Dave wrote:


> If you look at computer graphics texts

Heh. Those texts are full of nonsense and pompous meaningless phrases 
like "we apply a Hugh transform to the Zobel kernel processed...". In 
fact they meant something very trivial. Image processing is no science 
but a fairly primitive bit-byte manipulation. It has to be very fast; 
therefore it has to be simple.

> and the representations of
> colour, you'll find there are 3 common systems (probably more). One is
> RGB, another is HSL (Hue, Saturation, Level). Hue represents the
> colour on the colour wheel, while the other values affect the shade
> and  brightness.

Those representations have to do with the human perception and the human 
convenience. Being tied to the human features when doing an image 
analysis by a computer is a common mistake.

> I would suggest to looking colour conversions, ... or as you are more
> likely to find on the Internet "color conversions". A quick search for
> RGB and HSV or HSL gave the following:
> http://en.wikipedia.org/wiki/HSL_and_HSV

It doesn't matter what is the system of representation of the image. If 
the image is made of 3 basic colors then the coloration is a 2d 
parameter. If the OP stupident wants to do the image segmentation based 
on coloration but not the luminance, then he should normalize every 
pixel to the same brightness.



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

On Jun 15, 6:06&#4294967295;am, Nadav <nadav...@gmail.com> wrote:
> Hi,
>
> I am trying to segment an image according to color, i am trying to
> achieve that using mean-shift segmentation.
> Segmentation should be according to color and position ( not
> intensity )
> The Algorithm in a nutshell
> 1.The image is split to regions
> 2.For each region the kernel of each of it&#4294967295;s associated pixels is
> calculated
> 3.The region centroid is moved to the position maximizing the kernel
> value
> 4.Neighboring regions are merged
> 5.Convergence == centroid/s are not changing ( thresholded )
>
> In order to properly calculate the kernel color is to be presented
> using a single dimension, that is, the 3D RGB color space need to be
> converted to a 1D color space such as the color spectrum color-wheel
> (http://en.wikipedia.org/wiki/File:Computer_color_spectrum.svg)
>
> A.Is it really required to convert 3D RGB to 1D in order to
> efficiently segment the image using mean-shift?
> B.How do I convert 3D RGB color to a 1D ~color-wheel~ variable ?
>
> Any help will B appreciated
>
> Nadav Rubinsteinhttp://www.sophin.com

If you look at computer graphics texts and the representations of
colour, you'll find there are 3 common systems (probably more). One is
RGB, another is HSL (Hue, Saturation, Level). Hue represents the
colour on the colour wheel, while the other values affect the shade
and  brightness.

I would suggest to looking colour conversions, ... or as you are more
likely to find on the Internet "color conversions". A quick search for
RGB and HSV or HSL gave the following:
http://en.wikipedia.org/wiki/HSL_and_HSV

Cheers,
Dave

Nadav wrote:

> Hi,
> 
> I am trying to segment an image according to color, i am trying to
> achieve that using mean-shift segmentation.

[...nonsense skipped...]

Think of each pixel as a vector in 3d space (R,G,B). Then the difference 
  in coloration between two pixels X, Y is the angle between the 
corresponding vectors:

        X * Y
acos ---------
       |X| |Y|

Take it from there.

> 
> Any help will B appreciated

It depends. How much is the appreciation  ?

> Nadav Rubinstein

Wow. Nadav and Rubinstein at the same time?



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

Hi,

I am trying to segment an image according to color, i am trying to
achieve that using mean-shift segmentation.
Segmentation should be according to color and position ( not
intensity )
The Algorithm in a nutshell
1.The image is split to regions
2.For each region the kernel of each of it�s associated pixels is
calculated
3.The region centroid is moved to the position maximizing the kernel
value
4.Neighboring regions are merged
5.Convergence == centroid/s are not changing ( thresholded )

In order to properly calculate the kernel color is to be presented
using a single dimension, that is, the 3D RGB color space need to be
converted to a 1D color space such as the color spectrum color-wheel
( http://en.wikipedia.org/wiki/File:Computer_color_spectrum.svg )

A.Is it really required to convert 3D RGB to 1D in order to
efficiently segment the image using mean-shift?
B.How do I convert 3D RGB color to a 1D ~color-wheel~ variable ?

Any help will B appreciated

Nadav Rubinstein
http://www.sophin.com