HSL may have better precision than RGB
Notice how the workflow keeps getting more complex?
I suspect that the RGB encoded version is worse than the HSL encoded version. For one reason, he should have used R & B, not R & G because most of the Luminance information is also stored in G.
He should also have done a logarithmic µ-law encoding on Luminance, because that way you'd see more of the rays coming out from the stars and be able to do more accurate correction. Low intensity values will be overrepresented and high intensity values will have big truncation errors.
#1 - Lobby him for a 16-bit version of the plugin. This will solve most of your woes.
#2 - If your stars are white on a black field, you really need to erase to black and perform the anti-aliasing (feathering on a selection or a soft black brush).
Yes, I know all about Image-J. Your tax dollars at work. Prior to that, Wayne Rasband was working on NIH-Image, which had a bad combination of poor science and being free (he's a government employee). Thus, people used it and had no reason to pay for a *real* package that had the requirement of using good science. It also made it really difficult to compete in the early 90's for the low end of the market. Image-J has just been a sport since then. If you can fit it into you workflow, go for it. It is worth exactly what you paid for it. (If you don't consider that your tax money has been spent driving little guys like me out of business.)</rant>
FFTW is just another implementation. Whose implementation is used is largely irrelevant as far as the quality of the result is concerned.
Chirakov x Fovea (Round 1)
If you look at the attachments first, you'll think that the first should be in Doug's Cloud thread and the second is a very uninspired art work.. read on...
My final year project tutor taught me that when you want to discover something, devise an experiment - so this is what I did.
1) Take an "average" photo (Dunno what is average, but the golf balls ain' t);
2) Separate out Luminosity;
3) Expand canvas to some size multiple of two, fill in with 50% grey ;
4) Run through FFT, then imediately return through inverse FFT;
5) Crop back to the original size;
6) Compare with the original.
The noisy image posted is an amplified (Levels 0 - 8) picture of the difference using the Alex Chirakov filter (average difference 3.5), the clean "art" is the Fovea filter (average difference 0.3).
There was no significant difference between the RGB and HSB flavours of the Chirakov filter.
-The Chirakov version does introduce a considerable amount of noise, whereas the Fovea process does not introduce any. (There is a 1 pixel difference, but this is uniform and to me that's not noise).
- There's no difference between the two Chirakov filters, so you can use the easier (RGB) one without problems.
PS Chris - keep it coloured, otherwise the folks will think we're twins!
You both look like you need to crawl out of your cubby holes and hit the beach for a while. Ro, have you taken my advice yet?
Duv, message well taken
(Actually, this was a break for me -- I'm supposed to be writing code.)
The test was a good idea, Rô.
Later. Since we're not near the ocean, it's time to go to the YMCA.
Wow! That’s heavy stuff. I feel like I need the beach after trying to understand that lot.
I am not a professional retoucher. I don’t have hundreds of textures to remove. I hit this problem once or twice a year and just want to get the best I can from an odd picture.
Thanks for doing that comparison I think you are saying
FFT RGB is as good as FFT HSL But FFT Fovea is Better.
But if I’ve read correctly you are comparing the loss between the transfer
What interests me is the end results and how to achieve it. So what probably interests me more is “What do I do to the FFT to improve the Picture?”
The initial picture I supplied is a very poor example to use as a ‘Test picture’
What do you think of jcr6 cat picture – is that a good test picture?
I’ve read chapter 4 and pages 9 and 10 of the pdf.
Filter > IP*Fourier > FFT(Forward)
Filter > IP*Fourier > FFT(Inverse) OR Apply Filter and FFTInverse
The picture comes back different (darker the grey padding goes black)
Could you please explain this?
Also could you please clarify that FoveaPro should be working on a Luminocity Image?
And if so then do you agree that byros method is OK.
1) Duplicate original image, right click on title bar and click Duplicate (that's a new image, not just a layer):
2) Set background colour to grey 127,127,127 (for Padding)
3) Image > Canvas Size > 512,512 in Pixels (In this case - square with dimension equal to an exact power of two)
4) On the new image, get Luminosity (New Layer (<Ctrl><J>), Edit>Fill - 50% grey, color, 100%.);
5) Flatten all, <Alt><L><F> (FFT does not understand layers);
6) Filter > IP*Fourier > FFT(Forward)
7) Paint out the Stars in white with a soft brush
8) Filter > IP*Fourier > Apply Filter and FFTInverse
Does that sound OK?
My suggestion earlier about adjusting brightness and contrast to ‘see the stars easier’
This does not seem to spoil the final image. What do you think?
You did ‘cheat’ a little. You added the extra step.
First I 'healed' out the stars in the Blue Channel (figure1), then in the Green Channel (figure 2) and finally I 'cleaned' the 'composite' RGB (figure 3).
'cleaned' the 'composite' RGB (figure 3). Is not in the manual
Perhaps you could explain this step then we can find out how to apply.
Hi everybody ...
Just wanted to say this indepth thread made me go back and look at FFT, I had come across this filter in the past but misunderstood it, I had done the classic FFT then IFFT and see no result (of course..doh!) now reading this and going through old threads I have got it working. I still find it easier to use HSL as it shows the issue up clearer I think, I had some good success carefully cloning the vertical and horizontal lines which got rid of small highlights in that area that caused problems in the images.
So thanks for thrashing this subject again, this time it stuck for me
It's not much... but I hope this helps....
|Currently Active Users Viewing This Thread: 1 (0 members and 1 guests)|