Updated "Bayerized" tests w/AA filter

Started Oct 2, 2002 | Discussions
Peter G
Veteran MemberPosts: 5,944
Like?
Re: I have now retuned the AA filter
In reply to Thomas Ko, Oct 3, 2002

This is an approximation. His filter is an estimate of what the AA filter does. Nikons seem to be more conservative and seldom show as much moire as the Canons (and tend to looks softer):

Nikon:

Mikes Simulation:

Also while his Bayer interpolation algorithm is the equal of the Canon one, there are still likely some differences.

A cool bit of simulation work for the Foveon feeding frenzy.

Peter

Thomas Ko wrote:
This is excellence!
I really appreciate your effort!!

Looks almost the same (expect there is no moire color on the foveon
bayer). But how come there are no moire pattern? is there any
trick?

Mike Chaney wrote:

Cameras vary widely in their AA filters, yours seems a bit aggressive.

Could you do this clip for a more controlled test:

Since this is equivalent to a 6MP (because of the square
orientation) and because we know what 6mp resolution clips look
like, it can be used to tune or asses your comparson.

If this is a comparison, I would like to see it on something I am
very familiar with.

Peter, your idea was a good one! By using the above technique, I
did end up toning down the AA filter slightly. I simply changed
some of the pixel weightings to simulate a filter with slightly
less blur (smaller radius). I should note, however, that it didn't
change the overall "look" of the Bayer-with-AA samples very much at
all. I have updated all the Bayer-with-AA samples on the page to
reflect the new tuning though, and the new samples do look a hint
sharper with a little more detail.

In the process of performing this exercise, I can now say that I at
least have a loose proof of concept because that small crop does
match what a real Bayer sensor resolves pretty closely. See the
"Update" near the top of the page for more info.

http://www.ddisoftware.com/reviews/sd9-v-bayer

-- hide signature --
-- hide signature --

C.Wolf

Reply   Reply with quote   Complain
pocoapoco
Junior MemberPosts: 38
Like?
A theoretical question.
In reply to Mike Chaney, Oct 4, 2002

Suppose you had a picture that looked like a bayerpattern image taken of a pure white, no detail, flat subject. And you had some joe pro lens that nobody else has with perfect sharpness and absolutely 0% distortion and color aberation. And you had a bayer sensor camera to attach it to. You then shot a picture of this bayer pattern at a distance such that the colored squares on the picture you were shooting mathced the resolution of the camera's photosites exactly, but you shifted the camera one pixel width to the right or left. Would you get a perfectly black image doing this? I'm not trying to proveanything for or against any sensor technology. I'm only curious.

Also, Mike, do you think that given equal size raw image files, would it be better to have a cirtain resolution with full color at each pixel, or three times the resolution with only a third of the color information at each pixel? Nothing else considered like price of course.

Reply   Reply with quote   Complain
Kok Chen
Contributing MemberPosts: 740
Like?
Re: A theoretical answer.
In reply to pocoapoco, Oct 4, 2002

Assume the spectral resonse of your target matches the spectral response of the filter array AND there is no overlap in the response curves for R, G and B, AND, if you do not have an anti-aliasing filter, the result would be black.

The only qualifyer I did not include above is a tooth fairy :-).

  • kc

Reply   Reply with quote   Complain
Stephen Waits
Senior MemberPosts: 2,093
Like?
Re: Updated "Bayerized" tests w/AA filter
In reply to Jimmy Chen, Oct 4, 2002

Jimmy Chen wrote:

Some people say that 3MP is going to be about as useful as 4.5 to
5MP, others say 6, and some more aggressive people think 7! Then
there's some crazies saying 9, 10, 11, and I've even seen 14 [wow,
some people aren't "thinkers"].

LAMO, if you REALLY want a laught, you need to read the latest from
Michael Long . He thinks because the Bayer method, with one pixel
using surround 8 pixles, a 6mp Bayer CFA is really 54MP.
http://www.dpreview.com/forums/read.asp?forum=1000&message=3497637

That is hilarious! Actually, in Michael's defense if you read the post closely you'll see that he's correcting someone else's logic.

-- hide signature --

Steve

Reply   Reply with quote   Complain
Eric Eilebrecht
New MemberPosts: 11
Like?
Re: Updated
In reply to Mike Chaney, Oct 4, 2002

I am absolutely fascinated by the work you have done on this. Great stuff. Like everyone else here, I only wish we could have some way to compare the 3.5mp SD9 image with, say, the exact same image taken with a D60. Some people have suggested upsampling the image before "bayerizing" it, which you have rightly rejected. I wonder, though, if you might be able to get meaningful results by down*sampling the SD9 image.

What I mean is this: take the original SD9 image and downsample to 1.75 mp. Imagine the result is a 50% crop of a larger image from the SD9. Now scale it up to 3.5mp again. Compare the result to the bayerized original, which is also 3.5mp, which we can pretend is a 50% crop of an image from a D60. Viola!

Does that make any sense, or have I had too much to drink?

Eric

Reply   Reply with quote   Complain
dgrogers
Veteran MemberPosts: 6,650
Like?
More questions for Mike
In reply to Mike Chaney, Oct 4, 2002

Hi Mike, you obviously know what you're talking about so I have a couple of questions for you. Sorry if they were asked before, I may have missed a few threeads.

1) Isn't the bayer process in your examples at an unfair dissadvantage because it's interpolating an image taken with the SD9 versus the actual scene? Did you compensate for this?

2) How would the results differ, if at all, if you used Fuji's honeycomb method on the SD9 samples instead of the traditional bayer pattern? I would really like to know how effective Fuji's method is on real photos (versus resolution charts). What disadvantages does it have?

Thanks.

Mike Chaney wrote:

Taking Phils advice, and the advice of a few others, I have updated
my page to include samples that simulate a Bayer sensor behind an
AA filter.

http://www.ddisoftware.com/reviews/sd9-v-bayer/

-- hide signature --
-- hide signature --
Reply   Reply with quote   Complain
Ladehaug
Senior MemberPosts: 1,840
Like?
Re: Updated "Bayerized" tests w/AA filter
In reply to Mike Chaney, Oct 4, 2002

Mike Chaney wrote:

Taking Phils advice, and the advice of a few others, I have updated
my page to include samples that simulate a Bayer sensor behind an
AA filter.

http://www.ddisoftware.com/reviews/sd9-v-bayer/

-- hide signature --

You have done an excellent job showing the "bayerizing" effect! I'll not ask you have many hours you have spent working with this!

How about a simulated resolution chart? As my job is a CAD specialist, I have the possibility to draw high resolution technical drawings using AutoCAD. I have put together some elements on a sample chart:

The original tiff image is actually 12000x8000 pixels, that's enough to make a 600dpi print on 20 x 13.3 (51 x 34 cm). Here is a crop of the original:

http://home.online.no/~ladehaug/Foveon/res-sample-l1.gif

Geir

Reply   Reply with quote   Complain
Mike Chaney
Senior MemberPosts: 2,826
Like?
Re: Updated
In reply to Eric Eilebrecht, Oct 4, 2002

That might be an interesting test. The only problem I see with that one is that to simulate a 6MP sensor, you'd have to downsample by non-integer values which means that you start dealing with partial pixels and the efficiency of the downsampling interpolation algorithm.

-- hide signature --
Reply   Reply with quote   Complain
Mike Chaney
Senior MemberPosts: 2,826
Like?
Re: More questions for Mike
In reply to dgrogers, Oct 4, 2002

1) Isn't the bayer process in your examples at an unfair
dissadvantage because it's interpolating an image taken with the
SD9 versus the actual scene? Did you compensate for this?

I thought of this, but the differences should be slight. The main difference is probably in the relationship between the SD9's AA filter (if it even has one) and a Bayer sensor's AA filter. I think I took care of most of that correlation by tuning the AA filter to a known Bayer camera (D60).

2) How would the results differ, if at all, if you used Fuji's
honeycomb method on the SD9 samples instead of the traditional
bayer pattern? I would really like to know how effective Fuji's
method is on real photos (versus resolution charts). What
disadvantages does it have?

That starts to get a little more complicated. Technically, a Super CCD captures exactly the same amount of information as a Bayer CCD with the same number of pixels. The difference is that the Super CCD has its pattern oriented to maximize horizontal/vertical resolution while a typical Bayer sensor actually has better diagonal (45 degree) resolution than a Super CCD. In addition (and partly due to its focus on horizontal/vertical resolution), a Super CCD produces images that are very receptive to a 2x expansion (interpolation) in both X/Y directions, resulting in a sharp 4x resolution image.

Since our eyes (brain really) focuses more on horizontal/vertical lines than diagonal lines, all this means is that the visual advantage of an X3 sensor over a Super CCD will be less than for X3 versus Bayer. In other words, if (for a particular image) an X3 image appears 1.5 times as detailed as its Bayer counterpart, it may only look 1.2 times as sharp as its Super CCD counterpart... just as an arbitrary example.

-- hide signature --
Reply   Reply with quote   Complain
Ulysses
Forum ProPosts: 39,576
Like?
Re: Updated "Bayerized" tests w/AA filter
In reply to Ladehaug, Oct 4, 2002

I was HOPING that someone would be doing this.

Is there any reason that something like this would not be of great value to Mike and/or Phil?

Clearly with the quality of the instruments being produced, we need new charts and new ways of studying the images. Something like what you've prepared may be leading to filling that need.

Excellent work, Ladehaug.

It may need some fine-tuning, but I'm sure you'll get some tips here.

Ladehaug wrote:

Mike Chaney wrote:

Taking Phils advice, and the advice of a few others, I have updated
my page to include samples that simulate a Bayer sensor behind an
AA filter.

http://www.ddisoftware.com/reviews/sd9-v-bayer/

-- hide signature --

You have done an excellent job showing the "bayerizing" effect!
I'll not ask you have many hours you have spent working with this!

How about a simulated resolution chart? As my job is a CAD
specialist, I have the possibility to draw high resolution
technical drawings using AutoCAD. I have put together some elements
on a sample chart:

http://home.online.no/~ladehaug/Foveon/res-sample-900x600.jpg

The original tiff image is actually 12000x8000 pixels, that's
enough to make a 600dpi print on 20 x 13.3 (51 x 34 cm). Here is a
crop of the original:

http://home.online.no/~ladehaug/Foveon/res-sample-l1.gif

Geir

-- hide signature --

Ulysses

Reply   Reply with quote   Complain
Michael Long
Veteran MemberPosts: 4,448Gear list
Like?
Re: Updated "Bayerized" tests w/AA filter
In reply to Stephen Waits, Oct 4, 2002

That is hilarious! Actually, in Michael's defense if you read the
post closely you'll see that he's correcting someone else's logic.

Thanks, but I've come to realize that Jimmy sees what he wants to see and often takes things completely out of context.

 Michael Long's gear list:Michael Long's gear list
Sony a6000 Sony E 10-18mm F4 OSS Sony E PZ 18-105mm F4 G OSS Sony FE 70-200 F4
Reply   Reply with quote   Complain
Michael Long
Veteran MemberPosts: 4,448Gear list
Like?
Re: Updated "Bayerized" tests w/AA filter
In reply to Mike Chaney, Oct 4, 2002

I'd be interested in knowing what algorithm was used to "bayerize" the image. Many of the most sophisticated bayer algorithms do some pretty heavy processing to increase edge detection.

The underexposure of the D60 chart does, however, cause some issues with contrast, as black on white will definitely come off better than dark gray on gray.

I took the resolution chart crops into PS and played with them a bit. The differences are a bit less apparent if you reduce the contrast of the SD9 image to match that of the actual D60 sample, or increase the black/white points of the D60 to match those of the SD9, or manipulate both to meet in the middle.

Good work.

(Still waiting for a Canon 24MP full-frame Foveon 1Dsx though....)

 Michael Long's gear list:Michael Long's gear list
Sony a6000 Sony E 10-18mm F4 OSS Sony E PZ 18-105mm F4 G OSS Sony FE 70-200 F4
Reply   Reply with quote   Complain
Erik Magnuson
Forum ProPosts: 11,623Gear list
Like?
Re: Updated "Bayerized" tests w/AA filter
In reply to Michael Long, Oct 4, 2002

Michael Long wrote:

I'd be interested in knowing what algorithm was used to "bayerize"
the image.

Mike answered this in the non AA thread:

http://www.dpreview.com/forums/read.asp?forum=1027&message=3489574

-- hide signature --

Erik

 Erik Magnuson's gear list:Erik Magnuson's gear list
Canon EOS 5D Mark II Canon EOS 450D Sigma SD10 Sony Alpha NEX-5 Nikon D3200 +24 more
Reply   Reply with quote   Complain
Eric Eilebrecht
New MemberPosts: 11
Like?
Re: Updated
In reply to Mike Chaney, Oct 5, 2002

Understood. Tried it anyway. The two images (simulated 50% crops of 3.5mp Foveon and 7mp Bayer images, using the sample of the woman sitting on the stairs) were almost indistinguishable when printed at 8x10. If anything, the SD9 simulation was slightly better (especially in the shoe and watch detail), but it wasn't anything I'd probably notice if I wasn't looking REALLY close. Both images looked great to my eyes.

I used the standard bicubic algorithm in Photoshop to scale both ways. If anyone wants to try this themselves, just resize the original sample image to 1069x1604, and then print it and Mike's "Bayerized" image at equal sizes and compare.

Eric

Mike Chaney wrote:

That might be an interesting test. The only problem I see with
that one is that to simulate a 6MP sensor, you'd have to downsample
by non-integer values which means that you start dealing with
partial pixels and the efficiency of the downsampling interpolation
algorithm.

-- hide signature --
Reply   Reply with quote   Complain
Keyboard shortcuts:
FForum MMy threads