Aliasing and scanning resolutions

[Ed Richards]

I wonder if some of the conflicts about scanning resolutions and print quality arise from confusion about alaising and scanning. From what I remember about sampling theory, you have to sample at twice the frequency of the data you want to capture. If I want a 16 x 20 printed at 240 DPI, then I need a data file that is 3840 x 4800. This would be scanning a 4x5 at 768 DPI. Doubling this for alaising, I get 1536 DPI. If I want to put 300 DPI data down on the print, this goes to 1920 DPI. Thus in our thread about comparing 4x5 and 8x10, the 100 meg scan of the 4x5 was at too low a DPI to collect enough data to even get 240 DPI down on the print.

This would be consistent with the observation that good consumer scanners can produce 16x20 prints from negatives that are indistinguishable from drum scans, because a good consumer scanner can probably produce a good 2000/2200 DPI scan from a negative by scanning at 4800 DPI and downsampling. (DMAX and noise become a big problem for chromes.)

Above 16x20, the advantage of drum scans would become more obvious as the print increased in size. Having an optimum negative might push the best printable size with the consumer scanner up to 20 x 24. Of course, this assumes fine detail in the picture. Clouds over a pond might look the same in any size print. I think that is also why we see people saying that their 8 MP DSLR makes great 16x20 prints - it will if you have no fine detail and properly process the data to smooth out the pixalation.

Comments from some of our math oriented scanner mavens?

[Guy Tal]

Ed,

I believe that aliasing is handled internally by the scanner so a "dot" in the output file will already be the result of whatever interpolation the scanner is doing (depending on the scanner technology) from multiple samples.
You should also distinguish between resolution and dynamic range. You can't compare scanner performance based on resolution alone as the sibjective "quality" of the image you get is a combination of detail and tone, not just one or the other.

Guy
Scenic Wild Photography (http://www.scenicwild.com)

[paulr]

some general ideas ...

How many samples you need to accurately reproduce a line pair worth of detail is the subject of some discussion. The points of view I've read suggest that two samples is optimistic, and that four is overly conservative. A lot of people consider 3 to be a good estimate. So to record 10 lp/mm of resolution, you would want at least 30 pixels /mm sampling rate.

Avoiding aliasing is a more complex topic. One of the reasons desktop scanners tend to have double the sampling frequency of their actual optical resolution is to provide a simple answer to aliasing problems. Drum scanners provide more sophisticated controls, but making them work right takes a lot more skill.

[David Luttmann]

Nyquist theory is such that at LEAST twice the sampling frequency is needed to resolve a given resolution in line pairs. Convention places this figure at approx 2.3

Thus, for a 1200 dpi scanning rez, you should scan at least at 2700 dpi or above on a cheap flatbed. Higher quality scanners reduce this figure as it is already built into the scanning system software. On a flatbed like the Epson 4990, scan at least 2X your desired rez.

[Ben Diss]

... and then what? Use resize to get back down to your working resolution? Sorry if this is obvious to others, but the concept of scanning at a higher resolution is new to me.

-Ben

[Bruce Watson]

Rather than comment on the mathematics of sampling theory, I'd rather point out an often discounted advantage of drum scanning - that of wet mounting. Not just wet mounting to a flat surface, but wet mounting to the precision curved surface of a drum.

Mounting on a curved surface holds the film rigidly in correct focus position. Then the scanner fluid fills in the surface imperfections of the film such as small micro scratches, and coats and floats at least some of the dust. The resulting scans are visibly improved, even at small enlargement factors.

Even if you make just small enlargements like your example 16x20 print from 4x5 film (only a 4x enlargement), you'll be able to tell the drum scanner print from a flat bed print. The effect isn't negligible in my book, nor will the two scans be indistinguishable.

Is the improvement worth it on such a small enlargement? Only you can say.

[David Luttmann]

Ben,

you would then downsample to your desired resolution in Photoshop. Some people prefer to downrez using "bicubic" sampling, others with "bicubic sharper." It depends on the amount of pre-reduction USM you've applied to your raw scan. Film tends to need higher USM amounts to achieve the same accutance as a digital source. You can find this through experimation. As a starting point, prior to downsampling, apply a USM of about 300, .9, 2 for a 4x5 scan. See how this looks and work from there.

If you need anything further, email me offline & I can run through a routine with you.

Best of luck.

[Ron Marshall]

Dave, Just to make sure I understand what you are saying.

For example, if I scan a 4x5 negative in order to make a 4x5, 300 dpi inkjet print, then I should scan at at least 600 dpi, then downsample in Photoshop to 300 dpi before printing?

[Struan Gray]

The real problem is that nobody ever defines what they mean by "resolution".

[David Luttmann]

That's it Ron. Best to scan a little higher than 600, but that is the idea. Now that said, the higher quality scanner, the less important the 2X+ threshold becomes, but considering memory and HD space is cheap, it's good to scan at a higher rez. I always scan my MF work at 3200 DPI on my Imacon and store that as the master scan. I scan 4x5 on a collegues scanner (Imacon as well) at 3200DPI. I can always downrez, but it's best to avoid upsampling if at all possible. ( a small amount of upresampling, +15% or so wouldn't be noticable). This allows you to keep a master, unsharpened file. That way in the future, you don't need to rescan, and you can sharpen depending on your output size.

[Ed Richards]

A bit more detail for a 300 DPI output file,

You scan at 600 x size of the long dimension of the print in inches / long dimension of the negative. For a 4x5 negative and an 8x10 prints, this would be:

600 x 10 /5 = 1200 DPI

[John_4185]

This would be consistent with the observation that good consumer scanners can produce 16x20 prints from negatives that are indistinguishable from drum scans, because a good consumer scanner can probably produce a good 2000/2200 DPI scan from a negative by scanning at 4800 DPI and downsampling. (DMAX and noise become a big problem for chromes.)

What good 'consumer' scanner does 4800 "DPI"? Or do you mean SPI or PPI? In any event, please share the brand name and model so I can try the thing.

When did downsampling by 50% become nondestructive?

[Brian Ellis]

"Some people prefer to downrez using "bicubic" sampling, others with "bicubic sharper."

There was a lengthy discussion about this in one of the digital groups recently. The conclusion there seemed to be that "bicubic smoother" was the best way to go.

[Ed Richards]

> What good 'consumer' scanner does 4800 "DPI"?

Canon 9950, Epson 4990. If you mean, do they really do 4800 - no. But averaged out and processed, they probably do a good 2200.

> When did downsampling by 50% become nondestructive?

When you have oversampled so all you are losing is the noise.

> The conclusion there seemed to be that "bicubic smoother" was the best way to go.

Bet that was with digital or 35mm, but I could be wrong. At these scan levels grain is not an issue in large format.

[David Luttmann]

Brian,

While generally accepted that bicubic smoother is better for uprezzing, some people prefer it to the somewhat edge enhancing effects of the bicubic sharper routine. I'd say that as long as you're downsampling something to no less than 1/2 its original size, you may find the sharper routine to work better. I find the undue softness introduced by the smoother routine to be less pleasing to the eye.

Once again, this is really up to personal choice rather than a right or wrong. Whatever gives you the print you want is what I would suggest.

What I find works well in my workflow is to do a USM on the RAW file, and then downsample using the sharper routine.

[sanking]

OK, I have been following this thread closely and have a question. I would like to archive all of my 12X20 negatives in digital file. I have a Microtek 9800XL, which will scan at a maximum optical resolution of 1600dpi. Based on what has been said so far I have come to the following conclusion. To optiize the quality of my scans I should scan the negatives at 1600 dpi and hen down sample to 800 dpi using bicubic sharper. Does this sound reasonable?

[David Luttmann]

Sandy,

As it is so dependant on the quality of the original scan (low quality scanners benefit from downsampling more than higher quality scanners), you really need to see what provides a better output at your desired print size by testing various routines. If the idea is to just print at 12x20 or 24x40, than there may not be much of a benefit to downsampling.....the aliasing effects would be buried in the output and probably not create any undesirable effects. 12x20 though is outside my scanning realm. I've only had 4x5 & 8x10 chromes scanned....but mainly 4x5.

The main idea to downsampling is that if the scanner really doesn't have a true 1600 dpi ability, all you'd be throwing away is aliasing garbage and noise.

Also, you may try downsampling less. Make your 1600 dpi scan a 1000 or 1200 dpi scan & see how that works. You may need to do no downsampling at all. Sometimes it makes no difference at all whether left raw or downsampled.

As usual, a lot of experimentation is needed with any scanner as they're all different.

[Ed Richards]

> The main idea to downsampling is that if the scanner really doesn't have a true 1600 dpi ability, all you'd be throwing away is aliasing garbage and noise.

Absolutely. Consumer scanners do not get anywhere near 4800 DPI, so the overscanning and downsampling is all about getting rid of noise and artifacts. Are these color or B&W? If they are color, you probably will want to downsample a bit just so they are small enough to edit in Photoshop. Since color has lower resolution than B&W it will not matter, and since your negatives are HUGE, none of this will matter until you get to REALLY big prints.

[chris jordan]

Hi guys, I think you might be looking at the problem from the wrong end, so here's my brain dump about scanning.

If you make your scans based on the print size you want, you take the risk that your desired print size and resolution might increase sometime later, and then you will have to rescan and re-image the photograph. Instead, start off by making one "legacy" scan that captures all of the image detail that is on the film. Do your Photoshop imaging work on that scan, and save that as your "master" file. Then resize it for different sized prints. If you eventually want to make a big print, or a new print process is developed that prints at 2000 dpi or whatever, you can still use that scan. Scanning quality is high enough right now that a good scan will capture everything on the film, so unless your master file gets corrupted and the scan is lost, you will never need to rescan even if someone makes a "better" scanner in the future.

So then the question is, what dpi do I need to scan to capture everything on the film? That has been researched extensively by lots of very smart people and here is the result, assuming you are using the finest films available: For 35mm scan at 4000 dpi; for medium format scan at 3000 dpi; for 4x5 and 8x10 scan at 2100 dpi. That scanning resolution will capture all of the image detail that your lens could focus on the film. That resolution won't capture all of the film grain, which is superfluous information that only complicates the imaging and printing process so you don't want it in your scan anyway.

This means that the scanners that claim to scan at 4800 dpi are doing something that you will never need. Some drum scanners can scan at 10,000 dpi, which is also total overkill for any kind of film in the world. The thing that matters is each scanner's ability to capture sharp image details at the resolutions listed above. In other words, a 2000 dpi drum scan will be far sharper and better than a 2000 dpi scan using a cheap flatbed scanner, despite the fact that the flatbed scanner can supposedly scan at 4800 dpi. Wet-mounting also makes a difference in the quality of the scan and in the quality of the final print, so a wet-mounted drum scan at 2000 dpi will be sharper, smoother and cleaner than a non-wet-mounted scan made by the best Epson flatbed at 2000 dpi.

And by the way, when ressing up or down, DON'T use bicubic "sharper"; it is highly destructive to your scan, and doesn't accomplish anything that you couldn't do with USM.

[John_4185]

If you make your scans based on the print size you want, you take the risk that your desired print size and resolution might increase sometime later, and then you will have to rescan and re-image the photograph. Instead, start off by making one "legacy" scan that captures all of the image detail that is on the film.

That had better be drum-scan because flatbeds will probably get better.

The rule-of-thumb in computing is that the rule will change.

[David Luttmann]

Chris,

I think we're saying somewhat the same thing. As to the bicubic sharper, that one is hotly debated. I personally use regular bicubic for most downsampling. Then again, as most of my workflow is digitally based, I rarely downsample....I just print at the native resolution.

I agree, most of the problems associated with scanning could be reduced or solved by using a higher quality scanner. But as most are using the Epsons or Microteks, downsampling does provide a valid reduction in aliasing effects. I notice this far less though on my Imacon than friends with Epsons! I rarely downsample at all....even with film scans.

By the way Chris, I won that auction for your book that was mentioned in these threads on Ebay. I should be getting it any day now.

All the best.

[Leonard Evens]

Various interesting points have been raised in this discussion. Let me ask one other related question.

Think of each step in the process as a filter which affects the spatial frequency response of the image. That frequency response is described by an MTF curve giving percentage reduction in response as a function of frequency in line pairs per mm. Often that is summarized by a single number which presumably gives the highest usable frequency, but that is an over simplification. Sampling theory says that you can't resolve a frequency higher than twice the sampling frequency, for digital sampling. Aliasing in a related but different issue. It concerns what happens to higher frequencis, if they are not filtered out. They are reflected or aliased down as lower frequency artifacts.

The last step in this process is printing. If you print at 240 ppi, i.e., 9.45 pixels per mm (ppm), your maximum detected frequency would be half of that or 4.72 lp/mm. If you print at 360 ppi, that maximum increases to about 7 lp/mm.

Similar reasoning applies to scanning (and indeed all steps in the photographic process). The Epson 4990 can scan at 4800 ppi, or about 189 ppm, and that translates to about 94 lp/mm. In fact, most reports suggest it does not do nearly as well as that, with figures like 30-35 lp/mm appearing to be more realistic. Let's take 32 lp/mm for the sake of argument. Enlarging 4 X reducing that to 8 lp/mm. (However, that figure may represent something somewhat down from the top of the MTF range, which is what the Nyquist limit presumably gives you.)

Now there is no simple way to combine successive filters. Ideally you would multiply, point by point, the MTF values, but since those curves are seldom available, other rough rules of thumb are used. The important thing to note, is that you are always going to get resolution in lp/mm lower than that which the coursest filter provides.

This reasoning would suggest that even printing at 360 lp/mm is not going to yield much above 5 lp/mm, which is the commonly used estimate of what the human eye can detect at 10-12 inches from the print. It also suggests that the print resolution is the most limiting factor in the entire process.

Any comments?

[Ken Lee]

To me one of the upshots of Leonard's analysis is that the Epson 4990, when set to its maximum, appears to deliver somewhere in the neighborhood 1/3 of its promised resolution. Other scanners do a better job, but very few consumer or prosumer scanners deliver as advertised. The files they create may contain lots of data - but much of it is...junk.

So if I were Sandy, I would presume that the 1600 ppi setting is likely to deliver much closer to 1000 ppi at best, and that an enlargement of 3x would be the practical limit, if we want to retain critical sharpness.

That being said, one of Sandy's images in a 36x60 print, critically sharp etc. would be quite something to see.

[Ed Richards]

> This reasoning would suggest that even printing at 360 lp/mm is not going to yield much above 5 lp/mm, which is the commonly used estimate of what the human eye can detect at 10-12 inches from the print. It also suggests that the print resolution is the most limiting factor in the entire process.

Even photopaper is limiting in a lot of situations. One more factor is sharpening. By increasing the edge contrast, you increase the eye's ability to resolve the fine detail and thus the perceived sharpness. That is why some folks have noticed that a scanned and processed image can be sharper than a contact print of the same same image at the same size. Sharpening changes a lot of the issues for how sharp a print looks and is a key component of scanning. It does not create new data, but scans have more data in them than appears to the eye - sharpening restores this.

> So if I were Sandy, I would presume that the 1600 ppi setting is likely to deliver much closer to 1000 ppi at best, and that an enlargement of 3x would be the practical limit, if we want to retain critical sharpness.

The better microteks get to their stated resolution - Paul B found that his 1800 really did do 1800.

[John_4185]

Okay, so what's the bottom line here? I'll continue to print at the enlarger (or will when I replace the broken lens) because it looks like regardless of inkjet or enlarger we end up with the same so-called "resolution" and I have infinite degrees of gradation within the tones of concern, rather than a troublesome maze of steps and complications leading, hopefully, to the same outcome with scanning and printing. No? My eyes tell me Yes.

[Ed Richards]

Bottom line? Digital is different from analog silver. Scanning can give you sharper prints than you can get with an enlarger, but they will not be silver prints. If you want silver, and have the resources to use it, quit reading these digital threads and stop worrying about it.:-) If time and space were not an issue, I would print with silver myself. Since they are, I waste time trying to get the best results from what I use, just as I did when I used a darkroom.

[John_4185]

Scanning can give you sharper prints than you can get with an enlarger,

Not according to the posts I consider most credible. Seems the outcomes bottom out to within 8lp/mm.

... still working on my home-made pin register for unsharp masks.

[sanking]

"The better microteks get to their stated resolution - Paul B found that his 1800 really did do 1800."

What standard targets and/or procedures are people using to test this?

[Ed Richards]

Air Force standard on glass:

http://www.butzi.net/articles/scannersoft.htm

These sell for about $130. I wish the guys doing the scanner tests had one. They cut through the speculation pretty quickly.:-)

[John_4185]

www.butzi.net/articles/scannersoft.htm (http://www.butzi.net/articles/scannersoft.htm)

Impressive. I actually have silverfast software and never used it. (Thanks for that, page Paul!)

[sanking]

"Air Force standard on glass:
http://www.butzi.net/articles/scannersoft.htm

These sell for about $130. I wish the guys doing the scanner tests had one. They cut through the speculation pretty quickly.:-)"

What I am trying to figure out is if there are any standard procedures for using the standard Air Force test pattern, or other patterns?

I looked at Paul Butzi'a article on scanning in RGB and extracting just one of the colors in comparison to scanning in Grayscale. I a not surprised that he found more resolution in the blue channel image than with the other two color. However, I wonder if this is not due to the fact that in camera optical systems the diffraction limiting resolution of blue light is quite a bit higher than that of green and red light, rather than to differnces in specific scanners and/or scanner software?

Unfortunatley in my application, i.e. scanning 12X20" negatives at 1600 dpi, scanning in RGB is simply not an option as the resulting file is simply much too large to work with. Even in Grayscale a 16 bit file of a 12X20" negative is close to 700 mb.

[sanking]

"Unfortunatley in my application, i.e. scanning 12X20" negatives at 1600 dpi, scanning in RGB is simply not an option as the resulting file is simply much too large to work with. Even in Grayscale a 16 bit file of a 12X20" negative is close to 700 mb.

Actually a 16 bit Grayscale file of a 12X20" negative scanned at 1600 dpi is about 1.2 gig in size. The 700 mb size I mentioned is after downsizing from 1600 dpi to 1200 dpi.