dynamic range and exposure

[Ian Lyons]

Michael,

My tutorial should NOT be read as the way of obtaining the "ideal exposure" for anything other than capturing the profiling target. I only quoted it so that folk could see that even when they think the highlight has been blown it very likely hasn't.

For real world images use your camera and normal methods to obtain the "ideal exposure" - capture the scene and read the histogram - if the highlights fall short of the right-hand end point increase exposure, but not so much as to blow the highlights.

Pleass don't make ass of yourself waving a white/grey card around

[BJL]

the values returned by the sensor are linear, and need to have a (logarithmic) gamma curve applied to "look right" to human eyes, since human vision responds to light logarithmically rather than linearly.

I slightly disagree: gamma rescaling (which is a power law, not quite logarithmic) is not directly to correct for the eye's response, and in fact it gets undone on output: gamma transformed values are something like

recorded level = (relative luminance) raised to the power 1/gamma

["contraction"], and on output this is undone by raising the level to the power gamma ["expansion", e.g. monitor gamma]. The virtue is that low luminance levels are "encoded" at larger numerical values, and so any subsequent noise, discretization error and such are a smaller fraction of the signal level (better S/N), and this noise gets shrunk in the expansion step. (TV used it first to deal with good old fashioned transmission noise.)

I think this is similar to the "bias" used in analog audio: when recording, boost the high frequencies, which suffer most from noise; on playback, scale them back down, reducing any hiss that accumulated in between.

This does relate indirectly to the eye's "logarithmic response" by giving smaller absolute noise levels in the shadows, where the eye can detect a lower level of noise than in highlights.

[dlashier]

Bernard, the camera histogram is after the gamma has been applied which is why Michael says "if you do not use the right-hand fifth of the histogram for recording some of your image you are in fact wasting fully half of the available encoding levels of your camera." Without the gamma applied the rightmost stop would be half the histogram.

All this aside, although the article contains some truth in the value of getting data up out of the noise band (improving S/N), it's full of so much mis-information concerning DR (eg. it does *nothing* to increase your DR unless you're currently underexposing) that I don't think it even warrants discussion. Even the bit talk is misleading. Only 128 possible values for the lowest stop! Horrors! but when you think about it, at print time you've only got 256 values for the entire 5 or 6 stop range, so 128 is MORE than adequate for the bottom 1/5 of the image (which will be represented by 50 or so values unless you're wildly pulling shadows up).

- DL

[victoraberdeen]

Michael,

Please not so, I would like to know on what you base this opinion “Most readers of this board are sophisticated enough to know that prints are done in 8 bit mode” or do you believe you are above qualification?

Further to the article, I see no device engineers, lets hear from some one who understands how the device functions, not how it is treated in software. I am sure you know the difference!

Maybe the article falls because it is neither; an expert assessment of hew the sensor device works or, an advice to photographers on how to get the results they desire.

I have to wonder if you understand photography, art and science?

Victor

PS – It is the accepted practice to provide qualification for your (or others) opinions.

[KenW]

I'm fascinated by this discussion. Unfortunately, I recently joined this forum and did not read the original article by Michael. Can anyone point me in the direction of finding it? Thanks,
Ken Weissblum

[victoraberdeen]

Looks like we are making progress, however, a little confused about gamma in the histogram! How about a few examples...

[dlashier]

> My understanding of Gamma is a measure of contrast

Victor, I'm not an expert on these matters and was hoping someone would jump in, but lacking that I'll give you some layman's comments.

I wouldn't consider gamma a measure of contrast. Technically it is a transfer function (and strictly speaking a power function) but if you need an everyday word to describe it I think 'brightness' would be closer than 'contrast'.

> Now I have a question on how the histogram is showing gamma, surly it is just a measure intensity by tone. So if the contrast is there (high to low) it is not representative of the  contrast or gamma in the picture. Did I miss it?

The histogram doesn't directly show gamma although it certainly reflects it. 'Gamma' adjustments are made because of different characteristics of devices and generally the reason is to provide similar appearance on the different devices.

But the word 'gamma' is also sometimes used in a loose sense to mean brightness (midtone) adjustment represented by pulling the center of a curve up or down. This obviously affects contrast also but is not usually the primary aim of the adjustment. Sometimes it's used in an even more general way to refer to all tonal adjustments used to render an image appropriately.

Someone please jump in and correct me if I'm too far off base.

- DL

[jwarthman]

Bruce,
I hope all those gamma references don't just obfuscate the issue!  :-)  But seriously, considering the work you've done and your mathematical bent, I'm eager to hear more of your views on gamma.

Enjoy!

-- Jim

[nnmmaa]

Michael's interesting article on dynamic range and exposure with dSLRs does not seem to me to go far enough. I have noticed that even though my exposure is accurate enough so that the histogram is similar to those pictured in the article, often one of the color channels is blown out. This leads to unacceptable oversaturation that cannot be corrected in PS. So, the question is how do we gauge the exposure in situations where one color predominates? My solution, based on experience, is to reduce the exposure.

[michael canyes]

I always keep my grey card hidden under my raincoat.

[Jonathan Wienke]

I am not getting this, the image is captured on an array of sensors, each sensor has a specification equal to the other millions. Now that means the shadow detail of an image is captured at 12bits, so can an expert please explain why the captured image would only reproduce the limited 'number of levels'.

Because shadows are a very limited subset of the total number of values. So if you underexpose, then you are further limiting an already small set of values representing shadow tones, thus causing blocky shadows and posterization.

The rest of your post is pretty meaningless. Duh, the whole point if an image sensor is to measure light. And water is wet, too. Yes, each pixel in a Bayer array is color masked, but no, that doesn't mean it has only 4 bits of accuracy. Each pixel is a 12 bit light measuring device in a specific color band. Silicon has a linear response to light intensity, therefore the values returned by the sensor are linear, and need to have a (logarithmic) gamma curve applied to "look right" to human eyes, since human vision responds to light logarithmically rather than linearly.

Don,

I'm not interested in debating the article with you. It stands on its own. You are free to agree or disagree, of course.

But I'd like to take you up on your last point...

"...at print time you've only got 256 values for the entire 5 or 6 stop range, so 128 is MORE than adequate for the bottom 1/5 of the image "

Most readers of this board are sophisticated enough to know that prints are done in 8 bit mode. The point that you appear to overlook with the comment is with regard to the advantages of working in a larger bit space prior to printing. That's why an understanding of this issue is worthwhile. (The fact that 6 or 7 ink dithering simulates a wider space is another issue).

People who are new to digital image processing, and even many of us who aren't, need to understand the underpinnings of our craft. Thomas and I were trying to shed some light on how imaging chips work so that photographers can have an understanding of why certain things happen.

Indeed, I suggested in the article that people try a simple test themselves, not so much to give them a tool that they might use in the future (though that too) but so that they could see first-hand what's going on "behind the curtain". Based on the huge e-mail feedback that I've received thus far I believe that this has been accomplished.

Michael

No. I think not.

The article stands as is and is supported by expert opinion. I have no inclination to enter into an exhaustive debate.

If you find it unsatisfactory, so be it.

Michael

[jwarthman]

Lots of good information in this thread!

Schewe (Jeff?),
You're right - it would be nice to have the option to see a linear histogram vs. a gamma-adjusted one. But as long as we treat the (Canon) histograms as representing the gamma-adjusted scene, can't we make make judgements based on the histogram?

Something I'd like to learn more about is the process of adjusting the image in Camera Raw (or other RAW converter) to take advantage of the linear data, while minimizing noise. I see relationships between the "exposure", "lightness", "darkness", and contrast controls, but I'm still unsure what, exactly, they're doing to my data. Yes, I've tried manipulating the sliders and watching the histogram. I still can't generalize to understand what's happening, beyond noticing the changes to the image.

Ray,
You raise a good point about the interconnectedness of dynamic range and signal-to-noise. And I agree with your description of how to handle the case where the camera has less dynamic range (with acceptable signal-to-noise) than the scene.

As for your other example where the camera has more dynamic range than the scene, It seems that the "best" approach would be to overexpose the image, while being careful not to blow out the highlights. That way the majority of the pixels are up out of the noise. If I understand this correctly, more bits are used to represent the image data the farther to the right the data is in the histogram. This should be the case whether the histogram represents gamma-corrected data or linear data.

In Photoshop, (or RAW converter) I suppose one goal would be to adjus the image to look, as much as possible, like the original scene. (Yes, many photographers prefer to take liberties with the image, and that's okay too.) Whatever the goal, if you wish to portray lots of deep shadows with little or no detail, then moving the left slider to the right will accomplish that. If you want to preserve lots of detail at the bottom-end, then perhaps adjusting the middle slider would be more appropriate.

These steps would tend to "undo" the overexposure, and possibly "stretch" the image to fill the histogram. But it seems that one result would be to avoid the "noisy" left side of the historgams, possibly resulting in cleaner shadows.

Okay, I'm all ears if I've bolluxed any of this up!

Enjoy!

-- Jim

[dlashier]

> a little confused about gamma in the histogram! How about a few examples...

Ok, http://www.lashier.com/home.cfm?dir_cat=25027&gal_col=5

I'd estimate that the 'bars' on these histos are roughly 2/3 stop apart. In particular note the linear histo. Even though this shot was only 1/3 stop away from blowout the linear data barely reaches midpoint. The Canon RIC histo is probably similar to what's shown on the camera. Also note the compression on the ends of the 'non-linear' conversions. These represent the toe and shoulder rolloffs in the TRC's.

ps I've added another page showing the non-linear to linear and vice versa transformations.
- DLhttp://www.lashier.com/home.cfm?dir_cat=25028&gal_col=5

[dlashier]

> I take it EC stands for exposure compensation, but not sure about TRC.

Ray, "Tonal Reproduction Curve" - see http://www.lashier.com/home.cfm?dir_cat=14477&gal_col=4

> There's a price to pay. Does this make sense?

Yes. Trade-offs seem to abound in this game, which is what I was trying to point out about the 'expose right' theory. For me the trade-off usually isn't good. If there's any "rule of thumb" I've discovered it's "the less pushing and shoving the better". This means I generally aim for an exposure that requires the least adjustment over the range of the image. But the corollary is "if you must adjust it's better to adjust down than up" which means that in some cases you may expose for the shadows.

- DL

[victoraberdeen]

Yes Don, I never think about the histogram, expose for the subject and compensate for the scene,  sounds crazy but, I meter with my eye and then check with a spot meter.

The runners were shot on raw mode and I used two different converters (Fuji / Bibble), both the same magenta. I do think the software should do what film would and burn white here! What do y'all think?

Victor

[victoraberdeen]

I am not getting this, the image is captured on an array of sensors, each sensor has a specification equal to the other millions. Now that means the shadow detail of an image is captured at 12bits, so can an expert please explain why the captured image would only reproduce the limited 'number of levels'.

Now what I think is meant by this is the exposure value only turns the luminosity level to a particular number, but this will work on a binary scale based on  how bright the light falling on the sensor is. In this context the sensor is a measuring device! Except there is another limitation in that each sensor is masked with a color, so you are only measuring 33% of the light, is that 4 bits of the 12?!

Then the solution must be for the next generation of sensors to measure color temperature (being RGB) and luminosity at a analogue to digital conversion level that creates clean tone across the dynamic range!

[BJL]

Even getting up from the current 8 to 12 bit D/A conversion to full 16-bit per color accuracy would require sensor S/N ratios of about 65000:1, about ten times higher than anything available now in either film emulsions or digital camera sensors. The best performance so far is probably good negative film, equivalent maybe to up to 12 or 13 bits depth.

Most current digital sensors do not have good enough S/N ratio to give full accuracy down as far as 12 bits; the least significant bit is probably all noise in many DSLR's, as indicted be the fact that the dynamic range of such cameras is distinctly less than 12 stops.

If noise levels could be reduced enough to give 16 significant bits/color, it would give close to a 16 f-stop range and almost a hundred luminance levels per f-stop even way down in deep shadows ten stops below maximum brightness. From what I have read about the eye's capabilites, this would go well beyond what the eye can handle in both dynamic range and distinction of luminance levels. It would certainly far exceed the luminance range of any image display method; slide, video or print.

In summary, full 16-bits per channel luminance resolution would already be more than we could ever need, (though people who love having big numbers displayed on decals on their camera bodies might want and pay for more!)

[victoraberdeen]

If only cash flow was not a constraint!

Victor