The paradoxical evolution of sensor SNR over time

Sunday, November 16, 2008

Camera Article
Introduction | New pixel technologies vs. pixel pitch reduction | SNR and image quality evolution | Conclusion

To see the impact of new pixel technologies on the SNR, we need to normalize the SNR with respect to pixel pitch. Doubling the pixel pitch yields a gain of 6dB on the SNR 3. We can use this relationship to normalize all the different types of pixels to a given size, per the following formula (using a pitch of 10µm as an example):

SNR (normalized, 10µm-equivalent) ~ SNR + 20 LOG10( 10/ pixel pitch)

With the SNR thus normalized, we can then redraw the first graph:

SNR at ISO200, 10µm-eq, (dB)

What we observe is a 2 dB gain over the past 5 years. This means that the technology for a fixed pixel pitch has produced an improvement of +2dB (or 2/3 f-stop). This is the equivalent of saying that the image quality at ISO200 five years ago can now be achieved at ISO300.

Another thing to notice is that for a given release year, the dispersion across cameras is of approximately the same magnitude as the improvement over the past 5 years. This means that first, not all manufacturers use or have access to the same level of pixel technology, and second, the differences among current DLSRs can be larger than the overall technology gain over the past five years.

As an example, let’s compare the differences in technological improvements between two manufacturers, Canon and Nikon:

SNR at ISO200, 10µm-eq, (dB)

From the graph above, we see that over the last five years Canon has been, on average, significantly ahead of Nikon in terms of pixel performance. However, we can also see that the improvement over the same period for Nikon pixels is about 2.5 dB (as opposed to 1.5 dB for Canon), so in principal, the difference in performance disappears in 2008. And indeed, today the Nikon D3 has the best pixel technology (the only point in the graph above 34dB).

Notes

3 Doubling the pixel pitch by averaging pixels by groups of four yields a noise reduced by two. Hence the SNR is increased by 20 LOG10(2) ~ 6dB.