Pixel density - can the playing field be leveled???

Started Jun 6, 2009 | Discussions thread
Malcolm Practice Contributing Member • Posts: 593
Re: Per area read noise?

Steen Bay wrote:

Malcolm Practice wrote:

If you/we use the argument that the 'image level' (or 'per area') shot noise is independent of pixel density, doesn't it then follow that the per pixel read noise (in electrons) also must scale with the absolute level of per pixel shot noise for the resulting image level noise to still be the same? And since the absolute level (not the SNR) of per pixel shot noise is lower for smaller pixels, because they recieve less light/photons, isn't it then reasonable to assume that the absolute level of per pixel read noise also must be lower, so that the ratio between read noise and shot noise stays the same, in order to keep the overall per area noise independent of pixel density?

It's easy to get confused here, and I realise reading my last few posts, I've mixed up a few things, though got the result right. First, if we assume strict scaling of a pixel, the per-pixel (shot noise excluded, because that doesnt have anything to do with the pixel) SNR is invariant under scaling. The reason is that these are voltage noises, and all get referred to effective electrons through the same agency as the signal, i.e. the capacitance of the source follower transistor (in the cell for a CMOS sensor). Thus if the voltage noise remains constant under scaling (which it does, since the SF noise is determined essentially by the length to width ratio of the transistor, and the downstream noises don't change as the pixel is scaled) the SNR of the pixel is constant. This means that absolute read noise under strict scaling (which might or might not be a realistic assumption) decreases according to the area of the pixel (as does the saturation capacity - hence the constant SNR). If we ally that to the increase in noise density inversely to pixel spacing, we end up with a net read noise density which decreases with pixel spacing. Other assumptions change matters. If that source follower stays the same size (doesn't scale) the absolute level of the noise (input referred) stays the same, while the saturation capacity of the pixel reduces in relation to its area, so the SNR decreases.

I suspect what we are seeing, as I suggested above, is that manufacturers design a good source follower transistor and use the same design for a small range of pixels. Thus Canon is most likely using the same source follower design for its DSLR sensors, which have pixel spacing in the range of 7.2 μm to 4.7μm (although it's possible they've put a new, smaller one in the 50D and 5DII, which accounts for the lower read noise compared with the older cameras). However, it is highly unlikely that they use this design for the SX1, since that has 1.6μm pixels, and it is unlikely that the DSLR SF design would fit.

Just trying to move to a middle position here, you can see that under strict scaling, read noise improves as pixel size goes down. If the source follower remains constant in size, but the rest of the pixel scales, it gets worse as the pixel gets smaller. That obviously give designers another option, which is to take the SF from a tiny pixel and put it in a big one, with the result it will have good read noise. The downside is, in a CMOS sensor, that the sensor will saturate earlier than a pixel that size with a correctly scale SF would, however, it will be good at high ISO's, due to the low read noise. This, I suspect is what Nikon did with the D3 and Canon is doing with the G11.

On the other hand, if Canon built a FF sensor with 50D size pixels and a SX1 size source follower....


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