Mean Shift procedure for finding local maxima in a Gray Scale image

Hi,

I am considering usage of the Mean-Shift algorithm to classify local
maxima in gray scale image ( result of Hough Transform for lines ).
For mean calculation I use the center of mass formula ( having Luma as
the mass ).

The mean shift procedure:
1. Define a radius, e.g. 6
2. For each 12x12 box in the image
  a. Calculate the center of mass
  b. Shift the 12x12 box center to the resulting center of mass
  c. Goto step 2
3. Repeat step 2 until the shift is small enough, or, until a certain
amount of iterations has elapsed.

The problem with the above flow is that in most of the cases the
algorithm ~locks up~ traversing a group of pixels until �max
iterations� is reached, this, result in invalid convergence of the
algorithm.
I have also tried to increase the radius ( to a certain extent ) when
the shift is too small and to reduce it when the shift is too big,
this, improved the result, but, still, most of the iterations ended
locked up traversing a fixed group of pixels ( repeating the same
pixels ).

My aim is to have each pixel in the image converging to one maxima in
the image.

Have I got the mean shift procedure correctly?
Is there any known improvement to avoid the closed loop pixel
traversing?

Any help will be appreciated.

Nadav Rubinstein
http://www.sophin.com

Nadav wrote:

> Hi,
> 
> I am considering usage of the Mean-Shift algorithm to classify local

[..description of a problem skipped..]

> Any help will be appreciated.

Does your appreciation imply some real money, or it is purely imaginary?


Vladimir Vassilevsky
DSP and Mixed Signal Design Consultant
http://www.abvolt.com

take your meds