Is this chroma aliasing?

[Matt W.]

What's a half pixel?

Half the distance between pixels. If the pixel pitch is six micron, then three microns. Or moving the test chart or tripod so that, once focused, it moves the corresponding image on the sensor by three microns should do the same thing. I'm not being very clear I know. :/

Btw, I'm not pretending like I know what I'm talking about here because I know I don't know what I'm talking about when it comes to this stuff except for the very basics of sampling theory. I'm mostly just curious how it works. So any mistakes or lack of clarity with nomenclature are from ignorance on my part and I apologize if I tried to sound like a know-it-all before. I'd like to be one, but this is an area where I'm fairly ignorant.

 

[Graeme Nattress]

That actually makes some sense, but what I fail to understand is this: let's say we're in a perfect world with a perfect system between the sensor and the subject, everything else in the chain is 100% mtf exactly, and we have a 4k sensor grid and a 4k printed image of 2048 line pairs. Let's line them up perfectly. Then we have up to 100% mtf of 2048 line pairs. The captured image is the same as the printed one.

Now let's move the camera a bit so that the sensor is offset by just one half pixel. Now, we're suddenly seeing gray at every photo site. And we have 0% mtf.

Are you implying that by removing just one line pair from that image, this phenomenon will no longer occur in this way? It seems like it still would but the pattern would be wavy rather than being straight 100% mtf line pairs transitioning to 0% mtf (pure gray) across the frame no matter how exactly it's placed. It seems to me like you'd have to remove half the line pairs from that image so that the frequency of the image itself was low enough that this phenomenon didn't occur at all and you could approach 100%mtf capture at any frequency. How is this not the case? Or is aliasing simply when you get erroneous results of >100% mtf?

The sampling theory criteria is not less than or equal to, but less than. Also, those line pairs on your chart have to be sinusoidal.

Graeme

 

[Matt W.]

The sampling theory criteria is not less than or equal to, but less than. Also, those line pairs on your chart have to be sinusoidal.

Graeme

Sure. And if they're sine waves that makes sense, but when I think of a line pair at a given frequency I imagine it as black and white, as a square wave rather than a sine wave, which if I'm not mistaken is how they're printed on charts. In the case of square waves, I imagine there would be aliasing (which is imo why test charts show aliasing with most cameras but real world scenes rarely do; nature abhors a straight line etc.).

Correct me if I'm wrong, but doesn't a square wave reproduced at 50% mtf closely resemble a sine wave of the same frequency at 100% mtf if we're talking spatial sampling?

 

[Graeme Nattress]

MTF is not derived from line pairs though, but from sine waves. That's because a square wave is not bandwidth limited with an excess of harmonics above the fundamental.

Graeme

 

[Matt W.]

MTF is not derived from line pairs though, but from sine waves. That's because a square wave is not bandwidth limited with an excess of harmonics above the fundamental.

Graeme

Right, I remember that from when I dabbled in subtractive synthesis and how to build each waveform as a function of harmonics of the sine wave. This is all starting to make sense now. My question then is: aren't resolution test charts effectively printed in square waves, being high contrast black and white rather than being a (sine wave-shaped) gradient between black and white? In that case, shouldn't they be printed in gradients (sine waves)? Of course square waves have higher order harmonics and so you're totally right. I understand what you're saying now and it makes sense. (In theory any black and white line pair is of an essentially infinite frequency, or its overtones are.)

My other question is this: when I listen to audio, it's usually at 44.1khz or 48khz because I understand the limits of human hearing to be about 20khz and so consumer audio deliverables double that (my personal limit is about 15.4kz, I've been to too many concerts). It's delivered at twice the maximum perceptible frequency so as to prevent it from aliasing in playback..

But doesn't a sine wave (in audio) contain both a maximum and minimum amplitude (a 0 and an N)? And so doesn't a sine wave in audio correspond more to a line pair than it does with a line? If so, why must audio be presented at twice the maximum audible frequency to be free of aliasing whereas video can be delivered and captured at 1X the native resolution without substantial aliasing (if using a monochrome sensor)? 22khz audio sounds bad to me and I know I can't hear to 22khz. I can see the difference between 2k and 4k, though, and 4k deliverables don't have any inherent aliasing so far as I can tell, unless it's introduced in capture. But 2k deliverables derived from 4k sensors look much better to me. Is that simply to do with Bayer interpolation?

I'm confused. Is this because a sine wave only represents 0 to N but the entire sweep of it is 0 to N to 0 to -N to 0 so you need to double the sampling rate? Or because vision has such high resolution that even 4k doesn't really approach its absolute maximum and viewing distance is of course a factor so with video it's still just a trade off because we haven't reached the limits of human perception yet?

(Further confusing things, I tend to slightly prefer 192/24 audio to CD quality, but I recognize there's no math to support my preference there.)

 

[Graeme Nattress]

Some rez charts are printed square waves as that's often easier. But you can get sine wave charts too like the sine zone plate I use. If you have a sine zone, you can plot MTF just by dragging a line sample over the image from the centre out as a radius and plotting it. For other types of charts the MTF has to be derived and that's a more complex process.

So audio is typically where we find discussions on sampling theory and indeed you need a sampling frequency of at least twice that of the the highest freq you want to be able to hear without aliasing. And as you say we're dealing in full wavelengths that start at zero, go up, peak, down, through zero, trough, back up to zero again. That's the equivalent of a line-pair in imaging, hence this divide by two confusion because often in image resolution we talk lines rather than line pairs. So it's only 1x because we talk lines. If we were talking line-pairs we'd use the same 2x factor as audio. Sampling theory remains the same in both cases, but the units are changing in imaging from wavelengths == line-pairs to lines.

In imaging we don't have the low pass filters available to us in audio. We're stuck with simple optical filters and we can't construct them to be sharp like the equivalent of an audio brick-wall filter. That is generally why we have to allow for some aliasing in imaging, although we do seek to minimize it. With the filters we have available to get a zero aliasing condition we'd be using 8k sensors to achieve un-aliased VHS resolution Part of the compromise is we do impact the pass-band with our optical filtering, just not to the degree of making the resulting image being VHS resolution. We also need to take into acount resolution losses of the lens too.

There's actually not that much resolution loss from using a bayer pattern. The major factor is the nature of optical low pass filters.

With digitical audio the primary benefit from 192khz sample rates are the relaxed requirements on the anti-alias filtering. No longer do you need a sharp brick wall filter. What you're doing is using 8k to generate a VHS image (well, sorta...)

Graeme

 

[Humberto Rivera]

Thank you Graeme for your explanation, I think I’m getting closer to understanding. So FIRST we need a; “sine wave charts, like the sine zone plate you use.” SECOND we take; “Sampling theory remains the same in both cases, but the units are changing in imaging from wavelengths = line-pairs to lines.” THIRD; “That is generally why we have to allow for some aliasing in imaging, although we do seek to minimize it. “ FOURTH; “We also need to take into account resolution losses of the LENS too.” FIFTH; “There's actually not that much resolution loss from using a BAYER pattern. The major factor is the nature of optical low pass filters.” SIXTH that’s where you loss me; “What you're doing is using 8k to generate a VHS image (Well, sort of...)” Thank you, interesting conversation.

Humberto Rivera

 

[Graeme Nattress]

So with regards to where I loose you.... In audio-land, we need to preserve the 0-20khz pass-band with abolute integrity. Sampling at 44.1khz gives you a band of around 2khz between 20khz and 22khz for us to low pass filter in. That means we need a brick-wall filter to avoid audible aliasing (which can sound rather nasty). The other way to achieve a clean 0-20khz pass band is to sample at 192khz and now the filter band has been extended to 20khz-96khz and we have much more choice on our anti-alias filtering and can be very sure the 0-20khz pass band is uneffected by the anti-alias filtration. The visual analogue is using 8k resolution with a hyper-strong optical low pass filter to make a VHS resolution image.

Graeme

 

[Humberto Rivera]

Thank you Graeme, I think it makes sense, obviously is far more complicated, I come at this from a Filmmaker’s perspective, I really do not need to understand (but would like too); the “Why” just that we’re using the best “Camera Technology” to make our Images after it’s all said and done.

Humberto Rivera

 

[Graeme Nattress]

Thank you Graeme, I think it makes sense, obviously if far more complicated, I come at this from a Filmmaker’s perspective, I really do not need to understand (but would like too); the “Why” just that we’re using the best “Camera Technology” to make our Images after it’s all said and done.

Humberto Rivera

After wandering the Halls of NAB last week, I do (with obvious bias) think we're using the best cinema camera technology there is. And the good news from that is through constant learning and listening to feedback, we keep improving and I'm incredibly excited for the future!

Graeme

 

[Sabovic Adis]

Half the distance between pixels. If the pixel pitch is six micron, then three microns. Or moving the test chart or tripod so that, once focused, it moves the corresponding image on the sensor by three microns should do the same thing. I'm not being very clear I know. :/

That must be some pretty precise tripod we're talking about here - 3mics/cycle on the screw, or something. Are we even allowed to breathe? :biggrin:
But I get it, Matt. :thumbsup:

Btw, I'm not pretending like I know what I'm talking about here because I know I don't know what I'm talking about when it comes to this stuff except for the very basics of sampling theory. I'm mostly just curious how it works. So any mistakes or lack of clarity with nomenclature are from ignorance on my part and I apologize if I tried to sound like a know-it-all before. I'd like to be one, but this is an area where I'm fairly ignorant.

Well, it's always good to ask. So, I asked. :wink5:

 

[Matt W.]

Some rez charts are printed square waves as that's often easier. But you can get sine wave charts too like the sine zone plate I use. If you have a sine zone, you can plot MTF just by dragging a line sample over the image from the centre out as a radius and plotting it. For other types of charts the MTF has to be derived and that's a more complex process.

So audio is typically where we find discussions on sampling theory and indeed you need a sampling frequency of at least twice that of the the highest freq you want to be able to hear without aliasing. And as you say we're dealing in full wavelengths that start at zero, go up, peak, down, through zero, trough, back up to zero again. That's the equivalent of a line-pair in imaging, hence this divide by two confusion because often in image resolution we talk lines rather than line pairs. So it's only 1x because we talk lines. If we were talking line-pairs we'd use the same 2x factor as audio. Sampling theory remains the same in both cases, but the units are changing in imaging from wavelengths == line-pairs to lines.

In imaging we don't have the low pass filters available to us in audio. We're stuck with simple optical filters and we can't construct them to be sharp like the equivalent of an audio brick-wall filter. That is generally why we have to allow for some aliasing in imaging, although we do seek to minimize it. With the filters we have available to get a zero aliasing condition we'd be using 8k sensors to achieve un-aliased VHS resolution Part of the compromise is we do impact the pass-band with our optical filtering, just not to the degree of making the resulting image being VHS resolution. We also need to take into acount resolution losses of the lens too.

There's actually not that much resolution loss from using a bayer pattern. The major factor is the nature of optical low pass filters.

With digitical audio the primary benefit from 192khz sample rates are the relaxed requirements on the anti-alias filtering. No longer do you need a sharp brick wall filter. What you're doing is using 8k to generate a VHS image (well, sorta...)

Graeme

Thanks! I see where I was confused earlier. And now that I'm looking into it there are indeed both binary and sinusoidal zone plates and that all makes sense. The 8k to VHS analogy makes sense for audio oversampling, too. Thanks for clarifying.

 

[Matt W.]

That must be some pretty precise tripod we're talking about here - 3mics/cycle on the screw, or something. Are we even allowed to breathe? :biggrin:
But I get it, Matt. :thumbsup:



Well, it's always good to ask. So, I asked. :wink5:

Yeah, I didn't explain it very well but that is what I meant.

 

[Marcus Friedlander]

After wandering the Halls of NAB last week, I do (with obvious bias) think we're using the best cinema camera technology there is. And the good news from that is through constant learning and listening to feedback, we keep improving and I'm incredibly excited for the future!

Graeme

I think there are a handful of people on this forum who would agree with you

 

[Patrick Tresch]

Thanks a lot for your input. Very informative.

Some rez charts are printed square waves as that's often easier. But you can get sine wave charts too like the sine zone plate I use. If you have a sine zone, you can plot MTF just by dragging a line sample over the image from the centre out as a radius and plotting it. For other types of charts the MTF has to be derived and that's a more complex process.

So audio is typically where we find discussions on sampling theory and indeed you need a sampling frequency of at least twice that of the the highest freq you want to be able to hear without aliasing. And as you say we're dealing in full wavelengths that start at zero, go up, peak, down, through zero, trough, back up to zero again. That's the equivalent of a line-pair in imaging, hence this divide by two confusion because often in image resolution we talk lines rather than line pairs. So it's only 1x because we talk lines. If we were talking line-pairs we'd use the same 2x factor as audio. Sampling theory remains the same in both cases, but the units are changing in imaging from wavelengths == line-pairs to lines.

In imaging we don't have the low pass filters available to us in audio. We're stuck with simple optical filters and we can't construct them to be sharp like the equivalent of an audio brick-wall filter. That is generally why we have to allow for some aliasing in imaging, although we do seek to minimize it. With the filters we have available to get a zero aliasing condition we'd be using 8k sensors to achieve un-aliased VHS resolution Part of the compromise is we do impact the pass-band with our optical filtering, just not to the degree of making the resulting image being VHS resolution. We also need to take into acount resolution losses of the lens too.

There's actually not that much resolution loss from using a bayer pattern. The major factor is the nature of optical low pass filters.

With digitical audio the primary benefit from 192khz sample rates are the relaxed requirements on the anti-alias filtering. No longer do you need a sharp brick wall filter. What you're doing is using 8k to generate a VHS image (well, sorta...)

Graeme