Ad
Ad
Ad
Pages: [1] 2 »   Bottom of Page
Print
Author Topic: Fake detail, on a feather  (Read 5499 times)
ErikKaffehr
Sr. Member
****
Offline Offline

Posts: 7503


WWW
« on: July 14, 2013, 02:29:49 AM »
ReplyReply

Hi,

According to signal processing theory, we would get aliasing on any structures that were transferred with significant contrast (MTF) past the resolution of the sensor. This is mostly seen as "moiré". Colour moiré is most obvious and is normally removed by local desaturation of the offending colour pattern. In absence of colour moiré we often see fake detail, that may actually enhance the image by giving impression of detail, that is not actually there.

This is easy to reproduce on artificial subjects, but may not be that often seen on natural subject.

I took a feather I found on the seashore and took two photographs of it at 3.5 m subject distance with my Hasselblad 555 using Planar 80/2.8 and Sonnar 150/4, both at f/5.6. Focusing was as accurate I could make it using a Zeiss 3X extender on my PME-5 (about 9X image magnification).

Both images show artificial detail, clearly visible on first try! ;-)

Note cross hatch pattern on the left side image.


Note strains bending down on the right side image, they are straight in reality.



Full crops are here:
http://echophoto.dnsalias.net/ekr/Articles/MFDJourney/FakeDetail/20130714-CF043486.jpg

http://echophoto.dnsalias.net/ekr/Articles/MFDJourney/FakeDetail/20130714-CF043488.jpg

Best regards
Erik

« Last Edit: July 14, 2013, 02:32:19 AM by ErikKaffehr » Logged

BernardLanguillier
Sr. Member
****
Offline Offline

Posts: 8085



WWW
« Reply #1 on: July 14, 2013, 03:52:14 AM »
ReplyReply

Erik,

Yes, AA filterless sensors are known to produce moire and various other artifacts.

Cheers,
Bernard
Logged

A few images online here!
ErikKaffehr
Sr. Member
****
Offline Offline

Posts: 7503


WWW
« Reply #2 on: July 14, 2013, 04:16:37 AM »
ReplyReply

Hi,

Yes, it is nice to be able to demonstrate it. First try... doesn't take a lot of effort. ;-)

Best regards
Erik

Erik,

Yes, AA filterless sensors are known to produce moire and various other artifacts.

Cheers,
Bernard

Logged

BartvanderWolf
Sr. Member
****
Online Online

Posts: 3684


« Reply #3 on: July 14, 2013, 05:29:43 AM »
ReplyReply

Hi,

According to signal processing theory, we would get aliasing on any structures that were transferred with significant contrast (MTF) past the resolution of the sensor. This is mostly seen as "moiré". Colour moiré is most obvious and is normally removed by local desaturation of the offending colour pattern. In absence of colour moiré we often see fake detail, that may actually enhance the image by giving impression of detail, that is not actually there.

This is easy to reproduce on artificial subjects, but may not be that often seen on natural subject.

Hi Erik,

Thanks for the example/demonstration. Thank goodness the theory is not just theory, it can actually happen ... Wink

There are of course some requirements that need to be fulfilled for it to happen, and knowing them may allow to reduce the risk of aliasing showing up when we can least use it.

  • 1. The subject matter must have finer detail than the sensor can resolve, and it must be of adequate contrast. Unfortunately, part of the problem is that the aliases may have higher contrast than the micro-detail that causes it. We can reduce the risk of aliasing by choosing a large enough magnification factor, shoot closer with a given focal length or use a longer focal length from a given distance. Of course the angle of view may prohibit us from solving the issue that way.
    This is also one of the reasons that MF cameras, with a longer focal length  and larger image circle to cover a larger sensor, in general have fewer problems (but are not exempt) from aliasing despite the absence of an AA-filter, than smaller sensor cameras without such an Optical Low-pass filter (OLPF) an with a similar sensel pitch. A longer focal length together with a larger sensor, allow a higher magnification factor for a given FOV. That also suggests why shooting with a longer focal length and then stitching for a larger FOV, will produce a higher resolution that can help in reducing aliasing risks.
    Sometimes rotating the sensor can help to avoid aliasing, because the two sampling patterns do not align, or because the higher diagonal resolution better resolves the finest detail in that specific orientation.
  • 2. The detail must be sampled at regular intervals to produce aliasing artifacts. Then it most clearly shows in repetitive structures, although it also happens on all other fine detail, but we may not be able to notice it that clearly. Human vision is always trying to find patterns, to reduce having to process too many visual stimuli. That's why disruptions of predictable patterns stand out like a sore thumb. So anything that disrupts the repetitive nature of the subject matter can reduce the risk of aliasing becoming visible. An example could be atmospheric turbulence, or subject vibration during the exposure time.
  • 3. It can only manifest itself when the micro-detail has adequate contrast to be quantified by the sensor when it reaches the sensor. As resolution approaches the limiting resolution, and ultimately the Nyquist frequency, the contrast of high spatial frequency detail drops. Therefore, the subject, and the atmosphere, and the lens, must offer enough contrast and resolution (MTF) to allow aliasing to happen.
  • 4. Diffraction reduces the micro-detail contrast. Ultimately, at narrow apertures, it will kill any possible micro-detail resolution (and thus micro-detail causing aliasing) because there is not enough contrast left to be recorded by the senor's sampling density.
    The absolute physical resolution limit caused by diffraction only, for a perfectly circular aperture, is at: cycles/mm = 1 / (N x wavelength), where N is the F-number, e.g. 16, and the wavelength is in millimeters, e.g. 0.000555 mm (=555 nanometers, green light). A sensor with a sensel pitch of 6 micron (0.006mm) will have a Nyquist frequency of 1 / (2 x 0.006 mm) = 83.3 cy/mm, so we can calculate that it is impossible to produce green light aliasing with such a sensor with an F-number of (2 x 0.006) / 0.000555 = 21.6 or narrower,  (2 x senselpitch) / wavelength = f/#.
  • 5. In practice, because the lens aberrations are not perfectly corrected, there will already be some loss of contrast. Therefore the absolute physical resolution limit set by diffraction will not even be reached, although it may be approached by very good lenses. This also assumes that no anti-aliasing filters are used.
  • 6. Defocus also reduces the contrast of micro-detail. It can therefore prevent aliasing except for in the exact plane of focus. The further away from that plane of focus, the larger the blur, and the less aliasing will manifest itself. One could therefore try and use the acceptable DOF zone away from the plane of best focus to reduce aliasing risks. That, combined with some diffraction, may be enough to prevent the manifestation of aliasing, while deconvolution Capture sharpening may restore some of the micro-contrast/resolution loss.

As always, knowing your enemy is the best remedy to cope with the situation and have success in the end. It may save some retouching time/cost, and nasty surprises with a deadline approaching.
 
Cheers,
Bart
« Last Edit: July 14, 2013, 12:01:37 PM by BartvanderWolf » Logged
ErikKaffehr
Sr. Member
****
Offline Offline

Posts: 7503


WWW
« Reply #4 on: July 14, 2013, 09:15:45 AM »
ReplyReply

Hi,

Sometimes they are hard to explain, check step chart in the image below.

What I speculate a bit is that sometimes detail we see are actually aliases, so aliases may enhance an image, giving for instance structure to fur or feathers.

I would speculate that the reason that I often (almost always) find color aliases in test images and seldom in real images is that test images tend to have more high contrast detail. Also I tend to stop down a lot for DoF.

Best regards
Erik


Hi Erik,

Thanks for the example/demonstration. Thank goodness the theory is not just theory, it can actually happen ... Wink

There are of course some requirements that need to be fulfilled for it to happen, and knowing them may allow to reduce the risk of aliasing showing up when we can least use it.


As always, knowing your enemy is the best remedy to cope with the situation and have success in the end. It may save some retouching time/cost, and nasty surprises with a deadline approaching.
 
Cheers,
Bart
Logged

BartvanderWolf
Sr. Member
****
Online Online

Posts: 3684


« Reply #5 on: July 14, 2013, 10:29:51 AM »
ReplyReply

Hi,

Sometimes they are hard to explain, check step chart in the image below.

Hi Erik,

Well, given that for this kind of moiré to exist, there have to be 2 regular sampling structures with different sampling frequencies and/or at an angle , one is the camera's sensor so the other must be a non-continuous tone print technique, a raster or half-tone screen print. They state on their website that the target has a true physical resolution of 1500 dpi, that is dots-per-inch. Unclear is whether they mean a printing process with that fine a raster, or that they mean a measured effective resolution of 1500 samples per inch (or PPI, pixels per inch). Either way, there apparently is a raster structure that exibits interference with the camera's sensor grid (maybe Stefan Steib, or someone else who has that chart can confirm if there is a visible printed raster pattern that triggers the moiré).

The false color moiré is caused by the different sampling densities of Red/Blue versus Green, caused by the Bayer CFA, hence different sized aliases for those colors.

There is an other general comment that can be made about that specific test chart, and that it is paradoxically not optimally suited for accurate testing of discrete sampling systems such as digital cameras, or scanners (unless they specifically wanted to create a problematic test surface, not something to quantify resolution with). The sharp edges will unavoidably result in aliasing artifacts, just like reproduction of text documents would (for which purpose this target would be a good stress test). That's for example why I designed my star test target for resolution testing with a sinusoidal radial grating, and not bi-tonal sectors.

Also the 1951 USAF five (or six) bar patterns (if you count black and white as separate bars) obviously stem from the analog film days, some 62 years ago, for which it was good, but not as well suited for regular discrete sampling devices. Also Imatest's creator Norman Koren explains that quantitative MTF tests on bi-tonal bar-patterns need to be corrected by a factor of approx. 78.5% because of the influence of aliasing from the sharp bar edges.

Here is an example from Norman's site which visually clearly shows how bi-tonal bars
mis-behave much more than sinusoidal patterns at the same spatial frequencies:

Quote
What I speculate a bit is that sometimes detail we see are actually aliases, so aliases may enhance an image, giving for instance structure to fur or feathers.

Only if they seemingly correspond with the pattern they cannot really resolve. But at an angle, aliases often produce curved patterns where straight patterns are to be expected. That can look un-natural.

Quote
I would speculate that the reason that I often (almost always) find color aliases in test images and seldom in real images is that test images tend to have more high contrast detail. Also I tend to stop down a lot for DoF.

Yes, most tests attempt to push to the limits of e.g. resolution. Because low contrast detail vanished first, one often uses higher contrasts to come closer to the real physical limit (the Nyquist frequency) beyond which no detail exists and only aliasing will be created. The false color aliasing is due to the differing sampling densities of the different colors of the Bayer CFA.

Cheers,
Bart
« Last Edit: July 14, 2013, 12:06:21 PM by BartvanderWolf » Logged
opgr
Sr. Member
****
Offline Offline

Posts: 1125


WWW
« Reply #6 on: July 14, 2013, 12:20:17 PM »
ReplyReply


The false color aliasing is due to the differing sampling densities locations of the different colors of the Bayer CFA.
Logged

Regards,
Oscar Rysdyk
theimagingfactory
eronald
Sr. Member
****
Offline Offline

Posts: 4058



« Reply #7 on: July 14, 2013, 12:43:03 PM »
ReplyReply

The artefacting results from frequencies present in the image beyond sampling limit getting folded back during the reconstitution process. I guess even if there were just strong noise at high frequencies ie. intense random noise one would see artefacting occur. And in fact this is probably what one calls chromatic noise.

Edmund
Logged
BartvanderWolf
Sr. Member
****
Online Online

Posts: 3684


« Reply #8 on: July 14, 2013, 01:33:42 PM »
ReplyReply

Hi Oscar,

Not really the locations. The demosaicing fills in the missing sampling locations, and as my earlier experiments have shown, the reconstructed Chroma resolution is very close for all three R/G/B channels to the actual Luminance resolution.

So sampling position is not the issue.

Diagonally every Green sensel again samples Green. That is a sampling density of 1.4x the horizontal/vertical sensel pitch. The Red and Blue filtered samples have a diagonal sampling distance that is twice as large as that of the Green sensels. The horizontal sampling density is the same for all 3 channels. It's the less dense sampling in the diagonal direction that is going to affect limiting resolution (Nyquist), and thus how soon aliasing will set in, and how large the aliases will be. Red and Blue will aliase faster and larger for the same level of diagonal detail.

Sampling density is the issue.

Cheers,
Bart
Logged
BartvanderWolf
Sr. Member
****
Online Online

Posts: 3684


« Reply #9 on: July 14, 2013, 01:37:24 PM »
ReplyReply

The artefacting results from frequencies present in the image beyond sampling limit getting folded back during the reconstitution process.

Hi Edmund,

Exactly. Red and Blue will create aliasing faster because of the lower sampling density, and thus also larger (lower spatial frequency, due to the fold-back).

On 45-degree rotated sensel layouts like Fuji used, the horizontal/vertical versus diagonal sampling switches in densities, but the same channel differences remain. In fact only the higher diagonal luminance resolution is traded for a higher horizontal/vertical resolution (which makes logical sense in a gravity driven environment). On the newer X-trans sensors the whole false color artifacting situation becomes worse.

Cheers,
Bart
« Last Edit: July 14, 2013, 01:43:28 PM by BartvanderWolf » Logged
opgr
Sr. Member
****
Offline Offline

Posts: 1125


WWW
« Reply #10 on: July 14, 2013, 02:18:36 PM »
ReplyReply


So sampling position is not the issue.

Red and Blue will aliase faster and larger for the same level of diagonal detail.

Sampling density is the issue.

No, sampling density is not the issue since the sampling density for Red and Blue is equal. However, sampling position is different and that is why they alias at different positions. That causes false color. Their density relative to green is utterly irrelevant, because for a decent demosaic algo the green is equal to the detail signal and should already have been subtracted from the color signals prior to reconstruction.

You should also be aware that this is not your normal sampling theorem problem: the sampling occurs disjunct, which specifically for AA filter-less sensors, makes the problem slightly different. It is the disjunct nature of the sampling that causes most of these artifacts.
 



 
Logged

Regards,
Oscar Rysdyk
theimagingfactory
ErikKaffehr
Sr. Member
****
Offline Offline

Posts: 7503


WWW
« Reply #11 on: July 14, 2013, 02:30:19 PM »
ReplyReply

Hi,

My main interest here is aliasing in general, not color aliasing in general. That is the reason I posted BW images. There are tools to reduce color moire but i don't think monochrome aliases can be removed.

Best regards
Erik



No, sampling density is not the issue since the sampling density for Red and Blue is equal. However, sampling position is different and that is why they alias at different positions. That causes false color. Their density relative to green is utterly irrelevant, because for a decent demosaic algo the green is equal to the detail signal and should already have been subtracted from the color signals prior to reconstruction.

You should also be aware that this is not your normal sampling theorem problem: the sampling occurs disjunct, which specifically for AA filter-less sensors, makes the problem slightly different. It is the disjunct nature of the sampling that causes most of these artifacts.
 



 
Logged

opgr
Sr. Member
****
Offline Offline

Posts: 1125


WWW
« Reply #12 on: July 14, 2013, 02:43:17 PM »
ReplyReply

My main interest here is aliasing in general, not color aliasing in general. That is the reason I posted BW images. There are tools to reduce color moire but i don't think monochrome aliases can be removed.

Yes they can, at the expense of detail, but you already knew that.
Logged

Regards,
Oscar Rysdyk
theimagingfactory
eronald
Sr. Member
****
Offline Offline

Posts: 4058



« Reply #13 on: July 14, 2013, 03:08:52 PM »
ReplyReply

Hi Edmund,

Exactly. Red and Blue will create aliasing faster because of the lower sampling density, and thus also larger (lower spatial frequency, due to the fold-back).

Agree.

Quote

On 45-degree rotated sensel layouts like Fuji used, the horizontal/vertical versus diagonal sampling switches in densities, but the same channel differences remain. In fact only the higher diagonal luminance resolution is traded for a higher horizontal/vertical resolution (which makes logical sense in a gravity driven environment). On the newer X-trans sensors the whole false color artifacting situation becomes worse.

Cheers,
Bart

At Photokina, the Fuji marketing guy held a press presentation in the presence of the engineers. He claimed no aliasing. I later asked an engineer "And what if we have a signal at 2x the nyquist limit". The engineer replied "That frequency is suppressed by the lens" Smiley

Edmund

Logged
eronald
Sr. Member
****
Offline Offline

Posts: 4058



« Reply #14 on: July 14, 2013, 04:29:26 PM »
ReplyReply

You won't be able to equalize the disparities between G and (R,B) aliasing because there is no easy way to resample in the digital domain ....

Edmund
Logged
BartvanderWolf
Sr. Member
****
Online Online

Posts: 3684


« Reply #15 on: July 14, 2013, 05:48:45 PM »
ReplyReply

At Photokina, the Fuji marketing guy held a press presentation in the presence of the engineers. He claimed no aliasing. I later asked an engineer "And what if we have a signal at 2x the nyquist limit". The engineer replied "That frequency is suppressed by the lens" Smiley

LOL

Surely the lens isn't that bad, or even exhibits such a sharp MTF cut-off. What can help to hide the aliasing is the more non-uniform sampling by the slightly more random sampling positions relative to uniform signal variations. That will result in random noise instead of more obvious aliases. Presumably the engineer didn't understand that concept.

Cheers,
Bart
Logged
BartvanderWolf
Sr. Member
****
Online Online

Posts: 3684


« Reply #16 on: July 14, 2013, 06:44:00 PM »
ReplyReply

No, sampling density is not the issue since the sampling density for Red and Blue is equal. However, sampling position is different and that is why they alias at different positions. That causes false color.

Hi Oscar,

All aliasing is caused by under-sampling of higher input signal resolution, not by the sampling position. Check out section 7.1.3 about aliasing in this excellent PDF document. Non-uniform sampling position differences can reduce the obvious aliasing by replacing it with noise, but that's the only position related aliasing effect.

Bayer CFA demosaicing is relatively simple for normal subject matter, especially with something as relatively monochrome as Erik's feather example, or the grayscale patches, or the fabric surface of his test chart example. Chroma information has usually much less high frequency detail, just look at a Lab mode image, or an HSL channel separation. It is therefore relatively simple to interpolate color accurately. It's not specifically sensitive to the sampling position.

Aliased information cannot be distinguished from real information because both are recorded together in the same sensel position. Therefore, only different aliasing amounts can cause these false color issues, and all that the demosaicing algorithms can do is iterative reduction of the local color differences where RB and G channel luminances significantly differ.

Luminance aliasing can not always be removed easily because it involves multiple different spatial frequencies, folded back from higher spatial frequencies from beyond Nyquist and they show up as multiple lower spatial frequency aliases. It requires elaborate reconstruction, by using the lesser aliased channels to replace the more aliased ones if amplitude reversal doesn't help enough.

Cheers,
Bart
Logged
wildlightphoto
Sr. Member
****
Offline Offline

Posts: 652


« Reply #17 on: July 14, 2013, 08:00:45 PM »
ReplyReply

Yes, AA filterless sensors are known to produce moire and various other artifacts.

The feathers of some bird species are especially likely to produce these artifacts even when a camera with an AA filter is used; the genus Callipepla, for example i.e. California Quail and Gambell's Quail.
Logged
ErikKaffehr
Sr. Member
****
Offline Offline

Posts: 7503


WWW
« Reply #18 on: July 14, 2013, 10:58:31 PM »
ReplyReply

Hi,

I guess that the AA-filters used in cameras are a compromise, mostly intended to reduce color moiré.
u
The Hasselblad produces great amount of moiré/fake color, see image below.  I am not doing resolution test using a USAF test chart but testing focusing methods. The test chart is nice to focus on and also useful to quickly sort out of focus images.

I don't really much of that in real pictures. Funny enough, first time I saw colour moiré on my Alpha 77 (APS-C 24MP) was when I was shooting a picture of my Hasselblad.

Best regards
Erik

The feathers of some bird species are especially likely to produce these artifacts even when a camera with an AA filter is used; the genus Callipepla, for example i.e. California Quail and Gambell's Quail.
« Last Edit: July 14, 2013, 11:00:36 PM by ErikKaffehr » Logged

yaya
Sr. Member
****
Offline Offline

Posts: 1131



WWW
« Reply #19 on: July 15, 2013, 01:28:16 AM »
ReplyReply

What software did you use for processing Erik?
Logged

Yair Shahar | Product Manager | Mamiya Leaf |
e: ysh@leaf-photography.com | m: +44(0)77 8992 8199 | www.mamiyaleaf.com | yaya's blog
Pages: [1] 2 »   Top of Page
Print
Jump to:  

Ad
Ad
Ad