p.3 #1 · Occurance of Aliasing (Moire) and Postprocessing...
rico wrote:
I do have an even more nasty pattern in mind, so expect a future challenge (insert evil laugh here). Meanwhile, how does your algorithm fare on an image with no aliasing present. In other words, do you have to steer the process toward problem areas and away from the good? This is a question about detection of sample ambiguity, which theory says cannot be done.
Hi Rico
I could just hear that evil laugh in your words and it made me laugh. Bring it on
Hmmm.....in regards to a sample of ambiguity, I thought that just reading my posts would have satisfied you . Seriously, my background is engineering and am fairly familiar with digital signal processing. I must admit that even though I have taken classes in the subject and have a couple thick books on Image Processing I have to say I am not familiar with your terms "detection of sample ambiguity." You might have to clarify or point me to a link. I'm a fast learner. (well I like to think so)
I also try and be careful of theories and laws that say "it can't be done." Too often they are used much more broadly than for which that limitation applies. i.e why let the laws of physics get in the way of a good practical solution?
Sorry, just having fun. I will give an example. It two signals are combined, often one may think that they cannot be extracted from each other. If one knows the algorithm with which there were combined, the original signal(s) can be deconvolved. Not as easy when the frequency of input in a sampled system exceeds the Nyquist criteria. Yet not all is lost and here are some reasons why
- Demoasicing algorithms are known functions so you know how the signal is being processed. Very predictable and repeatable.
- There are many anti-aliasing artifacts that occur when you are right on the cusp or edge of the Nyquist criteria. The behavior is much more predictable in these situations.
- There is more information than one might think. Demosacing is not done a a pixel by pixel basis. For Bayer type sensors (RGGB 2x2 grids) the algorithm must include information from the 2x2 supergird (or even larger). Both luminosity and color has to be extracted from this RGGB 2x2 grid and converted to 4 full pixels of 16bit RGB. When the image signal frequency approaches this 2x2 grid frequency or higher you often (yet not always) get some pretty predictable behavior both in incorrect color as well as luminosity separation. There is relatively low frequency color/luminosity separation that one sees as striations. There are also some high frequency adjacent pixel separations as well.
- human intervention can often recognize the pattern and tune the process for a better result
So to answer your specific question, for your paper images, the only Layer Mask that I used were masking each of the paper interiors separately yet in their entirety. No layer masks were needed at a sub level within a given paper. Now there are single shot images where various areas need separate treatment. Basically if the image frequency that causes the Moiré striping is substantially different in different parts of the image then each must be addressed separately.
Now, it turns out that the PS approach I take for the adjacent pixel luminosity separation, while I do not use a Layer Mask, it effectively self selects those areas with the artifacts (too hard to explain in text and may I may do a video and post on a blog). Basically action is taken where that luminosity separates between adjacent pixels. So in that respect, action is limited to portions of the image, just not with a Layer Mask and it the case of your paper images, I did all 4 with the same processing pass (they did not need to be done individually).
So was this a good "sample" of my posting "ambiguity"?
There is no question that the techniques I mention do not work on all anti-aliasing problems. I have found that of the cases presented to me, it has worked on a good number.. Also, by "working" I mean it reduces the visibility of the artifacts. It does not recover the high frequency image signal. However, suppression of the artifacts has the potential for some images to provide a practical improvement when a re-shoot is impractical.
p.3 #2 · Occurance of Aliasing (Moire) and Postprocessing...
John Wheeler wrote:
... Basically if the image frequency that causes the Moiré striping is substantially different in different parts of the image then each must be addressed separately.
Thanks for the comprehensive response. By sample ambiguity, I mean a data sample that has been influenced by a high-frequency alias. You corrected the large-area sampling errors, but my next mission is to create a similar chromatic effect which is real. I don't expect to generate a real maze effect since its frequency exceeds the sensor geometry. Now scheming...
