Simon97 wrote:
> VNG doesn't not do this, but it's output is soft.
> HPHD has the maze effect too and it leaves more of other types of
> artifacts than EHAD.
> I tried adjusting some of the other settings too.
>
> I'm trying RAWhide now. It gives 6 different RAW conversion
> algorithms to choose. I read somewhere on the site that this "maze"
> pattern is a algorithm issue. The ACC algorithm does not "maze" and
> it handles false color in repeating patterns better, but darned if I
> can get it to remove remove the moire color from single line detail.
> RAW Therapee seems better there. (Or I should stop shooting with a
> super sharp lens - the 50/1.7 A at f/4)

Simon, you'll find most raw demosiacing algorithms produce this "mazing" artifact, including RawTherapee, Silkypix, Adobe Camera Raw (ACR), etc. It arises from the "smart" algorithms trying to "connect-the-dots" when there is finer detail than the algorithm can handle or noise in flattish textures. Types of algorithms that don't have this "mazing" are limited to types that produce their output by some sort of simple averaging between selected adjacent photosites, which is why VNG4 works better in that it just does bilinear interpolation and then corrects the edges according to the gradients above a certain threshold. Most of these averaging/linear interpolation algorithms tend to produce softer output unless one applies sharpening, as you have seen.

Removing edge fringing (not moiré) from single line detail is very difficult to do and likely impossible for all cases as in a single photosite width line that is just slightly off horizontal or vertical. The problem is that the Bayer pattern doesn't contain enough colour information from single line width details to properly determine the colour. Most automatic solutions to this problem are to try to detect single width lines and to turn those lines white in these situations, but these types of fix often produce other types of artifacts for other types of situations.

The "false colour in repeating (fine) patterns" is true colour shift moiré and is a natural artifact of digital sampling sensors. Completely removing it in all cases is impossible without blurring the image, but one can use the green channel where the moiré pattern is rarely very strong and where the green channel is strong enough relative to the other channels to correct the pattern in the red and blue channels. Algorithms that do this automatically tend to be quite complex, slow, and cause other artifacts when they go wrong. I prefer to do this moiré correction manually for those few cases when it occurs.

Bayer sensor pattern demosiacing is always a trade-off between the finest detail resolution, artifacts, edge accutance, complexity, and speed. I don't know that we will ever discover an algorithm that excels in all of these. The only good thing is that the megapixel race will gradually make true moiré impossible as the lenses will limit the highest frequencies in the image, edge accutance will no longer be a problem for the same reason, and we will only have to deal with a very slight edge colour fringing as single line detail will be impossible due to the lens limits. Speed of processing these very large raw files is going to become one of the most important characteristics of these algorithms.

Most of these complex algorithms and their problems came about because of the necessity of getting the maximum detail from barely adequate resolution sensors.

Regards, GordonBGood