Re: Best Method of Combining estimates
From: Peter Spellucci (spellucci_at_fb04373.mathematik.tu-darmstadt.de)
Date: 11/23/04
- Next message: Anu: "can someone help me with simplex LP from numerical recipes"
- Previous message: Jonathon: "Best Method of Combining estimates"
- In reply to: Jonathon: "Best Method of Combining estimates"
- Messages sorted by: [ date ] [ thread ]
Date: Tue, 23 Nov 2004 12:30:16 +0000 (UTC)
In article <11783d36.0411230026.796ccbd3@posting.google.com>,
jon@pidham.vispa.com (Jonathon) writes:
>I posted this at sci.math, and got no answers. Maybe it is too
>applied. Can anyone here help?
>
>
>I have inherited a processing scheme that looks for features of a
>certain size in an image. This is done by applying a series of
>filters, each of which has a maximum output at a certain feature size
>and whose output drops off at bigger and smaller sizes. At present,
>size detection is carried out by simply selecting the filter with the
>greatest output.
>
>The outputs of the filters are fairly smooth, and overlap
>significantly. The outputs are proportional to image contrast, which
>can vary significantly from image to image.
>
>I would like to try to get a bit more accuracy from the existing
>processing scheme. I thought of some method of looking at the ratios
>of the output of the 'maximum' filter to those of the filters on
>either side. This gives me two numbers that should allow some form of
>'interpolation' for a more accurate result.
>
>I could run the filters with features of different lengths and get
>enough information for a 2-D lookup table for each filter (1-D for the
>ones at the ends), but I can't help thinking that first, there should
>be a more elegant (and storage-efficient) way of doing this; second,
>it might be useful to be able to extend the scheme to more than just
>the ones either side, and that would mean a many-D lookup table; and
>third, someone must have done this before.
>
>Does anyone have any suggestions, references, links, hints, etc?
>
>Thanks in advance,
>
>Jon
sounds as if a least squares fit using radial basis functions and a variable
number of nodes (until the fit is sufficiently good) should be the right way
to go. the problem is that variable peak position makes this a nonlinear fit
which may be nasty, but this depends on the data. codes for interpolation
with radial basis functions is in http://www.netlib.org/toms/660
and .../790 but there seems to be no fit program ready for use
hth
peter
- Next message: Anu: "can someone help me with simplex LP from numerical recipes"
- Previous message: Jonathon: "Best Method of Combining estimates"
- In reply to: Jonathon: "Best Method of Combining estimates"
- Messages sorted by: [ date ] [ thread ]
Relevant Pages
|
