About those speckles...

Well, as you know - if you read some essential blog posts on DNGMonochrome - I discovered some problems with the two sharpest algorithms a while back, most noticeable on high ISO M8 files, leading to white or light gray speckles. So far I haven't seen them on high ISO M9 files, but I also didn't test a lot of high ISO M9 photos.

I discovered the speckles when I was experimenting with a downloaded 2500 ISO M8 DNG, and it took me a while to understand what was happening.

Essentially the two sharp algorithms stay very close to the pixel being interpolated, and the number of pixels looked at - to get to the right value for the pixel under consideration - is small. The advantage: the interpolation stays close to home and it leads to sharper results.

The smoother algorithm averages over a larger area, looks at more pixels, and loses sharpness that way. The advantage though is that noise (and dead pixels) gets eradicated. It leads to a smoother, less noisy, less sharp result and no speckles on high ISO images.

What happens with the sharp algorithms on high ISO is that the color noise starts to have a big impact on the outcome, in a rather negative way. This particular M8 file was riddled with color noise, especially the red channel was totally out of control, with big red blotches in the shadows. As a consequence, the information picked up from the red channel was incorrect. And when averaging based on only a few pixels, these out of control red pixels were too big an influence, causing wrong values for the pixel under consideration, leading to the speckles.

I managed to solve that today and the results are quite satisfying now. Still sharper results than with the smoother algorithm, without the speckles and without using a median filter.

Lower ISO images are untouched by this method. It only kicks in when necessary and on a per pixel basis.

I still need to transport this new approach to the red and blue filtering, and test if I also need to implement this for M9 high ISO files. Need to produce some very noisy photos first for some testing.

Obviously these improvements will be in the next release...