Why and how small pixels influence the image quality

Started Mar 31, 2012 | Discussions thread
bobn2 Forum Pro • Posts: 53,151
Re: Part 2.

DSPographer wrote:

The primary point I am arguing with is your insistance that only strict transistor scaling is used as you define in proposition 2 below. This is not an argument that you never made.

I never made it in the sense that only strict scaling should be used to produce pixels of different sizes, which seems to be the sense you wish to interpret it. As I say repeatedly, since however often I say it, it never seems to go down, if you wish to separate out the effect of pixel size as as a variable, and consider that by itself, then strict scaling is the only assumption that makes sense. Otherwise you get into an open ended set of speculative assumptions about which different configurations a pixel might have, and there is no point making the statement in the first place. Seems to me that your arguing with me because I choose to make a clearly defined statement, whereas you never seem to argue with those that say, without any foundation and without any specification of what they say, that small pixels have higher noise and less DR. The fact that you attack the well specified, well reasoned and evidenced statement and choose to leave alone the ill specified, unreasoned and unevidenced statement to the contrary is very puzzling.

How about just plotting the read noise per area versus pixel size for a wide selection of sensors. That will show which scaling rule really applies.

Hmm, a statistical observational approach to engineering. Interesting. The point is that the uncontrolled variables in that set are likely to dominate and scaling effect.

Well, you define a scaling effect that is not seen in practice.

I claim that it is seen in practice. As a trend, smaller pixels are clearly producing less read noise.

Please suggest a way to compare real-world sensor results to see how pixel scaling affects read noise per area.

Well one way would be to take a set of similar sensors and plot read noise against pixel size. These are the Canon sensors, the trend line is a linear regression, see how it falls as pixel size falls.

Of course, as we have discussed, Canon does not design by strict scaling. What they do is design in distinct generations, so here is the mean pixel size and mean read noise of their three generations. Funny how as the mean size gets smaller the mean read noise gets smaller, especially as we know that they are all fabricated using the same process.

If real sensors never show your scaling rule, then what use is it?

Because it predicts rather well that read noise falls with pixel size, as hown in the trends above. Still, I think your argument can be levelled against most pieces of engineering background. For instance, we discussed the noise characteristics of a FET. If you cannot demonstrate that is true by observing the behaviour of an op-amp, what use it it?

So back to what I do say rather than what I haven't said.

People make statements about the effect of pixel size on image noise and DR. These tend to take two forms.

Proposition 1: Small pixels means more noise and smaller DR. These statements are almost never evidenced. If they are, it is by reference to Roger Clark's rather obviously flawed site. Still, they are rarely challenged except be me and people who agree with me. What I say is this.

This is not what I said.

I know it's not what you said, but it is the statement that is made, I respond to, and you chip in saying what I say is not true.

What I said is that for a given process node the read noise tends to scale by the row read-out speed. So without process scaling, increasing the product of the row count and the frame rate tends to increase the read noise per area.

That is an interesting theory, and in no way incompatible with the idea that it is also dependent on pixel size. the problem indeed is that read noise is dependent on all sorts of things. Still, as the saying goes 'please suggest a way to compare real-world sensor results to see how row read-out speed affects read noise per area' and then 'if real sensors never show your scaling rule, then what use is it?' - sauce for the goose, and all that.

Proposition 2: If we want to make a statement about how size affects anything, then we need to isolate size as a variable. That means that the question we are asking is 'if we take a pixel, and then make the same pixel design smaller, what happens?

No, that would mean improving the minimum feature size of the process.

Possibly, possibly not, but whether it does or not is irrelevant. We want to know how pixel performance changes with pixel size. Practically, how you achieve the new pixel size is irrelevant. It's akin to you contesting that if we make an cast object larger by strict scaling it will get heavier. You say 'no, that would mean using a new mould'. Irrelevant and a straw man.

I would say the important question is "if you start with a given minimum feature size, what effect does pixel size have on saturation capacity and read noise per area?

That is a different question, also of interest, particularly to people constrained by a minimum feature size. You seem to think that there can only be one question, and it has to be yours.

Apparently only you may choose what a reasonable basis of comparing pixel sizes is: even when you never see your scaling rule with real sensors.

Actually, that is your position. I do what one always do in science when one wants to observe behaviour under the change of one variable. You insist that I cannot answer that question and raise different questions that you insist are answered. I do not deny that they are potentially interesting questions, but they are not the question of 'what happens to the performance of a pixel when you change only its size'.

Post (hide subjects) Posted by
Keyboard shortcuts:
FForum PPrevious NNext WNext unread UUpvote SSubscribe RReply QQuote BBookmark MMy threads
Color scheme? Blue / Yellow