The Nikon Z pdaf-pattern issue

[32BT]

Most of us are familiar with the concept of demosaicing or debayering. Demosaicing is a bit of a misnomer as it suggests the imagedata was fully intact and complete initially, and that this complete data was scrambled into a mosaic and only needs unscrambling to become a complete image again.

As we know, this is not what happens, instead we have a subset of samples from which we need to extrapolate additional samples to recreate something that resembles the original image. The subsamples are ordered in a regular grid which allows a demosaic algorithm to infer the remaining samples from partial data at each sample location.

Fuji has been creative with this concept in the past and once came up with something called a superCCD which basically interspersed smaller sensels between larger sensels to increase the dynamic range potential by having the less sensitive smaller sensels take over where the more sensitive larger sensel may already have clipped.

Which brings us to the new mirrorless offerings with PDAF sensels. They apparently tend to be slightly darker than the other sensels which may result in a visible pattern under some (currently rare and extreme) circumstances IF (and that's a big if for emphasis) the raw converter does not account for the difference.

If the raw converter however does account for the difference, it suddenly has some interesting additional data at its disposal when recovering the original image. This additional data namely extends the available dynamic range.

Therefore, contrary to what is suggested in this article over at dpreview, the dynamic range of the new Nikon Z isn't less than its brethren but actually has more dynamic range, provided of course that the raw conversion process makes use of the additional “information” or at the very least should compensate for it.

[Telecaster]

One of the early Canon DSLRs—the 20D, I think—had a sensor readout issue that resulted in obvious banding in monotonal regions of images. Once folks figured out the nature of the issue it proved easy to correct during the demosaicing process. I'll bet this patterning issue gets handled similarly. I can see the apparent reduced sensitivity of the PDAF photosites, assuming it's uniform or at least follows some discernable logic, being useful for blown highlight reconstruction if nothing else.

-Dave-

[JaapD]

You’re stating “IF the raw converter does not account for the difference”. I very much expect there is a gain adjustment on the PDAF pixels in-camera before the RAW data gets generated.

Regards,
Jaap.

[bassman51]

Assuming that the PDAF pixels actually collect less data (because they're "darker"), then there's less data.  Not more data.  When that pixel is in a low light region, it will stay at full black longer, even when the normal pixels have registered a dark grey.  Recovering these greys will require interpolation, or "creating" data that isn't there.

More generally, if this scheme ("darkening" pixels) worked, then manufacturers would have adapted it in the absence of PDAF on-sensor.

[JaapD]

  Recovering these greys will require interpolation, or "creating" data that isn't there.

Not correct. The PDAF pixels only require a gain correction, no interpolation. The only disadvantage here is a slightly worse s/n ratio but percentage wise it does not deteriorate the quality of the image in general.

Regards,
Jaap.