This article should be read by everybody

Started Dec 12, 2013 | Discussions thread
K E Hoffman
K E Hoffman Veteran Member • Posts: 5,103
Hard to get some to think Digital...

Nordstjernen wrote:

K E Hoffman wrote:

When I scale up.. If there is one red I a Sea of Gray.. it won't turn it into a big blog of 4 red pixels.

What if the single pixel is noisy, as noise is distributed at random? Then you might get a cleaner result with data from four pixels.

A 24 MP image from an APC of the same generation will start to get muddy when we go above ISO 6400.EVEN SCALED>... it will look much better on the FF 24MP.. same scaling.. difference? Larger pixels more signal per pixel and across the sensor vs the the pixel and across the sensor.

Now you are comparing sensors with different size, one capturing much more photons that the other one. The difference is NOT tied to pixel size, but sensor surface area and light-gathering power.

Small or large pixels - with modern sensors of THE SAME SIZE the amount of noise will be about the same. This is what the writer of the article is trying to tell you. And this is what is confirmed by test results and real world work.

I wanted to be DONE with this... But the problem here is you think Analog in a Digital Word.

Lets assume that the sensor media the reacts to light has the same noise per SQ micro meter.

Like it was a sheet of film. Fine.

When I slice it up I lose surface area. Micro lenses have helped a lot. Moving most of the circuitry to the back has helped to that you are not burying the each new sub divided pixel in support circuits.

So say if some how you moved it all to the back.. so the front of the sensor is just light reactive semi-conductor and insulator between. When you slice it up 24 million times that insulator eats at your total sensor area.

(We see this in the grocery store. The price stays the same but of the user looks the bottle of spagettit sauce loses goes tom 24 OZ to 23.5 Oz. Not much.. but over a million units that is 500K OZ of sauce saved real money)

You sensor when divided just lost some of it total light gather ability. No to my knowledge there are other losses of AREA depending on the design. The Micro lenses do help catch the light grabbing it from areas that are not light sensitive to and focusing it to the light catching semi conductor.

BUT smaller sensor area means lower possible voltage off the sensor site as it has less room to hold freed electrons. The smaller we get the close we get to absolute limits that even better sensor media can't improve on.

So when we doubled the Sensor Pixels.. we doubled the lost area of light sensitivity and reduced the overall electron holding capacity. But of course we reduced the media area that generates random in media noise. If we stopped there it would be close to a push.

But every single SENSOR site needs to have transistors etc to managed the reading of the voltage, circuit paths to carry the voltage to the A/D converters which are also a mass of transistiors etc.

All of that has a fixed amount of noise it can add that does not go down when the light collection area goes down. This is talked about in Astrophotography as BIAS Noise.. IT is what in older cameras often shots up as banding.. as the sensor is read out in rows etc. and electronic processes cycle in time etc. We will note that in the last 8 year complaints of banding noise have gone down.. because it is the place the design can reduce noise the easiest. Sony put the AD converters on the sensor to shorten the path to them when the value gets locked and electronic noise will no longer alter the value for a sensor site.

The key is NOISE does not happen at the SENSOR level. It happens PIXEL by pixel.

Sensor A has fewer MP and larger light sensing media. (arbitrary numbers for example)

Light reading range of sensor is 0-1000 electrons (number would be much larger)

Native media noise level is 0 - 50 electrons

Fixed circuit noise is 0-25 electrons.

That means my maximum noise level is 75 or 7.5% of the max signal. So if the light drops 2 stops from max.. I get 250 electrons and a range of 0 to 25% noise.. So a dark pixel could end up being 25% brighter than the ones next to it.

Sensor B has double the MP

Max voltage of the sensor area is now 0-480 electons (lost 20 for cost of subdividing and insulating)

Max native noise is 24 electrons

Fixed Circuit Noise is 0-25 electrons

That means the maximum noise is 49 electrons or 10% of Max signal

IF I capture light 2 stops less than max I am getting 120 electrons with a max noise of 49 electons

So a this pixel is now about 40% brighter than the other pixels.. ( you see this in the smaller sensor noise patterns the reds look brighter. So the same level of noise makes a brighter pixel in the new lower relative voltages.. And even when I scale that is MORE red to be averaged in to the other pixels.

Now with design.. can improve on the fixed noise.. and design can find ways to reduce the loss of smaller pixels so maybe in the next generation we can do 35 MP and still have max noise level of 10% but but that could have also been done as the same sensor and maybe taking the max noise per sensor down to 5%..

Problem is when I am marketing a camera.. I can up "36 Mega Pixels" on the box and ad or I can say 3% lower noise etc etc.

Keep I in mind as we raise ISO one thing that happens is the sensor gain is turned up amplifying the signal to the AD... this also amplifies noise.. the darker the area the more we see the noise.

This gets really made clear in a Maxium case of top ISO on 24MP a77 and 16MP A57 (not in normal light these cameras behave the same up to about 1600 ISO. the noise is brighter at the max

What about scaling?

Scaling will not fix this.. the ROSE and even the paper clips have lost detail that is not getting averaged back in.

Here is the scaled to 16 MP 66% of original PS CC 64 Bit Bicubic. sorry the scaled 24 MP image did not become as good as the scaled 16 MP image again this is the extreme end to deal with the over proof vs the fact that at lower ISOs scaling will help cover the per-pixel noise.

A77 ISO 1600 scaled to 66% to match the 16 MP camera

just for the DXO 8MP people.

Here it is both scaled to matching 8MP images sorry the 16 MP sensor does better.. at very high ISO where the noise cost of more MP is amplified.

Because in the end you can talk about surface area of the sensor and combined electron carrying capacity.. but in the end this is DIGITAL.. and noise happens not to the sensor but at each pixel..

Debayering does spread the noise.. around some which helps..

Scaling can help.. but when the voltage of the smaller sensor gets overwhelmed with noise image data is compromised beyond recovery. and it starts to affect the surrounding pixels as the debayering factors the bright pixel color into what the other pixels saw.

That's why over exposure (as long as you don't blow highlights.. helps because you get more electrons into each site to work against the noise base)

And if you look above you can see with your own eyes Scaling this will not help

Diagrams to help show what I am talking about where the smaller sensor has less signal vs fixed noise...

[DIAGRAM colors have nothing to do with noise colors.. just used to show sources]


Subdividing a sensor using similar technology and design DOES increase NOISE.. Just a fact. Smart people each year find ways to reduce noise so the smaller pixel has a better Signal to Noise Ratio and the noise cost of More MP is reduced..

Does the noise price undo the value of more MP?... in all but the extreme cases.. I don't think so. I am less worried about 1:1 pixel sensor noise than I used to be.. because scaling does help and PP software gets better. But you can't make up facts.. buy ignoring actual changes that happen in a sensor when it is subdivided and thinking of it as a theoretical collective when noise increase happens per pixel.

-- hide signature --

K.E.H. >> Shooting between raindrops in WA<<

 K E Hoffman's gear list:K E Hoffman's gear list
Canon EOS 450D Nikon 1 J1 Sony a6500 Sony E 18-200mm F3.5-6.3 OSS Sony E 50mm F1.8 OSS +3 more
Post (hide subjects) Posted by
Keyboard shortcuts:
FForum PPrevious NNext WNext unread UUpvote SSubscribe RReply QQuote BBookmark MMy threads
Color scheme? Blue / Yellow