Puzzled about in-camera histograms.

[kevin_krumwiede2]

I'm a bit confused by <a

href="http://luminous-landscape.com/tutorials/expose-right.shtml">this

article</a> on Luminous-Landscape.com. I understand it as far as the

table relating f/stops to pixel values.

 

<p>However, have a problem with the paragraph immediately following

the table. As the article states, CCDs are linear capture devices.

Am I incorrect in believing that the horizontal axis of the in-camera

histogram is also linear? The article is based on the notion that the

horizontal axis is logarithmic. In investigating this, I have learned

that at least some raw conversion software can display logarithmic

histograms. But most digital imaging software I have encountered uses

linear histograms.

 

<p>If the in-camera histogram <i>is</i> logarithmic... why? Isn't

that an unnecessary (and counter-intuitive, if you know how CCDs work)

throwback to film behavior?

 

<p>Thanks,<br>

Krum

[digitmstr]

The histogram is like an oscilloscope for sound waves, it simply tells you the content of the samples.

 

The "right bias" theory explained in the L.L. article is correct but, it is nothing new for those of us who work and/or know about digital audio conversion.

 

IN order to squeeze all of the bits out of an audio A/D converter (usually 24bit) the recording level has to be appropriately high (without clipping). If it isn't, you will incrementally lose bits and your recording will suffer. Same for digital imaging and, as far as I know, ANY sort of A/D conversion.

 

Some programs, such as C1 PRO, will also allow you to use non-linear conversions when "developing" your RAW images.

[digitmstr]

To sum it up in a more...palatable way: making a good A/D conversion is like making organge juice, if your juicer is not strong enough half the juice will be thrown away with the orange ;-)

[keith_urry]

Kevin, I'm not sure why you're confused. You say the article is based on the notion that the horizontal axis is logarithmic, but I don't see what diffence it makes - the recommendation of shifting everything to the right as much as possible still holds. As far is I know though, the scale is logarithmic to match the f-stop progression and the eye's perception of brightness. That seems the most intuitive behaviour to me - how many photographers want to think in terms of image sensor response?

[Brad_]

Actually, when loading an A/D converter, at least with signal (non-image) data, permitting

some clipping is actually an optimal strategy. The amount of permissible clipping

depends upon the statistics of your signal and the A/D's inherent noise level, expressed as

SNR or effective number of bits.

 

If you allow too much clipping, wideband noise/distortion is generated. If you play it safe,

and back your signal input off some to guarantee you never clip, SNR will suffer as you

never have a full scale signal, from which to maximize SNR.

 

So, instead, you want to load your A/D such that the clipping occurs occasionally, to the

degree that the extra clip noise generated is commensurate (you pick how close) with the

inherent noise level of the converter. Depending on your signal's statistics/crestfactor,

that can vary widely. Similarly, with image data, a few clips may be tolerable, depending on

the distribution. There's always that trade to consider when trying to maximize dynamic

range.

[digitmstr]

>>The amount of permissible clipping depends<<

 

you are splitting hairs in a photo forum here :) It is understood that "clipping" was meant as "digital distortion" as in a clipped wave/samples.

[kevin_krumwiede2]

Keith & Giampiero, thanks for your answers. I know that recording high can increase the SNR, because while the signal increases, at least *some* of the noise in the system does not. What I disagree with is the statement that the top fourth or fifth of the histogram represents "fully half" of the possible pixel values. If the horizontal scale is logarithmic, then the article is correct. But rarely have I seen a logarithmic histogram in digital imaging software, and I'd be surprised if that's how it's represented in the camera. Of course, I could be wrong!

 

<p>It really makes no difference, in the sense that you should still do what the article recommends doing. I'm just being picky about <i>why</i>.

[gordonr]

The horizontal axis is linear, but the values are gamma-encoded - a form of semi-logarithmic coding used on all computers for displaying images, and a source of endless puzzlement to those not "in-the-know".

<p>

See: <a href="http://www.photo.net/equipment/nikon/scanner/ls-1000/gamma">Color Depth and Monitor Gamma</a> (http://www.photo.net/equipment/nikon/scanner/ls-1000/gamma) and <a href="http://www.poynton.com/notes/colour_and_gamma/GammaFAQ.html">Gamma FAQ</a> (http://www.poynton.com/notes/colour_and_gamma/GammaFAQ.html).

[keith_urry]

After a bit of research, I think Gordon is correct about this. From what I have read it seems that the histogram is derived from the jpeg representation of the image - it's just a count of the number of pixels in each luminance band. The process of converting to jpeg involves an exponential function with an exponent of 2.2 (for the standard sRGB colour space). That means the horizontal axis of the histogram is logarithmic, but it doesn't match the f-stop progression exactly because that would require an exponent of 2.

<p>

None of this affects the advice in the article, but it should be pointed out that it only applies if you are shooting RAW.

<p>

There's a great article on this subject by Norman Koren, <a href="http://www.normankoren.com/digital_tonality.html">here.</a>