Nikon D800E vs Canon 5D Mark III: the Nikon D800E Optical Low Pass Blur+Unblur Filter Pack
Nikon D800, Canon 5D Mark III, optical low pass, anti-aliasing, sharpness, raw file
Regarding my D800E vs Canon 5D Mark III image sharpness comparison, a reader comments:
Very interesting to see how unreliable pure pixel count may be to conclude about perceived image quality. The explanation hypothesis is of course quite speculative. We don't know exactly the filter design in both cameras, we also don't know what both parties mean with RAW; it might also be pre-cooked a bit.
I think I remember that Canon had announced with the MkIII something like deconvolution image restoration, but I also don't know exactly.
Would be nice to have time to dig deeper in these things - yet the needs of practical picture taking with lenses are a bit distant from that.
DIGLLOYD: Agreed, it is hard to say what is the cause, and at some point one might just choose a camera brand one enjoys shooting, since the race for quality and megapixels remains highly competitive and thus that brand will eventually deliver the high-end ultra camera. Of course, one might reasonably ask whether Sony might get there before Nikon or Canon, say with a full-frame FNEX.
My hypothesis of the “glass sandwich” being involved is of course speculation, but with some basis for making the hypothesis: the Nikon D7100 with its much higher pixel density (56 vs 36 megapixels, equivalent density) does not seem to exhibit the D800E behavior.
The D7100 has no optical low pass filter, though Nikon does not go into details on exactly what glass is used on the sensor. See Zeiss 50mm f/2 Makro-Planar: Future Potential on ~56 Megapixel DSLR and How Does the Zeiss 135mm ƒ/2 APO-Sonnar Perform on a 56-Megapixel DSLR?.
Cooking the raw file
It seems likely that Nikon does something to “cook” the raw file, because the the dark tones are so clean in the near-black tones (as compared with Canon). Ditto for Sony NEX-7 and various Nikon DX cameras. But perhaps also with newer Canon EF-S cropped frame cameras.
Canon DSLRs in general have long been favored by astrophotographers because faint stars which might look like noise can be extracted by stacking 100 frames or more, something not possible with a camera raw that smooths away near-black pixels.
As for “cooking” the raw file with deconvolution, from what I see with many small sensor cameras I would be quite surprised if this is not already being done, and it’s possible that the Canon 5D Mark III employs such an approach also. Such an approach would explain what seems (to me) to be artificially high acutance on the Canon 5D Mark III images. But this too is speculation.
The D800E filter pack design
As an aside, my information is that Low Pass Filter 2 is actually bonded to the sensor, so that an infrared or full spectrum conversion effectively leaves a partial optical low pass filter in place, one that blurs in only the vertical direction.
As for the D800E “blur + unblur” filter pack, a filter that blurs followed by one that unblurs would seemingly have to be in absolutely perfect alignment for the claim to bear out—and any misalignment on the micron level would presumably make the camera perform somewhere between the D800 and the D800E, a sort of weakened optical low pass filter
Trond S writes:
I have read somewhere, Clarkvision maybe, can't find the source right now, that Canon adds a "DC bias" equivalent to digital 1024, before the ADC.
The thermal noise in the pixel readout amplifiers, and the AFE of the ADC will act in the same way as in dithered audio ADCs: increasing the effective resolution of the ADC.
The added white noise will "lift" the faint image details out of the noise, normally below one LSB, to above one LSB, as long as the noise is a little more than one LSB on the ADC.
Autocorrelation will recover these faint detail, while random white noise is filtered aug in the autocorrelation. This is why astro photographers favours Canon.
Nikon does not do the same, meaning that the noise is "clipped" at zero, so that negative part of the expectance value of the noise does not get into the raw file.
Clipping the RAW pixel value at zero effectively gives half a stop of noise cancellation since half of the noise is "filtered" in the clipping process away.
DIGLLOYD: all cameras have a black level setting, and RawDigger makes this explicit, but I’m unsure if this is the same idea.
As far as detail, the dark areas and ADC to digital conversion do not explain the results in very bright areas.