Hi all,

I have a bunch of pictures like this one:

(http://cid-1627b3c28e2dc281.spaces.live.com/photos/cns!
1627B3C28E2DC281!125 )

(A picture of rocks with soil between them)

I would like to detect and extract every single rock in the picture.
What would be the best way to do this? I have tried using Canny edge
detection to find each rock, but even though the contrast between rock
and soil is quite high, I don't get all the edges in the picture.

Also I would like to find the green box which is somewhere at the
center of the picture. How would you suggest that I do that?

Best regards,
Bjarne

Hi all,

I have a bunch of pictures like this one:

(http://cid-1627b3c28e2dc281.spaces.live.com/photos/cns!
1627B3C28E2DC281!125 )

(A picture of rocks with soil between them)

I would like to detect and extract every single rock in the picture.
What would be the best way to do this? I have tried using Canny edge
detection to find each rock, but even though the contrast between rock
and soil is quite high, I don't get all the edges in the picture.

Also I would like to find the green box which is somewhere at the
center of the picture. How would you suggest that I do that?

Best regards,
Bjarne
--------------------------------------------------------------------------------

On May 21, 4:35 am, stormogulen <bjarn... at (no spam) gmail.com> wrote:



Hi all,

I have a bunch of pictures like this one:

(http://cid-1627b3c28e2dc281.spaces.live.com/photos/cns!
1627B3C28E2DC281!125 )

(A picture of rocks with soil between them)

I would like to detect and extract every single rock in the picture.
What would be the best way to do this? I have tried using Canny edge
detection to find each rock, but even though the contrast between rock
and soil is quite high, I don't get all the edges in the picture.

Also I would like to find the green box which is somewhere at the
center of the picture. How would you suggest that I do that?

Best regards,
Bjarne

---------------------------------------------------------------------------­-----
Bjarne:
An edge detector depends on the window size so it will find large dirt
areas in addition to the dirt cracks between the rocks.  You might try
just simple thresholding and then maybe a morphological opening to
make the dirt connect in the narrow spots.  Or you could use the edge
detector but to have it not find the large dirt spots, you need to
check your areas against their average intensity and then thrown out
regions close to the dirt color.   Doesn't really look too tough to do
a reasonably good job. Find the green box with a color classifier, for
example just convert the RGB image into HSI and threshold on the hue
channel.
Regards,
ImageAnalyst- Hide quoted text -

- Show quoted text -

Hi all,

I have a bunch of pictures like this one:

(http://cid-1627b3c28e2dc281.spaces.live.com/photos/cns!
1627B3C28E2DC281!125 )

(A picture of rocks with soil between them)

I would like to detect and extract every single rock in the picture.
What would be the best way to do this? I have tried using Canny edge
detection to find each rock, but even though the contrast between rock
and soil is quite high, I don't get all the edges in the picture.

Also I would like to find the green box which is somewhere at the
center of the picture. How would you suggest that I do that?

Best regards,
Bjarne

Find the green box with a color classifier, for
example just convert the RGB image into HSI and threshold on the hue
channel.

On May 21, 9:48 pm, ImageAnalyst <imageanal... at (no spam) mailinator.com> wrote:

Find the green box with a color classifier, for
example just convert the RGB image into HSI and threshold on the hue
channel.

In my experience, that won't work because there will be numerous
unwanted green (or any) hue regions with S slightly greater than
zero. In other word, these unwanted regions look white or grey but
have green (or any) hue. Hue and saturation must be both examined to
properly detect an object by color.

Hi all,

I have a bunch of pictures like this one:

(http://cid-1627b3c28e2dc281.spaces.live.com/photos/cns!
1627B3C28E2DC281!125 )

(A picture of rocks with soil between them)

I would like to detect and extract every single rock in the picture.
What would be the best way to do this? I have tried using Canny edge
detection to find each rock, but even though the contrast between rock
and soil is quite high, I don't get all the edges in the picture.

Also I would like to find the green box which is somewhere at the
center of the picture. How would you suggest that I do that?

Best regards,
Bjarne

On May 21, 2:35 am, stormogulen <bjarn... at (no spam) gmail.com> wrote:





Hi all,

I have a bunch of pictures like this one:

(http://cid-1627b3c28e2dc281.spaces.live.com/photos/cns!
1627B3C28E2DC281!125 )

(A picture of rocks with soil between them)

I would like to detect and extract every single rock in the picture.
What would be the best way to do this? I have tried using Canny edge
detection to find each rock, but even though the contrast between rock
and soil is quite high, I don't get all the edges in the picture.

Also I would like to find the green box which is somewhere at the
center of the picture. How would you suggest that I do that?

Best regards,
Bjarne

First, I removed texture with bilateral filter,http://www.general-cathexis..com/images/bilateral.jpg

Then I generated mask by color detection and applied to original,
coloring outside region red,http://www.general-cathexis.com/images/segmented.jpg

Why is your file name so long?  It fosters Murphy's law.- Skjul tekst i anførselstegn -

- Vis tekst i anførselstegn -

On May 22, 6:14 pm, aruzinsky <aruzin... at (no spam) general-cathexis.com> wrote:

On May 21, 9:48 pm, ImageAnalyst <imageanal... at (no spam) mailinator.com> wrote:

Find the green box with a color classifier, for
example just convert the RGB image into HSI and threshold on the hue
channel.

In my experience, that won't work because there will be numerous
unwanted green (or any) hue regions with S slightly greater than
zero.  In other word, these unwanted regions look white or grey but
have green (or any) hue.  Hue and saturation must be both examined to
properly detect an object by color.

I found helpful in my project (on low-en cellphone camera) use YUV
based "saturation" and "hue". That is saturation=sqrt(U*U+V*V),
hue=atan(U,V). For some reason it works a lot better for low end
camera than HSV, probably because in HSV hue almost pure R , G and B
take most of the  range, and YUV  have more place for "transitional"
colors.

On May 22, 10:11 am, serg271 <serg... at (no spam) gmail.com> wrote:





On May 22, 6:14 pm, aruzinsky <aruzin... at (no spam) general-cathexis.com> wrote:

On May 21, 9:48 pm, ImageAnalyst <imageanal... at (no spam) mailinator.com> wrote:

Find the green box with a color classifier, for
example just convert the RGB image into HSI and threshold on the hue
channel.

In my experience, that won't work because there will be numerous
unwanted green (or any) hue regions with S slightly greater than
zero.  In other word, these unwanted regions look white or grey but
have green (or any) hue.  Hue and saturation must be both examined to
properly detect an object by color.

I found helpful in my project (on low-en cellphone camera) use YUV
based "saturation" and "hue". That is saturation=sqrt(U*U+V*V),
hue=atan(U,V). For some reason it works a lot better for low end
camera than HSV, probably because in HSV hue almost pure R , G and B
take most of the  range, and YUV  have more place for "transitional"
colors.

Why don't you combine and compare hue and saturation thusly?:

[ (U - U0)*(U - U0)  + (V - V0)*(V - V0) ]  <  threshhold

where Y0 U0 V0 is the reference color.- Hide quoted text -

- Show quoted text -

On May 22, 10:11 am, serg271 <serg... at (no spam) gmail.com> wrote:



On May 22, 6:14 pm, aruzinsky <aruzin... at (no spam) general-cathexis.com> wrote:

On May 21, 9:48 pm, ImageAnalyst <imageanal... at (no spam) mailinator.com> wrote:

Find the green box with a color classifier, for
example just convert the RGB image into HSI and threshold on the hue
channel.

In my experience, that won't work because there will be numerous
unwanted green (or any) hue regions with S slightly greater than
zero. In other word, these unwanted regions look white or grey but
have green (or any) hue. Hue and saturation must be both examined to
properly detect an object by color.

I found helpful in my project (on low-en cellphone camera) use YUV
based "saturation" and "hue". That is saturation=sqrt(U*U+V*V),
hue=atan(U,V). For some reason it works a lot better for low end
camera than HSV, probably because in HSV hue almost pure R , G and B
take most of the range, and YUV have more place for "transitional"
colors.

Why don't you combine and compare hue and saturation thusly?:

[ (U - U0)*(U - U0) + (V - V0)*(V - V0) ] < threshhold

where Y0 U0 V0 is the reference color.

On May 23, 7:47 pm, aruzinsky <aruzin... at (no spam) general-cathexis.com> wrote:





On May 22, 10:11 am, serg271 <serg... at (no spam) gmail.com> wrote:

On May 22, 6:14 pm, aruzinsky <aruzin... at (no spam) general-cathexis.com> wrote:

On May 21, 9:48 pm, ImageAnalyst <imageanal... at (no spam) mailinator.com> wrote:

Find the green box with a color classifier, for
example just convert the RGB image into HSI and threshold on the hue
channel.

In my experience, that won't work because there will be numerous
unwanted green (or any) hue regions with S slightly greater than
zero.  In other word, these unwanted regions look white or grey but
have green (or any) hue.  Hue and saturation must be both examined to
properly detect an object by color.

I found helpful in my project (on low-en cellphone camera) use YUV
based "saturation" and "hue". That is saturation=sqrt(U*U+V*V),
hue=atan(U,V). For some reason it works a lot better for low end
camera than HSV, probably because in HSV hue almost pure R , G and B
take most of the  range, and YUV  have more place for "transitional"
colors.

Why don't you combine and compare hue and saturation thusly?:

[ (U - U0)*(U - U0)  + (V - V0)*(V - V0) ]  <  threshhold

where Y0 U0 V0 is the reference color.

I don't have any reference color. I'm doing color classification - how
similar colors of different objects, and use both hue and saturation
for kind of hashing.- Hide quoted text -

- Show quoted text -

It just occurred to me that YUV is asymmetric:

Y = 0.2990*R + 0.5870*G + 0.1140*B;
U = -0.14713*R - 0.28886*G + 0.436*B;
V = 0.615*R - 0.51499*G - 0.10001*B;

As I recall, these coefficients (or was it YCbCr?) are determined by
the characteristics of CRT phosphors which changed from the 1950s.
Anyway, I bet the rationale for these coefficients is far fetched.

It would be better to use a symmetric version:

Y' = (R + G + B)/3.0;
U' = (2.0*B - R - G)/3.0;
V' = (2.0*R - G - B)/3.0;


R = Y' + V';
G = Y' - U' - V';
B = Y' + U';

In my experience, this works just fine.

I suspect internal camera format is YUV, before it converted to RGB
later by driver - it seems it's either yuv or jpeg. So working  with
asymmetrical YUV could be preferable.- Hide quoted text -