Tutorial: Illustrated Guide to B&W scan & processing

[robc]

Perhaps you would care to define for the benefit of the list just exactly what an excess data point is and also what makes scanning a random process.

[robc]

Perhaps you would care to define for the benefit of the list just exactly what an excess data point is and also what makes scanning a random process.

to add: the fact that the gif is not random is purposefully designed to emphasize the destructive process of downsizing. Just because a landscape image is relatively random doesn't mean that the destructive process of downsizing isn't taking place. It is, but it is much less evident to the eye than with a highly structured image.

So arguing against the validity of this is the same as saying it doesn't matter if fine detail is altered by the downsizing process. Well thats fine except if thats the case, then why bother with such a long drawn out process to achieve fine detail? I guess it depends on whether you are trying to achieve apparent fine detail or as close to real fine detail as possible.

[Larry Gebhardt]

Rob, scanning at 4800dpi and then averaging the 4 pixels should give you a 2400 dpi that has less noise than one scanned directly at 2400. Assuming the noise is random (which it must be to be noise) taking 4 samples and averaging will result in a more uniform signal than one sample. I have done this experiment with both resizing and multi-sampling on a drum scanner. It does work very well.

Also, the Epson scanner buze is using isn't capable of capturing much detail above 2400dpi, so scanning at 4800 dpi is only giving him 4 samples for each real pixel. I would argue that on my 4870 the real resolution is around 1600dpi so you could really get about 9 samples to average.

The problem with buze's method isn't noise reduction, but rather the claim that averaging 8 bit pixels when resizing will give you 16bit precision. In an ideal world it may not as the initial 8bit conversion will round all pixel values to one of a very few values in the dark blacks (whites in negative film). Averaging won't bring back the subtle differences if all the neighboring pixels were rounded to the same value as well.

[buze]

But, scanning at 16 bits is not going to give you 16 bits anyway. As I explained, you never reach 16 bits because for that you would have to have an image that "covers" the complete DMax of the scanner, and most of the time, you use just about a half . Here goes one bit.
Then, that remaining signal is "stretched" to fit the 16 bits space of the file, and photoshop remove another bit when opening it (because it is actually 15 bits internally). Here goes another one...

My method gives you 10 bits of clean signal, for a file that is massively smaller than your noisy 14 bits one. I'm just saying that the loss of quality incurred by using JPEG is acceptable to me, /especialy/ since my method further down ensure minimum loss of signal and banding (something that most people using complex curves, don't)
And again, this is a detail; if the software supported JPEG 2000, I would keep the "16 bits" and make this whole argument irrelevant.

The "interesting" bits of my tutorial are not about the source files anyway.

[David Luttmann]

But, scanning at 16 bits is not going to give you 16 bits anyway. As I explained, you never reach 16 bits because for that you would have to have an image that "covers" the complete DMax of the scanner, and most of the time, you use just about a half . Here goes one bit.
Then, that remaining signal is "stretched" to fit the 16 bits space of the file, and photoshop remove another bit when opening it (because it is actually 15 bits internally). Here goes another one...

My method gives you 10 bits of clean signal, for a file that is massively smaller than your noisy 14 bits one. I'm just saying that the loss of quality incurred by using JPEG is acceptable to me, /especialy/ since my method further down ensure minimum loss of signal and banding (something that most people using complex curves, don't)
And again, this is a detail; if the software supported JPEG 2000, I would keep the "16 bits" and make this whole argument irrelevant.

The "interesting" bits of my tutorial are not about the source files anyway.

You keep missing an important point....while you won't get true 16 bit data from the scan, you will get more than you will from an 8 bit scan. Thus, using your procedure with the 16 bit scanner setting will yield you more usable data. As well, saving to jpg is a non-starter....the results aren't as good....period!

The interesting point in your tutorial fall by the wayside with your erroneous capture and filing methods.

[tim atherton]

I'm just saying that the loss of quality incurred by using JPEG is acceptable to me, /especialy/ since my method further down ensure minimum loss of signal and banding (something that most people using complex curves, don't)
And again, this is a detail; if the software supported JPEG 2000, I would keep the "16 bits" and make this whole argument irrelevant.

So, for example, how often do you make 50" or 60" prints?

[robc]

Rob, scanning at 4800dpi and then averaging the 4 pixels should give you a 2400 dpi that has less noise than one scanned directly at 2400. Assuming the noise is random (which it must be to be noise) taking 4 samples and averaging will result in a more uniform signal than one sample. I have done this experiment with both resizing and multi-sampling on a drum scanner. It does work very well.

I've written this before on this list but I'll do it just one more time.
Dowsampling to reduce noise does average or perhaps a better term would be smooth out the noise. But in doing so it also reduces sharpness. Reducing sharpness equates to losing detail. Get it? It Loses detail as well as smoothing. i.e. smothing equals reduction in fine detail period. Now most people consider that 360dpi in the print is all you need because of the human eyes lack of ability to resolve detail past approx 7 line pairs per millimeter and that is with very high contrast line pairs. What I am saying is that instead of downsizing to reduce noise (which also loses sharpness and introduces downsizing artefacts) so that you can print at 360dpi, you don't downsize and you print at 720dpi instead. Result equals no loss of sharpness through downsizing. No downszing artefacts introduced to image. AND because the noisy pixels are now printed so close together, a group 4 pixels which would have been averaged by your downsize will npw be one line pair which cannot be resolved by the human eye at 720dpi. That means it will be blurred which means to all intent and purpose it is averaged. Get it? Downsizing is unecessary to achieve removal of noise. Get it? And you have not introduced any other aliasing artefacts and softening of the print to get to that point. And you can forget all the crap about signal noise ratios in the knowledge that there really isn't any need to worry about it if you print at a high enough dpi to render it irrelevant. And you won't need as much final sharpening in the print which also means less artefacts. Its the simplest workflow possible and the only limiting factor is whether you have a pc capable of processing the files. If not then get one.

[robc]

My method gives you 10 bits of clean signal

You're talking complete bollocks. The only signal is happening inside the scanner. Once that signal is converted in the analogue to digtal converter inside the scanner the term signal is no longer valid. You're stuck in analogue mode. Convert yourself to digital mode where you will understand that digital bits are no longer susceptible to noise or signal variations. If they were then all computers would fail.

[David Luttmann]

You're talking complete bollocks. The only signal is happening inside the scanner. Once that signal is converted in the analogue to digtal converter inside the scanner the term signal is no longer valid. You're stuck in analogue mode. Convert yourself to digital mode where you will understand that digital bits are no longer susceptible to noise or signal variations. If they were then all computers would fail.

Rob,

He doesn't get it because he has remained on the theoretical side as opposed to working to obtain high quality output in large prints. Theory and reality don't often mesh here. That is why certain sharpening methods that work best in theory, don't give us the best results in printing....ditto with interpolation algorythms. Until he actually compares prints like I have at large sizes, his theory of what is best holds no water with me....especially since I've compared in the past and what he states with scanning bit depth & file compression is plain wrong....and visible!

[Ed Richards]

> Dowsampling to reduce noise does average or perhaps a better term would be smooth out the noise. But in doing so it also reduces sharpness.

Only if you had 4800 real dpi that you were averaging. With these scanners you have about 1800 real DPI, so that when you average the data from 4800 to 2400, you are not losing detail because there was no real 4800 DPI (or, really, 2400 DPI) detail there in the first place.