Pages: 1 [2]   Go Down

Author Topic: A critical look at demosaicing, capture sharpening and noise rendering  (Read 5226 times)


  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 454
Re: A critical look at demosaicing, capture sharpening and noise rendering
« Reply #20 on: February 17, 2017, 07:55:05 PM »

To follow up, I've conversed with Prof Emil Martinec by email and he has very kindly allowed me to share his insightful comments about the AMaZE algorithm. Rather than try and paraphrase what he said, I'm copying them wholly in this post:

Yes, AMaZE useds a different method for interpolating diagonals and yes it sometimes produces artifacts of the sort you are seeing. Itís a different way of doing the Bayer demosaic that is invoked when there are edges which are more strongly diagonal than horizontal/vertical. The non-implemented aspects of AMaZE had more to do with other parts of the algorithm, not the diagonal interpolation, so would not have any effect on this particular issue. If I ever get back to tinkering with it, one of the things I want to look at is to fix the sort of zipper-type artifacts you are seeing.

AMaZE uses several different interpolation schemes, and then attempts to select the ďbestĒ according to some criteria. In the non-implemented parts of the original scheme, there were different interpolation methods, and a different evaluation algorithm (the latter was much more time-intensive).

BTW, the diagonal artifiacts occur when the diagonal interpolation is used. In the Bayer array, along one diagonal are all green pixels, and along the neighboring diagonal are red/blue. So if green and red/blue are substantially different strength or there is a high contrast gradient, one can get alternating one-pixel wide diagonal lines from the differing brightness of the green diagonals vs the red/blue ones. In standard horizontal/vertical interpolations, these effects donít occur because there is a tendency to make the interpolated missing green pixels more strongly the average of their neighbors, and then the red/blue information is interpolated based on that, resulting in a stronger correlation between luminance values at native green vs native red/blue sites. Itís all about attempting to infer edges at a resolution higher than that of the green pixels in the Bayer array, and if you make a mistake you get ringing artifacts at the pixel scale.

--Emil Martinec

So I think this clearly establishes that the diagonal artifacts are indeed present and are a known issue with the AMaZE demosaicing. As cuniek pointed out in the thread I created on the Raw Therapee forums, it may be possible to average the brightness of neighboring green pixels with red/blue pixels, like DCB does, and so avoid the ringing artifacts. However, the result of doing that is chunking into 2x2 pixel block jaggies on diagonals, which is what DCB, Capture One V10, Lightroom and other demosaicing algorithms suffer from.
Pages: 1 [2]   Go Up