Those Tiny, Noisy Pixels

[John Sheehy]

Here's a rough simulation of what it might be like if a 1.6x-crop DSLR had the pixel spacing of the Panasonic FZ50 (and the same electronics):

http://www.pbase.com/jps_photo/image/74020772/original


Now, I understand the Canon 10D has a fairly strong AA filter (takes almost 3 pixel lengths to go from 90% white to 10% black with an optically sharp edge), but this is the closest I can simulate; my largest-pixel-pitch DSLR with a sharp lens (Tamron 90mm Macro), vs my tiny-pixel-pitch FZ50.

I'd rather have the images from the tiny pixels.

Pixelation is a kind of "noise", IMO, and it actually exaggerates the noise that exists at the pixel level.

Obviously, not every lens made for the 35mm format or for APS DSLRs have the resolution to benefit from the smaller pixels, but many do.

[BJL]

Thanks for the interesting if extreme comparison. I admit liking it partly because it fits my expectations that "pixel density on the print" and "pixel size on the sensor" have opposing effects on print quality with images coming from an equally sized portion of a sensor.

It would be nice to see a comparison in which pixel size is changed as here, but keeping the sensor the same size, using the same type and level of sensor technology (and using the same lens, aperture, focal length, raw conversion and so on.)

Unfortunately it is very rare for sensor makers to offers sensors of the same size at the same time with significantly different pixel counts; instead, new higher pixel counts generally drive older, lower ones off the market.

Perhaps one opportunity is to compare the Sony 6MP and 10MP CCD's of the Nikon D40 and D80, or the Pentax K100D and K10D. The Sony 6MP sensor is in some sense an old design, but seems to have been incrementally updated over the years from the original version in the Nikon D200.

Otherwise, maybe a 10MP 15x22.5mm EF-S sensor image could be compared to a 15x22.5mm crop from a Canon 5D or 1DMkII, at least allowing lens equalization and rough sensor technology parity.

[Ray]

There seems to have been a view shared by many photographers that fewer but larger pixels are better than more but smaller pixels. I think this might have begun with the introduction of the 11mp 1Ds when some people began to notice resolution fall-off towards the edges and corners of the frame and felt the need to upgrade their lenses.

I've never subscribed to the view that 11 or 12mp on a FF sensor was sufficient resolution for most Canon lenses and that there was nothing more to be gained by greater pixel density. In fact, I was one of the few contibutors on this site who urged Michael to show comparisons between the 1Ds and the D60, using the same lens and same distance to subject. This, however, was not a priority for Michael and the comparison was eventually made by others.

We now have a DSLR in the form of the 1Ds2 which has the same pixel density as the earlier D60. Interestingly, when comparisons were made between the 1Ds2 and the newer 5D, some people found it difficult to discern any higher resolution from the 1Ds2 in real world shots, which goes to show that a relatively small percentage increase in pixel count, such as 25-30%, doesn't count for much.

I didn't find any significant increase in resolution when I upgraded from the D60 to the 20D (didn't really expect to). The major improvement for me was better performance at high ISO. I imagine that owners of both the 20D and 400D will also find it difficult to discern any resolution benefit of the 400D in real world shots. However, those who have upgraded from the 6mp D60 to the 10mp 400D should notice a worthwhile increase in resolution at base ISO because a 66% increase in pixel numbers does count for something.

Likewise, if there is any doubt about the resolution benefits of the 1Ds2 compared with the 5D, there should be no doubt about the superior resolving power of a FF sensor with the pixel density of the 400D (26mp) compared with the 5D.

If anyone happens to own a D60 or 1Ds2 as well as a 400D, it would be interesting to see comparison shots with same lens (same f stop and shutter speed) at same distance to subject, at ISO 100. I get the impression that image quality in terms of dynamic range and noise at base ISO does not change much in Canon DSLRs with successive models, comparing images with the same number of pixels.

[jani]

I recall some older discussions on this board (2004, or earlier?), when people were discussing the possible merits of pixel binning.

I've largely come to subscribe to the view that as long as the light is actually captured and read out fast enough, the more pixels the merrier.

That is, the optimum number of pixels is when the pixel density exceeds the performance of any lens, theoretical or real, or possibly some point beyond that. That's when there's absolutely nothing more to be gained from increasing the pixel count.

Sure, we are aware that quadrupling the pixel count has a theoretical maximum benefit of doubling the resolution, but in reality has even less effect, but I don't feel that you have to pretend that this should quadruple your print area, or anything near it; just improve on what you've already got.

The main problem I see, is with the core physics of the sensor wells actually "catching" the photons (or waves, if you prefer). Is there a cut-off size when the noise per sensor area in square millimeters increases because of greater pixel density, and if so, what is that size? How does it affect highlight performance and shadow performance?

Given the same displayed size (print or on-screen), e.g. 24x36:

Would a 264 megapixel "full frame" sensor yield better results in any circumstances than a 66 megapixel one?

[John Camp]

Erwin Puts has an interesting comparison on his website between a ~10mp Leica M8 and a ~40mp Hassy H3D, in which he more-or-less quantifies the "quality difference" between the two by suggesting that the Hassy is about 1.2x better than the M8, with 4x as many pixels (he says, by the way, that the pixels are the same size in both sensors.) Puts does some interesting tests and has some unusual ideas, and I like reading him, but I'm sure some people would scratch their heads at the ultimately subjective 1.2x...

See here:
 
http://imx.nl/photosite/leica/M8_6/m8_6.html

Cute, huh?

Whatever you think of the Puts comparison, it's interesting to speculate where Nikon and Canon will go with their next sensors. Nikon people are generally expecting FF and about 20-22mp; that should result in a distinct increase in quality, with more pixels on much larger real estate. Canon people seem to be expecting 22-24. But a lot of people who had the 1Ds said that resolution differences beween that camera and the 1Ds2 were almost imperceptible, and some people with the 5D say the same...and also claim that the 5D is better at high ISOs. So...what is there to be gained at all by going from 16 to 24? Maybe not much...and since it'll be more pixels in the same FF space, will the new camera actually lose high ISO quality compared to the current one? Or has the technology improved enough that both resolution and high-ISO quality will increase?

(I know that Canon has no choice if Nikon goes up, because, despite what everybody says, the mp/pr war continues.)

JC

[Ray]

As Jani implied in his post, there comes a point where pixel density is so great that AA filters can be dispensed with. I don't believe that either the M8 or H3D employs AA filtering, just as the Kodak 14n didn't.

The removal of AA filters itself should result in a noticeable increase in resolution, but still with some aliasing artifacts at the lens' sweet spot and sweet aperture, except in the case of really high pixel count.

[jani]

But a lot of people who had the 1Ds said that resolution differences beween that camera and the 1Ds2 were almost imperceptible, and some people with the 5D say the same...and also claim that the 5D is better at high ISOs. So...what is there to be gained at all by going from 16 to 24? Maybe not much...

This isn't at all surprising to those of us who consider resolution in terms of angular resolution, rather than the rather ... unique ... view that "twice the megapixels = twice the resolution".

The 1Ds MkII has an angular resolution that's only 23% higher than that of the 1Ds, and with no other differences, that's the amount of extra detail that you might get.

The 5D has an angular resolution that's only 12.5% lower than that of the 1Ds MkII.

Going from the 1Ds MkII to a MkIII with 24 megapixels will again yield about 22% increase in angular resolution.

This is part of the point that Erwin Puts makes, although perhaps not in so few words.

and since it'll be more pixels in the same FF space, will the new camera actually lose high ISO quality compared to the current one? Or has the technology improved enough that both resolution and high-ISO quality will increase?

That might depend on how you measure high ISO quality. Will you measure it per pixel, or on a print of the same size as that you got out of your 1Ds MkII?

It is likely that per-pixel noise will increase (though perhaps not necessarily so, if other improvements are made) with smaller sensor wells.

[John Sheehy]

Would a 264 megapixel "full frame" sensor yield better results in any circumstances than a 66 megapixel one?
[a href=\"index.php?act=findpost&pid=99935\"][{POST_SNAPBACK}][/a]

Well, the FZ50 has the pixel density of a 223MP 36x24mm sensor.  Even at ISO 1600, you can see pixel-level detail, and if you resample to 50% and back again, the image loses a lot of detail (but still looks better than the 10D upsampled to the same subject size).  The sharpest Canon L lenses, such as the 300/2.8, 500/4, 180/3.5 Macro, 85/1.2 all perform admirably with stacked TCs, so there is no reason to think that they wouldn't do as well as the Leica lens on the FZ50.

For people shooting with Canon wide-angle lenses; there is less hope, as they are not particularly sharp (in the context of such resolution).  Finer pixels would probably make CA correction easier, however for FF cameras.

[John Sheehy]

As Jani implied in his post, there comes a point where pixel density is so great that AA filters can be dispensed with. I don't believe that either the M8 or H3D employs AA filtering, just as the Kodak 14n didn't.

The removal of AA filters itself should result in a noticeable increase in resolution, but still with some aliasing artifacts at the lens' sweet spot and sweet aperture, except in the case of really high pixel count.
[a href=\"index.php?act=findpost&pid=99973\"][{POST_SNAPBACK}][/a]

While the smaller pixels would result in higher noise at the pixel level, going beyond the resolution of the lens also means that it becomes much easier to distinguish noise from signal.  If it takes 7 pixels to go from black to white, any pixel that is much different than its neighbors is clearly doing this because of noise, and not because of signal.  Post-capture Low-pass filtering near the nyquist would kill a lot more noise than signal.

Speed and storage are the issues with Super-MP cameras, not noise per se.