Welcome to our community

Be a part of something great, join today!

Log in
  • Hey all, just changed over the backend after 15 years I figured time to give it a bit of an update, its probably gonna be a bit weird for most of you and i am sure there is a few bugs to work out but it should kinda work the same as before... hopefully :)

Leica M Monochrom

No, it's better than a stacked sensor without OLPF if you do a proper demosaic and downsample from a properly optically filtered source - you get more rez and less aliasing.

Graeme
 
A full coverage stacked pixel sensor is basically perfect, you cannot improve over that, its equivalent to a bayer sensor of infinitely high resolution downsampled to that particular resolution. This is what the leica monochrom tries to achieve for B&W, it does not have an olpf since you only need it to get the individual color pixels to sample their entire area to prevent aliasing caused by missing information due to the color filter array. Either B&W or stacked color can sample the full pixel area without olpf.
 
Kenneth, I don't know what you're talking about, but it's not supported by sampling theory or analysis of cameras such as the Sigmas which use an approach similar to what you're suggesting.

Graeme
 
Remember my first comment: 'An 8k bayer sensor used for 4k would be as sharp as a 4k B&W'

Are you claiming an 8k bayer sensor could produce a better 4k image than a perfect 4k full color sensor? This would also mean you think the leica monochrom does not achieve near optimal results for its output resolution?
 
Well yes it would because it would be properly optically filtered at 8k, and hence through proper use of a downsampling filter have less aliasing than your 4k sensor. That is unless your 4k sensor used proper optical filtering to minimize aliasing, and then the 4k sensor would not measure 4k in resolution. Fact is, no matter what the kind of sensor, if you want the same resolution out as pixel count you will get aliasing, and the practical example here is the Sigma and it's Foveon where they omit the OLPF and get significantly strong aliasing which corrupts images of fine detail.

Graeme
 
Graeme Nattress said:
Kenneth, I don't know what you're talking about, but it's not supported by sampling theory or analysis of cameras such as the Sigmas which use an approach similar to what you're suggesting.

Graeme
Click to expand...
I am talking about aliasing caused by pointsampling or not sampling the full area of the pixels share on the image sensor, similar to what you get from lineskipping which RED is against for good reason, bayer sensors get this for the individual colors without olpf simply because three quarters of the information is missing for the reds and blues and half for the greens, the olpf is there to send light from the missing areas to the pixels.

I think you are talking about a different kind which is caused by detail coming in and out of alignment with the pixel grid, i am not sure i agree that this is something that should be eliminated, it is not wrong information, it is simply what you get by using a resolution which is too low, blurring detail that is aligned with the pixel grid does not seem like an improvement to me.
 
Kenneth, aliasing is aliasing and yes it needs to be minimized. An area sample basically alters the MTF of the sensor, lowering it, whereas a point sample increases MTF. THe higher the sensor MTF the stronger any optical filtering needs to be to minimize aliasing. You're right that line skipping is more akin to point sampling, hence higher sensor MTF, hence more propensity to aliasing. The only way to get high resolution and low aliasing is to have more pixels than you need for that target resolution, and properly optically low pass filter the image. Then, when you measure the resolution it will be high (because you have more than enough pixels) and the aliasing will be low (because you properly filtered).

"blurring detail that is aligned with the pixel grid does not seem like an improvement to me" well it's better than false detail corrupting the image and it's more like our eyes see things. And it's also because you're thinking about "pixel resolution" which is the wrong way to see these things. The goal is high uncorrupted resolution, not resolution at a pixel level which violates the basics of sampling theory. If you do want that "crunchy" look, then start high on your resolution and downsample. If you want it really crunchy, use a really poor downsample filter. Actually this contradicts what you said about the C300 where the averaging of the greens is actually a filter operation which reduces aliasing compared to a single green extraction (ignoring the second green). If you measure both approaches (single green and green average) the single green approach is actually sharper with more MTF. It also has more unwanted aliasing, leading to a "crunchier" image.

Graeme
 
Graeme Nattress said:
Well yes it would because it would be properly optically filtered at 8k, and hence through proper use of a downsampling filter have less aliasing than your 4k sensor. That is unless your 4k sensor used proper optical filtering to minimize aliasing, and then the 4k sensor would not measure 4k in resolution. Fact is, no matter what the kind of sensor, if you want the same resolution out as pixel count you will get aliasing, and the practical example here is the Sigma and it's Foveon where they omit the OLPF and get significantly strong aliasing which corrupts images of fine detail.

Graeme
Click to expand...
I dont know about the quality of the foveon sensor, do the pixels cover their entire area? They may be pretty poor compared to an ideal sensor? So it seems you think what leica is doing is flawed in theory?

I am interested to know what you think the ideal way would be to downsample a very high resolution image to a lower resolution. Say you have a 16k bayer or rgb sensor and you want to create a 4k image from that, the olpf if chosen for the 16k resolution would be insignificant in size for the 4k output resolution. If you would make the 4k image by combining 4x4 pixels into one you would get a result very similar to an ideal full coverage sensor which you consider to be flawed, would you downsample the image by blurring subpixels from neighboring areas together to achieve the equivalent of an olpf? I would like to see what you consider to be the ideal solution.
 
Sigma pixels are very good on their fill factor.

What Leica is doing is flawed because they're omitting the optical low pass filter. There are some circumstances where there's other limits on MTF int he system that make an OLPF not as necessary - when you're shooting wide open all the time and hence not much is in focus or when you're diffraction limited by stopping down too far. If you're in a position where the lens is passing you enough fine detail and the sensor sees that in focus, you can get aliasing. When there's a fine repeating pattern you'll see this as moire, either luma moire for a black and white or foveon type sensor, or a mix of luma and chroma moire on a bayer type sensor. For the most part chroma moire is easy enough to deal with via processing, but luma moire is "darn hard" approaching "impossible" to deal with adequately, especially in motion.

No, combining pixels together is basically approximating a "box filter" which is a very poor filter. That kind of filter will allow through way-to-much aliasing. Sinc, or it's lanczos approximation are the "best" filters to achieve high resolution and low aliasing in the final downsampled image.

When you're downsampling there are three qualities we're looking to achieve: low aliasing, high resolution and low ringing. No filter will achieve all three. Let me explain further....

Let's, for sake of simplicity, take our theoretical black and white sensor. The fill-factor on this sensor is very high. An optical low-pass filter on this sensor will not introduce any ringing in the image. That is because physical optical filters don't have (can't have without inventing darkons, or -ve light) negative lobes. So no ringing, that's good, but it does mean there will be either aliasing or significant resolution loss. So pick one - you can't have high resolution or low aliasing with an optical filter.

So we know that any aliasing we allow through can't be fixed later, so we pick low aliasing and accept a resolution loss. But say we're aiming for 4k resolution and start with a 5k image. Our 5k image won't measure 5k because of the OLPF, it will measure around 4k.

Now we use our electronic downsample filter to go from 5k to 4k. The electronic downsample filter can have negative lobes, so it can produce high resolution and low aliasing. It will allow through some ringing, but good filters (I like the single lobed lanczos) will keep ringing to a minimum and you'll end up with a high resolution, low ringing, low aliasing image. But at 4k rather than the sensor's 5k. But hey, we got what we wanted.

Graeme
 
Graeme Nattress said:
"blurring detail that is aligned with the pixel grid does not seem like an improvement to me" well it's better than false detail corrupting the image
Click to expand...
Detail from full coverage pixels is never false detail. It is simply detail that goes in and out of alignement with the pixel grid resulting in detail appearing and disappearing either in a spatial pattern or over time with moving images. But the detail that is shown when in alignment is not false. I think this is very different from aliasing from pointsampling which shows wrong luminance or color depending on what point in the pixels area it happens to sample, it does not show the average color of the whole area hence corrupting the image. Full coverage pixels always show the correct color and brightness so there is never any wrong information in there. Yes i am sure you know all this but im not quite sure why you seem to suggest they are both equally wrong, they are very different.

Graeme Nattress said:
And it's also because you're thinking about "pixel resolution" which is the wrong way to see these things. The goal is high uncorrupted resolution, not resolution at a pixel level which violates the basics of sampling theory. If you do want that "crunchy" look, then start high on your resolution and downsample. If you want it really crunchy, use a really poor downsample filter. Actually this contradicts what you said about the C300 where the averaging of the greens is actually a filter operation which reduces aliasing compared to a single green extraction (ignoring the second green). If you measure both approaches (single green and green average) the single green approach is actually sharper with more MTF. It also has more unwanted aliasing, leading to a "crunchier" image.

Graeme
Click to expand...
I am not interested in the crunchy look you are suggesting, i hate pointsampled images. My idea of an ideal image is simply getting closer to full coverage sampling which is what the c300 type pixel binning and leica B&W gets closer to. I know the type of aliasing you talk about is also a potential issue and i actually am undecided what the best way to deal with it is, i am open to the possibility that different ways of blurring with neighboring areas may be better and i will look into all the things you mentioned in your next post so i will have some reading to do... :)

Also i am unsure about the terminology but i think what you are talking about may be called moire rather than aliasing?

Thanks for taking the time to write all these detailed replies.
 
Well, no, basic sampling theory says that when you put too much detail in, you get aliases which are false detail. They fold back to lower frequencies making them more visible and they corrupt such low frequencies that they're practically impossible to blur out without destroying the whole image by blurring it to death. That applies to area sample as well as a theoretical point sample, although as I mention, the MTF of the sensor as you approach a point sample increases making aliasing worse. Although an area sample is better, it's a still a sample, still must obey sampling theory and still gets aliasing issues. "But the detail that is shown when in alignment is not false" - yes it is false. "Yes i am sure you know all this but im not quite sure why you seem to suggest they are both equally wrong, they are very different." - nope, both aliasing. Both violations of sampling theory.

"Also i am unsure about the terminology but i think what you are talking about may be called moire rather than aliasing?" - aliasing is what you get when you violate sampling theory. Aliasing manifests itself visually as moire. Moire is how aliasing looks when it occurs on a repeating pattern whereby you see the high frequency detail that cannot be correctly sampled folded back into the lower frequencies to manifest as visible bands. Aliasing can also manifest as stair-stepping on lines, and as "crunchy" detail. In all cases, the cause if the same - aliasing, but the visual manifestation has different names. Similarly chroma moire is a manifestation of aliasing on just the red/blue on a Bayer CFA. Fortunately in the presence of un-aliased green, this kind of chroma moire is reasonably easy to correct for, but luma-moire (or luma aliasing) is inherently hard to fix.

Graeme
 
I understand that even with full area sampling higher frequencies can cause random patterns that do not accurately represent the patterns in the original image but the brightness and color of the pixel is still accurate for its location, with pointsampling you are basically collecting random information with completely random patterns and colors arising in the image. Those clearly are not directly comparable.

Graeme Nattress said:
Aliasing can also manifest as stair-stepping on lines, and as "crunchy" detail. In all cases, the cause if the same - aliasing, but the visual manifestation has different names. Similarly chroma moire is a manifestation of aliasing on just the red/blue on a Bayer CFA. Fortunately in the presence of un-aliased green, this kind of chroma moire is reasonably easy to correct for, but luma-moire (or luma aliasing) is inherently hard to fix.

Graeme
Click to expand...

These are all caused by pointsampling and simply do not exist with full coverage pixels.
 
Last edited:
The fact that on a bayer sensor it is generally considered necessary to have an olpf while on the foveon or B&W sensors it is assumed by many to be unnecessary seems to suggest that i am not the only one who thinks that there is a clear difference between these two types of aliasing.
 
"These are all caused by pointsampling and simply do not exist with full coverage pixels." - that is incorrect.

"The fact that on a bayer sensor it is generally considered necessary to have an olpf while on the foveon or B&W sensors it is assumed by many to be unnecessary seems to suggest that i am not the only one who thinks that there is a clear difference between these two types of aliasing." - it's still necessary. The difference being in those cases you only get luma moire, not chroma moire. Go look at the Foveon white papers - they mention that they still need an OLPF to avoid aliasing, but of course, Sigma don't want one because then they'd not get that "crunchy" aliased look their users "like".

Here's the appropriate quote from the Foveon white papers: "Using a direct image sensor such as Foveon’s X3 technology, reduces the requirement of a blur filter because aliasing occurs in all color channels equally giving no brightly colored aliasing fringing." - in other words, they get luma aliasing only as aliasing occurs equally across all three channels. Notice how they say "reduce" rather than "eliminate". But remember the OLPF is not there to reduce chroma aliasing in a Bayer CFA, it's to reduce all aliasing, both chroma and luma, and in a good demosaic, chroma aliasing can be effectively removed especially when used with a proper OLPF to effectively minimize the luma aliasing.

Graeme
 
Last edited:
Just as a point of interest, Leica has been hinting that the upcoming M10 will have video. No idea how that will work in a rangefinder.
 
What advantages/differences might a native black and white sensor image have over dialing out the color in an RGB sensor? Will the native black and white sensor not be aligned better to simulate/equal a panchromatic emulsion response?
And as such, would it not also provide for classic B&W filtering methods? A red or orange filter on a color sensor (with the color dialed out) is not the same as applying these to a B&W emulsion. A light green or yellow filter will subtly, and often beautifully, alter skin tones....
However, I can get a very decent black and white image out of RGB, with a lot of control over each color channel, but a true panchromatic black and white response would be OUTSTANDING.

Greame - others: any thoughts on pros and cons of a dedicated B&W Vs Desaturation?
 
Yes for a B/W camera you'd need to use classic filtering techniques. Pros and cons have been pretty well covered above. As I said, it's something I'd love to have on a RED, but it's not practical or really affordable to do :-( But used as is, the desat method makes for great images, so I'm not too worried.

Graeme
 
My 2 cents... straight DeSat is a poor representation of black and white film. B & W cellulose is rather Red and Yellow sensitive, so an equalized desaturation gives you a false representation of how colors are mixed to tones by monochromatic films. If you use a B & W adjustment layer in Photoshop, it gives you the options to bias for Red or Yellow which feels very naturalistic. You can also just use the channel mixer out to monochrome (weighting RGB to something like 60/20/20 percent). Not sure how you would go about this in RCX or resolve.

G-Meister?
 
You must log in or register to reply here.
Back
Top