DETECTING AN INTEREST POINT IN AN IMAGE USING EDGES

• US 20120281872A1
• Filed: 05/03/2011
• Published: 11/08/2012
• Est. Priority Date: 05/03/2011
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Zitnick, Charles Lawrence III
• Granted Patent: US 8,538,077 B2
• Time in Patent Office: Days
• US Class Current: 382/103
• CPC Class Codes: G06V 10/451 with interaction between the filter responses, e.g. cortical complex cells

Abstract

Technology is described for detecting an interest point in an image using edges. An example method can include the operation of computing locally normalized edge magnitudes and edge orientations for the image using a processor to form a normalized gradient image. The normalized gradient image can be divided into a plurality of image orientation maps having edge orientations. Orientation dependent filtering can be applied to the image orientation maps to form response images. A further operation can be summing the response images to obtain an aggregated filter response image. Maxima can be identified in spatial position and scale in the aggregated filter response image. Maxima in the aggregated filter response image can be defined as interest points.

Claims

1. A method for detecting an interest point in an image using edges, comprising:

computing normalized edge magnitudes and edge orientations for the image using a processor to form a normalized gradient image;

dividing the normalized gradient image into a plurality of image orientation maps having edge orientations;

applying orientation dependent filtering to the image orientation maps to form response images;

summing the response images to obtain an aggregated filter response image;

identifying maxima in spatial position and scale in the aggregated filter response image; and

defining maxima in the aggregated filter response as interest points.

2. The method as in claim 1, wherein forming a normalized gradient image further comprises normalizing edge magnitude by computing the average gradient in a local spatial neighborhoods for pixels in the image.

3. The method as in claim 1, wherein defining maxima further comprises identifying pixels with an intensity above a defined threshold.

4. The method as in claim 1, wherein maxima are identified in proximity to edge foci to detect interest points based on the aggregated filter responses near the edge foci.

5. The method as in claim 1, further comprising applying the orientation dependent filtering using linearly separable filters.

6. The methods as in claim 1, wherein dividing the normalized gradient image into a plurality of image orientation maps further comprises dividing edge magnitude of pixels into separate image orientation maps based on a gradient orientation of pixels and on a linear interpolation between at least two image orientation maps with a relevant orientation for the pixels.

7. The method as in claim 1, further comprising applying orientation dependent filtering to image orientation maps at multiple scales and aggregating filter responses at multiple scales to provide interest points at multiple scales.

8. A method for detecting an interest point in an image using edges, comprising:

computing normalized edge magnitudes and orientations for the image to form a normalized gradient image;

applying image filters to the normalized gradient image to form filter response images;

identifying maxima having spatial position and scale in aggregated filter response images; and

defining maxima in the image as an interest point.

9. The method as in claim 8, wherein applying image filters further comprises applying orientation dependent filters, wherein an orientation dependent filter corresponds to orientations of image orientation maps derived from the normalized gradient image.

10. The method as in claim 9, wherein applying orientation dependent filters further comprises applying a series of linear filters on the image orientation maps.

11. The method as in claim 9, wherein applying orientation dependent filters further comprises applying a Laplacian filter parallel to a normalized edge to determine the normalized edge'"'"'s scale.

12. The method as in claim 11, further comprising applying a series of Gaussian filters to edges of predicted positions of edge foci on a side perpendicular to the edges.

13. The method as in claim 12, further comprising summing together responses for the Gaussian filter for edges to obtain a filter response.

14. The method as in claim 11, further comprising applying the Gaussian filter perpendicular to the edges to reinforce the filter response for edge foci.

15. The method as in claim 11, wherein the Gaussian filter is a summation of two Gaussian filters centered at −

σ and

σ

.

16. The method as in claim 8, further comprising applying orientation dependent filters to a plurality of image orientation maps using linearly separable filters.

17. The method as in claim 8, wherein a pixel in the image is a local maxima if the pixel'"'"'s intensity response is higher than a response threshold as compared to pixel neighbors in a 3×

3×

3 neighborhood as identified using a 3D (three dimensional) quadratic function.

18. A system for detecting an interest point in an image using edges, comprising:

a memory device for storing the image;

an normalization module to compute locally normalized edge magnitudes and orientations for the image using a processor to form a normalized gradient image;

a plurality of image orientation bins representing edge orientations into which the normalized gradient image is divided to create image orientation maps;

an orientation dependent filtering module to apply orientation dependent filters to the image orientation maps to form response images using a processor;

a summation module to sum the response images in order to obtain an aggregated filter response image; and

an interest point identification module to identify local maxima in the aggregated filter response image in spatial position.

19. The system as in claim 18, further comprising applying the orientation dependent filters to orientation image maps using parallel processing units.

20. The system as in claim 18, wherein the interest point identification module is configured to apply orientation dependent filtering to image orientation maps at multiple scales and aggregating the responses at multiple scales to identify interest points at multiple scales.