Here's a good page for comparison of various demosaicing algorithms: http://www.rawtherapee.com/RAW_Compare/
From that page, it seems clear that RawTherapee's EAHD is the best, but of course these things vary from photo to photo. Unfortunately you can't just pick up the algorithm, though. In many cases they're proprietary, or perhaps even patented.
With the Article on Raw Therapee -- I couldn't download the original source, so I can't do an independent assesement. I not sure what the default settings were left or not.
Anyone have a link to the original RAW source (the one on the site is broken)?
I have tried as many algorithms as I can find, and find the traditional AHD to be the best overall algorithm, because it doesn't really cause this edge issue (which comes have as wierd lines or sparkling of the image is sharpened). For example, according to the article, the EAHD algorithm seemed to do the best, but I'd like to see how it works overall with a lot of images.
I'm glad Curt posted that above, because it did make me think of a couple things. I can see why most editors don't offer a selection of algorithms -- you really need to know a lot do that, and it ends up complicating things for people who don't look at it that seriously. I do, for sure, but a lot of people who use Sagelight don't -- I think I am finally figuring out how to strike a balance with all that, though. I hope. ha.
Also, I wouldn't necessarily agree with one point, though -- I don't think that all editors that don't offer the choice of demosaicing algorithms necessarily use the same one. Certain cameras identifiably don't work with certain algorithms; it's just that the editor wants to pick for you based on certain criteria.
The nice thing (with the proprietary issue) is that a lot of these algorithms are under GPL license. So, I can just write a .DLL to use them without having to give away the source to Sagelight.
Rob, I wonder if your work on color spaces could be leveraged here. I wonder if the non-proprietary non-patented algorithms might work better when they're operating in alternate color spaces.
You know, I was thinking the same thing. I can't say I am familiar with how all of these demosaicing algorithms work, but I was thinking that it would be interesting to look at it as a C*I*E LAB-interpreted image, for example -- this may be the basis for some of these algorithms that "blur" the image, the idea being that you'd construct a grayscale image from the pattern and then average (in some way or another) just the colors.
Well, it makes me excited to try it all out.