Better than ETTR ?

[marcmccalmont]

The article on ETTR got me to thinking. My father and his chief engineer (Arnold M. McCalmont, Matthew Slate, Technical Communications Corporation) patented a signal processing technique  for voice encryption that might have an application in digital imaging. Their technique is clever and it improved recovered voice quality on HF SSB radio links which is unheard of with analog crypto gear, intelligibility usually gets worse. The idea is simple, high pass and low pass the analog signal, invert the lower half and read it out at twice the frequency (for the analog encryption they encoded the analog segments using time division multiplexing (TDM).
So ETTR is trying to put as much data as possible into the most significant bits, in effect lowering the noise floor.
My thought is to take the analog signal, high pass and low pass filter it, invert it, digitize all 4 segments (highlights. upper mid, lower mid and shadows), store each segment as a 16  bit word. I'm thinking if digitized at 2x the frequency the information would be stored in the most significant byte? In this way all four segments of the signal would have higher resolution. Your thoughts? I haven't thought it out thoroughly so don't be too critical just throwing the concept out to see if it has merit. Oh, all rights reserved Marc McCalmont 2011 just in case it might have commercial merit.
Marc

[jeremypayne]

As I understand it, the point of ETTR is to collect the maximum number of photons. 

[marcmccalmont]

good point! the captured photons are the captured photons no mater how they are recorded
marc

[dreed]

So ETTR is trying to put as much data as possible into the most significant bits, in effect lowering the noise floor.
My thought is to take the analog signal, high pass and low pass filter it, invert it, digitize all 4 segments (highlights. upper mid, lower mid and shadows), store each segment as a 16  bit word. I'm thinking if digitized at 2x the frequency the information would be stored in the most significant byte? In this way all four segments of the signal would have higher resolution. Your thoughts? I haven't thought it out thoroughly so don't be too critical just throwing the concept out to see if it has merit.

This sounds like details that would be relevant to the analogue->digital converter on the sensor or in the camera regardless of exposure. The final step in the above needs to be converting those four 16bit values into one 16bit value or the size of a raw data file is going to quadruple - not something that most people would enjoy.

Oh, all rights reserved Marc McCalmont 2011 just in case it might have commercial merit.

Find an intellectual property lawyer and ask them what rights and protection you have for a design that you've detailed in a public forum.

[Guillermo Luijk]

As I understand it, the point of ETTR is to collect the maximum number of photons.

The point of ETTR is to maximize SNR. The best way to achieve this is by maximizing the number of photons, but if you cannot increase aperture and shutter and still have insufficient exposure, you can push ISO to maximize SNR (not all cameras will benefit equally; Canons are the ones which take more advantage of pushing ISO, the new Sony sensor found on Pentax K5 and Nikon D7000 gets a negligible improvement from pushing ISO).

In the following two images from a Canon 350D, the same number of photons were collected (they were shot at the same aperture and shutter, only changing ISO):




So the technique exposed by Marc would be of no advantage here. ETTR is not about number of levels (they are always sufficient, no matter how much you expose, because noise always dithers posterization), ETTR is only about improving SNR.

Regards

[marcmccalmont]

This sounds like details that would be relevant to the analogue->digital converter on the sensor or in the camera regardless of exposure. The final step in the above needs to be converting those four 16bit values into one 16bit value or the size of a raw data file is going to quadruple - not something that most people would enjoy.

Yes it is an analog technique to increase the "resolution" of the digitized signal, the increase in file size might be worth it if the IQ improved enough

[marcmccalmont]

one might argue that the shadows don't have enough levels and more in the lower greys might help too.
Marc

[Guillermo Luijk]

one might argue that the shadows don't have enough levels and more in the lower greys might help too.

More levels could help in what? in displaying noise more accurately?.

Look at these images, left is 8-bit, right is only 5-bit:



However the extra levels in the first image are useless, because noise is greater than the gap between two 8-bit adjacent levels:

[marcmccalmont]

I'm not proposing not to ETTR. ETTR does 2 things, lowers noise and for each stop of ETTR doubles your usable levels correct?
As sensors get better and lower noise I'm sure the middle greys where eyesight is most sensitive could be improved with more resolution (levels).
Granted more data below your noise floor is useless but most of your data in the highlights not in the midtones doesn't seem logical to me just the result of binary math?
Are more levels important? I see a slight improvement in print with a 16 bit work flow and a 16 bit print driver, small but noticeable.
Marc

[michael]

Dissing having more levels may be fine from a theoretical perspective, and yes, noise is the main reason to use ETTR. But...

In the real world (rather than the theoretical one) it is frequently the case that one wants to manipulate the image (open up shadows, for example, to reveal  nuanced detail). If you've used ETTR and then "normalized" the image you now have many more tonal level in the shadows than you would otherwise have had.

It doesn't take much time in front of the screen to tell that the benefit of doing this is quite real. And, as with most such things, maybe not in the day to day, but most usefully in the extremes.

Michael

[Graeme Nattress]

More tonal levels (I normally talk "code values") don't help if they're just more code values representing noise. Say you have a sensor, and you normally have a 12bit DAC on it, but you swap to a 16bit DAC. You now have vastly more code values in the shadows, but you don't necessarily get a better picture unless the sensor has low enough noise to actually put image data (rather than noise) into those newly added code values.

Graeme

[bjanes]

Dissing having more levels may be fine from a theoretical perspective, and yes, noise is the main reason to use ETTR. But...

In the real world (rather than the theoretical one) it is frequently the case that one wants to manipulate the image (open up shadows, for example, to reveal  nuanced detail). If you've used ETTR and then "normalized" the image you now have many more tonal level in the shadows than you would otherwise have had.

It doesn't take much time in front of the screen to tell that the benefit of doing this is quite real. And, as with most such things, maybe not in the day to day, but most usefully in the extremes.

Michael

Michael,

You have to move away from the number of tonal levels affecting image quality and start considering noise and signal to noise. A high number of levels is advantageous in producing smooth tonal gradations, but the advantage is lost when the image noise exceeds the quantization step as shown earlier in this thread by Guillermo and discussed at length by Emil Martinec in his excellent treatise Noise, Dynamic Dange, and Bit Depth. The brightest f/stop of a 12 bit raw capture contains nowhere near 2048 levels but it does have a high SNR.

Exposing to the right increases the SNR over the entire exposure range and the improved SNR is often most appreciated in the deep shadows where noise limits the dynamic range of the capture. According to Emil, raw data are never posterized, since noise exceeds the quantization step, but posterization may be brought by processing (for example, application of a median filter).

In most cases, ETTR improves shadow image quality by increasing the SNR and not by increasing the number of encoded levels. Insufficient levels results in posterization. My experience with Nikon DSLRs (D700, D200, and D3) is that shadow image quality is usually affected by noise and not by posterization.

I encourage you to read Emil's essay which discusses your previous ETTR article in this section and get back to us on Emil has screwed up things. However, I will not hold my breath while awaiting your reply.

Regards,

Bill

[michael]

I've read Emil's treatise, and have great respect for his work. And, I also understand the theory.

But, in practice, I am confident that I see smoother tonal gradations on ETTR "normalized" shadow areas rather than native ones when I need to strongly open up such tonalities.

So, who do I believe? As the old joke has it, "The experts, or my lying eyes"?

Michael

Ps: Bumble bees can indeed fly, and prove it to themselves every day.

[Graeme Nattress]

Michael, the question is not do you see an improvement in ETTR, but is that improvement due to more code values or lower noise? If you actually saw banding / posterization in the shadows then the answer is "ETTR is better because of more code values for shadows", whereas if you don't see banding / posterization, it's due to "ETTR lowers noise in the shadows".

I've experimented with bit-depth in RAW processing and indeed, if you reduce the bit depth precision too much (and at the wrong point in the processing) you can indeed induce banding / posterization. I could certainly see that happening if you put a DAC on the sensor that lacks sufficient precision for the task at hand - say you put a 10bit or 8bit DAC on a modern sensor.

Graeme

[deejjjaaaa]

But, in practice, I am confident that I see smoother tonal gradations on ETTR "normalized" shadow areas rather than native ones when I need to strongly open up such tonalities.

may be the software is just not doing a great job, that's it... 2x2 = 4, but if you software will display 5 - are you going to believe it ?

[fdisilvestro]

if you reduce the bit depth precision too much (and at the wrong point in the processing) you can indeed induce banding / posterization. I could certainly see that happening if you put a DAC on the sensor that lacks sufficient precision for the task at hand - say you put a 10bit or 8bit DAC on a modern sensor.

Graeme

Exactly!, that's the key issue. As long as you have sufficient bit depth, all that matters is to maximize SNR

[Eric Myrvaagnes]

So do I conclude that I'm right to be trying to use ETTR, even though I may be doing it for the wrong reason??? 

[ErikKaffehr]

Hi,

Yes, ETTR is a bright idea, but avoid clipping. I had done some ETTR experiments lately and got bad clipping.

Best regards
Erik

So do I conclude that I'm right to be trying to use ETTR, even though I may be doing it for the wrong reason??? 

[michael]

Graeme,

You may very well have hit the proverbial nail on the head.

I don't just use one or two high-end cameras. I use a dozen or more cameras a year, many of them low-end consumer grade for testing.

So, let's just say that since I have observed this phenomenon from time to time, I like to tip the odds in my favour, even if doing so may not always be necessary.

Once again for clarity. I most definitely understand the science. It just doesn't always jibe with my experience.

Michael

[digitaldog]

Yes, ETTR is a bright idea, but avoid clipping. I had done some ETTR experiments lately and got bad clipping.

Then that wasn’t proper ETTR, it was over exposure. ETTR is about optimal exposure for the data. If you clip and didn’t wish to, that was an exposure error.