Simple Sensor Design Tweak

[Tim Gray]

Would there be a price to be paid in terms of ISO sensitivity?  Any hole at all in the bucket sees like it would reduce low light sensitivity?  Maybe you could activate the high DR option based on user input.

[Jonathan Wienke]

Would there be a price to be paid in terms of ISO sensitivity?  Any hole at all in the bucket sees like it would reduce low light sensitivity?  Maybe you could activate the high DR option based on user input.

It would reduce base ISO in terms of the exposure level required to clip the highlights. Shadow exposure would only be marginally affected, but the big gain would be a substantial increase in non-clipped highlight exposure levels.

The one factor you have missed is that your gamma curve would be a funtion of time and charge, not just charge in the photosite as your article alludes.

I didn't miss that, it's implicit in the design. For longer exposures and varying ISO settings, the resistor would need to be cycled on and off to achieve the target gamma curve.

Second problem is that as the excess charge is passed through the resistor it is converted to heat. You run a large danger that for intense illumination you will cause the resistor to heat to such a level as to damage the chip.

That's pretty much a non-issue, as more intense illumination is directly related to shorter exposure times. The only way you increase the net illumination of the chip is to lower the ISO setting. And the charge accumulated by the photodetectors is only a small fraction of the total power consumed by the chip anyway, so it would be fairly irrelevant. In current chips, that charge is dissipated in the A/D converter and preamplifier circuitry, so spreading out the area where that charge is dissipated over the whole chip instead of a small area of the support circuitry would probably be a net good thing anyway.

For long exposures where there is a blurring, then you will end up with some interesting exposure artifacts due to different illumination at the start and end of the exposure.

No you wouldn't. The charge levels accumulated by the sensor would simply reflect the accumulation gamma curve.

[Jonathan Wienke]

Interesting idea, but why not just drain to a tiny capacitor, then nothing is lost and you've effectively increased the well size. I'm sure there are practical considerations why this has not been done.

The photodetector is in effect a small capacitor (among other things); that is how the charge is stored until read by the A/D converter. The technical issue involved would be making a bank of capacitors (1 per pixel) off-sensor, and routing the millions of individual connecting wires from sensor to capacitor bank. This would not solve the linearity problem anyway; all it would do is flatten the slope of the exposure/charge voltage graph. You'd still have the issue of deficient sensor tonal resolution in shadows and excessive tonal resolution in highlights inherent to linear sensors, along with the accompanying limitation to dynamic range due to the photodetector capacity filling unnecessarily quickly.

[dlashier]

> he's proposing a continuous leakage of the voltages generated by each pixel during exposure so the final reading will correspond to the brightest highlight in the scene

but the novel part of it is that the leakage is intensity dependent effecting changing the gamma. In addition to extending DR this also cures the coding level thing. I think the concept is great but the question is how to implement it.

- DL

[Jonathan Wienke]

Jonathon's just saying that the earlier it happens the better as it eliminates some of the obstacles to obtaining better dynamic range.

Exactly. And doing it in the analog charge accumulation phase of exposure prior to A/D conversion is about as early as one can get.

Instinctively, I'm not happy about the possible bad consequences of throwing information away.

What's being thrown away is the orders-of-magnitude precision overkill in the highlight values inherent to linear sensors, so that for a 12-bit sensor, instead of values-per-stop ranging from 32-2048 with 7 stops usable, values-per-stop might range from 250-300 with 12 stops usable.

[dlashier]

> The current flowing through a resistor (measured in amps) is directly proportional to the voltage applied to it and its resistance value in ohms. By connecting a resistor in parallel to a photodetector during exposure, the photodetector's behavior can be changed from a linear response to light intensity to logarithmic, thereby applying a gamma curve to the photodetector's analog output prior to digitization.

Jonathan, despite having a degree in mathematics and despite the fact that my father was an EE and Physics professor (or maybe because of it), I'll admit to having a major blind spot re electromagnetic theory, but where is the power function that would apply the gamma change? Yes the "leakage" would be proportional to voltage but this is still a linear relationship.

- DL

[Jonathan Wienke]

Lets imagine a simplified example, an experiment you can try for yourself if you have some basic lab equipment. Take a 1000 microfarad capacitor, a 10 volt power supply with a 1 milliamp current-regulated output, and a 1000 ohm resistor. If you connect the power supply to the capacitor for 1 second, the capacitor will have a charge of 1 volt. But if you connect a 1000 ohm resistor in parallel with the capacitor, charge for 1 second, and disconnect the power supply and the resistor simultaneously, the capacitor will have less than 1 volt of charge, because as the voltage of the capacitor charge increases, proportionately more of the current flows through the resistor instead of the capacitor. If you were to leave the power supply connected for more than 1 second, over time the charge of the capacitor would very closely approach, but not quite reach 1 volt. Basically it's an inverted version of a typical RC voltage decay curve, except that it's happening during charging instead of discharging. Given a fixed charging current, a the charge voltage of capacitor with no resistor in parallel will increase linearly until the maximum voltage output of the power supply is reached. But with a resistor in parallel to the capacitor, the same fixed charging current as before will increase the voltage of the capacitor following an inverted voltage decay curve. That's how the gamma is applied to voltage before A/D conversion.

In a photodetector, the charging current is controlled by the number of photons per second striking the detector, and the capacity of the well is determined by the detector size and chip fabrication techniques. All of the same principles apply, but are simply scaled down to microscopic size.

[Ray]

I'll continue to play devil's advocate. Here's a quote from Roper Scientific.

High-performance CCD imagers have extremely good linearity. Deviations from linearity are often less than a few tenths of a percent for over five orders of magnitude. This is far superior to video CCDs and other solid-state imagers, which can exhibit nonlinearity of several percent or more. For quantitative imaging, linearity is a stringent requirement. CCDs must be linear in order to perform image analysis such as arithmetic ratios, shading correction, flat fielding, linear transforms, etc.

Essentially, whatever arrangement you have, linear or gamma corrected, accuracy is crucial. I doubt whether a single transistor could have the required accuracy.

Secondly, we should look at the way the charge is collected from the photodetector. As I understand, photons knock off electrons in the silicon substrate of the sensor chip, ie 'electron hole-charge pairs' are created.

However, the different wave lengths of light penetrate the silicon to different depths, red light penetrating to the greatest depth. These electrons have to be 'collected'.

Can a single transistor 'collect' them or 'drain the charge' at various depths in an accurate manner? There's more going on here than the simple explanations available on the web.

[dlashier]

> Basically it's an inverted version of a typical RC voltage decay curve

Gotcha, that does indeed introduce a non-linearity.

> believe that a CCD is charged before exposure and that the actual exposure is supposed to generate a current (not a charge) which is what is detected and converted to a digital signal.

I believe a CCD is precharged positive but according to the reading I've done the exposure is (at some point anyway) stored as a charge in a mini-capacitor, then read out, so the principle should apply. The real question is, is it actually feasible to tap off between the photodiode and well?

> But more importantly: why would you want to create non-linear behavior? Does the dynamic range increase really weight up against the colorinfidelities?

This shouldn't affect color. The reason for non-linear behaviour is to increase dynamic range.

[DiaAzul]

I have had a chance to read through Jonathans post again and would like to clarify some of the aspects of his proposal.

The first is that the photodetector is a current source and generates current proportional to the amount of illumination falling upon it. It does have a capacitance, but in normal design the capacitance of the photo-detector is minimised to improve its responsiveness to changes in illumination. The capacitance of the pixel is determined by the parasytic capacitance of a number of elements (which does include the photodetector, but also the junctions of transistors as well).

The photo-detector can operate in two configurations. Reverse biased (polarity of the voltage is negative) and forward biased (polarity of the voltage is positive). In the reverse biased configuration the current generated by the photodetector is linearly proportional to the illumination falling on it; this is the typical operating mode in a digital camera. In forward biased mode the photodetector is non linear and is the mode which is typically used in a video camera to increase dynamic range. In the forward biased mode the photodetector is in series with a resistor and the voltage across the resistor is propotional to the instantaneous illumination falling on the photodetector. The capacitance of the configuration is minimised to ensure that the detector has a responsiveness to prevent smearing of the image between frames. The disadvantage of this configuration is that it is noisy compared with the reverse biased configuration which is described next.

In the reverse biased case additional capacitance is engineered into the solution. A reset transistor (t1) is closed for a short period of time to charge this capacitor to maximum potential. The potential on the capacitor (Cpx) holds the photo-detector in reverse bias and the current generated by the photo-detector slowly discharges the capacitor for the duration that the circuit is illuminated to light. The charge remaining on the capacitor at the end of the exposure is proportional to the integral of the illumination falling on the photodetector during the exposure period. The advatange of this design is that the measurement of light is more sensitive and the influence of noise is minimised.

The following diagram shows a typical three transistor active CMOS pixel design (though there are others) with the modifications suggested by Jonathan in the green box - making a total of four transistors.



The circuit works as described in the preceding paragraph. When the camera control circuit wishes to read the voltage on the capacitor it raise the voltage on the Row line, opening transistor (T2) and reading the voltage from the column line through (T3). Once the value has been read then the pixel can be reset by charging the capacitor up to full voltage again prior to the next exposure.

Jonathan's modification suggests introducing a resistor into the design as shown R to provide an RC decay curve during exposure of the photo-detector. Whilst Jonathan described this as being parrallel to the capacitor I have drawn it here in series, however, the net effect is the same. When the exposure commences transistor (t4) is opened for the duration of the exposure. The photodetector generates current and reduces the charge (in an exponential decay curve) until the voltage on the capacity reaches a certain voltage. This voltage is equal to the current flowing through the photodetector at the given illumination multiplied by the value of resistance R, once we reach this voltage the capacitor will be neither charging or discharging. Therefore, the ultimate settling voltage on the capacitor is determined by the illumination on the photodetector, and because the current is linearly proportional to the illumination the final voltage is linearly proportional to the illumination as well. Therefore, whilst we will get an exponential decay curve, that decay curve will be different for each level of illumination on the photo-detector. For this particular design the photodetector will need to be illuminated for a minimum period, defined by the RC decay curve, and the final voltage will still be linear with respect to illumination.

The following graph shows the voltage on the capacitor versus exposure time at two levels of illumination. The blue curve has twice the illumination of the red curve. For very low illumination (which Jonathan was hoping we could get a boost in signal as per a gamma curve) the signal becomes very weak (very low final settling voltage), except for very long exposure; for larger signals we need a long exposure to retain accuracy in the measurement (because of the effect of the RC curve and the need to allow the voltage to settle).



The end result is that we end up with a linear version of the forward biased design, but with worse performance in terms of sensitivity and smearing between adjacent symbols.

Other downsides to the modification are:
1/ Additional components in the pixel area reduce the fill factor (in this case by perhaps 10-15%) reducing sensitivity of the system.
2/ Dark current noise is sensitive to temperature and doubles with every 8 degree C rise in temperature. Introducing a heat generating source directly into the pixel (the resistor) will only make dark current noise performance worse.
3/We have no certainty of what voltage will be produced on the capacitor for a given exposure - except through complex measurement and mathematics.

A final note:
The dynamic range of the pixel is determined by the maximum voltage swing across the capacitor that can be measured. This is limited by transistor T1 which has a voltage drop of approx (0.8v) so that the capacitor cannot be charged to the full supply voltage, and the output of T3 which is 0.8v less than the value of the charge on the capacitor. The impact of this is that as the transistor size is made smaller then the safe operating voltage of them is reduced. However, for maximum dynamic range we want as large a voltage as possible. Therefore, even though there is opportunity to make smaller components on the sensor due to improvements in lithography this may not always be to the photographers advantage. However, as most people are aware the larger the transistor the lower the fill factor of the pixel and the lower its sensitivity. So we have a trade off (of sorts) between dynamic range and sensitivity. As the pixel size gets smaller then this trade off becomes more accute. The question I have is that even though Nikon now have a 12Mpix camera with potentially as good noise and sensitivity as canon, have they sacrificed dynamic range? This may not be the case, as there are tricks and design changes from the above to increase the voltage swing that can be measured on the capacitor, but ultimately smaller pixels will cramp the performance that can be obtained.

Well, I hope the above is useful. As everyone is fessing up to technical backgrounds I am a Member of the Institute of Electrical Engineers (UK equivalent of the IEEE), hence the background in technology. Its a while since I dealt at this level of design directly, so apologies for any inaccuracies, but on the whole it should be correct.

ciao ciao

[Jonathan Wienke]

The end result is that we end up with a linear version of the forward biased design, but with worse performance in terms of sensitivity and smearing between adjacent symbols.

You raise some very good points; with my original concept, while the graph of voltage over time during exposure changes along the RC decay curve instead of linearly, the final voltage at the end of exposure is still linearly proportional to the illumination level. Here's a slight rework, replacing the transistor + resistor with a just a transistor using the charge level of the capacitor to gradually pinch off the photodiode as the exposure progresses.



T4 is gated so that it is at maximum conductance immediately after reset when exposure begins. As exposure progresses and the charge of Cpx begins to deviate from Vcc, the conductance of T4 gradually decreases to zero.

Let's assume that the conductance of T4 is designed such that inserting it into the circuit would reduce the voltage swing from 99% (an exposure level just below clipping in the standard configuration) to 63%. Doubling the illumination intensity or exposure time would raise the voltage swing to ~86%, and doubling it again would raise the voltage swing to ~98%, assuming a linear relationship between gate voltage and conductance in T4.

Unless I'm missing something, this would deliver a solid two stop increase in usable highlight headroom without negatively affecting shadow response to any significant degree, and would also change highlight response from an ugly abrupt cutoff when clipping happens to a more film-like smooth transition to pure white.

P.S. If you fabricated T4 conductance such that a given exposure caused a ~30% voltage swing instead of a 63% voltage swing (1/4 TC instead of 1 TC, assuming an exposure that would cause a 99% voltage swing without T4), you could  get a highlight rolloff point (4 TC) approximately 4 stops above the standard clipping exposure level without compromising midtone performance too much. With a 12-bit sensor, this would put the standard clipping exposure level around level 1227. 4 stops above that would work out to level 4013, 5 stops above would be at level 4055 or so, still enough to maintain some highlight detail. 5 stops below standard clipping level would be ~37 or so, so on paper at least, it looks like a sensor with a 6-stop usable linear DR could be modified to capture 10-11 stops by adding a properly configured T4. And the most beautiful thing is that sensor response would roll off gradually in both the highlights and the shadows, just like film. Instead of exposing for the highlights, one could expose for the midtones, centering the histogram to avoid digging too hard into either shadows or highlights for detail.

[jani]

I tend to think that the guys in the research laboratories at Kodak or Canon have already thought of such ideas as this and not implemented them because of some serious trade-off in performance that is too difficult or expensive to work around.

While it's possible that the researchers in Kodak, Sony or Canon laboratories have considered this idea, it's also possible that they haven't.

Remember, if everything at a given point of time had been considered and tested, then there would be no further progress.

Granted, it's unlikely that anyone who isn't directly working in the field will come up with a solution that's better than "real" researchers do.

It's even more unlikely that you'll get a response out of the laboratories of Kodak, Sony or Canon if you ask them if they've tried, that would be a trade secret.

I'd suggest that if Jonathan hasn't already been in touch with techs in either of the above mentioned companies, he should try to, perhaps over a beer or something.  (That's something us techies usually appreciate.)

And if one of the companies has already thought of the idea, and are going to patent it, it's better if the idea is publicly well-known before that happens.  

[Jonathan Wienke]

Also posted at http://visual-vacations.com/Photography/se...improvement.htm:

The biggest challenge facing digital camera sensor designers today is dynamic range—the exposure difference between the brightest highlights that can be captured without clipping and the darkest shadows that can be recorded at an acceptably low noise level. The greatest obstacle to improving dynamic range is the fact that all current sensor designs are linear; a doubling of the intensity of light striking the sensor doubles the numeric value output by the sensor. Since human vision responds to light in a logarithmic fashion, this means that most of the tonal detail recorded by a digital sensor is in the brighter values; in order to obtain acceptably smooth shadow tonal response, the sensor has to have ridiculous overkill in the precision with which it records highlights. A typical DSLR sensor has a 12-bit output. This means that the sensor uses 2048 values to describe the brightest stop of tonal range it can record, but only 32 values to describe shadows 6 stops further down the brightness scale. As you can see, this is a huge mismatch, and a lot of bits are simply wasted describing brighter tones with a completely unnecessary overkill of precision.

Standard color spaces solve this problem by applying a gamma curve (usually 1.8 or 2.2) to the image data values, so that there is a logarithmic relationship between the numeric values of the image data and the associated luminance when the image is displayed or printed. This means that the same number of bits can be used to describe highlights and shadows with similar levels of precision, and has proven so effective that all commonly-used editing color spaces have a built-in gamma curve. The challenge that no has addressed so far is to design a sensor with the same innate behavior, so that there is not a gross mismatch between the precision of highlights and shadows. I believe that I have devised a very simple way to accomplish this.
The magic ingredient is one of the simplest electronic components: the resistor. A resistor passes allows electricity to pass through it in a restricted fashion. One of the forms of Ohm's Law is I = E/R; I is current, E is voltage, and R is resistance. The current flowing through a resistor (measured in amps) is directly proportional to the voltage applied to it and its resistance value in ohms. By connecting a resistor in parallel to a photodetector during exposure, the photodetector's behavior can be changed from a linear response to light intensity to logarithmic, thereby applying a gamma curve to the photodetector's analog output prior to digitization.

A photodetector accumulates a charge as it is struck by photons. You could think of it as a bucket that fills up as it is exposed to light. When the "bucket" is full, and additional light striking the photodetector is simply wasted. Connecting a resistor in parallel to the photodetector is like poking a small hole in the bottom of the bucket. When the bucket is nearly empty, very little water (or charge) will leak out, so shadow performance will be minimally affected. But when the bucket is nearly full, water (or charge) will leak through the hole quite rapidly, allowing considerably more water to be poured into the bucket before it fills completely. When exposure is complete, the hole on the bucket is plugged (the resistor is disconnected from the photodetector) and the charge level is measured and digitized. Because flow through the hole in the bucker (the resistor) is faster when the bucket is full than when it is empty, a gamma curve has been applied to the relationship between the rate at which water flows into the bucket (or the intensity of the light) and the level of water in the bucket (or level charge accumulated in the photodetector) when exposure is complete. By carefully choosing the size of the hole in the bucket (or the resistor value) usable dynamic range could be as much as doubled (assuming an optimum native sensor gamma of 2), from the current 6-7 stops to 12-14 stops without requiring a radical redesign of the photodetector itself or any change whatsoever to the A/D converter circuitry.

This would require some changes in the way RAW data is processed in-camera. Connecting a resistor in parallel with each photodetector during exposure would reduce the effect of dark current leakage and the need for dark-frame subtraction, but the reality of chip fabrication dictates that each photodetector in the sensor would have a slightly different native gamma due to variations between individual photodetector and resistor pairs, which means that the sensor would have to have a lookup table with the native gamma of each individual pixel so that they could be converted to a common gamma value before RAW conversion. But overall, the potential performance benefits would be well worth the additional processing steps.

[DiaAzul]

I would doubt that this would fly in the real world. The one factor you have missed is that your gamma curve would be a funtion of time and charge, not just charge in the photosite as your article alludes. Witness the Jack and Jill story, that by the time the reached the bottom of the hill all the water had leaked out of the bucket. Same would be true for your design, that the gamma curve would change dependent upon the amount of time that the photosite was exposed. For long exposures where there is a blurring, then you will end up with some interesting exposure artifacts due to different illumination at the start and end of the exposure.

Second problem is that as the excess charge is passed through the resistor it is converted to heat. You run a large danger that for intense illumination you will cause the resistor to heat to such a level as to damage the chip.

The forum is probably better off trying to stick to what we can do with existing technology, rather than trying to second guess what Canon, Fuji, Nikon, Sony, etc al.. should or could be doing in camera design.

[dlashier]

Interesting idea, but why not just drain to a tiny capacitor, then nothing is lost and you've effectively increased the well size. I'm sure there are practical considerations why this has not been done.

Another idea that has been patented by Canon is a feedback mechanism that dynamically controls an electronic ND filter pixel by pixel.

Further, another object of the present invention is to provide an image sensing apparatus using an optical modulation element which has controllable light transmission characteristics so as to widen the latitude of the solid-state image sensing device by each pixel or by each predetermined part in order to overcome a problem caused by a narrow dynamic range of the solid-state image sensing device, and the problem caused by unevenness in sensitivity features of pixels due to manufacturing processes.

- DL

[Ray]

Might as well offer my 2 cents worth on this. First, increased dynamic range is always going to involve greater exposure. In the days of film, using in-camera automatic exposure, I used to underexpose by about 1/2 a stop when using slide film, to avoid blown highlights, and overexpose by around 1 stop when using negative film to capture the greater dynamic range that negative film is capable of.

If I've understood Jonathan's idea, he's proposing a continuous leakage of the voltages generated by each pixel during exposure so the final reading will correspond to the brightest highlight in the scene.

I've got no idea how practical such an arrangement would be. It sounds like a clever idea. "The camera with leaky photodetectors!"  

I'd be concerned about any extra noise generated by such a system. Transistors usually need lots of feed-back and cancellation processes to reduce noise.

[Ray]

but the novel part of it is that the leakage is intensity dependent effecting changing the gamma.

This raises another question. Does changing the gamma not normally result in adding values to the darker parts of the image, through a sort of interpolation? Jonathan's process involves throwing information away on the basis that the eye cannot discern the vast number of closely spaced different levels normally produced in the upper zones/f stops.

I'm not sure about this, but I would suspect the eye's ability to differentiate between this vast number of brightish tones will depend upon final print size. The bigger the print, the easier it will be to see the tonal variation in the bright parts of the image.

Instinctively, I'm not happy about the possible bad consequences of throwing information away. I know this is what is done in jpeg compression, but we don't edit images in jpeg format.

[dlashier]

> Does changing the gamma not normally result in adding values to the darker parts of the image, through a sort of interpolation?

Not really. Values get remapped and in a limited bit world this does indeed open up new possible values, but from a technical standpoint there's no interpolation or "new" image values. Jonathan's method actually would open up new "values" in darker tones that otherwise would be mapped to the same value because of limited bit encoding.

> Jonathan's process involves throwing information away on the basis that the eye cannot discern the vast number of closely spaced different levels normally produced in the upper zones/f stops.

In a sense, but this happens anyway with a linear sensor because ultimately the same gamma change has to occur in order to view the image. Jonathan's just saying that the earlier it happens the better as it eliminates some of the obstacles to obtaining better dynamic range.

> I'm not sure about this, but I would suspect the eye's ability to differentiate between this vast number of brightish tones will depend upon final print size.

I'm sure this is true, but the fact is that 16 bit (or even 12 bit) linear data contains way more granularity in this region than required under the most demanding requirements while the dark end is just the opposite. Also iirc the eye is much more adept at detecting gradient changes on the dark end than the light end.

> I'm not happy about the possible bad consequences of throwing information away

It's not throwing away any information, at least the gamma part isn't. What it's doing is using it in a more efficient way so there's actually an increase in meaningful information (which was the point of the exercise).

If you're talking about throwing away electrons, yes, the idea gives up some speed but I'd gladly give up a stop to double DR.

- DL

[Jonathan Wienke]

For long exposures where there is a blurring, then you will end up with some interesting exposure artifacts due to different illumination at the start and end of the exposure.

No you wouldn't. The charge levels accumulated by the sensor would simply reflect the accumulation gamma curve.

Actually, after analyzing this carefully, I realized I miscalculated. What would actually happen is that some kinds of motion blur would be significantly reduced.

Imagine a typical night exposure where a car is driving through the frame with its headlights on, and the background is dark. Say exposure is 1 second, so that the headlights are near the left edge of the frame when exposure begins, and are near the right edge of the frame when exposure is complete. The image will show a streak of light going from left to right, but the intensity of the streak will not be the same all the way across as it is with film or traditional linear sensors. On the left side, the photodetectors illuminated briefly by the headlights would have some time to bleed off some of the charge dumped into them as the headlight light passed over them, while on the right side the photodetectors would not have any time to bleed off charge. This means that the streak would start out at its full brightness on the right side, and quickly fade away as it goes to the left.

Another effect of this phenomenon would be a huge difference between the effect of first-curtain and second-curtain flash. If exposure time is significantly longer than flash duration, second-curtain flash will result in a significantly higher exposure level than a first-curtain flash, even if both flash pulses are exactly the same duration and intensity.

An additional side effect would be that exposure would have to be controlled by the chip rather than a focal-plane shutter like Canon's current DSLRs, or overall exposure would be quite uneven when high shutter speeds were used--the portion of the chip exposed first would be darker than the part exposed last--unless the bleeder resistor cutoff signal was independent for each row of pixels, and each row of pixels was cut off (the bleeder resistor disconnected from the photodetector) as the second shutter curtain passed it.

But if shooting action with second-curtain flash, ghosting would be less visible than a shot taken with a linear sensor camera using the same exposure settings and flash power.

[Scanner Darkly]

I don't know much about electronics, but wouldn't it be easier to just record the the times it takes for the individual pixels to fill up. If every pixel would be configured so that when full, release charge. The arrival time of the released signal would then be recorded for every pixel and converted to exposure information. The pixels that would still be accumulating charge at the end of the exposure, would be read in the usual way. The dynamic range would maybe be limited only by the accuracy of the clockchip and the calibration.

Should I file for a patent or is this idea full of it?  :p

-SD