News and discussions about image sensors
If I understand correctly, their "FCP" pixel is a 4T with a "soft reset".So they trade higher read noise (kTC) for a larger full well.The idea helps in shot noise limited conditions only, but suffers in low light. the demo does not reveal the shutter speed or lens f-stop. how can any comparison be made?
No, FCP is not soft reset. Soft reset leaves image lag - this would kill our solid state image stabilization. Feedback controlled reset suppresses kTC noise and has no image lag - in that sense it's similar to well-designed CDS.The videos were shot agout half a year ago, but from the top of my head, the low light image stabilization demo uses 0.7-0.8s exposure, split into many short sub-frames which are re-aligned on the sensor to compensate for camera shake. The scene illumination is about 20-30 Lux. The lens F# is 3.1.Vladimir Koifman,CTO,Advasense
Do you mean that there is no true CDS to take out the KTC and the KTC is only "suppressed" by the feedback control?The large FWC indeed improve the high illumination side DR but the low lux performance is still limited by RO noise.
True, there is no CDS. However, kTC noise is suppressed by a huge factor, x30 or more. Instead, there are other noise sources dominating, mainly pixel amplifier transistor noise. This is very similar to a well-implemented CDS pixel where source follower transistor dominates the noise.Talking about the low-Lux performance, our photodiode design can be greatly improved, resulting in higher QE and lower color crosstalk. Since FCP does not require the full pinning, the doping of the photodiode can be much higher than regular 4T. Also, the depth of the photodiode can be improved with no care about charge pockets, potential barriers, image lag and other challenges making 4T photodiode design so difficult. Improved QE and crosstalk result in better low-Lux mode.I would gladly add more details, if there is an interest here.Vladimir Koifman,CTO,Advasense.
Could you provide us with some experimental data that gives pixel size, FWC, and measured read noise? You have given me some numbers privately but I am inviting you to present them here if you wish.I am sure the data will be of interest.
Eric,Sure. The demo talks about 5MP sensor with 1.4um pixels made in 0.11um process. Full well capacitance is 15Ke.The read noise is 8e, while the full kTC noise is about 35e. The readout noise number will be much lower in the next version.Vladimir Koifman,CTO,Advasense
if the scenery is in motion when you take pictures with the stabilisation on, this stabilisation will even make it worse, right?
Advasense on-sensor image stablization is like optical image stabilization - it allows to compensate for camera shake for long exposures. It does not prevent the moving subject blur.The image stabilization that you see in the demo uses gyro to get the information about camera shake. Another possibility is to analyse the fast sub-frames data electronically. In the later case the image stabilization can lock on the features rather than global picture. For example, if kid is moving across the picture, the image stabilization engine can follow his face, making everything else on the picture blurred.We used to have this electronic image stabilization engine in the older 2.2um sensor. In 1.4um 5MP sensor we opted for a cheaper gyro-based solution. We might or might not bring this feature back in the future sensors.Vladimir Koifman,CTO,Advasense
A few other companies have demonstrated the use of feedback to suppress KTC noise. How is this different than the similarly named company AltaSens? They do it on a 3T. Add a 4th transistor to avoid the contact on diode.
You are right about name similarity with Altasens. When we were chosing the name of the company, our intention was to abbreviate Advanced Sensor Technology. Initially we thought about Advasent, but this sounded more like a name for deodorant company. So we ended up with Advasense. We missed the similarity with Altasens at the time.There are many companies doing kTC noise reduction in 3T pixel. Altasens and Fairchild (through Avago and Pixel Devices) have current products based on that. Cypress-Smalcamera and Fujitsu had such products in the past. In research world JPL-Caltech, BNL, KAIST, Hokkaido University worked on that as well.The difference between 4T and 3T pixel noise reduction is that in 4T one needs to compensate 2 sources of noise: one upon RST turning off and another one upon Tx gate turning off. To the best of my knowledge, Advasense is the only company that suppresses both of these sources.Having 4T pixel instead of 3T dramatically reduces dark current and allows smaller pixel size through pixel sharing.Vladimir Koifman.
Vladimir, This sounds like nice work. It will be interesting to hear about the improvements that you make in read noise with this approach.
All comments are moderated to avoid spam.