"A backside-illuminated complementary metal-oxide-semiconductor (CMOS) image sensor with 4.0 μm voltage domain global shutter (GS) pixels has been fabricated in a 45 nm/65 nm stacked CMOS process as a proof-of-concept vehicle. The pixel components for the photon-to-voltage conversion are formed on the top substrate (the first layer). Each voltage signal from the first layer pixel is stored in the sample-and-hold capacitors on the bottom substrate (the second layer) via micro-bump interconnection to achieve a voltage domain GS function. The two sets of voltage domain storage capacitor per pixel enable a multiple gain readout to realize single exposure high dynamic range (SEHDR) in the GS operation. As a result, an 80dB SEHDR GS operation without rolling shutter distortions and motion artifacts has been achieved. Additionally, less than −140dB parasitic light sensitivity, small noise floor, high sensitivity and good angular response have been achieved."
Interesting... but I don't know if I trust the table, the dynamic range calculation seems to be black magic.
ReplyDeleteAm I missing something?
Low gain
@ GS
DR = 20*log10(40000/30) = 62.5 dB
@ RS
DR = 20*log10(40000/13) = 69.8 dB
High gain
@ GS
DR = 20*log10(7000/4) = 64.9 dB
@ RS
DR = 20*log10(7000/2.1) = 70.5 dB
Agree... It seems that the DR calculation is taking the best case of QSat (40000e- low gain) versus the best case of noise (4e-, high gain) which gives us those wonderful 80dB.
ReplyDeleteKind of black marketing magic?
Anyway I guess this needs further clarification
It is a perfectly fine DR calculation. Both low gain and high gain are captured in 1 image. One gain or the other will be chosen based on light level so yes, 80dB is no black market magic but reality.
DeleteDR number alone does not tell the full story. Just as an example, let's assume a pixel with 1e- FWC and 0.001e- noise. One can boast 60dB DR in that case, but any real-world picture would look like junk.
ReplyDeleteThis is an exaggerated example, but it should explain that DR is indicative of the sensor real-world performance only when taken in combination with other parameters.
I disagree Vlad. DR should be taken from SNR=1 at lowest light to SNR=1 at saturation (or some other condition if you want linear FW etc.) In your example, one gets a very small DR using this approach. Using the noise floor directly is only a valid shortcut when the read noise is higher than 1e- rms.
DeleteI'd guess there is no common definition. EMVA 1288 defines DR in a way close to yours. However, Brillnics defines it as "DR = Saturation level/ Noise"
Deletehttps://www.brillnics.com/technology/high-dynamic-range-image-sensors
Taking your definition, if we have 2e- read noise, the SNR=1 would be achieved at the signal level of 2.5e-. Then, if FWC is, say, 2,000e-, the DR is 58dB. To my eye, it looks a bit unusual.
The EMVA standard 1288 requires that the FWC and noises be measured at the same gain setting ...
ReplyDeleteYou are right about EMVA, but the paper does not mention that they calculate/specify the DR according to the EMVA 1288 ....
DeleteYes, same gain over the full sensor. But if you can choose the gain autonomously per pixel then EMVA is still valid. I think you confuse interscene with intrascene DR definitions...
Delete