AJSJones wrote:
AA filters don't necessarily lose huge amounts of resolution if implemented as desired. Canon's descrptions always imply that the filter is deigned to split an incoming ray into 4 distinct rays at the corners of a square with sides that are the pixel pitch - ensuring each ray is ampled by 4 sensels. It creates 4 separate overlaid images - not the same as blurring one image out so eachvray covers 4 sensels. I once was curious to see the effect, mimicked in PS, of this compared to a blur. The image below shows what the AA filter does* - and how Canon's suggested "undo" for the AA filter performs (200 or 300%, 0.3 radius, 0 threshold USM) the 1 pixel Gaussian blur is waaaay worse while the AA filer effect is indeed largely undone by the mild, subpixel USM. The AA effect is actually quite mild and largely recoverable.
* the original image copied , moved one pixel right, 50% opacity, flatten, rinse repeat with one pixel up.
I agree that the worries about huge losses from this filtervare overstated or over worried! The examples I've seen if before and after filter removal were underwhelming, to say the least. ...Show more →
Thanks for sharing, very interesting! I tested your suggested sharpness levels - not sure what is meant with 0 threshold USM, but 200% and 0.3 radius was only a tiny bit sharpening. Doesn't look like this is the mentioned 15-20% blur loss coming from an AA filter. But who knows - maybe the blur loss is really a lot less than often mentioned elsewhere.
The original Photoshop had three parameters for USM and the third one was threshold.
The subpixel, large amplitude USM undoes some of the AA filter effect without adding sharpening artifacts to non neighbour pixels. It doesn't really sharpen, but then again, the 1 pixel AA effect is also quite subtle. However, it is extremely difficult to ascertain how big the sides of the square formed by the 4 rays actually is in the camera - it is proportional to pixel pitch but it may well be more than the 1 pixel I mimicked here. I have a couple of other posts I'll link to when I get to a computer - tough on an iPhone
KaaX wrote:
There's quite a limited selection of the geometric shapes which would cover a 2D plane without gaps. And moire is a function of sampling high frequencies, going to, say, hexagonal sensels would give you different moire but not get rid of it altogether. And I still don't understand what does Poisson distribution have to do with all this.
<mind CONTRAAAAaaacting>
Well, leave it that way if you prefer.
For the rest of the readers, we can consider a Poisson distribution of non-square, non-uniform, differently shaped, differently sized sensels, just like the human retina, which astute readers will remember I made extensive reference to earlier in the thread. It will come within a decade or two, I think. The hardware/firmware/software will be taught at the factory to function, just like the human visual system teaches itself how to function, starting from a blurry zero as a newborn infant.
retrofocus wrote:
I tested your suggested sharpness levels - not sure what is meant with 0 threshold USM, but 200% and 0.3 radius was only a tiny bit sharpening. Doesn't look like this is the mentioned 15-20% blur loss coming from an AA filter. But who knows - maybe the blur loss is really a lot less than often mentioned elsewhere.
The Anti-Aliasing filters on the 5D series of cameras (and all modern DSLRs) are weak. Thus aliasing can show up in unexpected places:
