| Re: fitting mixed gaussians [message #44114] |
Fri, 20 May 2005 07:10  |
btt
Messages: 345
Registered: December 2000
|
Senior Member |
|
|
Craig Markwardt wrote:
> Ben Tupper <btupper@bigelow.org> writes:
>
>
>> Rob wrote:
>>
>>> Thanks for the plug, David. Yep, PAN is just a big GUI wrapper for
>>> MPCURVEFIT. In fact the kind of fitting described in the original post
>>> is done routinely at our neutron scattering facility where the whole
>>> model function can be composed of many Gaussians, Lorentzians,
>>> Lognormals, etc.
>>>
>>
>>>> Just draw and adjust the initial curve parameters on
>>>> the data itself, then go fit it with a click of the button.
>>>> Of course, it is simply a wrapper for MPFIT. The tutorial
>>>> Rob has provided is helpful for getting started.
>>>>
>>
>> Thanks David and Rob,
>>
>> Yes I looked at these - the trouble is that I want to perform this on
>> literally thousands of images (each image has one object in it).
>> Manual seeding is not practical. I think what I am after is described
>> here...
>>
>> http://tinyurl.com/9horr
>>
>> I have started to translate the MatLab code to IDL - but it is clearly
>> over my head.
>
>
> Long ago I had somebody write to me, asking about how to do this kind
> of thing. I don't think he ever succeeded.
>
> You are going to have problems like:
>
> 1. uniqueness - there are essentially an infinite number of ways to
> add gaussians to reproduce the data; for example, why not have
> one gaussian per sample?
>
> 2. robustness - the problem is so unconstrained that there is
> significant potential for screwed up fits.
>
> I would recommend constraining the analysis as much as possible based
> on your problem domain, for example if you know that peaks must be
> positive, or the natural widths of the peaks, etc.
>
> One technique would be to find the tallest peak, or perhaps the N
> tallest peaks in the data, and fit gaussians to those. Then subtract
> that fit and see what the next highest peaks are, and fit another set
> of gaussians. You keep doing that until you reach some noise
> threshold (i.e. the errors on the amplitudes are comparable to the
> amplitudes). Sometimes this is known as "CLEAN" in image space.
>
> But be prepared for the fitting to fail or give whacked results at
> least 30% of the time.
>
Thank you. Yes there do seem to be a lot of pitfalls. As long as the
peaks are distinct, it seems feasible. Rob Dimeo has shown me how to do
that using his GET_PEAK_POS routine. It really works well until the
peaks are either close together or the second peak has a very small
amplitude. I see the latter case when my images contain objects that
are 'all edge'. That is when the interior of the objects (plankton in
this case) have transparent interiors. The
fit-the-biggest-peaks-then-remove works pretty well in this case.
Thanks again,
Ben
|
|
|
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
