Chi squared two sample test to measure image correspondence
in [Matlab]
|
From: ImageAnalyst on 22 Feb 2010 10:09 Randall: What exactly are you going to be comparing? The histogram distributions? Or some kind of pixel differences? If you're just going to compare the histograms, that's a really lousy way to compare how alike two images are. As you can imagine, lots of images can have similar, even identical, histograms and look totally different.
From: Randall Flagg on 22 Feb 2010 13:55 ImageAnalyst <imageanalyst(a)mailinator.com> wrote in message <87102c4c-2d6b-4c38-b36a-1b5638c5c36d(a)o3g2000yqb.googlegroups.com>... > Randall: > What exactly are you going to be comparing? The histogram > distributions? Or some kind of pixel differences? If you're just > going to compare the histograms, that's a really lousy way to compare > how alike two images are. As you can imagine, lots of images can have > similar, even identical, histograms and look totally different. Not sure how to understand that question. Currently I'm comparing images with the Pearson correlation coefficient. My question was whether the same can be done by a Chi squared two sample test, as I would like to have a more robust measure. I guess you're right and it wouldn't work like that. So there's no way to use chi squared for that purpose? If so, does anyone know of another measure I might use?
From: ImageAnalyst on 22 Feb 2010 14:24 Well, like I said PSNR. But there are a few more on this page: http://en.wikipedia.org/wiki/Comparison_of_video_codecs (where it says PSNR is the "most widely used video quality metric during last 20 years (used approximately in 99% of scientific papers and in 20% of marketing materials). " Maybe one of the others will work well as well, or better. Or just Bing/Goole "image quality metrics" or "compression quality measures" - stuff like that. The reasoning is that you do something to an image (such as compress it) and you want to measure how much it's like your original reference image, so you can see why it's big in compression research.
From: Randall Flagg on 22 Feb 2010 16:53 ImageAnalyst <imageanalyst(a)mailinator.com> wrote in message <ee1f5dac-c4cb-4380-9953-07690112a80b(a)g10g2000yqh.googlegroups.com>... > Well, like I said PSNR. But there are a few more on this page: > http://en.wikipedia.org/wiki/Comparison_of_video_codecs > (where it says PSNR is the "most widely used video quality metric > during last 20 years (used approximately in 99% of scientific papers > and in 20% of marketing materials). " As I said the image doesn't come from a compression, it has a completely different scaling that's unknown beforehand. A value that's 1 in the original can be 1e15 in the second one. That's what prevented me from using PSNR so far and made me consider rank correlation. So you're suggesting to just squeeze both into an identical range (say [0;1]) and use PSNR?
From: ImageAnalyst on 22 Feb 2010 19:05
On Feb 22, 4:53 pm, "Randall Flagg" <randall1.fl...(a)gmx.de> wrote: > So you're suggesting to just squeeze both into an identical range (say [0;1]) and use PSNR? -------------------------------------------------------------------------- Sure, why not? Give it a try. |