[biz-engineer]

It makes about as much sense to debate that as where this thread is heading.

It's heading towards image processing. Signal processing architecture does matter for best performance.
I highly suspect that all cameras store the data read out from the image sensor directly (with help of processor) into buffer. The image processor takes data from the buffer memory , process it and store the data into memory cards (via a onboard or separated memory card controller). I couldn't find image processing diagram for Pentax, but I found something from Panasonic here: Signal Processing in a Digital Camera | Digital Camera Know-Hows | Digital Camera | Digital AV | Support | Panasonic Global


Based on the assumption that sensor data go into buffer memory before becoming a usable image files, pixel binning (or accelerator chip doing it's own Pentax things) done by a pre-processor changes the image processing performance, without needing more buffer memory. You are always better off shrinking data to useful data early on in the signal chain... less work down the signal chain.

[UncleVanya]

Ok. I went back and looked at the dpreview links provided; two things stand out:

1) the k1ii shot is darker even at iso 100. I’m not sure what this suggests but it means to me that the actual light received appears to not be the same which makes no sense to me. Any thoughts on what caused this difference?

2) I can’t see how to compare the images at 3200 to those at 100 directly. But even comparing only the pair at 3200 suggest I don’t see a lot more detail in one vs the other. I do see noise. It’s hard to fully compare as the initial iso 100 exposures don’t look identical either. I wonder what caused the slight differences in darkness in these images. I’m sure the “settings” of exposure (shutter speed, aperture and actual light levels) were held constant, but clearly one camera was producing a brighter base image.

[luftfluss]

I'm a pixel-peeper myself, but at some point one has to wonder if the reason they are pixel-peeping so much is simply because the rest of the image is not that interesting...

[Rondec]

The accelerator chip is a hardware based digital low pass filter (FIR filter) designed on a standard PLD (Programmable Logic Device). And it takes data processing load off the main processor, before the data are stored in buffer memory, that's why it's called "accelerator". The impact is big. For example, if you shoot zone I or II shadow frames in burst high mode with the Pentax K1 at ISO800, the buffer is full after 11 exposures. If you shoot the same shadow frames in burst high mode with Pentax K1 II, the buffer is full after more than 30 exposures , the a big step up in data efficiency.



Yes, there is a way to bypass the accelerator, set max ISO to 500 and shoot raw, and do auto level or curves tool in you favorite PP tool, if you want to chose yourself the noise/detail tradeoff.

---------- Post added 08-06-23 at 11:22 ----------


Noise reduction comes with detail reduction, as you can see when you compare ISO100 and ISO3200 exposures, whether noise is reduced by an accelerator chip or camera processor, or Silkypix, it's the same. By principle (of physics) unless using AI, pixel level noise can't be distinguished from pixel level image detail. There is another exception which is if the signal (or image signal) would be periodic (like telecom carrier signal), where you can use auto-correlation function (correlate signal with itself) to weed out noise while reinforcing the signal. Image being 2D non-periodic signal, it is not possible to auto-correlate, unless capturing a stack of exposures, lining them up, and them taking the median (same as performing auto-correlation on a periodic time sampled signal).

That is true in the K-1 II, but my understanding with the K-3 III is that the accelerator is doing something, even at iso 100.

To be clear, I'm not arguing that the accelerator doesn't reduce detail a small amount. Of course it does. I just don't think in the real world, where you didn't shoot a K-1 and K-1 II side by side and can compare the images directly to notice the difference, that you would be bothered by the softening.

If I am shooting landscapes, I always use iso 100, try to use electronic shutter, tripod, and often pixel shift. If I am doing snapshots at iso 1600, odds are good that fine detail isn't as crucial on the images and I can deal, either with my K-1 or K-1 II's output without any problem.

[biz-engineer]

I'm a pixel-peeper myself, but at some point one has to wonder if the reason they are pixel-peeping so much is simply because the rest of the image is not that interesting...

Pixel peeping is interesting to see how the noise looks like. With the K1 II, I may shoot pixel shift at ISO500 (instead of 100) or ISO800 (instead of 500), just to generate different noise look (like film grain). When I shoot PS at ISO500, the noise is very fine grained. If I want coarse noise I shoot ISO800 + PS enable because the noise become correlated by the accelerator which reinforces that noise, still false colors eliminated. If I want more granular noise look, I disable PS at ISO100 set shutter speed to underexpose and pull shadow in post, the noise look more like sand-paper. Some people play with film for film grain. It's also possible to play with ISO/exp. compensation to intentionally produce digital noise.

[Wheatfield]

That is true in the K-1 II, but my understanding with the K-3 III is that the accelerator is doing something, even at iso 100.

To be clear, I'm not arguing that the accelerator doesn't reduce detail a small amount. Of course it does. I just don't think in the real world, where you didn't shoot a K-1 and K-1 II side by side and can compare the images directly to notice the difference, that you would be bothered by the softening.

If I am shooting landscapes, I always use iso 100, try to use electronic shutter, tripod, and often pixel shift. If I am doing snapshots at iso 1600, odds are good that fine detail isn't as crucial on the images and I can deal, either with my K-1 or K-1 II's output without any problem.

It seems to me that if one is shooting at high ISO, there is going to be detail lost to noise. I don't see moderate noise reduction as removing detail, since the detail is already lost.

Obviously heavy handed NR can turn an image into a hot mess, but from what I've seen, that isn't what the accelerator is doing.

[BigMackCam]

Apologies to all for drifting even further OT, but...

Pixel peeping is interesting to see how the noise looks like. With the K1 II, I may shoot pixel shift at ISO500 (instead of 100) or ISO800 (instead of 500), just to generate different noise look (like film grain). When I shoot PS at ISO500, the noise is very fine grained. If I want coarse noise I shoot ISO800 + PS enable because the noise become correlated by the accelerator which reinforces that noise, still false colors eliminated. If I want more granular noise look, I disable PS at ISO100 set shutter speed to underexpose and pull shadow in post, the noise look more like sand-paper. Some people play with film for film grain. It's also possible to play with ISO/exp. compensation to intentionally produce digital noise.

Have you used G'MIC-Qt in GIMP? There's a really excellent "Add Grain" filter in the Degradations group... It's highly configurable, and with some experimentation you can get some very convincing grain effects. I'm not suggesting you use it instead of different ISO settings on your camera, but perhaps as an additional tool. I think you'd really like it...

[HoundFrog]

. I don't see moderate noise reduction as removing detail, since the detail is already lost.

What we (annoying, always complaining people) don't undertsand is, why on Earth this must be done to RAW files? You have your JPEG engine, do whatever you want there. Why mess with the source? Some brands apply lossy compression to RAWs. This is also annoying but at least justifiable, it saves space and maybe makes things faster.

But doctoring RAWs is just insulting.

[gatorguy]

The biggest problem here is not the noise in images, it is to understand what I wrote. It seems , even after 3 times trying to explain that the accelerator chip of the K1 II reduces noise and details as much as the image processor of the K1 reduces noise and details, still everyone is trying to demonstrate that the accelerator chip doesn't noticeably reduce details any more than it would without one.

Fixed.

So can we all now please get back to the original topic? It's fine to start a new discussion on noise reduction and the connection to detail if anyone feels driven to. It seems to be an area of some interest,
Not in this thread please.

[HoundFrog]

if we combine few pixels, say 4, to create a larger "pixel", we can increase the sensitivity of the sensor right ?

Well.. The sensitivity, truly, is a fundemental constant of a silicon (or any* solid-state) detector and cannot be changed. What we increase by this approach is the Signal/Noise ratio.


Now, the "sensitivity" in modern digital cameras is not really a sensitivity. It's gain, applied later in the pipeline. Usually it is done in software, i.e. it's just a multiplier. The greater the gain, the more visible noise becomes. With less noise we can apply greater gain ("ISO") while keeping results in the ballpark of "reasonable". So I would say yes, the approach increases what's camera mfgs are naming sensitivity.


Somewhat similiar trick is used in audio. There's a kind of Analog-to-Digital converters that are just a 1-bit but very-very fast ones. So they produce a high-rate stream of bits, which then is downsampled to a high-resolution (16-24 bit) stream with a "normal" sampling frequency.



* except for the avalanche diodes.

[lotech]

Well.. The sensitivity, truly, is a fundemental constant of a silicon (or any* solid-state) detector and cannot be changed. What we increase by this approach is the Signal/Noise ratio.


Now, the "sensitivity" in modern digital cameras is not really a sensitivity. It's gain, applied later in the pipeline. Usually it is done in software, i.e. it's just a multiplier. The greater the gain, the more visible noise becomes. With less noise we can apply greater gain ("ISO") while keeping results in the ballpark of "reasonable". So I would say yes, the approach increases what's camera mfgs are naming sensitivity.


Somewhat similiar trick is used in audio. There's a kind of Analog-to-Digital converters that are just a 1-bit but very-very fast ones. So they produce a high-rate stream of bits, which then is downsampled to a high-resolution (16-24 bit) stream with a "normal" sampling frequency.



* except for the avalanche diodes.

Thanks for the simple and clear explanation in 'English' ha ! when you say SN ratio I understand better.