> I always thought that down sampling consisted of some kind of averaging > (of > samples). > I thought bicubic and bilinear and such terms could as well be related to > down sampling as they could to upsampling. > Now I wonder: how does downsampling work? > Does it exist of sampling only one of the pixels in the previous larger > image for each pixel in the new image? > It more or less explains why I have some grainy images that retain a lot > of > their graininess when downsampled. But why is downsampling often called > better than downsizing? > > Thank you all in advance, > > Jerry > > > -----Original Message----- > > From: Shough, Dean [SMTP:[EMAIL PROTECTED]] > > Sent: Wednesday, December 06, 2000 4:38 PM > > To: '[EMAIL PROTECTED]' > > Subject: RE: filmscanners: RE: Film Scanners and what they see. > > > [cut] > > I expect you are right except perhaps for the Epson 1200 and 1600 series > > scanners. I am not sure if they use a custom CCD with smaller pixels or > > if > > they are micro-stepping with an ordinary 600 and 800 dpi array. But, > now > > that I think about it, if you use a scanner at 1/2 or 1/4 of its full > > resolution, then the pixel size remains the same but the Nyquist limit > is > > much lower. Sounds like a recipe for alaising and another good reason > to > > always scan at higher resolution and average down (not down sample). > > > [cut] > I created some test files so everyone could see what I am talking about. This first image shows a section of a 512 by 512 pixel zone plate. <<...OLE_Obj...>> The original file is symmetric about the bulls eye. The fringes seen in the image increase in frequency as one goes from the center to the edge of the image. The edge of the zone plate is exactly at the Nyquist frequency. Anything in the image outside the inscribed circle is alaised (In the original image, of which only a small section is shown above. If I reduce the size of the image by a factor of 4 in both directions using Photoshop's bicubic interpolation, I get: <<...OLE_Obj...>> This is very similar to what a 128 by 128 pixel scanner would see when looking at the above image. Because of the averaging effect, the bulls eye pattern fades away about 1/4 o the way out from the center. Some slight modulation is still apparent out at the edge o the image - this is alaised information whose magnitude has been greatly reduced due to the low pass filtering of the pixels finite size. If I reduce the size by using Photoshop's nearest neighbor algorithm (the same as down sampling) I get: <<...OLE_Obj...>> Outside the central 1/4 of the image, the strongly alaised image no longer bears any resemblance to the correct image. Further more, any filtering I apply to this image will not produce correct image - not even close. You could say that this alaised image is sharper than the non-alaised image, but I don't think anyone interested in obtaining an accurate image would like it. In a real image, the same type of effects will take place, although not nearly as dramatic. There are two reasons for this: 1) No real scanner is going to be able radically alias the image to the extent shown here. Unless of course you tell the scanner to do a low resolution scan and the scanner uses down sampling. 2) No real image will have such strong and regular features above the Nyquist limit. I hope that I made the images small enough to not trip the email filters yet large enough to demonstrate the point. ---- Dean Shough [EMAIL PROTECTED]
