SD9 Resolution prediction

[Peter G]

Thanks Mark,

This is what I have been witnessing, thanks for posting the theory behind what I had seen in practical reality.

Peter

Imagine a single wire stretched across a piece of white paper..
Taken with a Bayer CFA sensor the wire would be at least two pixels
wide all the way across the image (plus some alias to keep the line
smooth). With an X3 sensor the line would be one pixel wide.

Not exactly correct; if it were, the theoretical limit of a Bayer
sensor would be 1/2 the one-dimensional nominal resolution, whereas
in fact it approaches 1/sqrt(2).

In two dimensions, Bayer's resolution approaches 1/2 the nominal
resolution, rather than 1/4 as this post infers.

This is because the 2x2 averaging required by Bayer is done on
overlapping groups; each pixel's output is used in the calculation
of four averages. If each sensor were only used in the calculation
of a single average (i.e. the 2x2 groups were non-overlapping) this
post would be correct.

Because of the overlapping 2x2 groups, each output pixel has a
correlation of 0.50 with the value of each adjacent output pixel.
Since the input array and the output array have the same number of
elements, the 0.50 correlation indicates that a certain amount of
information has been discarded.

Although Bayer would never correctly render an array of evenly
spaced 1-pixel wide lines, the minimum spacing isn't quite as wide
as 2.0 pixels. That is to say, there would be (at the extreme
limit of resolution) a great deal of aliasing, but a Bayer sensor
is able to resolve an arbitrarily oriented grating where each line
approaches 1.414 pixels in width.

 

[Ken Phillips]

Dana,

You must be a young one! The Intel powered PC is more than twenty years old, and much of the software written for it (on the 8088, DOS 1.0) will still run on a P4 XP machine! WAY before the Motorola PowerPC/DEC Alpha/IBM RISC stuff. (At a local USAF base, we've replaced several RISC machines with PCs!)
I live with only 4.1 megapixels in my 1D ... which I can do, because I'm old!
Ken

--

All kinds of old camera and motion picture bodies, lenses, tripods, enlargers, mostly gathering dust, because digital is immediate! NO Canon 1200mm f/5.6.

 

[Veniamin Kostitsin II]

Great series of posts Don, thank you. that kind of posts became a rarity lately ...

1/sqrt(2) is a good approximation for the lumina, however we still have to remember the colour part of the equasion, which makes life a bit more complicated, espacially when we come to contrast-rich patterns which might introduce moire as well which will give the foveon imager yet another advantage in this - not very lifelike - situation.

on the gut level i would have estimated 1.5 performance factor for the foveon as well. another thing to consider is needs of the BW photographers - foveon would be quite interesting for those who like to use different RGB channels for the BW work. my hope wouuld be, that foveon would provide images with cleaner channels and less noise - but 3 MP foveon even with a 1.5x quality bonus won't cut it resolution-wise ... unfortunately.

regards

veniamin kostitsin
http://www.digitalimage.at

i wonder, how foveon would help me here ... ;-)

What if it were a black line running over a red, green or blue
background?

White is too easy

Yeah, color complicates it, as does (I think) contrast. You'll
notice in my example that the almost-single-line detail worked a
lot better in example1 where it was basically grey, and low
contrast.

I would love to have a fovean sensor. I just am not ready to accept
the compromises that the Sigma entails. And as resolution continues
to climb I'm not sure it'll matter any more from a practical
standpoint. History is rife with examples of superior technology
that missed it's window of opportunity.

• DL

--
Veniamin Kostitsin II
http://www.digitalimage.at/

 

[zettlers #105592]

This is turning from a pro photographers forum to a geek forum. You guys need a space of Your own. Make it happen Phil!

Regards
Leifilund

 

[John Burns]

It's that Askey guy going all techy on us. Why don't you E-mail Phil and see if you can get him banned. God forbid a pro might actually learn how all those pretty images end up on their memory card.

This is turning from a pro photographers forum to a geek forum. You
guys need a space of Your own. Make it happen Phil!

Regards
Leifilund

--
John

 

[Anders Borum]

What a great (and tack sharp) image. I'm impressed!

--
with regards
anders lundholm · [email protected]
the sphereworx / monoliner experience

 

[Phil Askey]

It was a D60 image reduced in size by 50% (that's a 6 MP image down to 1.5 MP).. it should be sharp!

What a great (and tack sharp) image. I'm impressed!

--
with regards
anders lundholm · [email protected]
the sphereworx / monoliner experience

--
Phil Askey
Editor / Owner, dpreview.com

 

[Phil Askey]

No, this is wrong. I resized the 220 x 150 down to 156 x 106 and then back up again to 220 x 150 and then up to 440 x 300 nearest neighbour to show a 'zoomed' view.

My interpolation of the same thing, but using Mitchell. No
sharpening or any processing done other than a resize to 312x213
then back to 440x300, just like you said you did. Of course I used
your JPEG image, did the processing, then recompressed, so there
may be a few minor losses there.

 

[Jimmy Chen]

Agreed, but how do you represent 1.414 pixels?

Huh? We're talking frequency here. One inscribes lines that are
1.414 times as wide as the pixel pitch taking into account the lens
system magnification.

I think the question is how to you display the imformation of 1.414 pixels when each pixel can only stores one piece of information. If we stay with b/w, how do you display 0.414 pixel being black and the other 0.586 pixel being white as any shade pf gray would have to use up the whole pixel.

--
jc
Sony F707
http://www.reefkeepers.org/gallery/f707
http://www.reeftec.com/gallery

 

[Anders Borum]

Hello!

It was a D60 image reduced in size by 50% (that's a 6 MP image down
to 1.5 MP).. it should be sharp!

I thought it was a straight-from-the-camera EOS 1D JPEG?! If the case is, as you states, that the image in fact was a 50% reduced D60 image, then yes, it should be sharp!

--
with regards
anders lundholm · [email protected]
the sphereworx / monoliner experience

 

[Z (is real)]

Nothing to do with it.

I'm not sure what this means, but you might want to read Alvy Ray Smith's paper "A Pixel Is Not A Little Square, ..." :
http://www.cs.princeton.edu/courses/archive/fall00/cs426/papers/smith95b.pdf

Consider this exercise: If you have a field of lines spaced at 1.414 times the Nyquist frequency (implied by the sensor) sampled with a Bayer pattern sensor, one should be able to downsample (resize) the image by a factor of 1.414 and get relatively crisp "1 pixel" lines. (Without downsampling, it should still look reasonable on screen, but if shown 1:1, the lines will be varying greys.) If the spacing gets any tighter, replacing 1.414 with a smaller number above, the result will have artifacts. This is of course a theoretical result (e.g. it likely depends on ideal reconstruction algorithms. Orientation can be figured into the above as well by adding a rotation into the resampling operation.)

But hey, I could be wrong or missing the point.

-Z-

 

[tim martin]

Phil,
If you join in the argument you will lay yourself open to shouts of
"bias" when you do a real test on the Sigma if it comes out better
than Canon/Nikon etc
Better keep to the side lines and let time win or loose the
argument for you.
Looking forward to your indepth reviews on all the new cameras.

This would only be true if a person was incapable of saying "I was
wrong".

Phil can say whatever he wants now, and later he can either say
"I was right", or "I was wrong". No big issue here.

And beyond that, my experience with seeing what Phil often has
to say is that he's often right. At least as far as digital photography
is concerned.

 

[Phil Askey]

You must have got confused somewhere.. This text was RIGHT ABOVE the crops:

"This is ONLY a simulation, the source image is actually a downsampled D60 image."

It was a D60 image reduced in size by 50% (that's a 6 MP image down
to 1.5 MP).. it should be sharp!

I thought it was a straight-from-the-camera EOS 1D JPEG?! If the
case is, as you states, that the image in fact was a 50% reduced
D60 image, then yes, it should be sharp!

--
with regards
anders lundholm · [email protected]
the sphereworx / monoliner experience

--
Phil Askey
Editor / Owner, dpreview.com

 

[Phil Askey]

Exactly. Which then effects surrounding pixels.

My point being that with an X3 sensor there's no need to 'disturb' surrounding pixels.

Agreed, but how do you represent 1.414 pixels?

Huh? We're talking frequency here. One inscribes lines that are
1.414 times as wide as the pixel pitch taking into account the lens
system magnification.

I think the question is how to you display the imformation of 1.414
pixels when each pixel can only stores one piece of information. If
we stay with b/w, how do you display 0.414 pixel being black and
the other 0.586 pixel being white as any shade pf gray would have
to use up the whole pixel.

--
jc
Sony F707
http://www.reefkeepers.org/gallery/f707
http://www.reeftec.com/gallery

--
Phil Askey
Editor / Owner, dpreview.com

 

[Don Lashier]

Phil, I think Anders was referring to the image that I posted ("examples").

• DL

You must have got confused somewhere.. This text was RIGHT ABOVE
the crops:

"This is ONLY a simulation, the source image is actually a
downsampled D60 image."

 

[Dana Jacobsen]

No, this is wrong. I resized the 220 x 150 down to 156 x 106 and
then back up again to 220 x 150 and then up to 440 x 300 nearest
neighbour to show a 'zoomed' view.

Ah, I see. I didn't read your text as saying that, but I see what is going on now, thanks. We certainly live in interesting times, don't we!

 

[Don Lashier]

I think the question is how to you display the imformation of 1.414
pixels when each pixel can only stores one piece of information. If
we stay with b/w, how do you display 0.414 pixel being black and
the other 0.586 pixel being white as any shade pf gray would have
to use up the whole pixel.

Here's three lines, the top one is 1 pixel wide and the bottom one is 2 pixels wide. How did I get the center one and how many pixels wide is it? To better illustrate it I rotated the image in PS after creating the lines.

• DL

 

[Dana Jacobsen]

Dana,
You must be a young one! The Intel powered PC is more than twenty
years old, and much of the software written for it (on the 8088,
DOS 1.0) will still run on a P4 XP machine! WAY before the Motorola
PowerPC/DEC Alpha/IBM RISC stuff. (At a local USAF base, we've
replaced several RISC machines with PCs!)
I live with only 4.1 megapixels in my 1D ... which I can do,
because I'm old!
Ken

I didn't mean before Intel! I was already into computers when the 8088 came out though (in high school I think). I don't recall much frenzy about 68k vs. x86 regarding MHz, but I do remember it in the workstations. Then Intel started playing in earnest, and AMD's competition has really spurred them on. Remember, before the Pentium, x86 machines were pretty pathetic in floating point, making them not a player at all in the workstation market. Well, yes, I had a 486 I did software development on, but it was over 2x slower on FP than the Sparcs, whcih were no speed demons compared to the HP and IBM machines. Cheap though.

My point was actually that people went over this whole MHz rules thing before Intel started playing that game in the consumer market. Not before Intel started making CPUs! (which would be before Sun was incorporated, before MIPS, etc.).

 

[Don Lashier]

Ken, I think Dana's point (which I made earlier also) is that superior technology frequently doesn't win out over brute force. I wouldn't be surprised but what we see the same thing with foveon vs bayer. This would be unfortunate.

• DL

Dana,
You must be a young one! The Intel powered PC is more than twenty
years old, and much of the software written for it (on the 8088,
DOS 1.0) will still run on a P4 XP machine! WAY before the Motorola
PowerPC/DEC Alpha/IBM RISC stuff. (At a local USAF base, we've
replaced several RISC machines with PCs!)

 

[u19925]

Resolution depends on lots of things besides pixels. True that at 3 MP, you are limited by pixel count, but at 6 MP+, you will also have to think about:

1) Lens quality
2) Camera shake
3) Bayer CCD vs Foveon.
3) In extreme cases, even the heat waves in the atmosphere.

When you look at all these factors, it turns out, there isn't much point going beyond 6-7 MP for P/S digital cameras and 10-12 MP for SLRs.

So, once you start looking at Foveon next generation at 6MP or more, you won't be able to beat its resolution just by increasing pixel count and if the Foveon claims are true and if it scales to 6MP+, then either Canon/Nikon etc will follow the suit or they will have to come up with some radically different technology.

Once resolution is not a factor, Foveon wins hands down (assuming their published claims are true).

Note on color bandwidth: While it is true that objects have more luminance bandwidth than chrominance, it is not as easy to exploit this for photos as it is for video (it is much easier to exploit for video, since in video, the scenes are changing fast and your eye cannot feel the difference. in photos, you can concentrate on a part of photo and can detect the fault much easily).

--dhiraj

I was just over in the "News Discussion" forum, and I can't quite
believe the arguments for/against the Foveon sensor.

Believe me, I am just as anxious for the state of the art in sensor
design to advance as the next guy; and I am VERY interested in
seeing indepentend tests of the SD9... but when I see messages such
as "it's the same as a 10Mp Bayer CCD for resolution" I just groan.

Here is a solid prediction:

The SD9 will have:

Horizontal LPH
Vertical LPH

Why?

Because there is NO way around Nyquest... or, to explain it to
those readers who never heard of Nyquest... think of alternating
black & white lines. Even if they are perfectly aligned with the
photo receptors it will take one photo receptor to read "white" and
one to read "black".. so you CANNOT resolve (on a rectangular
layout) more than 1/2 the number of line pairs as you have photo
receptors on that axis.

It is true that if you are trying to resolve Green/Black, Red/Black
or Blue/Black lines the bayer sensor will be at a disadvantage;
however as colors are rarely pure primary colors the de-bayering
algorithm can compensate to a certain extent.

If the Foveon proponents merely pointed out that the Foveon will
have a more accurate spatial color resolution, I'd have no argument

• but arguing that a 3Mp Foveon sensor will resolve as much detail

as a 6Mp Bayer is a MUCH weaker argument.

If you are interested in the LPH/V of S2/D100/D60 look at Phil's S2
review at the bottom of
http://www.dpreview.com/reviews/fujis2pro/page19.asp

Now a 6Mp Foveon sensor capable of ISO1600 would be a wonderful
thing... but by the time that arrives, Nikon/Canon/Fuji will be
shipping 12Mp-16Mp bayer sensors.

You just HAVE to love a tech "arms" race

Personally, I'd rather be out there shooting than arguing the
merits of existing sensor tech vs. yet unreleased sensor tech. The
fact is any of the current DSLR's are capable of producing great
results - so enjoy photography!

Best Regards,

Bill

--
http://whphotography.com/
http://www.cpureview.com/