[BigMackCam]

@reh321 and @normhead - thank you both, you're very gracious

We mods try to keep things level-headed for everyone. I'll be the first to admit I don't get it right all of the time, but all of us do try our very best...

[MJKoski]

If Sony had sold 42MP sensor to Ricoh, there would be no need for artificial enhancements considering high ISO. But that did not happen.

Mk2 is a bit questionable release when not looking through Pentax-glasses. ISO100 gave already superb quality with K-1 and ISO3200 (practical limit of the 36MP sensor) was on par with other brand offerings. Time move on and leave old sensors behind.

[biz-engineer]

Time move on and leave old sensors behind.

With all good camera we have available, I think it's time to use and enjoy what we have.

[awscreo]

With all good camera we have available, I think it's time to use and enjoy what we have.

Amen!

[xandos]

my point was that "outside the sensor" it is impossible for further noise to be added. Once the data is available in digital form -- and it is outside a Sony sensor -- then there cannot be any further "sources of noise".

Sorry for the late reply. As you seem to be quite knowledgeable about these things, I have a question for you. Specifically because I think the reverse is true: as long as the data lives as small voltages inside electronic systems, there is always a noise source. Johnson-Nyquist noise immediately comes to mind. How important that is depends on exactly how the data is digitized. Do you have solid reasons to assume such noise can be completely neglected?

---------- Post added 04-29-2018 at 09:06 PM ----------

But what about information-preserving or even information-enhancing NR? An NR algorithm that removes dark current, corrects for pixel-to-pixel sensitivity variations, or removes cross-talk both reduces noise and enhances information in the sense that the resulting RAW file is a more accurate representation of the light levels in the original scene.

This is a very interesting point. However, in practice I don't think such a miraculous algorithm exist. Could you give a concrete example of any algorithm that enhances information? Or even one that reduces noise while not decreasing information? I think it can probably be mathematically proven that any NR algorithm that does not rely on magic can at the very best preserve the information. And with that I actually mean that the amount of detail/information lost by the very best NR algorithm might be nonzero but negligible. I'd be very happy to be wrong though. I'll try to see if I can come up with some kind of simple proof of this statement.

To slightly formalize this, lets take the definition of noise to be the additional random amplitudes added to the signal due to things such as poissonian shot noise. Things that are perfectly reproducible on each shot (such as the different sensitivities of pixels) are not noise according to this definition as they are perfectly predictable, and should obviously be compensated for.

[stevebrot]

Based on these 2 images, to my eyes the K1II is obviously better---the dust fibers are showing up better. But there's also to me some obvious exposure difference. The color is also different....

What I am not seeing with my eyes is a problem with detail being erased....

I have been looking a the DNG for the ISO 800 images and there is a reason for the color and contrast differences. I won't go into details (would be major thread derail), but the differences do make it hard to assess whether point-size elements are being removed by noise-reduction. I could not find anything obvious.


Steve

[stevebrot]

OFC, if I knew back then that underexposed ISO100 would have given another kind rendition I would have used it.

You would not have liked it...color shifts and with slide film, massive loss of shadow detail...


Steve

[derekkite]

That would be completely fine, of course.

Note that I consider the possibility that the accelerator chip only performs information-preserving transformations. The jury is still out.


In theory, that's a possibility. Given the nature of Sony's "all-in-one" sensors, it is highly unlikely. Of course one can consider all kinds of possibilities but I'd love to see some shred of evidence which lends credibility to the idea that the accelerator chip has some kind of privileged access to the sensor. I'll eat my words, if anyone comes up with such evidence. To be clear, I'm not talking about data that could very well be packaged up for post-processing (such as a dark frame).

If Ricoh wants to be taken seriously regarding "kind of signal processing which cannot be obtained by just software processing mechanism" then they must mean "cannot be replicated by post-processing in principle" not just "you cannot do it in post-processing because we withhold the data to do so from you".


That is a possibility.

However, the better approach would then be to make the in-camera processing optional and provide some way to deliver the data to the user as part of the image or separately.

As an analogy, let me point out that the long exposure NR is not mandatory on the K-1. There are cameras (IIRC, even from Pentax) that make a dark-frame subtraction mandatory. Instead of enforcing this kind of processing, Ricoh allows you to take an alternative path (e.g., take your own dark frames and substract them in post-processing) because there are sometimes advantages to this path.

In the same spirit, I believe any kind of image processing that may imply any kind of disadvantage should be optional. Again, I do not know that the accelerator processing is destructive in any way and I'm not making claims either way.


That is most certainly not true. There has been a time when demosaicing algorithms made big strides and re-processing files certainly made sense. Regarding the subject at hand, if the NR is destructive in any way (and the OP "knows" that it is, doesn't he?) then it is almost certain that future NR methods could do a better job with the unmodified data.


True, hence my suggestion to provide the calibration data separately (instead of baking it in somehow).

We agree that if there is only one (non-destructive) way the calibration data can be used to optimal effect then applying the respective algorithms to the RAW data is just fine. I have my doubts that this is the case with the K-1 II's processing, but I'm open to any kind of outcome of future investigations. As I said before, I don't care that much; my K-1 is all the camera I need and want. I would prefer Ricoh not to make any kind of destructive processing mandatory in future cameras, but to be frank, I think there is next to 0% influence we have on the matter. I'm confident, though, that Ricoh will make competent decisions.


Yes, that's why I suggested it would be ideal to have a PixelShift image for comparison. That wouldn't amount to the "truth" either, but it would be much better than comparing a pushed ISO 100 shot with an ISO 800 shot.


Yes, DxOMark detected certain noise-reduction (smoothing) RAW cooking using such analyses.


Quite possible. I can't say.

I think it is time for me to bow out of this discussion. I'm really not that bothered what the exact K-1 II story is (it is a fabulous camera either way) and before proper comparisons are made, there cannot be any real progress.

So what happens to make the same chip produce higher iso? Iso is essentially amplifying the signal. Amplification makes flaws worse.

There is something going on in that chip.

And I wouldn't disregard the values of dsp. Audio engineers have worked for decades to produce a microphone that reflects what we hear with our ears reliably in almost any environment including noise, wind, high volumes, feedback etc.

I suspect the results will be similar to what I experience with the K1. Some situations are amazingly clean at high isos, others are awful at lower isos. I'm expecting the K1 II to be awful in some situations and amazing in others. I intend to find out.

[stevebrot]

That is bad performance and shows up in A3 sized print.

Have you done the comparison or are you just saying this? I understand your complaint and personally don't really like the K-1II's approach to noise reduction, but large print often acts as the great equalizer in that pixel-level detail is long-gone and tonal progression is as much a property of the printer driver, papers, and inks as the best edit on an additive display.


Steve

[stevebrot]

With the k1 sensor I really don't know why they would be doing mandatory NR at iso800

I agree and would prefer that they provide an ON/OFF setting or some provision to designate the low boundary for NR.


Steve

[normhead]

I agree and would prefer that they provide an ON/OFF setting or some provision to designate the low boundary for NR.


Steve

Ya, I guess you have to just point out the engineers at Pentax found reason for doing it the way they did. That not every one can comprehend what that might have been is understandable. But that just doesn't mean the Pentax engineers were wrong. It just means some people don't get it.

[stevebrot]

Ya, I guess you have to just point out the engineers at Pentax found reason for doing it the way they did.

It is possible and perhaps even likely that making the feature configurable was part of the original specification, but that the capability was not not "mature" enough by release time.* With any luck, something will sneak into a future firmware update.

It just means some people don't get it.

...or there are differences in taste and as we all know, there is no accounting for taste. There is also the matter that some people's taste is beyond satisfaction.


Steve

* Deferred features the norm when development schedules are tight and post-sale modification is part of the business model.

[MJKoski]

Have you done the comparison or are you just saying this? I understand your complaint and personally don't really like the K-1II's approach to noise reduction, but large print often acts as the great equalizer in that pixel-level detail is long-gone and tonal progression is as much a property of the printer driver, papers, and inks as the best edit on an additive display.


Steve

I have so large monitor that I can simulate full sized A2 print on it in natural size and observe how it would look like when viewed at different distances. No, I have not yet printed the rabbit. I have a good idea what happens when it is printed as I print large quite often and A3 is the standard tryout size before going for the big one. The larger the print, the messier the outcome from this kind of filtering. Especially when lens used gives sharp outcome @ 100% crop size. Ink-printing does smooth out the outcome but it seems to happen with slightly oversharpened edges, not smeared textures.

[reh321]

I agree and would prefer that they provide an ON/OFF setting or some provision to designate the low boundary for NR.


Steve

How would you feel if this were a Sony project, so they had the capability to build all this into the sensor?


I'm not an electronics engineer - nor do I play one on TV, but I'm guessing that a switch would add noise back into the image

[normhead]

I have so large monitor that I can simulate full sized A2 print on it in natural size and observe how it would look like when viewed at different distances. No, I have not yet printed the rabbit. I have a good idea what happens when it is printed as I print large quite often and A3 is the standard tryout size before going for the big one. The larger the print, the messier the outcome from this kind of filtering. Especially when lens used gives sharp outcome @ 100% crop size. Ink-printing does smooth out the outcome but it seems to happen with slightly oversharpened edges, not smeared textures.

My monitor is my 4k TV. It's 55 inches, and 3840 x 2160.

The only actual research I've seen done on this issue suggest people can't tell the difference be tween 100 DPI and 300 DPI at a normal viewing distance, so you're going to have to be more specific to make an impression here.

What size comparison prints have you done, and did you have a blind test to show which image was preferred by a number of people. I'm not really interested in generalizations.