Re: Size of Gaussian filter for best subsampling
- From: Martin Leese <please@xxxxxxxxxxxxxxxxxxxxxxxxxx>
- Date: Thu, 21 Apr 2005 11:28:59 -0600
Jonathon wrote:
I have a fairly large image that I need to subsample for analysis. I believe that, rather than subsample the original image, it is better to low-pass filter the image and subsample the filtered image. First question, is this correct?
Yes. You must filter to avoid aliasing.
I would prefer to use a Gaussian filter, as I can implement this with two 1-D passes, rather than a full 2-D convolution. Second question, how do I choose the proper standard deviation and extent of the Gaussian filter if I want to subsample by any given factor?
To work correctly, your anti-aliasing filter must eliminate high frequencies above half your new sampling frequency. What you are trying to do is to satisfy Nyquist. When your original image was created this was (hopefully) done using a low-pass analogue filter.
So, if you are sampling by 2 then you need to halve the frequency content. If you are sampling by 5 then you need to reduce the frequency content to one fifth of the original.
Look at how your Gaussian filter behaves in the frequency domain to choose a sensible value for the width in the spatial domain.
-- Regards, Martin Leese E-mail: please@xxxxxxxxxxxxxxxxxxxxxxxxxx Web: http://members.tripod.com/martin_leese/
.
- Prev by Date: Re: Size of Gaussian filter for best subsampling
- Next by Date: Re: Size of Gaussian filter for best subsampling
- Previous by thread: Re: Size of Gaussian filter for best subsampling
- Next by thread: CFP: Special session on Image and Volume Segmentation (VIIP 2005)
- Index(es):
Relevant Pages
|