Saturday, October 16, 2010

IsInvariant Proposes New Sensor Technology

I was given a presentation from IsInvariant - a new start-up company proposing its own way to make sensors:

Value Proposition

IsInvariant™ is introducing a pixel preamplifier to increase dynamic range while decreasing power consumption, chip real-estate, and noise

Dynamic Range of Camera, Comparison of Systems

  • Ordinary Camera – 48 dB
  • High End Camera – 96 dB
  • Logarithmic Camera – 120dB
  • Humans – 80dB with biasing 200dB

Noise in Cameras

  • Logarithmic cameras suffer from high NEP (Noise Equivalent Power) that arises partly from biasing the pixel PN junction
  • Humans have the remarkable characteristic of being able to sense single photon events, as if the visual apparatus operates with no NEP

Ideal Camera Characteristics

  • Unbiased PN Junction
  • Logarithmic from one photon to ten billion photons per second
  • Insignificant power consumption

Pre-Amplifier in Image Sensor

Use of Pre-Amplifier

  • A PN junction converts photons into electron/hole pairs
  • A preamplifier converts pairs into usable measurements

An Ideal Pre-Amplifier

  • Converts input light intensity into logarithmic output signal
  • Covers theoretical 200 dB range without introducing or amplifying noise

Current Technology Limitations

Biasing of PN Junction

  • Johnson noise and consumption of DC current, reduces effective range
  • Dynamic ranges of pre-amplifiers bounded by absolute “noise floor” and “saturation”

Crux of Problem

  • Camera image planes fail to exploit full dynamic range of PN Junction sensors
  • Available preamplifiers force an artificial saturation
  • Johnson noise

IsInvariant's Solution:

IsInvariant Pre-Amplifier

  • Exploits full dynamic range of PN Junction
  • Enables low-noise transduction of the entire photosensor dynamic range
  • Consumes less power

Pending Patents

  • 20050036655 - Imaging system (Continuation in part)
  • 20050104632 - Geometric remapping with delay lines
  • 7796173 - Imaging system (Issued September 14, 2010)

Few performance slides:


The traces apparently relate to a pixel structure shown below:




Advantages of IsInvariant's Solution

  • Dynamic range increases over current preamplifiers (from 120 dB to 200 dB)
  • Power requirements will decrease (no DC current)
  • VLSI masking complexity will decrease (no OpAmp)
  • Chip die size will decrease (small silicon footprint)
  • No camera flash needed (no bias-related noise floor)
  • Time-to-market is driven by manufacturer remasking (no retooling needed)

24 comments:

  1. I am the first to admit I am not a circuits guy so maybe that is why I am having trouble understanding this pixel circuit and how it is all biased. Anyway, I am sure the these guys have this all figured out so I am hoping they can explain more details here. Maybe a paper at IISW?
    e.g. what does D1 do relative to the Q1 E-B diode?
    I thought a p-gate JFET J1 should have its gate reversed biased relative to S and D? (I understand you can get away with small positive bias).
    How does this circuit not draw dc current?
    What is the time response of this circuit

    I like the idea of using JFETs and bipolars to reduce circuit noise (e.g. RTS perhaps) but I am not sure this works well for, say, 100 e-/sec current in the detector.

    ReplyDelete
  2. Actually, I've asked for more explanations too. The response was: "We are still in internal discussion to figure out what details can be disclosed without signing non-disclosure agreement. "

    ReplyDelete
  3. This is very similar to Carver Mead's log photoreceptor circuit ... It's interesting to see how this tuff can work.

    ReplyDelete
  4. They have signed a NDA with the patent office too?

    ReplyDelete
  5. Funny information! Just having taken a look this guy's CV. How can a software engineer invent soundable pixel design and state that the noise can be eliminated?

    ReplyDelete
  6. They should ask money from DoD, because 200dB = 1:10^10 and a 200dB sensor can cover from 100K lux to 10uLux. Just remind that the heavily covered starlight night has 100uLux.

    ReplyDelete
  7. << Funny information! Just having taken a look this guy's CV. How can a software engineer invent soundable pixel design and state that the noise can be eliminated? >>

    This is poor, unconstrutive comment that Image Sensors should remove. People from different backgrounds have different experiences and perspectives on approaching problem solving - This is the key to innovation and creativity.

    ReplyDelete
  8. @ "This is poor, unconstrutive comment that Image Sensors should remove."

    I'm not going to remove it. However, I agree with you that good ideas can come from everybody, no matter what his/her background is. The only difference is the frequency: professionals in the field tend to generate good ideas much more frequently, in my practice, anyway.

    In any case, I would not reject any idea based on the originator's background.

    ReplyDelete
  9. do you hire this guy for CMOS sensor design position?

    ReplyDelete
  10. I whish them the best of luck reconstructing the image haha =D

    ReplyDelete
  11. Since when is the noise floor for a log sensor 5 orders of magnitude better than for a linear sensor? Is this marketing physics?

    ReplyDelete
  12. if you claim 200dB dynamic range, you have to decrease at least several orders of magnitude the noise floor. It's necessary that this gentleman show the details to image sensor community this invention.

    ReplyDelete
  13. @ "This is poor, unconstrutive comment that Image Sensors should remove."

    If someone comes tell you that he has invented a bicycle capable to go to the moon with. What is your reaction?

    ReplyDelete
  14. @ "If someone comes tell you that he has invented a bicycle capable to go to the moon with. What is your reaction?"

    What would be your reaction if some well-known professional in the field says the same thing?

    @"do you hire this guy for CMOS sensor design position?"

    At least, I would interview him.

    ReplyDelete
  15. "If someone comes tell you that he has invented a bicycle capable to go to the moon with. What is your reaction?"

    How about: "Show me. A moon rock will suffice."

    I'll wait for images and data.

    ReplyDelete
  16. "At least, I would interview him. "
    Let's interview this gentleman together here !

    ReplyDelete
  17. @ "Lag"

    I’ve been always wondering why most images sensor designs are based on measuring charge from a photodiode.

    It has always seemed obvious to me that a sensor could be using photo-resistors and be measuring resistance.

    Or, use photodiode at the PN junction, like in this design here.

    As another posted said, though, it seems that both the photo-resistor design and the PN junction design suffer from lag – and that’s why these sensors are not practical.

    ReplyDelete
  18. @ "It has always seemed obvious to me that a sensor could be using photo-resistors and be measuring resistance."

    Photoresistors usually have higher noise than photodiodes, not to say about dark current and nonuniformity. Photodiodes in photovoltaic mode also have uniformity challenge and their noise is somewhat higher.

    @ "if you claim 200dB dynamic range, you have to decrease at least several orders of magnitude the noise floor."

    Since IsInvariant's PD does not have reset, I'd speculate that it's effective integration time varies depending on illumination. At low light the integration time goes up, so the SNR degradation is less obvious. However, longer integration can be interpreted as a lag. Also, the PD should be of very high quality, virtually defect-free to extend sensitivity to a significant degree.

    ReplyDelete
  19. For normal CMOS sensor after integration the accumulated image charge is stored for conversion to digital numbers. The theoretical dynamic range is ratio of the number of full-well electrons to that of the noise electrons. However, if you check any CMOS sensor datasheet form the major suppliers, they never show the number of noise electrons. Because it is the most difficult part to control. The next problem is to get a good, say, 16-bit ADC with minimum additional noise. None of the avove problems are easy to solve. The wide dynamic range CMOS sensor MT9M033 released by Aptina recently relys on two or three exposures with different shutter time to achieve wider dynamic range as high as 120dB. It tells us something.

    ReplyDelete
  20. 1) Noise due to DC current significantly raises the noise floor.

    2) Agreed, the old precursor pixel circuit shown in the blog draws DC current.
    3) The undisclosed pixel circuit we use now draws no DC current.
    4) Details on undisclosed new circuit will be published in patent application.
    5) Lag and parasitic capacitance are indeed issues when no reset is used.
    6) Agreed, doping uniformity affects per-pixel photon conversion efficiency.
    7) Photodiode changes gate charge pressure on FET changing S/D resistance.
    8) Reconnecting photodiode N lead to ground doesn't change response magnitude per decade.

    ReplyDelete
  21. 1. The read noise comes mostly from 1/f noise in the output transistor. Probably a JFET would be better in the output stage.
    2. right
    3. Just like every other CMOS image sensor pixel.
    4. I hope so
    5. How will you avoid kTC noise on reset? I guess the effective input photon noise is worse for a logarithmic pixel.
    6. ok
    7. Indeed. In photovoltaic mode you can get by with the gate slightly positive relative to the source.
    8. Photovoltaic mode is of course not new. A big problem with this kind of pixel (besides those already pointed out previously) is that images taken under low contrast (e.g. cloudy conditions outdoors) don't come out well when companded by the logarithmic-response device. And, if you try to stretch the contrast, all the non-uniformity warts come out to bite you on the butt.

    With the modifications described, it will be difficult to patent this pixel due to prior art. In my opinion the commercial viability of this approach is not as rosy as you may think because there is no new nor compelling advantage to this approach.

    Nevertheless, best of luck. I think the money for you may be in the system, not in the pixel.

    ReplyDelete
  22. I should add that I think the device physics at a few hundred electrons per second incident on the photovoltaic device might be interesting. Probably equivalent circuit models are not too good for the time-domain for this device due to low signal statistics. That is, current flow in the photodiode may have a high shot-noise component. Maybe the low background IR astronomy guys have looked at this...

    ReplyDelete
  23. Wouldn't it be nice if a ring-geometry JFET worked out well, so you could use the ring gate as detector (PV mode)? Possibly patentable (at least up until now) but it has all the issues described above not to mention the pesky microlens issue that Olympus encountered with its CMD device.

    ReplyDelete

All comments are moderated to avoid spam.