> From: Preston Earle > > I've been thinking more about this 8-bit vs. 16-bit question, and one > thing puzzles me and has generally been ignored in this discussion. > Someone (Arthur, Austin, Laurie, ????) brought up the question of > "noise" in image data, but that issue has been bypassed in these > discussions in favor of other comments. Yet it seems noise is the reason > high-bit data is superfluous. What I'm thinking: > > 1. High-bit data is very small compared to low bit data. The ninth-bit > is only 0.25% of the value of the full tonal range of 0-100%. > > 2. All visible files are the product of a final resize/pixel-combination > of some sort, at least until we get 2800x4200 or larger video screens. > > 3. When scanners measure and assign digital values to image elements, > adjacent pixels are given discrete values that are generally different > by more than 0.25%, that is, the precision of the measurement is less > than 8-bits for adjacent pixels. > > 4. Image editing steps which spread existing pixel ranges over larger > ranges do not create more precise intermediate values than the starting > value's precision. If an intermediate pixel value must be created > between two pixels whose values are 128 and 130 (8-bit), the value won't > be more precise if the original values are 127.504 and 129.504 (16-bit). > > I don't know how typical CCD scanners scan at lower resolutions than > their maximum. Whether by averaging pixel values along the CCD array and > making larger steps along the film movement, or by some other way, they > still end up with adjacent pixel values that differ by more than 1 unit. > Knowing these values to .5-unit precision doesn't change the average > values reported. > > For 16-bit processes to be relevant, wouldn't adjacent pixels have to be > identical to more than 8-bit precision?
I was the one who brought it up. A couple of points: 1) When you reduce an image by averaging the pixels, you reduce the noise. Cutting the linear resolution in half, for instance, effectively averages a square of four pixels together, which reduces the noise by a factor of the square root of four, which is two, or 6db, so effectively allows for one more bit of useful resolution. 2) To avoid all banding in 8-bit data, the signal-to-noise ratio must be poorer than you might think. I've found from experimentation that you need five or six lsbs of noise, peak-to-peak, to break up the banding. With only two or three lsbs, you won't see outright posterization, but you'll see bands of more noise alternating with bands of less noise. (This is only if you apply a really drastic curve.) I suspect, though, that most good film scanners have plenty of S/N, to where more than eight bits really could be useful, if the film had a quiet enough image. The question, then, is whether film grain itself supplies enough noise to make the extra bits useless. In my experience, scanning Kodachrome 25 and E6 100, I've always seen plenty of film grain noise to render the extra data useless, even for B&W. If I take a clear blue sky, which is about as noiseless an image source as you can find, convert to 8-bit B&W, and then apply a drastic curve to stretch the gradient out to the full black-to-white range, I see no banding at all, just lots of noise. I've not done much negative scanning--it may have less noise. I've also done some pretty rigorous tests on my two digicams, the Minolta DiMage 7 and the Canon 10D. In raw mode, the DiMage 7 has absolutely no useful information in the extra four bits, under even the best circumstances, but the 10D does. -- Ciao, Paul D. DeRocco Paul mailto:[EMAIL PROTECTED] ---------------------------------------------------------------------------------------- Unsubscribe by mail to [EMAIL PROTECTED], with 'unsubscribe filmscanners' or 'unsubscribe filmscanners_digest' (as appropriate) in the message title or body
