Reader Question: Sensor Pixel Density, Oversampling
Related: Bayer, Bayer matrix color filter array, digital sensor, display pixel density, oversampling, pixel shift, reader comment, shutter, tripods and support, true-color sensor or image
See my Sony wish list and Sigma mirrorless wishlist and tripods and mounting wishlist.
David K writes:
I have a question that has been bothering me for some time about the physical size of the individual pixels per sense-size in cameras—
Is there an optimum density MP size, in relation to the sensor size?
(12 or 50 MP for FF? 40 or 100MP for MF?)
(Considering the present day Bayer matrix type sensor.)For example, one factor could be the decrease in sharpness because of increase motion sensitivity from having smaller and smaller sized individual pixels on highly dense MP sensors.
^ I don't know if this is a true statement. (It's why I'm asking this question.)(Leaving aside the 'blow-up' printing factor and it's need more MP the larger the print.)
Carried to the absurd, will we all have to carry 'granite-tripods' in order to get a sharp photo (because the pixels are too small and densely packed!). Obviously everything is relative, especially in this question but please enlighten us from your practical hiking photo experience.
DIGLLOYD: in general, 36/42/50 megapixels is optimal on full frame as of today. But only because of the current approach and current technical limitations—more on that below under the Oversampling discussion.
See how the 50MP Canon 5Ds R can beat the 24MP Leica M Monochrom. Accordingly, I claim that any 36/42/50-megapixel camera when downsampled to 24 megapixels can trounce any 24MP camera on a per-pixel basis, the only exception being extremely high ISO.
Any kind of movement during exposure causes blur. Let’s just use specific numbers to make that clear: assume ~5 micron pixels as in the Nikon D810. Then assume camera movement of 2.5 microns (tiny!)—that’s a large amount of blur equivalent to half a pixel. But if the pixels are 2.5 microns, then that 2.5 micron movement is an entire pixel, so yes indeed smaller pixels do matter in terms of the care needed to forestall camera movement. Camera phones are particularly at risk, but image stabilization compensates enough to make shooting feasible on all cameras that offer it.
The Sony A7R had that much and more blur from shutter vibration, causing a severe loss of image quality at certain shutter speeds and/or with longer lenses (since the lens moves too, not just the sensor).
No, we don’t need “granite tripods” and the proof of that is any Sony RX100 model or Micro Four Thirds camera of 16 or 20 megapixels—very high pixel densities and these cameras can make sharp pictures. But they do have an electronic first curtain shutter (EFC shutter), which means no shutter vibration, as well as image stabilization. On a tripod with an EFC shutter, there is zero camera-generated vibration (at least until and unless the curtain closes), so the main risk is vibration from wind, a passing train or truck, footsteps on a wooden floor, etc—all such things can cause micro vibrations that can affect sharpness.
BUT all that said, remember that for any given camera, if the sensor could be swapped out, the same projected image would be seen by the sensor; it’s just a matter of sampling frequency (equating to megapixels for any given sensor size).
On the flip side, larger and heavier cameras generally reduce high frequency movement, a fact easily seen with binoculars if nothing else. Handholding technique with mass-coupling is critical to sharp images at lower shutter speeds regardless of camera, which is one reason camera phones suck (arm’s length shooting, lots of movement).
True color sensor
A true color sensor, like the sensors in the Sigma dp Merrill, Sigma dp Quattro and Sigma sd Quattro cameras cannot necessarily resolve more spatial detail on black and white test charts, but out in the real world, such a sensor can deliver much higher real detail because there is no demosaicing process, that is, no guessing about the actual color and detail as with a conventional Bayer marix sensor. That is why Sigma can legitimately claims equivalence to a far higher pixel count than the nominal image resolutions imply, even if Sigma’s claims exceed (on average) what I’d rate as the difference (about 50% better in terms of pixel count). So I deem a 24 megapixel Sigma sensor to be roughly equivalent to a 36 megapixel Bayer sensor—a little less at times, and significantly more for strongly monochromatic color images where a single color dominates.
The implication is this: with fewer photosites needed to accurately capture color, the demands on lens resolving power are reduced since fewer photosites mean lower pixel density. This is a hugely important distinction when even on 50 megapixels one can see the limitations of even Zeiss Otus lenses start to creep in.
Oversampling
Oversampling means capturing an image at a much higher resolution than needed for the end result. Oversampling offers terrific promise in obtaining a result free of digital artifacts and with higher per-pixel quality.
One of many proofs of this that I have shown is how the 50MP Canon 5Ds R can beat the 24MP Leica M Monochrom. Accordingly, I maintain that extremely high quality 72-megapixel images would come from a 144-megapixel DSLR.
Thus, more resolution is not really the goal; higher per-pixel image quality has become more important in my view. The main limit on image quality today stems from limitations of the Bayer matrix demosaicing process: color and spatial moiré, jagged staircase edges, spurious resolution and false color, and a general inability to resolve color and textural detail anything close to what sensor resolution would suggest.
The Sigma sd and dp cameras offer self-evident proof of just how much quality is lost when using a Bayer matrix sensor (that is, by NOT using a Bayer sensor). If Bayer is to be used, oversampling is a mitigating solution. For example, a sensor of 144-megapixel outputting 72 or 36 megapixel raw images. Higher resolution (oversampling) used not for more resolution, but for much higher per-pixel quality.
- Oversampling
- Rumors of High Megapixels True or Not, Oversampling for Image Quality Makes Sense: 109 Megapixel Tulips via Sony RX100
- Sony A6300: Oversampling for 4K Video Quality
- Oversampling: More Pixels Always Better, at Least at Lower ISOs
- iPhone 5s vs Nokia Lumia 1020 Camera Phones (Oversampling and Downsampling)
- Olympus OMD E-M5 Mark II: Just How Good is the Hi-Res Sensor Shift Mode?
- Pentax K1 SuperRes Pixel Shift Mode
- Sony A7R II: Where is Pixel Shift?
The Nyquist-Shannon theorem says that the sample rate must be double the desired resolution. That is, if we want detail at 200 lines per millimeter, sampling must be done at 400 lines per millimeter (200 lp/mm). This is the idea behind oversampling. Add on the fact that a Bayer matrix camera samples red and blue in only 1/4 of the pixels and green in 1/2 of the pixels, and a lot more than 2X sampling is needed for accurate image capture with color images. The Pentax K1 solves the color sampling problem neatly with its Super HiRes pixel shift mode, but the Nyquist-Shannon theorem still applies to spatial resolution.
Oversampling is not a panacea: as pixels grow smaller, the balance of resolution, color discrimination, dynamic range and noise all change, with those latter items degrading as the pixels grow smaller. Sensor tech keeps improving however. Moreover, special pixel shift modes offer major gains in quality that ought to make even 72 megapixel images of extremely high quality without using oversampling (Pentax K1, Olympus pixel shift, as per above). Consider that with pixel shift technology, a 144MP sensor could approach the quality of a 36 MP sensor on a per-pixel basis (4X the pixels, but 4X the exposure). Maybe not right now, but within a few years. While pixel shift requires a still subject, there are many good use cases for it.
Accordingly my quick answer on the “how many megapixels” question is that the Sony RX100 V pixel quality is very high at ISO 100 and the pixel density of its sensor would work out to 144 megapixels on a full frame sensor. Thus the Sony RX100 V is proof that a 144-megapixel full-frame DSLR is a very reasonable answer, and 200 megapixels is not unrealistic. The problem: Sony has never offered a 35mm full-frame sensor with more than 42 megapixels. There are manufacturing challenges in scaling up a sensor, and clearly that goal will be elusive for a while.
