Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Posting-Version: version B 2.10.2 9/18/84; site rochester.UUCP
Path: utzoo!linus!philabs!cmcl2!seismo!rochester!sher
From: sher@rochester.UUCP (David Sher)
Newsgroups: net.graphics
Subject: Re: FFT of image in sections ?
Message-ID: <10900@rochester.UUCP>
Date: Tue, 6-Aug-85 18:15:25 EDT
Article-I.D.: rocheste.10900
Posted: Tue Aug  6 18:15:25 1985
Date-Received: Thu, 8-Aug-85 01:08:13 EDT
References: <360@ur-laser.uucp>
Reply-To: sher@rochester.UUCP (David Sher)
Organization: U. of Rochester, CS Dept.
Lines: 25
Keywords: FFT convolution
Summary: FFT only requires a row or collumn at a time.

In article <360@ur-laser.uucp> nitin@ur-laser.uucp (Nitin Sampat) writes:
 ...
>We know that processing small images takes less time.  Well, how does 
>one go about breaking up a large image and process it in sections,
>and then most importantly, how does one put all these sections back 
>to get the FFT of the original image ?
>
>				nitin
>		{seismo,allegra}!rochester!ur-laser!nitin.uucp

It turns out that on a 2d image applying FFT only requires that a row
or collumn of data need be processed at a time.  If the image is reflected
along a diagonal between the row and collumn processing the paging should be 
minimized fairly well and the entire transformation is applied.  Of course
there is still a log n factor that is lost from the nature of the FT.  
However I would suspect that for the small images (1Kx1K) that are worked
on today it would cost more to do the gluing operation that glues together
the results of FFT's on partitions of the image than to do the entire FFT.

On the topic of FFT vs convolution, my latest TR on doing template matching
on the WARP and the Butterfly addresses this issue for these architectures.
However it is not classy to push ones own work so enough said.
-David Sher
sher@rochester
seismo!rochester!sher