Sunday, May 04, 2014
Omnivision Proposes a Compact Way to Enhance DR in 4T Pixel
Omnivision's patent application US20140103189 "Compact In-Pixel High Dynamic Range Imaging" by Gang Chen, Duli Mao, Hsin-Chih Tai, and Howard Rhodes proposes quite a minor modification in 4T pixel which dynamically changes its conversion gain based on the signal level. The proposal is to add few Vth-adjust implants under the RST gate, so that various gate areas go to inversion mode depending on the floating diffusion voltage:
Quite nice idea, but it will make the reset transistor longer I guess. Unfortunately the characterisitic (output versus input) will strongly dependent on the VT of the various parts of the reset transistor, and that will result in a large amount of fixed-pattern noise. Very similar to the piece-wise linear response characteristic that can be generated by changing the gate voltage during the exposure of the sensor (see the work of Bart Dierickx at the time he was with FillFactory).
ReplyDeleteBut nevertheless, it remains surprising how much work that is already done on a pixel making use of only 4 transistors.
So, how about a tapered reset gate, so that the potential slope under the gate is defined by geometry rather than implants? Would also remove any actual knee points, although matching could still be a concern, but perhaps less so.
Delete... or maybe even more ? Could be a very challenging subject for a MSc thesis ! In the good old days at Philips, when CCDs were still considered as image sensors (Eric understands what I mean), we used a tapered reset gate, But that was for other reasons than shown in this patent. I do not know whether my former Philips'colleagues still use the same trick.
DeletePerhaps I've misunderstood this, but I don't see how a tapered reset gate with a uniformly doped channel underneath would allow you to control the conversion gain of the pixel, Could you please elaborate on this?
DeleteOne thing it would do though, is reduce charge injection onto the floating diffusion node at reset (assuming the narrow end was narrower than a rectangular-gate reset MOS would otherwise have been),
Well, hard to say for sure if this would work but I just asked someone in my lab to model this as he learns how to use TCAD tools (so might take a while).
DeleteBut the idea is that the potential under the reset gate not only depends on the usual factors of oxide thickness, doping under the gate and gate voltage, etc., but it also depends on the width of the channel - a 2D effect if you like.
So, make the channel wider at the FD end, and narrower at the reset drain end, and the potential will therefore be deeper at the FD end than at the RD end.
(If the RG is remote from the FD this will still work).
This is similar to the effect described by the OVTI patent app above. It is also, unfortunately, opposite to what you describe. This will maximize reset charge partitioning into the FD on reset (and so will the OVTI concept).
The reason I say this may not work is two-fold. First, as Albert points out, the onset of the effect depends on the threshold of the RG. So, there will be FPN for sure. But, second, and perhaps more important, is that it will be difficult during reset to remove the charge from under RG if the deep end of the RG is by the FD. If you turn RG off "hard", the charge might just go anywhere or everywhere. Still it might work. Hard to say.
DeleteUpdate. Visiting Prof. Atsushi Hamasaki looked at this as he was learning how to use our Synopsys TCAD simulation tools. While at much larger dimensions, the narrow channel effect results in perhaps a 100 mV potential different from one end to the other, at competitive dimensions, a tapered reset gate geometry did not result in any significant potential gradient, much less one that would be useful for this effect. Perhaps additional investigation would result in something but at this time I would say this idea of tapering the geometry does not work well. We did not look at the implant approach that Omnivision proposes so no comment about that.
DeleteThe fixed pattern is not a problem if is the same for all pixel. In this case, the processing can take this into account and produce good images.
ReplyDeleteThe problem that I see is that the DR is not controlable. You have a non linear characteristic that cannot be modified by programming. This is needed to adapt the HDR characteristic to the scene.
One HDR image may have 70dB or 120dB, depending the scene. If a 120dB compression is used with an scene of 70dB, the resulting image will have a poor contrast.
This is a problem like using the log compression. Even the lin-log is too extreme in the log region.
Rafa, what do you mean by "if the FPN is the same for all pixel" ? If a noise component is the same for all pixels, then there is no longer any FPN ....
DeleteThere WILL be a significant pixel FPN due to variations in the implant diffusion, in much the same way that there are always Vt variations.
DeleteThe multiple knees will also move around from pixel to pixel creating a stitching nightmare.
You are right. I refer to fixed pattern, but not noise. The fixed pattern can be removed by processing. Even if is random (noise) can also be removed, but requires more memory and processing resources.
Deletehmm, fixed pattern in a spatial noise; there is no such thing as fixed pattern without noise. Of course, image processing can be used to remove FPN, but then there is no such thing as free lunch. You trade high dynamic range for computational power, residual FPN and other artefacts. Better to nip such effects in the bud: in the pixel; thats the whole point.
DeleteThe curve "out" versus "in" shows several knee-points, which will be different for the various pixels due to their dependence on the threshold voltage of the reset transistor. Will be not easy to compensate or correct. Wishing you good luck with it.
DeleteWhy not just add a few RST transistor in serial, does the same trick and do not require process change. We had a patent application about this a year ago.
ReplyDeleteAdd several reset transistor has the drawback that several control lines are needed, one for each reset transistor. This may be a big problem in small pixels.
ReplyDeleteIn reality, the charge injection of RST transistor will be very high. Since the left portion will be cut off first, so all the charge in the right side will be injected into the FD. This will low the dynamic range of FD.
ReplyDelete-yang ni
By incorporating another standard reset transistor this problem could be circumvented. You could actually remove the drain of the large reset transistor so that you would have only one small reset transistor and an additional capacitor providing a piecewise voltage vs capacitance curve.
ReplyDeleteIn this case, you should have 2 control lines.
Delete-yang ni
You are right, but if you have a BSI sensor with multiple metal layers this shouldn't matter.
DeleteIs this just another idea or has OVT actually made a test chip?
ReplyDelete