About bit-depth and ISOless, wall of text.

Started Apr 8, 2012 | Discussions thread
ForumParentFirstPreviousNextNext unread
Flat view
Peksu
Junior MemberPosts: 28Gear list
Like?
About bit-depth and ISOless, wall of text.
Apr 8, 2012

Here is some food for thought on the common subject of bit depth, at a file and at a practical level. These issues are brought up all the time, and I wanted to create an easy reference for future. This started as I was explaining these things to a friend, and after all that writing I thought I could just format it a bit better and post it for anyone to see. (Maybe it got out of hand as I was writing).

I could be wrong with anything, please discuss. Also, I apologize for my English beforehand, it's not my first language.

A JPG file has only 8 bits/color/pixel, and you often hear that that is what's required from the camera/RAW file. However, the camera RAW stores linear bits, and a JPG stores nonlinear bits. If the JPG is encoded in sRGB color space, we could assume that that nonlinearity should follow gamma 2.2 from the sRGB standard. Gamma 2.2 uses more information for the brightest luminosity levels, that being almost exactly twice* as much as it would were it linear instead, thus you need nine (9) linear bits to create a "perfect" JPG from RAW.

( This is calculated from a graph, not from a table or formula, to save time. If you google "gamma 2.2 curve", you can find the curves with encoded values (0-255) on the horizontal axis and luminosity (0-1) on the vertical axis. In practice you can multiply the luminosity by 255, because that's what the actual output is usually going to be, the steps become of equal size, and you can see that the slope at the highest luminosity is almost exactly 2.)

The above assumption is only valid if the JPG and RAW have the exact same dynamic range. In practice, you would almost always end up cropping some of the darkest/brightest/both immensities off when you turn a RAW into a JPG (or any similar format), whether in-camera or not. Thus, let's ad another (10th) bit, and now we can crop the RAW into half of it's full dynamic range as we turn it into a JPG, without inducing banding. Such an extreme change is quite rare, so the tenth bit leaves some further room for editing.

Another common subject is the RAW-encoding being done either at a 12 bit depth or at a 14 bit depth. The data is originally an analog signal from the sensor, and a sampling chip measures the signal and turns it into a digital value at either sampling accuracy. In practice the real accuracy at witch the signal can be measured could be anything, and doesn't need to be anything even close to the stored sampling accuracy. If we go by the DxO Mark measurements (witch seem well done), even the camera with the highest measurements so far, Nikon D800, can only distinguish 9.63 bits of luminance data. That is at the ISO 74 (100 by manufacturer), and drops as ISO is increased. Another examples, Sony Nex-C3 peaks at 8.55 bits, Canon D600 at 8.53 bits and Canon 5D Mark II at 9.19 bits. There is no harm in storing extra bits, but no more information can be deducted from the file than can actually be sampled from the sensor.

If we look at color data instead of simple luminance, the numbers are even harsher. For each output pixel you can have either 36 (12+12+12) bits or 42 (14+14+14)* bits of color data. Even D800 only distinguishes 23,7 bits of color data at base ISO, and Canon D600 distinguishes 22,1 bits.

( You can add the bits up, because the combinations of tones for a given pixel come from the multiplication of the defining factors, and adding binary data is much like doing multiplication in the normal number system.)

My conclusion is that it makes no practical difference whether the RAW file is stored at 12 or at 14 bit accuracy, both exceed the actual sampling accuracy by many fold (2 bits is 4-fold, 4 bits is 16 fold). The reason for such accuracy being used might be as preparation for future, to ensure format compatibility, or maybe just marketing. There must be a reason why not all manufacturers chose to use 14 bits.

Thank you for reading, please discuss if you disagree, I could have made wrong assumptions anywhere.

I'll post some thoughts on the issue of so called ISOless sensors in the following post. Thank you.

 Peksu's gear list:Peksu's gear list
NEX5R Tamron SP AF 60mm F/2 Di II LD IF Macro Samyang 35mm F1.4 AS UMC Sigma 19mm F2.8 EX DN Samyang 8mm F2.8 UMC Fisheye
Canon EOS 5D Mark II Nikon D600 Nikon D800 Sony Alpha NEX-C3
If you believe there are incorrect tags, please send us this post using our feedback form.
ForumParentFirstPreviousNextNext unread
Flat view
ForumParentFirstPreviousNextNext unread
Keyboard shortcuts:
FForum PPrevious NNext WNext unread UUpvote SSubscribe RReply QQuote BBookmark post MMy threads
Color scheme? Blue / Yellow