Newton Raphson based template matching

Message-ID:<3ea48544-98d4-4c24-9bdb-f46e06bd73e3@s14g2000vbp.googlegroups.com>
Subject:

Newton-Raphson based template matching

Date:Thu, 29 Jan 2009 17:50:09 +0100


Basic template matching by cross correlation uses an exhaustive
search. I'm trying to reduce the number of iterations needed to find a
match between a template image and a search image by adopting a Newton-
Raphson based iterative method. I will have a reasonable initial
estimate from where I can begin searching.

I have doubts as to how I can expand the Newton-Raphson method for use
with images. My understanding is that it is usually used to match
linear/nonlinear curves (visualized by sliding the template curve
across the search curve to find best match). Should I treat my problem
to be a matching of surfaces since the images have x-y coordinates and
brightness information at each pixel?

I came across the paper "An Iterative Image Registration Technique
with an Application to Stereo Vision" by Lucas and Kanade that has an
explanation for this (http://citeseer.ist.psu.edu/
lucas81iterative.html). Section 4.5 shows the generalizations to
higher dimensions, but I am not able to get the matrix sizes to
agree.

If any one of you have done something similar, I'll be glad if you can
share your thoughts.

Message-ID:<4981e93d$1@news.arcor-ip.de>
Subject:

Re: Newton-Raphson based template matching

Date:Thu, 29 Jan 2009 18:36:59 +0100


perhaps you will find the script "least squares matching tutorial"
on the course site "http://cobweb.ecn.purdue.edu/~bethel/603_05not.htm"
interesting...

greets

Message-ID:<1187a8f2-1b70-4bed-9700-84b805a9b7aa@y23g2000pre.googlegroups.com>
Subject:

Re: Newton-Raphson based template matching

Date:Tue, 3 Feb 2009 02:58:37 +0100


On Jan 30, 12:50=A0am, Bala L  wrote:
> Basic template matching by cross correlation uses an exhaustive
> search. I'm trying to reduce the number of iterations needed to find a
> match between a template image and a search image by adopting a Newton-
> Raphson based iterative method. I will have a reasonable initial
> estimate from where I can begin searching.
>
> I have doubts as to how I can expand the Newton-Raphson method for use
> with images. My understanding is that it is usually used to match
> linear/nonlinear curves (visualized by sliding the template curve
> across the search curve to find best match). Should I treat my problem
> to be a matching of surfaces since the images have x-y coordinates and
> brightness information at each pixel?
>
> I came across the paper "An Iterative Image Registration Technique
> with an Application to Stereo Vision" by Lucas and Kanade that has an
> explanation for this (http://citeseer.ist.psu.edu/
> lucas81iterative.html). Section 4.5 shows the generalizations to
> higher dimensions, but I am not able to get the matrix sizes to
> agree.
>
> If any one of you have done something similar, I'll be glad if you can
> share your thoughts.

If you are intended to save iterations , one way is to do pre-search,
such like image pyramid

best regards