Thanks to RP for the info!
Tuesday, February 04, 2020
Fairchild Imaging Announces sCMOS 3.0 Sensor with 0.5e- Noise
BAE Fairchild Imaging presents a new ultra low light HWK4123 sensor featuring a BSI sCMOS 3.0 technology. It is said to have "advanced the state-of-the-art in low light imaging by lowering the read noise to 0.5 electrons while improving the broad-spectrum quantum efficiency." The 4/3” 9MP HWK4123 enables night vision and surveillance cameras to image at less than 0.001 Lux (starlight). The HWK4123 is sampling now.
Thanks to RP for the info!
Thanks to RP for the info!
0.5e- read noise at 120fps is amazing! what is the temperature condition to achieve this read noise figure?
ReplyDeleteAmazing, really? What is the current record for read noise in CMOS image sensors?
DeletePlease take a look at: http://ericfossum.com/Publications/Papers/2018%20Laser%20Focus%20World%20December.pdf See esp. Fig. 5 which is an image taken at very very low light and denoised via computational imaging techniques.
This is a product, Eric. Among the available products in the market this is amazing achievement.
DeleteOK. Technically speaking you can buy a pathfinder camera from Gigajot from which the above referenced image was taken. Is that a product? Well, not a production product, but still a product. I understand your point however.
DeleteSaw the demo. Not exactly 120 fps. More like 0.5 fps. And computational imaging doesn't help framerate...
DeleteQIS has been championed as a much cleaner solution for single-photon imaging.. the reality is that people do not pay attention to the hidden elephant, that is computational imaging overhead. This overhead itself can kill lot of performance parameters needed.
DeleteHardly hidden and hardly an elephant. One could have argued that ISP was a hidden elephant for digital color imaging years ago, but digital processing advances at a much faster rate than image capture. In any case, for just emulating SCMOS, the computational imaging part is trivial. It is the new capabilities that QIS+computation bring that have some overhead but even this is far from optimized or implemented in an ASIC. Please provide the list of "lot of performance parameters" that computational imaging overhead "can kill" and we can discuss them one by one. BTW, I guess the demo referenced above was multibit QIS (3b?) using the slow analog output port of the Dartmouth research chip. The analog port was added to allow device characterization and this is not a real technology limitation.. The digital port was the 1b QIS 1 Mpixel @ 1040fps @20 mW. This was described and demonstrated in the published Optica paper a couple of years ago (https://doi.org/10.1364/OPTICA.4.001474) and would be more comparable to SPAD-based QIS devices than the Dartmouth CIS-based QIS.
DeleteFound this on the datasheet :
ReplyDelete"This product is not available for scientific applications. BAE Systems reserves
the right to terminate sales to a customer if BAE Systems determines that
the customer is using the product in such scientific applications."
wtf ? Why ? i mean... why ?
typicallyTthis is because some other scientific customer paid for this development and requested exclusivity..
DeleteIs characterizing their sensor considered scientific application? I guess you cannot verify their 0.5e- claims then. :D
ReplyDeleteIt is amazing, if what they claim is totally true. 0.5e- at 11 bits mean 3500 FWC, then it looks like some gain is applied. Maybe, for complete range it needs to go 60fps. It would be good to know the dark noise distribution, since high gains usually show long tails... anyway it looks great, but I still miss details...
ReplyDeleteWell, theoretically to collect 3500 e- in FW, it can go very slow. What matters is 0.5e- noise at 120fps, since your front-end is going to dominate the noise at high speed.
ReplyDelete