Toshiba Starts Mass Production of 1080p 2.7um Pixel HDR Sensor
Business Wire: Toshiba starts mass production of the previously announced 1/3-inch 1080p60 TCM3232PB sensor at the end of April. The PR stated the sensor HDR mode provides 108dB DR, while the flyer talks about 100dB:
The normal SNR itself is problematic as response saturates, as the regular SNR tends to blow up. But, light referred SNR behaves "normally" and drops off. S is the light level, and N is the noise-equivalent light. In saturation the amount of light (delta-light) required to obtain a certain delta-signal equal to the noise level becomes large and drives the light-referred SNR to a steadily decreasing value, and eventually to 1. Light referred SNR is equal to the regular SNR in the linear region. But, for non-linear response, esp. near saturation, it is the more sensible metric, at least in my opinion. I note that a discontinuity in the derivative of the response slope is problematic for light referred SNR (e.g. true piecewise linear response).
This is called iSNR in ISO and i called it rate of information transfer in my book. Don't forget that SNR<1 also exists at various points in the SNR curve and these gaps must be investigated.
SNR-H is described in this paper: E.R. Fossum, Modeling the performance of single-bit and multi-bit quanta image sensors, IEEE J. Electron Devices Society, vol.1(9) pp. 166-174 September 2013. Digital Object Identifier 10.1109/JEDS.2013.2284054 in section IV. Pretty brief though. The non-linear dynamic range definition is described on slide 24 of this talk: http://ericfossum.com/Presentations/2014%20Stanford%20Oct%201.pdf That may be the first place it is documented.
Dynamic range = ... TOSHIBA CALCULATION METHOD -> so no meaning at all
ReplyDeleteHarsh. So if they put no note next to it, it would have more meaning?
DeleteEveryone should measure HDR the way I do: from SNR=1 in low light, to SNR=1 in saturation, input light referred.
SNR=1 in saturation? can you explain more on this? thanks
DeleteThe normal SNR itself is problematic as response saturates, as the regular SNR tends to blow up. But, light referred SNR behaves "normally" and drops off. S is the light level, and N is the noise-equivalent light. In saturation the amount of light (delta-light) required to obtain a certain delta-signal equal to the noise level becomes large and drives the light-referred SNR to a steadily decreasing value, and eventually to 1. Light referred SNR is equal to the regular SNR in the linear region. But, for non-linear response, esp. near saturation, it is the more sensible metric, at least in my opinion. I note that a discontinuity in the derivative of the response slope is problematic for light referred SNR (e.g. true piecewise linear response).
DeleteThis is called iSNR in ISO and i called it rate of information transfer in my book.
ReplyDeleteDon't forget that SNR<1 also exists at various points in the SNR curve and these gaps must be investigated.
Eric, do you have a paper describing your method? I'm working on a new version of my book and would like to include the extra reference.
ReplyDeleteSNR-H is described in this paper:
DeleteE.R. Fossum, Modeling the performance of single-bit and multi-bit quanta image sensors, IEEE J. Electron Devices Society, vol.1(9) pp. 166-174 September 2013. Digital Object Identifier 10.1109/JEDS.2013.2284054
in section IV. Pretty brief though.
The non-linear dynamic range definition is described on slide 24 of this talk:
http://ericfossum.com/Presentations/2014%20Stanford%20Oct%201.pdf
That may be the first place it is documented.
Showing the graph of SNR would be a nice practice. Or at least give the value of the SNR drop at the switch between odd and even row...
ReplyDeleteSince the odd or the even row is chosen, shouldn't they reduce the claimed sensor resolution to 1MPixel??