Velocity Reviews - Computer Hardware Reviews

Velocity Reviews > Newsgroups > Computing > Digital Photography > The future of the AA filter

Reply
Thread Tools

The future of the AA filter

 
 
TheRealSteve
Guest
Posts: n/a
 
      02-12-2012

On Sat, 11 Feb 2012 08:19:50 -0900, http://www.velocityreviews.com/forums/(E-Mail Removed) (Floyd L.
Davidson) wrote:

>TheRealSteve <(E-Mail Removed)> wrote:
>>On Fri, 10 Feb 2012 15:33:18 -0900, (E-Mail Removed) (Floyd L.
>>Davidson) wrote:
>>
>>>TheRealSteve <(E-Mail Removed)> wrote:
>>>>On Wed, 08 Feb 2012 10:17:26 -0500, nospam <(E-Mail Removed)>
>>>>wrote:
>>>>
>>>>>In article <(E-Mail Removed)>, Rich
>>>>><(E-Mail Removed)> wrote:
>>>>>
>>>>>> Do you honestly want to see that comparison?
>>>>>
>>>>>bring it on. you obviously don't understand what it is you're looking
>>>>>at. if you remove the aa filter, you *will* get artifacts unless there
>>>>>isn't much detail to begin with (e.g., a solid colour wall).
>>>>
>>>>Nope. You will only get artifacts when the spatial frequency of what
>>>>you're sampling is greater than 1/2 the spatial resolution of your
>>>>sensor, given that your lens can also resolve to that level of detail.
>>>
>>>Whoa. Lets stop right there. You have said virtually
>>>the same thing he said. He isn't wrong. I am not
>>>getting the idea that you have a perspective on what it
>>>means that allows understanding the significance of what
>>>he said (or what you've written either).
>>>
>>>>It takes a lot more than a solid wall to do that and it doesn't happen
>>>>very often in real photos. Of course you can find examples when you're
>>>>looking for it. But you have to look through a lot of shots to find
>>>>those few examples.
>>>
>>>It happens in *most* real photos. That is because such
>>>detail exists in real life objects.
>>>
>>>What you have to look hard to find is *not* where
>>>aliasing has happened, but where the aliasing produces
>>>something you can visually identify as aliasing. Hence
>>>if the fine detail is the result of multiple parallel
>>>lines, or even worse if there are two sets of parallel
>>>lines that have different offsets, the result in the
>>>image will be moir, which can easily be identified.
>>>
>>>But if the fine detail is anything else, it is virtually
>>>impossible to identify the resulting detail as the
>>>result of aliasing unless you have a second image that
>>>is exactly the same and does not contain the aliased
>>>detail. Then it is very obvious too.

>>
>>What you think of as "aliased detail" in most cases is real detail
>>that's missing from the image that had the aa filter. The detail that
>>exists in real objects is exactly what the aa filter removes.

>
>Do you understand what "aliasing" is? Spatial
>frequencies that are higher than the Nyquist Limit (1/2
>the sampling frequency) are *folded* down in frequency.


No, they're not *folded* down. They still exist at the original
frequency but due to ambiguity in reproduction, the are duplicated.

>Which is to say that your opening sentence is pure nonsense.


No it's not. It's very easy for you to mistake additional detail in an
image as an artifact when it might not be.

>> They show up in the image as something that did
>>not exist at all in the original scene. But you
>>will get absolutely zero aliasing anywhere in an image where the
>>detail has not exceeded the nyquist frequency. With today's sensors
>>that have several times the spatial resolution of sensors of only a
>>few years ago, there's much less real life detail that exceeds the
>>sensor's nyquist frequency. Of course it can happen, but much more
>>rarely.

>
>"Much less real life detail"?? That is more nonsense.


Actually, it's not nonsense. It's axiomatic. Given that the real world
that we photograph hasn't changed all that much in the past few years
or so and given that sensor spatial resolution has increased by
several factors over the last few years or so (say a decade), then
it's self evident that there's much less real life detail that exceeds
the nyquist frequency of today's sensors than those of a decade ago.
If you're going to argue against an obvious truth then there's not
much point in continuing the discussion

>>>>It's simple sampling theory that a lot of poeple just don't seem to
>>>>understand.
>>>
>>>Do you?

>>
>>Yes, I do.

>
>So far you've said very little that suggests you do.


And you're arguing against the indubitable. So it's in my favor that
you think I know little about sampling theory.

>>>If you think visible moir is the only artifact produced
>>>by aliasing, or that other aliasing artifacts are not
>>>detrimental to the quality of an image, then perhaps you
>>>don't yet really understand sampling theory.

>>
>>Yes, moire' is the only artifact produced by undersampling.

>
>That is an astoundingly ignorant statement to make
>immediately after you claim to understand sampling
>theory.


Thank you for continuing your trend.

>>But the
>>moire is most visible with large areas of the same spatial frequency
>>that is higher than nyquist, such as parallel lines of a brick wall or
>>fabric texture. Here, the low frequency moire artifact is constant and
>>creates a "beat" that is constant and visible across a large portion
>>of the repeating pattern.

>
>But that is hardly the only artifact produced.


Spoke too soon. See below:

>>In other areas of spatial frequency greater than nyquist that do not
>>have the same constant spatial frequency, such as a field of grass,
>>the low frequency artifacts and the resultant moire' are too random to
>>be easily picked out by your eye as something different than the
>>randomness of the original field of grass. The moire' patterns
>>produced are too localized to a small area to be noticed unless you
>>pixel peep. Over the larger area it's too random to be noticed as a
>>moire', but it is moire' nonetheless.

>
>No it is not moire', it is simply aliasing distortion.


Apparently you don't know what moire' is. It's simply interference
between, in this case, the original content of an image and alias
artifacts. There's no said limit on how big or small an area of
interference must be in order to be considered moire'. It doesn't have
to be large areas. It's doesn't have to be parallel patterns. It can
be curved, blobby, large or small. In some cases, moire' may even look
like the fringes you get when you over-sharpen. It's only a special
case of interference, or moire', that creates the more familiar
patterns that you might exclusively refer to as a moire' pattern.

>>The main point is that with today's high resolution sensors coupled
>>with lens resolution increases that have not kept up with sensor
>>resolution increases, aliasing is much less of a problem than it used
>>to be. A body with the aa filter removed is a useful tool in the
>>hands of a photographer mindful of it's limitations.

>
>A body with a appropriate AA filter is a much more
>useful tool.


Maybe, maybe not. Depends on what you're trying to do. It also depends
on what the definition of "appropriate" is. Some bodies came with a
much too strong AA filter. The D200 is a notable example.

I'm through debating this with you. Go ahead and have the last word if
you like.

Steve
 
Reply With Quote
 
 
 
 
nospam
Guest
Posts: n/a
 
      02-12-2012
In article <(E-Mail Removed)>, TheRealSteve
<(E-Mail Removed)> wrote:

> >>What you think of as "aliased detail" in most cases is real detail
> >>that's missing from the image that had the aa filter. The detail that
> >>exists in real objects is exactly what the aa filter removes.

> >
> >Do you understand what "aliasing" is? Spatial
> >frequencies that are higher than the Nyquist Limit (1/2
> >the sampling frequency) are *folded* down in frequency.

>
> No, they're not *folded* down. They still exist at the original
> frequency but due to ambiguity in reproduction, the are duplicated.


nonsense. yes they are folded down. do you even understand what
aliasing *is* ?

> >Which is to say that your opening sentence is pure nonsense.

>
> No it's not.


your sentence above definitely is nonsense.

> It's very easy for you to mistake additional detail in an
> image as an artifact when it might not be.


actually, the opposite is what happens. many people mistake alias
artifacts as additional detail, which is why they erroneously assume
removing an aa filter is a good idea.

since you don't seem to understand what aliasing actually is, it's not
too surprising you want to remove the aa filter.

> >> They show up in the image as something that did
> >>not exist at all in the original scene. But you
> >>will get absolutely zero aliasing anywhere in an image where the
> >>detail has not exceeded the nyquist frequency. With today's sensors
> >>that have several times the spatial resolution of sensors of only a
> >>few years ago, there's much less real life detail that exceeds the
> >>sensor's nyquist frequency. Of course it can happen, but much more
> >>rarely.

> >
> >"Much less real life detail"?? That is more nonsense.

>
> Actually, it's not nonsense. It's axiomatic. Given that the real world
> that we photograph hasn't changed all that much in the past few years
> or so and given that sensor spatial resolution has increased by
> several factors over the last few years or so (say a decade), then
> it's self evident that there's much less real life detail that exceeds
> the nyquist frequency of today's sensors than those of a decade ago.
> If you're going to argue against an obvious truth then there's not
> much point in continuing the discussion


in other words, the nyquist frequency of more recent sensors is higher.
you still need an aa filter, just one with a higher cutoff frequency.
 
Reply With Quote
 
 
 
 
TheRealSteve
Guest
Posts: n/a
 
      02-12-2012

On Sat, 11 Feb 2012 16:56:20 -0500, nospam <(E-Mail Removed)>
wrote:

>In article <(E-Mail Removed)>, TheRealSteve
><(E-Mail Removed)> wrote:
>
>> What you think of as "aliased detail" in most cases is real detail
>> that's missing from the image that had the aa filter. The detail that
>> exists in real objects is exactly what the aa filter removes.

>
>wrong. what the aa filter removes is detail the can't be resolved by
>the sensor because it's too high. if the aa filter wasn't there, that
>detail would alias. with the aa filter, there wont be any aliasing.
>
>that's a bit simplistic, but it makes the point. real world aa filters
>aren't perfect so there's a bit of a tradeoff as to how strong it needs
>to be.


And some are way too strong, which is the point here. Obviously it
won't improve images much removing an AA filter if the AA filter is
not too strong. But that's not always the case.

>> But you
>> will get absolutely zero aliasing anywhere in an image where the
>> detail has not exceeded the nyquist frequency.

>
>actually, aliasing starts to become a problem a little below nyquist.


Really? Care to explain that which goes against everything ever
understood about sampling?

>> With today's sensors
>> that have several times the spatial resolution of sensors of only a
>> few years ago, there's much less real life detail that exceeds the
>> sensor's nyquist frequency. Of course it can happen, but much more
>> rarely.

>
>several??? more like double. a 24 mp sensor, such as the nikon d3x, has
>double the linear resolution than that of a 6mp sensor, such as the
>nikon d70.


Now try the comparison with a D1 against a D800. Everyone can play
numbers games. And it's the D800 at 36mp that may be available from
the manufacturer without an AA filter.

>> Couple that with lenses that cannot resolve to level of detail that
>> the sensor can, the lens itself is acting as an anti-alias filter. No
>> need to mush it up further with another filter.

>
>maybe *your* lenses don't resolve that well. you obviously need new
>lenses, ones which can resolve what the sensor can resolve. and a book
>on signal theory.


Just having a book on signal theory isn't good enough. You may need to
crack yours open to the part where it explains how aliasing is a
problem with frequencies "a little below nyquist" and tell us what it
says.

Also, check the MTF of your lenses and see if they really will resolve
lines down to the level of a D800 sensor. You may be surprised what
you find.

Steve
 
Reply With Quote
 
nospam
Guest
Posts: n/a
 
      02-12-2012
In article <(E-Mail Removed)>, TheRealSteve
<(E-Mail Removed)> wrote:

> >> What you think of as "aliased detail" in most cases is real detail
> >> that's missing from the image that had the aa filter. The detail that
> >> exists in real objects is exactly what the aa filter removes.

> >
> >wrong. what the aa filter removes is detail the can't be resolved by
> >the sensor because it's too high. if the aa filter wasn't there, that
> >detail would alias. with the aa filter, there wont be any aliasing.
> >
> >that's a bit simplistic, but it makes the point. real world aa filters
> >aren't perfect so there's a bit of a tradeoff as to how strong it needs
> >to be.

>
> And some are way too strong, which is the point here. Obviously it
> won't improve images much removing an AA filter if the AA filter is
> not too strong. But that's not always the case.


in a perfect world, we could have a perfect aa filter. in the real
world, there are tradeoffs. most aa filters are well matched.

> >> But you
> >> will get absolutely zero aliasing anywhere in an image where the
> >> detail has not exceeded the nyquist frequency.

> >
> >actually, aliasing starts to become a problem a little below nyquist.

>
> Really? Care to explain that which goes against everything ever
> understood about sampling?


really.

how about you explain why no aliasing at all will occur below nyquist,
which is what you seem to be saying.

> >> With today's sensors
> >> that have several times the spatial resolution of sensors of only a
> >> few years ago, there's much less real life detail that exceeds the
> >> sensor's nyquist frequency. Of course it can happen, but much more
> >> rarely.

> >
> >several??? more like double. a 24 mp sensor, such as the nikon d3x, has
> >double the linear resolution than that of a 6mp sensor, such as the
> >nikon d70.

>
> Now try the comparison with a D1 against a D800.


what for? you said 'sensors of only a few years ago', not those of 13
years ago.

but if you insist, let's compare the d800 with the nikon d1. since the
d1 actually had 10.8 megapixels (but output 2.7 mp), the d800 has just
under twice the linear resolution (1.8x) of the d1 sensor. so much for
'several times'.

> Everyone can play numbers games.


yes, i can see that.

> And it's the D800 at 36mp that may be available from
> the manufacturer without an AA filter.


actually, the d800e has an aa filter and an anti-aa filter.

<http://www.dpreview.com/previews/nikond800/images/d800e-olpf.jpg>

in any event, the d800 at 36 mp is twice the linear resolution of a 9mp
sensor. there weren't any 9mp dslrs, however, there were 10 mp (nikon
d80 & d200 and canon 40d) and 8 mp (canon 30d) dslrs about 5 years or
so ago.

> >> Couple that with lenses that cannot resolve to level of detail that
> >> the sensor can, the lens itself is acting as an anti-alias filter. No
> >> need to mush it up further with another filter.

> >
> >maybe *your* lenses don't resolve that well. you obviously need new
> >lenses, ones which can resolve what the sensor can resolve. and a book
> >on signal theory.

>
> Just having a book on signal theory isn't good enough.


true. you ought to enroll at an engineering school and have someone
explain it to you.

> You may need to
> crack yours open to the part where it explains how aliasing is a
> problem with frequencies "a little below nyquist" and tell us what it
> says.


it says you can get aliasing.

> Also, check the MTF of your lenses and see if they really will resolve
> lines down to the level of a D800 sensor. You may be surprised what
> you find.


no, actually i won't be surprised at all.
 
Reply With Quote
 
TheRealSteve
Guest
Posts: n/a
 
      02-13-2012

On Sat, 11 Feb 2012 23:05:51 -0500, nospam <(E-Mail Removed)>
wrote:

>In article <(E-Mail Removed)>, TheRealSteve
><(E-Mail Removed)> wrote:
>
>> >> What you think of as "aliased detail" in most cases is real detail
>> >> that's missing from the image that had the aa filter. The detail that
>> >> exists in real objects is exactly what the aa filter removes.
>> >
>> >wrong. what the aa filter removes is detail the can't be resolved by
>> >the sensor because it's too high. if the aa filter wasn't there, that
>> >detail would alias. with the aa filter, there wont be any aliasing.
>> >
>> >that's a bit simplistic, but it makes the point. real world aa filters
>> >aren't perfect so there's a bit of a tradeoff as to how strong it needs
>> >to be.

>>
>> And some are way too strong, which is the point here. Obviously it
>> won't improve images much removing an AA filter if the AA filter is
>> not too strong. But that's not always the case.

>
>in a perfect world, we could have a perfect aa filter. in the real
>world, there are tradeoffs. most aa filters are well matched.
>
>> >> But you
>> >> will get absolutely zero aliasing anywhere in an image where the
>> >> detail has not exceeded the nyquist frequency.
>> >
>> >actually, aliasing starts to become a problem a little below nyquist.

>>
>> Really? Care to explain that which goes against everything ever
>> understood about sampling?

>
>really.
>
>how about you explain why no aliasing at all will occur below nyquist,
>which is what you seem to be saying.


It's not just me saying that no in band aliasing will occur below
nyquist. Of couse out of band aliasing will occur but that's not what
we're talking about. You don't have to be anywhere near nyquist for
that to occur.

>> >> With today's sensors
>> >> that have several times the spatial resolution of sensors of only a
>> >> few years ago, there's much less real life detail that exceeds the
>> >> sensor's nyquist frequency. Of course it can happen, but much more
>> >> rarely.
>> >
>> >several??? more like double. a 24 mp sensor, such as the nikon d3x, has
>> >double the linear resolution than that of a 6mp sensor, such as the
>> >nikon d70.

>>
>> Now try the comparison with a D1 against a D800.

>
>what for? you said 'sensors of only a few years ago', not those of 13
>years ago.
>
>but if you insist, let's compare the d800 with the nikon d1. since the
>d1 actually had 10.8 megapixels (but output 2.7 mp), the d800 has just
>under twice the linear resolution (1.8x) of the d1 sensor. so much for
>'several times'.


Wrong. The only reason the D1 had 10.8 MP was for increased
sensitivity and SNR. The pixels were grouped together which did
nothing to increase the resolving power, which is what we're talking
about. So the D800 has 3.6x the linear resolution of the D1.

>> You may need to
>> crack yours open to the part where it explains how aliasing is a
>> problem with frequencies "a little below nyquist" and tell us what it
>> says.

>
>it says you can get aliasing.


One cite would suffice. I'm willing to learn. So please find one
reputable cite that says you'll get in-band aliasing artifacts (i.e.,
spurious frequencies between 0 and nyquist frequency) if you sample a
little below nyquist.

Steve
 
Reply With Quote
 
David J Taylor
Guest
Posts: n/a
 
      02-13-2012

"TheRealSteve" <(E-Mail Removed)> wrote in message
news:(E-Mail Removed)...
[]
> Wrong. The only reason the D1 had 10.8 MP was for increased
> sensitivity and SNR. The pixels were grouped together which did
> nothing to increase the resolving power, which is what we're talking
> about. So the D800 has 3.6x the linear resolution of the D1.


Nikon D1: 2000 pixels in 23.7 mm - 11.85 um pixel spacing

Nikon D800: 7360 pixels in 35.9 mm - 4.88 um spacing

2.4 times the linear resolution in pixels per mm. 3.68 times the
resolution in pixels per picture width. Using data from:

http://www.dpreview.com/reviews/nikond1/
http://www.dpreview.com/products/nikon/slrs/nikon_d800

[Interesting comment on the D1 page, dated November 2000: "The D1 was
Nikon's answer to Kodak's domination of the professional SLR's market."

> One cite would suffice. I'm willing to learn. So please find one
> reputable cite that says you'll get in-band aliasing artifacts (i.e.,
> spurious frequencies between 0 and nyquist frequency) if you sample a
> little below nyquist.
>
> Steve


Many of the resolution tests showing a step-wedge chart apparently show
aliasing at just less than Nyquist, but this is using a square-wave light
pattern which has harmonics. Perhaps this explains what is being
reported?

Possible example:
http://a.img-dpreview.com/reviews/Ca...84-ACR-003.jpg
From: http://www.dpreview.com/reviews/canoneos600d/page11.asp

David

 
Reply With Quote
 
RichA
Guest
Posts: n/a
 
      02-13-2012
On Feb 11, 11:51*pm, (E-Mail Removed) (Floyd L. Davidson) wrote:

> >It also depends
> >on what the definition of "appropriate" is. Some bodies came with a
> >much too strong AA filter. The D200 is a notable example.

>
> Too strong for one job, but absolutely correct for a job where it is
> indeed the "appropriate AA filter".


Hey maybe they should offer and AA filter selection set, for different
scenes?
Has anyone been inconvenienced because their camera had a "too light"
AA filter?
 
Reply With Quote
 
TheRealSteve
Guest
Posts: n/a
 
      02-14-2012

On Mon, 13 Feb 2012 15:13:49 -0000, "David J Taylor"
<(E-Mail Removed)> wrote:

>
>Many of the resolution tests showing a step-wedge chart apparently show
>aliasing at just less than Nyquist, but this is using a square-wave light
>pattern which has harmonics. Perhaps this explains what is being
>reported?
>
>Possible example:
>http://a.img-dpreview.com/reviews/Ca...84-ACR-003.jpg
>From: http://www.dpreview.com/reviews/canoneos600d/page11.asp
>
>David


That's entirely possible. Using the audio analogy, a frequency even
right at nyquist will have no in-band aliasing. When it's reproduced
it will be reproduced as a square wave, which has the fundamental plus
all of the odd harmonics, each diminishing in amplitude as you go up
in frequency. However, the reconstruction filter will take out those
odd harmonics and reproduce the original sine wave at nyquist. The
original signal *must* have been a sine wave for it not to have any
content above nyquist if the fundamental was right at nyquist. If your
original signal was not a sine wave, then it's frequency content above
nyquist will cause aliasing. This is the case you mentioned above.

While frequencies just below nyquist will have an alias just above
nyquist and frequencies close to 0 will have an alias close to the
sampling frequency, those aliases exist only mathmatically and they
are not reproduced during reconstruction. That's because
reconstruction only reproduces the lowest frequency of all the
possible aliases, of which there are an infinite number.

The problem with aliasing when sampling above nyquist is exactly due
to the fact that reconstruction only reproduces the lowest of the
aliases, which in this case is the wrong one. If all the content is
below or even right at nyquist, there will be no aliases reproduced
during reconstruction.

"nospam" may also be getting confused by what reconstruction filtering
is doing. It has to "smooth the jaggies" produced because you have
steps in the reconstructed signal but the original was probably
smooth. The high frequency content in those steps has nothing to do
with aliasing and they can be removed with a good reconstruction
filter while aliasing cannot be removed by filtering. That's because
the steps' frequency content is well out of band so you can filter it
with a lowpass filter. But the aliasing produced by undersampling is
in band so filtering it won't work because you'll also filter out part
of your original signal.

So in summary: There are no aliases reproduced from digital content if
all of the original sampled signal is at or below the nyquist
frequency. Not even a frequency "a little below nyquist" will cause
aliases to be reproduced. You're going to have to search long and hard
to find a reputable DSP textbook that says otherwise.

Steve
 
Reply With Quote
 
nospam
Guest
Posts: n/a
 
      02-14-2012
In article <(E-Mail Removed)>, TheRealSteve
<(E-Mail Removed)> wrote:

> So in summary: There are no aliases reproduced from digital content if
> all of the original sampled signal is at or below the nyquist
> frequency. Not even a frequency "a little below nyquist" will cause
> aliases to be reproduced. You're going to have to search long and hard
> to find a reputable DSP textbook that says otherwise.


wrong.
 
Reply With Quote
 
David J Taylor
Guest
Posts: n/a
 
      02-14-2012
"nospam" <(E-Mail Removed)> wrote in message
news:140220120838118504%(E-Mail Removed)...
> In article <(E-Mail Removed)>, TheRealSteve
> <(E-Mail Removed)> wrote:
>
>> So in summary: There are no aliases reproduced from digital content if
>> all of the original sampled signal is at or below the nyquist
>> frequency. Not even a frequency "a little below nyquist" will cause
>> aliases to be reproduced. You're going to have to search long and hard
>> to find a reputable DSP textbook that says otherwise.

>
> wrong.


Perhaps you might provide a citation to support your assertation?

 
Reply With Quote
 
 
 
Reply

Thread Tools

Posting Rules
You may not post new threads
You may not post replies
You may not post attachments
You may not edit your posts

BB code is On
Smilies are On
[IMG] code is On
HTML code is Off
Trackbacks are On
Pingbacks are On
Refbacks are Off


Similar Threads
Thread Thread Starter Forum Replies Last Post
"rec.photo.digital.txt" Filter File Posted Online (for newsreadersthat can import a list of e-mail addresses to filter out) SMS 斯蒂文• 夏 Digital Photography 2 11-25-2007 11:00 AM
Polarising filter with UV filter? Stimp Digital Photography 23 11-17-2006 11:51 AM
to filter of not to filter Ken Digital Photography 2 12-23-2005 12:45 PM
UV Protector filter vs. Skylight filter? john Digital Photography 8 06-26-2004 03:44 PM
Kodak "future" DX6490 Filter Attachment ?? Tom Rowsell Digital Photography 3 01-04-2004 04:23 PM



Advertisments