Morfeus wrote:
looks good after sharpening. What did you do to resize?
Heinz, I grabbed the raw file as well as I got interested in the problem. I run it through ACR and got a TIF i opened in Photoshop. Downsizing to a width of 1000 pixels using "Bicubic (best for smooth gradients)" and then sharpening after that worked fine.
Morfeus wrote:
looks good after sharpening. What did you do to resize?
I was afraid you would ask that. I opened the image in Matlab, performed a discrete 2D Fourier transform, threw away the high-frequency contents, applied windowing to my liking, and finally took an inverse 2D Fourier transform.
The first attempt was the best I could do with bicubic resampling in Paint Shop Pro X3.
Toothwalker wrote:
I was afraid you would ask that. I opened the image in Matlab, performed a discrete 2D Fourier transform, threw away the high-frequency contents, applied windowing to my liking, and finally took an inverse 2D Fourier transform.
oups. I do have Mathlab 13 installed on one of my computers but I am a total idiot in working with it.
Toothwalker wrote:
I was afraid you would ask that. I opened the image in Matlab, performed a discrete 2D Fourier transform, threw away the high-frequency contents, applied windowing to my liking, and finally took an inverse 2D Fourier transform.
Yeah, that's the hardcore way to do it. The problem though with low pass filtering is that you're also nuking the fine detail in the image. A 2D discrete wavelet transform might work better though. You could identify the precise component of the aliasing and remove just it, hopefully not killing too much detail along the way.
Here I selected (before resize) the offending metal blinds in the circle and applied a Gaussian blur (r=2.5 px) to it. Then I used the same resize script: http://peltarion.eu/img/comp/moire/B_preblur.jpg
Much better than my first version, wouldn't you say?
Not quite as good as nuking the high frequency components, but at the same time it doesn't destroy any other detail.
denoir wrote: Yeah, that's the hardcore way to do it. The problem though with low pass filtering is that you're also nuking the fine detail in the image. A 2D discrete wavelet transform might work better though. You could identify the precise component of the aliasing and remove just it, hopefully not killing too much detail along the way.
Here I selected (before resize) the offending metal blinds in the circle and applied a Gaussian blur (r=2.5 px) to it. Then I used the same resize script: http://peltarion.eu/img/comp/moire/B_preblur.jpg
Much better than my first version, wouldn't you say? ...Show more →
Not bad. Not bad at all. This technique might be less suitable for a batch job though
Not quite as good as nuking the high frequency components, but at the same time it doesn't destroy any other detail.
Well, there are always trade-offs. I notice now that the window that I used is still too steep: there is some ringing in my FFT image.
I've been meaning to make a change to it, but I have not gotten around to do it. A Carsten pointed out, it's not good to have even steps between the resizes. So the 3000->2000->1500->920 should preferably be something like 3001 -> 2003 -> 1493 ->920 or something similar (i.e using prime numbers)
By the way, I've forgotten to say - great shots guys. I've really enjoyed the images you've posted. It looks like you had a great time
denoir wrote:
Much better than my first version, wouldn't you say?
Your first and second version have in common that the blinds end up greenish, whereas I get gray blinds. I suppose that is a bug/feature of the change to Lab color mode?
Probably just different color profile for the 5DII RAW development. The first shot is just resized and all sharpening operations are done on the L channel alone which means that colors can't be affected.
