METHOD AND APPARATUS FOR FILTERING, CLUSTERING, AND REGION FITTING BY MEAN SHIFT OF IMAGES USING KERNEL FUNCTION VALUES

• US 20080013784A1
• Filed: 03/22/2007
• Published: 01/17/2008
• Est. Priority Date: 07/13/2006
• Status: Active Grant

• Current Assignee: Kabushiki Kaisha Toshiba (Toshiba Corporation)
• Original Assignee: Kabushiki Kaisha Toshiba (Toshiba Corporation)
• Inventors: Ida, Takashi, Takeshima, Hidenori
• Granted Patent: US 7,792,382 B2
• Time in Patent Office: Days
• US Class Current: 382/100
• CPC Class Codes: G06F 18/2321 using statistics or function optimisation, e.g. modelling of probability density functions
• CPC Class Codes: G06V 10/267 by performing operations on regions, e.g. growing, shrinking or watersheds
• CPC Class Codes: G06V 10/763 Non-hierarchical techniques, e.g. based on statistics of modelling distributions

Abstract

A path connecting a target sample of an arbitrary N-dimensional vector to a plurality of candidate samples in a sample set in an N-dimensional space having γ-dimensional positions and an η-dimensional feature quantity is set. On the path, the feature quantity is integrated to obtain an integration value. Using the integration value as a variable value of a preset kernel function, a kernel function value corresponding to the target sample is calculated.

Claims

1. An image filtering method comprising:

setting a path connecting a first pixel of an arbitrary N-dimensional vector (N is a natural number) to a plurality of second pixels in an N-dimensional space having γ

-dimensional positions and an η

-dimensional feature quantity (γ

is a natural number equal to 2 or more and n is a natural number equal to 1 or more);

obtaining an integration value by integrating the feature quantity on the path;

calculating a kernel function value corresponding to the first pixel using the integration value as a variable value of a preset kernel function;

calculating a mean shift vector by dividing the sum of products of the kernel function value and the feature quantities of the second pixels by the sum total of the kernel function values;

moving the first pixel according to the mean shift vector; and

replacing the feature quantity of the first pixel with the feature quantity of the moved first pixel.

2. The method according to claim 1, wherein the integration value is the sum total of the difference between the feature quantity of the first pixel and each of the feature quantities of the individual pixels existing at a plurality of positions set on the path.

3. The method according to claim 1, wherein the integration value is the sum total of the difference between the feature quantities of adjacent pixels existing at a plurality of positions set on the path.

4. The method according to claim 1, wherein the kernel function is a normal distribution function.

5. The method according to claim 1, wherein the kernel function is a function which, if the variable value exceeds a predetermined value, outputs 0 and, if not, outputs a predetermined constant.

6. An image clustering method comprising:

setting a path connecting a first pixel of an arbitrary N-dimensional vector (N is a natural number) to a plurality of second pixels in an N-dimensional space having γ

-dimensional positions and an η

-dimensional feature quantity (γ

is a natural number equal to 2 or more and η

is a natural number equal to 1 or more);

obtaining an integration value by integrating the feature quantity on the path;

calculating a kernel function value corresponding to the first pixel using the integration value as a variable value of a preset kernel function;

calculating a mean shift vector by dividing the sum of products of the kernel function value and the feature quantities of the second pixels by the sum total of the kernel function values;

moving the first pixel according to the mean shift vector;

replacing the feature quantity of the first pixel with the feature quantity of the moved first pixel; and

allocating the same cluster to a plurality of pixels whose similarity of feature quantity is equal to or higher than a predetermined threshold value among the moved first pixels.

7. The method according to claim 6, wherein the integration value is the sum total of the difference between the feature quantity of the first pixel and each of the feature quantities of the individual pixels existing at a plurality of positions set on the path.

8. The method according to claim 6, wherein the integration value is the sum total of the difference between the feature quantities of adjacent pixels existing at a plurality of positions set on the path.

9. The method according to claim 6, wherein the kernel function is a normal distribution function.

10. The image clustering method according to claim 6, wherein the kernel function is a function which, if the variable value exceeds a predetermined value, outputs 0 and, if not, outputs a predetermined constant.

11. An image region fitting method comprising:

obtaining an image composed of a plurality of pixels having two-dimensional positions and a one-dimensional feature quantity;

setting an initial first region and an initial second region schematically showing a subject region and a background region respectively in the image;

setting a path connecting a first pixel of an arbitrary N-dimensional vector (N is a natural number) to a plurality of second pixels in the image;

obtaining an integration value by integrating the feature quantity on the path;

calculating a kernel function value corresponding to the first pixel using the integration value as a variable value of a preset kernel function;

obtaining a first probability density function value by adding a first kernel function value at each pixel belonging to the initial first region among the kernel function values;

obtaining a second probability density function value by adding a second kernel function value at each pixel belonging to the initial second region among the kernel function values;

comparing the first probability density function value and the second probability density function value and obtaining region information indicating which of the subject region and background region a target pixel set to a pixel in at least a part of the image belongs to; and

outputting the region information.

12. The method according to claim 11, wherein the integration value is the sum total of the difference between the feature quantity of the first pixel and each of the feature quantities of the individual pixels existing at a plurality of positions set on the path.

13. The method according to claim 11, wherein the integration value is the sum total of the difference between the feature quantities of adjacent pixels existing at a plurality of positions set on the path.

14. The method according to claim 11, wherein the kernel function is a normal distribution function.

15. The method according to claim 11, wherein the kernel function is a function which, if the variable value exceeds a predetermined value, outputs 0 and, if not, outputs a predetermined constant.

16. An image mean shift filter comprising:

a setting section which sets a path connecting a first pixel of an arbitrary N-dimensional vector (N is a natural number) to a plurality of second pixels in an N-dimensional space having γ

-dimensional positions and an η

-dimensional feature quantity (γ

is a natural number equal to 2 or more and η

is a natural number equal to 1 or more);

an integrating section which obtains an integration value by integrating the feature quantity on the path;

a first computing section which calculates a kernel function value corresponding to the first pixel using the integration value as a variable value of a preset kernel function;

an extracting section which extracts the feature quantities of the second pixels;

a storage section which stores the feature quantity;

a second computing section which calculates a mean shift vector by dividing the sum of products of the kernel function value and each of the feature quantities of the second pixels by the sum total of the kernel function values; and

a replacing section which moves the first pixel according to the mean shift vector and replaces the feature quantity of the first pixel with the feature quantity of the moved first pixel.

17. A region fitting apparatus comprising:

an obtaining section which obtains an image composed of a plurality of pixels having two-dimensional positions and a one-dimensional feature quantity;

a first setting section which sets an initial first region and an initial second region schematically showing a subject region and a background region respectively in the image;

a second setting section which sets a path connecting a first pixel of an arbitrary N-dimensional vector (N is a natural number) to a plurality of second pixels in the image;

an integrating section which obtains an integration value by integrating the feature quantity on the path;

a first computing section which calculates a kernel function value corresponding to the first pixel using the integration value as a variable value of a preset kernel function;

a first adding section which obtains a first probability density function value by adding a first kernel function value at each pixel belonging to the initial first region schematically showing the subject region of the image among the kernel function values;

a second adding section which obtains a second probability density function value by adding a second kernel function value at each pixel belonging to the initial second region schematically showing the background region of the image among the kernel function values;

a comparing section which compares the first probability density function value and the second probability density function value and obtains region information indicating which of the subject region and background region a target pixel set to a pixel in at least a part of the image belongs to; and

an output section which outputs the region information.