Having SNR of 35dB at 40% gray level means that this 40% signal is at least 3100e-. Then the 100% signal can reach 7.7Ke-. If true, it's quite an achievement for that small pixel.
Correction: As written in comments, the gray level is closer to 20% (19.8%, according to Poynton). Then the full well should be beyond 15Ke-, hard to believe. The flyer only mentions on-chip bad pixel and lens shading corrections, says nothing about noise reduction.
The production schedule seems to be delayed as compared with the September announcement. The flyer says that the S5K2P1 is sampling now with volume production scheduled for Q1 2012.
It IS 18% gray level, then 100% signal can reach 15Ke- if the image was RAW. I bet not, but fully processed.
ReplyDeleteI'm impressed seeing all these details (fps, Int. time, Again) in additional to standard lux level. If lens f/# was also presented, would this be the 'best' single image metric of sensor performance for vendors to publish? A question for ISW readers.
ReplyDeleteThe omitted factor is the color accuracy, delta-E. Depending on how accurate the color is, the SNR can change a lot. Also, illumination type should be specified (daylight, incandescent, etc.). Since the SNR is given off-center, color shading needs to be factored in, and it depends on the lens CRA.
ReplyDeleteSo, there is no single number that says it all. Nokia SNR10 is still the best figure of merit - note it's defined for camera modules including lens.
There are many ways to calculate SNR. If I were in the marketing I would calculate it straight from the image pixel values, because it gives much higher readings than actual linear sensor SNR. And I suspect Samsung has done so as well.
ReplyDeleteThis image is obviously processed with sRGB gamma or something very similar.
This has the effect that in the processed image the 18% signal is amplified but noise is amplified much less. For 18% linear signal we would get 2.56x amplification for the signal and 1.2x amplification to the noise.
This translates to -6.57 dB correction to the reported SNR.
Thus we get linear SNR =28.5 dB at 18% signal = 708 e- -> full well 3933 electrons, a figure that's much more credible.
But, as Vladimir stated, the image looks quite saturated so probably cross-color matrix has been applied which has the effect of amplifying the noise.
My point is that if sensor does not have gamma, sRGB or other processing functions integrated on chip, the SNR data should not include them. However, the presented SNR data is so good that it's hard to believe that it was not processed in one way or another. The processed SNR figures say nothing about the pixel quality, especially if processing is unspecified.
ReplyDelete