[lotech]

As far as I understand the basic principle of digital photography, for a given sensor size, FF for example, sensor of higher resolution (pixel density) has smaller pixel size, and each pixel receives less light, than a lower resolution sensor of larger pixel size, so that if we combine few pixels, say 4, to create a larger "pixel", we can increase the sensitivity of the sensor right ? it seems not that simple, I don't know of any camera has such option to adjust between resolution and sensitivity, reduce resolution does not change the sensor usage, but simply downsizing, or down sampling the image, so why we can not do that, or how can we do that ?

[pschlute]

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

No. Combining 4 pixels to make 1 pixel is downsizing. It does not change the exposure. 4 pixels downsized to 1 pixel will have the same brightness as each individual pixel (assuming a uniform exposure of the original 4)

[BigMackCam]

As far as I understand the basic principle of digital photography, for a given sensor size, FF for example, sensor of higher resolution (pixel density) has smaller pixel size, and each pixel receives less light, than a lower resolution sensor of larger pixel size, so that if we combine few pixels, say 4, to create a larger "pixel", we can increase the sensitivity of the sensor right ? it seems not that simple, I don't know of any camera has such option to adjust between resolution and sensitivity, reduce resolution does not change the sensor usage, but simply downsizing, or down sampling the image, so why we can not do that, or how can we do that ?

This article might be useful reading:

Binning

Note the difference between binning with CCD and CMOS sensors.

The math is beyond me, but the principle makes sense...

[lotech]

This article might be useful reading:

Binning

Note the difference between binning with CCD and CMOS sensors.

The math is beyond me, but the principle makes sense...

Thanks for that, I can barely follow ! I think the current sensor technology is good enough to produce high quality image at very high ISO, so developers won't bother to do binning so we have not heard about that.

[beholder3]

As far as I understand the basic principle of digital photography, for a given sensor size, FF for example, sensor of higher resolution (pixel density) has smaller pixel size, and each pixel receives less light, than a lower resolution sensor of larger pixel size, so that if we combine few pixels, say 4, to create a larger "pixel", we can increase the sensitivity of the sensor right ? it seems not that simple, I don't know of any camera has such option to adjust between resolution and sensitivity, reduce resolution does not change the sensor usage, but simply downsizing, or down sampling the image, so why we can not do that, or how can we do that ?

The answer to your question depends on the degree of sillyness in use cases:
a) if you look at full photos the only thing that has meaningful impact is sensorsize as this defines the amount of light in total captured on the full photo. Pixelcount and pixelsize are utterly irrelevant.
b) if you stare at minuscule bits and pieces of the photo AND you do so with different magnifications depending on pixelcount ("100% crop") then, yes you will see differences. Then you best buy a 4 pixel Camera for lack of noise.

Keep in mind your eyes do not resolve more than about 6 MPx when watching a photo from a 43mm on FF FoV.

[biz-engineer]

Image quality and sensitivity don't change when grouping pixels. Pixel binning is good for data processing efficiency and storage. Meaning, at low ISO noise is low, signal content is high, so it makes sense to process all pixels. At higher ISO, noise become predominant in small pixels, and it is not efficient to feed the image processor with data that mostly contain noise. By grouping pixels first , image data fed into image processor are more information rich. Binning is one method. The accelerator chip of Pentax achieve the same aim of data efficiency, by cutting noise on raw data before it reaches the image processor, while still having all pixels albeit with less noise and less detail.

[mlag]

averaging 4 pixels might improve/eliminate noise a bit , but the electronics of each pixel will not get more signal...(become more sensitive).., this is more playing on noise reduction than increasing sensitivity....

in theory a larger pixel catches more light , so it would be easier to develop a more sensitive sensor using larger sensor pixels, but unfortunately the fine chemics of sensor design do not scale linear..., so it is not safe to assume that a 4x bigger pixel in a certain sensor production technique just like that will be 4x better, it might even not even work .... (like you might also scale up also noise levels or interference to neighbours...), so a new sensor design would be needed.
but most new sensors designs zre nowadays developed to yield a certain , typically higher resolution for a given minimum sensitivity allowed by the proction technique. There are not many low resolution, high sensitivity designs if you ask me....market demand is not going that direction.

[Bob 256]

Astro cameras using binning all the time. It can be selected in the software that drives the camera (2x2, 4x4, etc.) Most regular cameras don't have that ability but you accomplish pretty much the same if you combine pixels in post-processing through resizing the image (changing the ppi value for a given sized image).

Larger pixels gather more light which results in less noise. That can amount to a "sensitivity change" since the signal can be amplified more without changing the noise floor, but the "sensitivity" of the sensor really doesn't change as biz-engineer mentioned. Combining pixels through binning or in post-processing does almost the same thing. Both decrease resolution for a given lens since larger effective pixels mean there will be fewer sample points (binned pixels also reduce sample point count), and the noise is averaged to a lower value.

Interestingly, in astro cameras, the reason for binning is to reduce the effects of oversampling point images (stars) which can cause artifacts in the image. If a star covers only one pixel, slight position changes can cause large fluctuations in its recorded brightness. By using several pixels as one (binning), that issue is reduced, at the cost of actual resolution (two really close stars may appear as one with binning). Also, with binning, file sizes are reduced from the git-go, so memory goes further (a 2x2 binning cuts the memory needs by a factor of 4).

[gatorguy]

and less detail[/B]

I've seen that claimed over and over. What I've never seen is photographic evidence sourced from a users image. I have cameras both with the accelerator and without, and even those with aren't applying it across the entire ISO range (unless it's from Canon). I don't find any obvious detail loss. If it's there it's gotta be darn subtle, to the point it can't be demonstrated except in some very extreme or patently oblique circumstance.

Have you seen this for yourself in your own images, or basing the comment on what you've been led to believe by "reviewers", or perhaps just accepted as common sense without needing proof?

[James O'Neill]

Put simply yes, it's called binning and there's quite a technical article on it already linked to.

Let's say you have a 4000x6000 image and you downsize it to 2000x3000 using a very simple algorithm.
The simplest would be to ignore the odd numbered rows and columns and keep the rest.
The next simplest would be at each even row & column add that pixel and those 1 to the right, 1 below, and 1 below and right, and divide the result by 4. What happens if you DON'T divide by 4? That's binning.

Or putting it the other way, if all the sensor readings are low numbers and you can't make out the details because everything is dark, you might multiply all the numbers by let's say 4. Now you can see the details but the random variation between neighbouring pixels gives you something quite noisy. If you average groups of 4 pixels you'll average the variation - reducing the noise.

If any group of four pixels is greater than 1/4 of maximum allowed brightness when they are multiplied, they'll blow out in the second case, in the first case having done the addition one can set the brightest and darkest point.

IIRC one of the early phase one cameras had this built in.

[gatorguy]

Put simply yes, it's called binning and there's quite a technical article on it already linked to.

Let's say you have a 4000x6000 image and you downsize it to 2000x3000 using a very simple algorithm.
The simplest would be to ignore the odd numbered rows and columns and keep the rest.
The next simplest would be at each even row & column add that pixel and those 1 to the right, 1 below, and 1 below and right, and divide the result by 4. What happens if you DON'T divide by 4? That's binning.

Or putting it the other way, if all the sensor readings are low numbers and you can't make out the details because everything is dark, you might multiply all the numbers by let's say 4. Now you can see the details but the random variation between neighbouring pixels gives you something quite noisy. If you average groups of 4 pixels you'll average the variation - reducing the noise.

If any group of four pixels is greater than 1/4 of maximum allowed brightness when they are multiplied, they'll blow out in the second case, in the first case having done the addition one can set the brightest and darkest point.

IIRC one of the early phase one cameras had this built in.

Pixel binning is of course why a smartphone has a sensor advertised as 48mp, but produces images at 12mp.

[biz-engineer]

I've seen that claimed over and over. What I've never seen is photographic evidence sourced from a users image. I have cameras both with the accelerator and without, and even those with aren't implying it across the entire ISO range (unless it's from Canon). I don't find any obvious detail loss. If it's there it's gotta be darn subtle, to the point it can't be demonstrated except in some very extreme or patently oblique circumstance.

Simple. When there is no accelerator chip, it's the image processor doing the noise reduction job at higher ISO settings. So if you compare image from K1 and K1 II, there is little to no difference (beside the difference between the hardware filter response in the accelerator chip, and the software filter by the processor). In K1 II the accelerator takes the noise reduction work load off the main processor, hence its name "accelerator". To find out, take the same picture, raw, at ISO100 (underexposed 5 stops) and ISO3200, then pull the 100 ISO image 5 stops in your raw developper, disable any noise reduction/sharpening, compare images for noise. Normally, you'll have more noise and more details in the ISO100 image, evidence that the ISO3200 capture went through a noise reduction filter, even when ISO NR is disabled, with a K1 (without accelerator).

[gatorguy]

Simple. When there is no accelerator chip, it's the image processor doing the noise reduction job at higher ISO settings. So if you compare image from K1 and K1 II, there is little to no difference...

Bingo! So it's as I expected; you can't see any loss of detail yourself.

You need to be more careful in claiming the accelerator chip is noticeably reducing detail. That's how assumptions get portrayed as fact and take on their own life. Maybe in some edge case it might... maybe... but if you can't find where it's been the source of any loss of detail in any image of yours, and being the professional you are I'm certain you've looked for it, then in practical terms it's not observable, and you have no evidence for it. I know that's been my experience.

[Lowell Goudge]

averaging 4 pixels might improve/eliminate noise a bit , but the electronics of each pixel will not get more signal...(become more sensitive).., this is more playing on noise reduction than increasing sensitivity....

in theory a larger pixel catches more light , so it would be easier to develop a more sensitive sensor using larger sensor pixels, but unfortunately the fine chemics of sensor design do not scale linear..., so it is not safe to assume that a 4x bigger pixel in a certain sensor production technique just like that will be 4x better, it might even not even work .... (like you might also scale up also noise levels or interference to neighbours...), so a new sensor design would be needed.
but most new sensors designs zre nowadays developed to yield a certain , typically higher resolution for a given minimum sensitivity allowed by the proction technique. There are not many low resolution, high sensitivity designs if you ask me....market demand is not going that direction.

the method of pixel binning as described for systems using it, with CMOS sensors is that the combination of pixel putput happens before the A-D and electronics, and therefore does NOT include the noise of the electronics for each pixel,. BUT that is a very special camera, not a DSLR

[biz-engineer]

Bingo! So it's as I expected; you can't see any loss of detail yourself.You need to be more careful in claiming the accelerator chip is noticeably reducing detail.

Maybe I wasn't clear. The accelerator causes loss of detail, but not more than when it's the image processor filtering the noise (and details). When you compare images with and without accelerator, it doesn't mean that there is no loss of detail. It's like if you heat a glass of water with the microwave and with a kettle, you notice the temperature with microwave heating is the same as the temperature with kettle heating, that doesn't mean microwave didn't heat the water. Less detail at ISO3200 vs ISO100, with or without accelerator, it's the same, the camera reduced noise and by reducing noise it also reduced details. Same with pixel binning, noise is reduced after binning, detail is also reduced. It's so simple I suspect you are playing to trigger more responses.