filtering said digital numeric data along a plurality of directions in said domain space, and for each scale aggregating the filtered digital numeric data over the plurality of directions in said domain space.

5. The method according to claim 4, wherein said aggregating the filtered digital numeric data comprises summing corresponding data over the plurality of directions.

6. The method according to claim 1, wherein determining the feature region descriptor comprises computing a histogram of gradient orientations or sum of gradients.

7. The method according to claim 1, wherein said obtaining digital numeric data comprises providing an image.

8. The method according to claim 7, wherein said computing the plurality of scale-space data comprises:

creating a plurality of filtered images by filtering said image in one or a plurality of directions using said Cosine Modulated Gaussian filterbank or said N^th-order Gaussian Derivative filter bank to compute a Cosine Modulated Gaussian response or N^th-order Gaussian Derivative response for each direction; and

computing a Cosine Modulated Gaussian Summation or N^th-order Gaussian Derivative Summation response for each filtered image representing a scale.

9. The method according to claim 8, wherein determining said feature regions comprises extracting of each Cosine Modulated Gaussian Summation or N^th-order Gaussian Derivative Summation response a maximal/minimal point corresponding to a scale and location of a feature region.

10. The method according to claim 7, further comprising:

performing edge suppression and shape detection on each of said feature regions;

performing interpolation to fine-tune determined location and scale; and

assigning an orientation to each of said plurality of feature regions.

11. The method according to claim 1, wherein the digital numeric data comprises a plurality of images, and the method further comprising:

detecting scale-invariant image features of each of said plurality of images based on the determined feature region descriptors;

selecting an anchor image from said plurality of images;

detecting common image features between said plurality of images and said anchor image; and

aligning said plurality of images based on said detected common image features.

12. The method according to claim 11, wherein aligning said plurality of images comprises performing a scaling, rotation and translation on said plurality of images.

13. The method according to claim 11, wherein selecting the anchor image from said plurality of images comprises selecting an image from the plurality of images comprising a largest number of image features.

14. The method according to claim 11, wherein selecting the anchor image from said plurality of images comprises selecting an image from the plurality of image featuring the least optical errors.

15. A device for detecting features in digital numeric data, the device comprising:

an input means for obtaining digital numeric data, the digital numeric data comprising values corresponding to a plurality of sampling points over a domain space having at least one dimension;

a processing means adapted for;

computing a plurality of scale-space data, each of said plurality of scale-space data comprising data over said domain space at a corresponding scale in said domain space, in which the computing of the plurality of scale-space data comprises filtering said digital numeric data using a filter bank,

determining a plurality of feature regions in said plurality of scale-space data, each feature region corresponding to a local extremum in scale and location of the scale-space data, and

determining a feature region descriptor for each of said plurality of feature regions; and

and an output means for outputting said feature region descriptor for each of said plurality of feature regions;

wherein said filter bank is a Cosine Modulated Gaussian filter bank in which the standard deviation parameter of the Gaussian equals

16. A computer program product for, when executing on a programmable computer, performing a method, the method comprising:

obtaining digital numeric data comprising values corresponding to a plurality of sampling points over a domain space having at least one dimension;

computing a plurality of scale-space data, each of said plurality of scale-space data comprising data over said domain space at a corresponding scale in said domain space, wherein the computing of the plurality of scale-space data comprises filtering said digital numeric data using a filter bank;

determining a plurality of feature regions in said plurality of scale-space data, each feature region corresponding to a local extremum in scale and location of the scale-space data; and

determining a feature region descriptor for each of said plurality of feature regions;

wherein said filter bank is a Cosine Modulated Gaussian filter bank in which a standard deviation parameter of the Gaussian equals