Bayer with NO anti-aliasing (Kodak Pro 14n)

Certainly not when you use a DSLR with a blurry sensor.

That's worth $2500. Why not buy 6 pro lenses instead?

Not with a blurry sensor they don't.

With a blurry sensor? Irrelevant.

That's because your 10D is completely blind to features less that 4 pixels
wide.

And you forget the main reason 10D owners buy L glass: It's a fashion
statement.

 

You really are excelling yourself now Georgey boy.
Go on post a link to another of your horrid pictures again, I like it when
you do that.
How about the blue/yellow tiger one?
Or the portraits with weird skin tones.
Or something with some near horizontal or vertical lines in it.

 

Click the links I provided first.

14n:
http://www.steves-digicams.com/2003_reviews/pro14n/samples_new/C7SG0367.jpg

Compare to medium format film quality (unfortunately heavily JPEG'd,
originally): http://www.pbase.com/image/23733001

Canon 10D images (hard to find, full size) are awash in Bayer color
artifacts too: http://www.pbase.com/image/23908485

You'll see how poorly these outrageously expensive Bayer (good camera
bodies, shame they only have an interpolated sensor inside) cameras really
perform. Image quality from Bayers is very poor, awash in color artifacts,
but given 13.72M sensors and massive downsizing (to 3.43MP), some of the
artifacting can be averaged out. Bayer is still no where even close to
color 35mm film, though when imaging a B&W target (and after being digitally
converted to B&W to eliminate any color moire) in can approach the B&W
resolution of 35mm color film, but probably still not 35mm B&W film.

Sigma DSLRs (only) approach medium format color film, when enlarged less
than 40 inches... http://foveon.com/faq_technology.html#FAQ_tech_13

 

Lets not confuse resolution with detection. By that reasoning, we
cannot see stars, because the human eye cannot resolve a feature size
anywhere near the angular size of a star.

Contrast of image of subpixel object is reduced, but if original
contrast is high enough, you will see it- camera is not 'blind' to it.

 

Well, you might see something, but you won't see "it." If a thin line of
color, say razor thin CA in this case, falls as a single pixel wide line on
the sensor, the SD9 will resolve its color perfectly and display it
perfectly (even if it is less than a pixel wide as long as it is the
dominant color), while a Bayer sensor will show it as a blurred collection
of mosiac color artifacts, as each sensor alternates from red to gren to
blue as the thin line falls across the sensor, the rainbow of these randomly
colored artifacts (in that, they have no correlation to the real color) then
get partially "borrowed" into adjacent pixels that didn't even sense it,
bluring anything less than 4 pixels wide unrecognizably. It would simply be
protrayed as mostly random color artifacts blurred partially into
surrounding pixels.

 

This analyzis is totally correct. And this shows a theoretical advantage
of Foveon over Bayer. I don't really think that anyone can argue that you
are wrong here.

But ... there are some things to take inte account

* A sampling system shall use an anti alias filter. Therefore, no
thin one pixel lines will exist on the sensor. I know you don't agree
(yet?), but it is a fact IMHO.
* A good Bayer interpolation algorithm shall do hue smoothing. Therefore,
the multi colored mosaic will not either exist in the final picture.
In practice, there might be some residue of color artefacts.
* Real world scenes (usually) are much nicer to the Bayer sensor than
the example you described.


Roland

 

Hello

Most of the images on the Dpreview of the Canon 300 have purple/green
artifacts, just like the SD9. Th Canon ones are just blurred away.
You can see the artifacts, where some sort of sharpening has taken
place. It is a exactly similar, but not to the same degree as the SD9.

It has to be said that at 100% viewing it is not noticable. The images
just lack detail.

The choice seems to be apparent sharpness with jaggs(SD9) or no jaggs
and blurriness(Canon et al).

As far as I can see, they all have to do better.


Mike Engles

 

Not totally. The word "perfectly" doesn't fit reality. Since the color
sensitivities of the Foveon sensors are determined by the depth of
silicon between them, they can't be tailored very well to make color
separation effective--hence lousy, not perfect, color resolution.

No free lunch. Even if real color filters could be introduced into a
stacked-element sensor, the lower ones would still be limited to
whatever got through the upper ones. Not insurmountable, in that film
manages to deal with the same problem. But nobody has tried it so far.

I just did, but who am I?

It is not just your opinion, humble or otherwise. It is a mathematical
fact that data at beyond-Nyquist frequencies, if allowed to reach the
sampler, will produce below-Nyquist artifacts that cannot be filtered
out afterward.




Rodney Myrvaagnes J36 Gjo/a


"In this house we _obey_ the laws of thermodynamics." --Homer Simpson

 

Anyone else like the irony of George posting this about ANYBOBY.

 

Err, it's not Sweaty Betty.

 

As long as it doesn't land in the gaps between the sensors....

 

Yep, the bayers kick the aqss of the foveon again. Not too bad ISO 1600 10D
shot, but I spose you hoped we wouldn't notice that and assume it was ISO
100.

 

As has been pointed out *many* times, the lack of an anti-aliasing
filter introduces aliasing artifacts which can be much worse than just
jaggs - e.g. low frequency signals created by the sampling process
which cannot be filtered afterwards.

/JB

 

I just tried to be nice to George

You are entirely correct BTW.

My point was that George's reasoning was theoretically sound.
The Bayer filters has the theoretical problem that he points out.
In practice, this theoretical problem is not all that large.
In practice, the SD9 does not have a perfect RGB sensor either.


Roland

 

Not really.

The SD9 artefacts cannot be removed.

The Canon et al "blurriness" can be improved, using sharpening.
For some reason, Phil at DPReview chooses not to apply any
post processing before publishing pictures. This is not fair to
cameras where you are assumed to do post processing for best result.


Roland

 

There you go again, Steve, thinking that you and the 'whole world' are
synonymous. Didn't you learn anything when we schooled you in
comp.sys.amiga.advocacy?

 

In message <>,

The texture of entire objects are screwed up by the SD9; not just edges.
Look at any distant trees, or brick walls, or shingled roofs; if they
are in focus, they have false texture.
--

 

That's true that the 10D AA filter has been set to remove most of the
chromatic artefacts, but not all. And at the expense of the
resolution, as you point out. This is not not the case for most of the
Bayer cameras: generally, vendors prefer to have a higher resolution
and more chromatic artefacts.

No, you make a confusion between aliasing and stair effects. True
aliasing results in *low* frequency artefacts: there are very probably
some SD9 images with such artefacts, but most of time I don't see such
ones.

What is always visible on SD9 images are the stair effects along sharp
edges which are not perfectly horizontal or vertical. I agree that
this is not very esthetic, but it is only a perceptual problem, which
is the consequence of coding high frequencies with a reduced amount of
pixels. Try upsampling the SD9 images to 6Mpix: if done properly the
upsampling doesn't destroy any information, and it removes most of the
stair effects.

 

(pehache) wrote in

Hmmmm ... yes it is true that aliasing results in low
frequency artefacts. But - it is not as easy as that.
You also have to take the phase into account. Phase errors
(due to aliasing) will give stair effects and other
funny stuff that affects only some few pixels.


Roland

 

No. Still the same confusion.

A frequency is either aliased (resulting in a low frequency
artefacts), or unaliased. This is completely on/off effect: if
unaliased (lower than Nyquist), it is perfectly coded from the
information theory point of view, including its phase.

The point is that it is not because a frequency is perfectly coded
that it will be nicely displayed. Frequencies close to Nyquist (but
still inferior to) are actually very poorly *displayed*, this is a
known effect in signal processing, resulting in stair effects and
other funny things as you said. Usually, it is recommended, to have
nice displays, to have a Nyquist at least twice the dominant frequency
of the signal. This can require upsampling prior display.

That's basically why the Sigma sofware offers the possibility to
output upsampled files. IMO this should be the default behavior. It
does not add any information, but the files are much nicer to observe.

The upsampling can be done afterwards, but to upsample properly
frequencies close to Nyquist, one need a very good algorithm (bicubic
may be not enough efficient). It is probably better to make the
upsampling stage inside the Sigma software (I assume that they have
chosen an appropriate algorithm).