5 Replies Latest reply on Mar 5, 2009 6:28 PM by MadManChan2000

    Details about BayerGreenSplit and its value

    Level 1
      Does anyone know anything more than the few lines written in the DNG documentation about this tag, BayerGreenSplit?

      How is the green pixels divergence measured? How can one decide what BayerGreenSplit value corresponds to his/her camera?

      The 0 - 5000 range looks quite strange. Maybe the phenomena described has also another name?

      Any help is appreciated.

      Thank you.
        • 1. Re: Details about BayerGreenSplit and its value
          sandy_mc Level 3
          There is an option in Adobe's dng_validate program:

          -b4 Use four-color Bayer interpolation

          This generates a four plane output file (aka a plane for each green channel, rather than one green plane that incorporates both). You can look at that, and decided whether your green channels are mismatched.

          But I have absolutely no idea how to set the value; the dng source code refers to it as "dimensionless", and I've never needed to set it to anything other than 0.

          BTW, green channel mismatches are notorious for triggering maze (or zipper) patterns with AHD demosaicing algorithms. Which is, at least as far as is publicly known, what camera raw uses.
          • 2. Re: Details about BayerGreenSplit and its value
            Tim Lookingbill Level 1
            Sandy,

            I've been intrigued by the demosaicing process and conducted a wikepedia search to find out more on the subject.

            Stumbled upon this pdf documenting a clinical study to improve demosaicing results:

            http://www3.ntu.edu.sg/home5/CHAN0069/JEI013003.pdf

            Was wondering if you'ld found anything similar that was more current and easier to digest with better image samples.

            I'ld be much appreciated if you could provide a link.
            • 3. Re: Details about BayerGreenSplit and its value
              sandy_mc Level 3
              Tim,

              Unfortunately not - for what its worth, the general belief is that most commercial demosaicing algorithms are variations of AHD with more sophisticated homogeneity calculations that take into account local variations in the image. It's is known, for example, that DxO extends their computation out to a greater distance than the classic AHD algorithm. But what the details of are hasn't been publicly disclosed, so far as I know.

              But what we really need is for Thomas or Eric to tell us what's inside the Adobe version.............:).............but I wouldn't hold your breath.
              • 4. Re: Details about BayerGreenSplit and its value
                Tim Lookingbill Level 1
                OK, I exhaled. ;)

                Used to think the weird clay texture ACR rendered high frequency detail like blades of grass and foliage in the distance in my Pentax PEF's was something undesirable. But after comparing ACR's ease of use, quality and speed of its editing tools and simpler interface compared to other raw converters, the odd texturing is tolerable.

                And besides that it can only be seen at 200% zoomed view anyway and can't be seen even on 20x30in. print enlargements. This texture also produces some very clean results when applying aggressive sharpening as well, so I'm thinking whatever algorithm Adobe's engineers are using to produce such a texture must be intentional.

                I'm just fascinated by all the variations achieved just with demosaicing algorithms.
                • 5. Re: Details about BayerGreenSplit and its value
                  MadManChan2000 Adobe Employee
                  Tim, Sandy, Luluta,

                  RGGB CFA demosaic methods generally assume the two G pixels have similar statistical properties. This makes it easier to detect image features. When the assumption fails, however, the dependent image feature detection methods tend to break down ... leading to artifacts of the type that Sandy described.

                  While I cannot go into detail about CR/LR's demosaic method, I can safely say that it is not AHD.