Scan at Maximum Optical or Stated Resolution?

[Ted Harris]

I know there is a school that believes in scanning at a slightly higher resolution than the real optical resolution of your scanner but i don't see any real advantage in doing this .... nor have any tests I have done shown there to be any advantage. I am also not sure that I understand the interpolation issue as it relates to the scan.

Using the 4990 as an example, when you scan at the stated resolution of the scanner, 4800, all you are doing is capturing blank information. The scanner is designed to theoretically scan at a resolution of 4800 but all most of the CCD's are doing is capturing blank info. The real world resolution of this scanner is ~2200 so if you scan at that or near that (2400) you will capture all the available info (depending on your settings) and will keep your file size much more reasonable. There are even some that argue that scanning at the higher resolution will create 'noise' in the image but I have no proof of that.

The point here is that the hardware's real resolution, when we are talking about consumer flatbed scanners, and the manufacturer's claimed resolution differ by a factor of 2-4 depending on the scanner in question. Yes, scan at the hardware's optical resolution; it's real resolution not its theoretical resolution as claimed in the specs.

As discussed in many other threads, these issues do not exist with the high end scanners which actually deliver as promised; in part because there are pinting industry standards (mostly set by Seybold) to whcih the manufacturers of these machines test and write their specs.

[robc]

Using the 4990 as an example, when you scan at the stated resolution of the scanner, 4800, all you are doing is capturing blank information.

The scanner is designed to theoretically scan at a resolution of 4800 but all most of the CCD's are doing is capturing blank info.

what do you mean capturing blank info? Are you suggesting that the software is dropping scan lines. If so I would like some evidence for this. The reason I don't beleive it (in the case of Silverfast) is because:

If the scan was done at minus 2 factor of the hardware then I think that would would work fine. But what if you set the the scan res to say 1378. 4800 is not divisible by 1378 exactly so which scan lines does the scanner drop? Some more sophisticated algorythm for processing is required to resolve that little problem, which, I suggest will result in a less than optimum result. i.e. if the scan res is set to anything other than a number which when divided into the hardware resolution, results in a whole number, then scan quality will likely be reduced. This is theoretical on my part but logically if this isn't the case then image processing must take place which will degrade image quality compared to the method I (and others) have suggested.

The real world resolution of this scanner is ~2200 so if you scan at that or near that (2400) you will capture all the available info (depending on your settings) and will keep your file size much more reasonable.

what do you think is physically happening when you scan at 2400ppi on a scanner which has 4800 sensors per inch of scanbed width such as in a 4990? Actually there are 3x more than that because of 3 colours.

The point here is that the hardware's real resolution, when we are talking about consumer flatbed scanners, and the manufacturer's claimed resolution differ by a factor of 2-4 depending on the scanner in question. Yes, scan at the hardware's optical resolution; it's real resolution not its theoretical resolution as claimed in the specs.

Yes the resolving power of the optical system is not upto the hardware resolution but you need to be careful about assuming that the optical resolution should dictate the software resolution setting for the reasons I have already suggested. i.e. by software limiting the resolution when you set a scan resolution to less than the hardware resolution, you are telling the software to do something about reducing the scan from the hardware resolution to whatever you set it to. That means either dropping scans lines or downsizing both of which loose information which can be useful. i.e. you rely on the software to be as good as you could do in PS which I don't think is acceptable for all images.

Again, Silverfast is pretty good at this but some images would benefit from a more tailored resize and sharpen rather than the generic hard coded settings in the software.

n.b edit is underlined.

[Ted Harris]

roc, no argument about setting the resolution at a factor easily divisable. In fact, generally, I recommend that you use the software setting that most closely equals the real optical resolution of the scanner (e.g. 1800 for real resolution of 1800, 2400 for real resooution of 2200, etc.).

By capturing blank information I do NOT mean they are dropping scan lines. I mean that the sesnors are there, they are just physically arrayed in a such a manner that they are not physically capable of doing any good beyond permitting the manufacturers to claim a hugher resolution. The chips used in these machines are so inexpensive now, such a small portion of the materials cost of the machines that it is a small cost to just add a sensor and claim higher resolution when there is no real gain.

[Sheldon N]

This thread has gotten me curious about which approach to take with scanning on my Epson 4870. I just ran a little test and thought I'd share the results with you all so that you can make the decision for yourself.

The scan is off a 6x6 negative shot with a Hasselblad 500cm and 80mm Zeiss Planar lens. It was taken on a tripod with mirror lockup and the self timer release on the shutter, shot at f/5.6. Film is Tri-X 400, exposed at 400 and developed in Xtol. I chose the Hasselblad shot since I figured it would have high enough resolution on the film to test the scanner fairly. The crop portion is from the actual plane of focus in the shot and represents about 1/10 of the width/height of the full frame. That means if you are looking at the crop at 6 inches wide on your monitor, it represents a 60 x 60 inch print - from a medium format negative.

Methodology was to set the default exposure, turn sharpening off, and scan at 16 -> 8 bit grayscale with Silverfast SE (Epson package software). I scanned the same area at 2400 dpi, 4800 dpi, and 9600 dpi. I then took the 2400 dpi crop into Photoshop and applied sharpening to my tastes (150, 1.5, 1). The 4800 and 9600 crops were also taken into Photoshop and downsampled to 2400 dpi, then sharpened the same amount (150, 1.5, 1) after the downsampling. All were saved as level 10 jpg's for the web. I've attached the three crops in order, 2400, 4800, 9600.

When viewing the originals in Photoshop, there is relatively little difference to my eye. I would have to give a slight edge to the 4800 dpi scan, but don't see any benefit to scanning at 9600 dpi. For most purposes, a 2400 dpi scan would be just fine.

I'm not sure the differences will even show up in these web jpg's, but you can be the judge.

[squiress]

Each 4-pixel cluster is averaged down to one pixel, so the effects of random noise are theoretically reduced by 75%. I don't see that kind of reduction in real life, but it does seem to help a bit.

Noise reduction is as the inverse of the square root of N. For four samples going to one you get a 50% reduction of noise which may be closer to what you actually see in real life ;-)

Stew

[Leonard Evens]

I think there may be some possible confusion in the terminology. 'Optical resolution' as specified by the manufacturer refers to the number of samples per unit length actually collected by the scanning hardware. This may be different in the vertical and horizontal directions, with the latter typically being half the former. Usually we take the 'optical resolution' to be the smaller of the two. It might be better to call this the sampling resolution and describe it in samples per inch, or spi.

Then there is the ability of the scanner to resolve fine detail as measured in line pairs per unit length, usually mm. There is no standard term to describe this. Perhaps we should call it the effective resolution. lp/mm and spi are related by digital sampling theory as follows. The maximal lp per inch obtainable is half the spi. (Roughly, you need at least two samples per line pair.) But no scanner will achieve this theoretical maximum. Higher quality scanners will do better. Epson flat bed scanners often yield one quarter of the spi or less. To convert from lp per inch to lp/mm, just divide by 25.4. You can also double the effective lp per inch to get a number which would tell you the optical sampling frequency of an equivalent ideal scanner which shows no losses. Thus, my Epson 3200, which samples at 3200 samples per inch yields the same effective resolution as a theoretically perfect scanner sampling at less than 1600 spi which would give 800 lp per inch or about 31 lp/mm.

As best I can tell, all scanners sample at the optical sampling resolution frequency at the hardware level. This is then downsampled in the firmware or scanning software to prouduce a scan at a smaller pixels per inch count. Here we are no longer talking about samples per inch, so this is best measured as pixels per inch or ppi. Or you can downsample instead in a photoeditor such as Photoshop. Which works out best would depend, I think, on how each step in the process does the downsampling. My feeling is that you get a better result downsampling in your photoeditor than by allowing your scanning software to do it, but it probably doesn't make a big difference either way.

There are also other factors. You can sharpen at any stage of the process. Also, there is the issue of grain aliasing. Aliasing takes place when there are spatial frequencies in the source which lie above the sample frequency. These can't be resolved by the scanning process, so they are reflected down in a distorted form at lower frequencies. Grain usually involves variations too small to be detected and would contribute such aliasing. Hence we may use the term 'grain aliasing' to refer to the general process. Ideally, the scanner should build in a sharp filter to cut off all spaital frequencies above the sampling frequency, and that would eliminate grain aliasing a a complication, but most scanners don't do that. Another way to deal with it is to sample at a frequency much higher than needed and then to downsample. That will act as a filter to eliminate spatial frequencies higher than the desired upper limit and thereby minimize grain aliasing'.