[Insight-users] Circle detection in itk

[Richard Beare]

Hi everyone,

I think Gaetan is likely to be right, but it would be good to see an
example image. When you are dealing with small objects, such as 3-10
pixel radii, it is often not very useful to assume too much about the
shape - basically every natural object of that size will look like a
circle to a human observer until they zoom in.

The critical issues I'd want to know about before tackling the problem are:
1) What do you want to measure - do you want to count the particles,
measure their brightness, look at spatial distribution or something
else entirely. If you only want to count them then you could think
about stereological approaches that are very fast. If you want spatial
distribution then you could look at peaks. If you want brightness
information then you need to segment them and use a method that
corresponds well enough with what an expert observer would do.

2) How well controlled is the problem? Is there a nice clean
background or are there other objects that you want to ignore? Do the
objects cluster together in groups that will make separating them
necessary. These factors are extremely critical - people are really
good at ignoring distractors but it is often much harder to design
procedures that deal with such problems.

3) Don't worry too much about speed, especially when getting started.
In most cases the performance will depend on the image size and
possibly the size of filter kernels, rather than the number of items
in the image.

If the background is smooth, and there are no distractors or problems
with touching objects then Gaetan's suggestions are definitely a good
way to start.


On 1/7/07, Gaëtan Lehmann <gaetan.lehmann at jouy.inra.fr> wrote:
>
> Hi Ali,
> Hi Luis,
>
> I guess this mail reply to some of the question I wanted to ask after your
> last mail about particle detection - great :-)
>
> Le Sat, 06 Jan 2007 14:10:53 +0100, Ali - <saveez at hotmail.com> a écrit:
>
> > Hi Luis,
> > Thanks for the reply, I thought the itk community was dead in the new
> > year.
>
> some of us may have survived to their holidays ;-)
>
> > I need to detect the position and diameter of randomly distributed
> > particle-images where it is assumed each particle-image has a Gaussian
> > profile.  > 1) How fast do you need the circle detection to be ?
> > The circle detection algorithm must be fast, it is applied to an
> > ensemble of particle-images with Gaussian profile. Each image may
> > contain tens to hundreds of particle-images.
>
> Are there only circluar particles in your image or also other shapes ?
> In the case there are only circular particules, a simple thresholding and
> a labelisation of the connected components may be enough. If the particles
> are on a non uniform background, you can first apply a top-hat to remove
> it. If the intensity of the particules is not uniform from image to image,
> it may be difficult to find a good threshold value, but some automatic
> thresholding methods may give good results.
> If there are other shapes, what are they ? How big are they compared to
> the particules ?
> It would help if you can give a link to an example.
>
> Gaetan
>
>
>
> >> 2) How big are your images ?
> > About 1k X 1k. Each particle-image could have a diameter from 3 to 10
> > pixels.
> >> 3) Are you looking for full circles ?> or only for circumferences ?
> > The circle (particle-image) centre and diameter (Gaussian profile
> > diameter) are desirable to be detected.
> >> > 4) Do you expect to find multiple circles per image ?
> > As described, yes.
> >> > 5) Do you have any priori information about the circle ?> (eg.
> >> expected positions, expected radius, expected> intensities)
> > Positions are random, diameters are variable (3 to 10 pixels) the
> > intensities are variable too. The only prior knowledge could be the
> > assumed Gaussian profile of each particle-image. Correlation-based
> > methods may work, but I am sure there are faster methods around. The use
> > of the existing code in the library is preferred.
> >>> 6) Have you profiled the time that it takes to run the> Hough circle
> >>> detection in a release (optimized) build ?>
> > No, but the following is a quote from 'Practical Algorithms for Image
> > Processing', Seul et al, Cambridge University Press 2000, 4.10.2:
> > '..., and it is for this reason that the Hough transform is far less
> > popular for geometrical objects than the line'.
> > _________________________________________________________________
> > Be one of the first to try Windows Live Mail.
> > http://ideas.live.com/programpage.aspx?versionId=5d21c51a-b161-4314-9b0e-4911fb2b2e6d
>
>
>
> --
> Gaëtan Lehmann
> Biologie du Développement et de la Reproduction
> INRA de Jouy-en-Josas (France)
> tel: +33 1 34 65 29 66    fax: 01 34 65 29 09
> http://voxel.jouy.inra.fr
> _______________________________________________
> Insight-users mailing list
> Insight-users at itk.org
> http://www.itk.org/mailman/listinfo/insight-users
>