Ben/Pom writes:I hear a lot about downsizing to match noise but have never seen the truth of it in the real world myself,
You've got company: DPReview Blog
Unfortunately, DPR understands very little about image processing, and many other technical aspects of what they do in their testing. This blog post has been thoroughly debunked in a couple of threads over at DPR:http://forums.dpreview.com/forums/read.asp...essage=30176643
Moreover, the DPR blog post isn't even a test of what it claims to be. For the test they should have done, seehttp://forums.dpreview.com/forums/read.asp...essage=30190836http://forums.dpreview.com/forums/read.asp...essage=30211624
Bottom line: downsizing eliminates fine scale structure from the image. If there is no fine scale noise (the "grain" is coarse; noise reduction has been performed), then there is nothing in terms of noise for downsizing to eliminate -- it has already been taken out of the image. This is why often one sees little effect on noise from downsampling an image; that is certainly what happened in the examples studied in the DPR blog post. When I ran a controlled study of downsampling a 50D test image to the dimensions of a 40D image (the second set of links above), I found results consistent with theoretical expectations -- the std dev of noise reduced by roughly the ratio of pixel dimensions of the original and downsampled image.
BTW, it is not so much that downsampling reduces noise. Rather, when one pixel peeps one is looking at a coarser scale in the image, and the noise at that scale. Since noise rises at finer scales (unless it has been scrubbed out by NR), one removes that part of the noise spectrum with the downsample and what is left over is the lower level noise at the coarser scale. But that noise was always there prior to downsampling, and wasn't changed by the downsampling. It was simply made into the "pixel level noise" by changing what "pixel level" means. So in this sense downsampling is not reducing noise; on the other hand, it is reducing noise since one is throwing away the high frequency noise (well, one is throwing away high frequency everything, including image detail) via the downsampling.