Epson and Other Scanner Matters
My first contact with flatbed scanners was in 1996, when the computer project group at school got a shiny new HP Scanjet 4c, with a screamin' Pentium 133 with 32 megs of RAM to go along with it. This HP could already push a megabyte per second over the SCSI bus. Now the first two scanners in our family weren't quite as high-tech – first a Mustek ScanExpress 12000P, later a Plustek OpticPro UT24. Cheap, slow, plasticky things (the UT24 isn't remotely dustproof, real awfully cheap build). This only changed many years later when I was looking for an A3 sized scanner (much easier handling of schematics) and ended up with a used Epson Expression 1640XL – a huge, heavy, nearly bulletproof pro-grade monster. And so the story begins…
Note that this page is still in beta and may be a little rough around the edges.
The turn of the century was a pretty wild time as far as flatbed scanners were concerned. They really broke into the consumer mainstream around this time, with people wanting to scan both documents and family photos (often negatives as well). CCD sensor resolution downright exploded, with semi-pro models going from 1200 dpi in 2000 to 4800 dpi in 2003, ultimately topping out at 6400 dpi in the mid-2000s when people realized that optics had become the limiting factor and weren't likely to improve much further. Dynamic range handling and ADC resolution also improved steadily. Consumer models tended to use the parallel port, later USB, and often were quite "dumb" and slow (far from the limits even of an ECP/EPP port or USB 1.1, which can transfer about 1 MB/s each). More upscale ones would use a SCSI interface and had more internal processing power and speed to offer; in the early 2000s, IEEE 1394 (Firewire) and USB 2.0 found use instead.
Epson (Seiko Epson Corp.) had been making scanners for quite some time at this point. From this page it would appear that their oldest models went EOL/EOS (?) in 1992, so those must have been introduced in the mid-late '80s. (The oldest model I could find at the well-curated Epson America support website was the ES-300C a.k.a. GT-6000, a 300 dpi color model introduced in mid-1990.) Basically they've been around for just about as long as HP with their Scanjets, with similar roots in the business sector (mere mortals would have been unable to afford anything but a little hand scanner at the time, a woefully flawed concept btw). You can find a lot of custom ASICs and gate arrays in their scanners, pointing to high levels of vertical integration. (That does not appear to apply to CCD sensors and A/D components though, which in the service docs I've seen were sourced from Sony and either Burr-Brown or AD, respectively.)
Let's start in 1999. This is the year that saw USB peripherals take off. By that time, Epson had some really nice scanners to offer – the A3-sized, 800dpi Expression 836XL / GT-12000 behemoth (SCSI + parallel, w/ focus adjust), its A4-sized Expression 800 / GT-9600 companion (SCSI only), remaining stock of the older 600 dpi Expression 636 / GT-9500 that was still being offered (SCSI + parallel), and the then-new Perfection 636 / GT-7000S (SCSI) as the most consumer-oriented option. Transparency units and partly ADFs were available for these. Below that it gets a bit sketchy but I think there was remaining stock of the older GT-5500 400 dpi model (SCSI or parallel), and the GT-2200 (SCSI or parallel). They had dabbled in the consumer market with the Perfection 600, a rebadged UMAX, in early '98, but had quickly run out of stock.
Epson Flatbed Scanner Timeline
I have taken the liberty to dig through the online resources available and figure out the product lineup from about 1998 onwards into the recent past. Older models are not likely to be too interesting since they usually won't even connect to something resembling a modern PC (who'd still want to put up with non-PnP technology like a SCSI scanner, or who still has a matching SCSI host adapter, for that matter?), and even older ones still did color separation by light source, i.e. switch lamps. Note that there may be some overlap that is hard to convey in such a table, I have usually included a note though. Availability refers to Germany.
Along with each model, some basic specs are given: Always-installed extra features like transparency unit (TPU) or automatic document feeder (ADF), connectivity, sensor resolution, bits per color channel (take x3 for full color spec), Dmax (maximum optical density), speed at 24 bit color, and available options. Note that speed is given for the sensor resolution stated unless specified otherwise – it is not a value that is directly comparable between different resolutions (the same time per line at a higher resolution means that the device is actually processing more quickly). The value is also given for "high-speed mode", which usually is not what you are going to use in practice unless running a prescan.
|Pro A4||Cons 1st||Cons 2nd||Cons 3rd|
(UMAX Astra 1200S)
P. V750 Pro
- Cons 1st, 2nd, 3rd
- Consumer class, 1st, 2nd, 3rd tier
- Perfection [model name]
- Expression [model name]
- Transparency unit. "Trk TPU" denotes one with a lamp tracking the scanner carriage, rather than just a lit film strip area.
- Automatic Document Feeder. Duplex capabiliy is noted.
- Digital ICE technology for film. Scans in infrared where film is near-transparent to identify dust, scratches and whatnot, then uses that information to clean up the scan.
- Digital ICE technology for prints. Identifies dust by illuminating the print with two different lamps – the stuff that gives different shadows must be surface dust.
- SCSI / USB
- There were two versions, one SCSI-only and one USB-only
- USB 1.1
- USB 2.0
- IEEE-1394, a.k.a. Firewire
- Parallel Port (IEEE-1284)
- (Automatic) Focus adjustment
- Dual Focus (focal point shift for transparencies by inserting a piece of glass in the optical path).
- Dual lens –device switches focal point for transparencies (~3 mm above glass vs. ~1 mm or so) by actually using two different lenses in the scanner carriage.
- ~ 2003B
- Model remained available up to (and including) the second half of 2003
- Build and performance in the consumer class tend to be best in the 1st-tier models, as you'd expect. The 2nd-tier models still tend to do a good job though, with only minor limits (more plasticky build, external supply that occasionally fails, typically less DOF, but still good IQ and same carriage movement reliability rating). Things start to get a bit rough in the 3rd tier, performance and/or build wise – but by then we're talking about the 100…150€ class. The Perfection 1660 seems to be the mandatory exception (same build as 2400). Note: The more plastic, the more likely you'll see a haze on the inside of the scanner glass.
- All of the pro and 1st-tier consumer models introduced up to 2005 use an internal power supply, as do Perfection 636/U, 610 and other oldies. Starting from 2001, the 2nd-tier and lower models came with plug pack type supplies only, extending to all A4 models by 2006.
- The Perfection 636/U and 1200 seem to share the same chassis, and the 1640SU still is very similar. While lacking a power switch and being more basic electronically, the 610 doesn't seem to be far off either, as it weighs the same. (That one has to be the best-built cheapie scanner ever.) Similar family ties exist in the 1650, 1660 and 2400, and the 1240U isn't too far-off either. The same also goes for 2450 through 4870, the 4990 then evolved from that. Likewise, the Expression 836XL and 1640XL seem to be very closely related, and the 1600 and 1680 virtually identical (as you might expect from an update after only a year).
- The Expression 10000XL is about one quarter lighter than its (admittedly hefty) predecessors and appears to be a total redesign. Sales price was a smidgen lower, too (but no more than that). The same can be observed in case of the GT-15000, which appears to share the same basic chassis much the same as the older models did. Since then, however, weights have remained about the same or even increased slightly.
- There are occasional reports of the scanner glass being rather prone to scratches, on 2nd-tier (2400) and sometimes even 1st-tier (4870) consumer models. Not sure why that is – glass is cheap and you don't need any unusual shapes, so I don't think it would make much sense to use acrylic or something.
- The 4490 has been noted to be pretty flimsy. The usual warnings not to put anything heavy on top seem particularly justified with this model. Looks like similar issues affect the 3170 as well (and thus probably the 4180, too).
- There are some obvious cases of "hand-me-down" sensors and electronics.
- Looks like the 1200 dpi sensor of the Perfection 1200 ended up in the 1240U a year later and finished its life in the lowly 1250U, which I presume is not a match for the 1200 optically (let alone speed or build wise).
- Then the sensor of the Expression 1600, no longer good enough for a pro graphics art scanner, found its way into the Perfection 1640SU.
- The Perfection 2400 seems to have inherited sensor and electronics of the short-lived 2450 (minus Firewire).
- The Perfection 1660 appears to be merely an updated 1650 using the same sensor. That year the model palette was unusually diverse (there were two lower-end models in the 1260 and 660, though the latter was rebadged), and it looks like effectively the 2nd tier was split between the 2400 and 1660. A good time for scanner buyers, as the 1660 Photo only cost about 160€ with a (small) transparency unit already included (the long-lived 2400 would reach the same regions about half a year later).
- Then it looks like the Perfection 3170 got the 3200 sensor, with less fancy electronics though.
- You better didn't buy based on resolution back in the day! If you'd read a glowing review of the old Perfection 1200 and then got yourself a lowly 1250U, chances are you would have been sorely disappointed. There are very few cases of a newer model being preferable to an older one of the same resolution – the 1660 comes to mind (again), offering USB 2.0 speeds over its predecessors (and actually beating out its 2400 dpi brother), though image noise and color accuracy would not be expected to be a match for a 1640SU or Expresson 1600, let alone a 1640XL.
- You will note that some models remained available unusually long. In some cases this would be because they could slot into the new lineup easily (like Perfection 2400 between 3170 and 1670, or 4490 below the V500 before the V300 came out), but there also were some that I believe to be related to getting some more mileage out of a case style running out – like the 1640SU whose non-Photo versions continued to be available for two more years, or the 2400, or the 4490.
- There seem to be three distinct lines as far as command language goes:
- The higher-end models use ESC/I-B8 or -B7 (-B6 in very old ones). (-Ax was found in monochrome scanners, I think.) These work with SANE in Linux out of the box (well, usually…). The Russian Epson support website has command specs for both '90s models and the Perfection 1200.
- The lower-end ones use ESC/I-Dx – ESC/I-D1 for the Perfection 610 and 640U, and ESC/I-D2 for the 1250 and all 2nd/3rd tier models from 3170 and 1670 up. Only the former ones work with SANE out of the box. Apparently 1250/1260 use a NatSemi LM9832 chipset? Seems to be less capable hardware in any case – apparently there's 512K of RAM to work with in these, while the ASIC in the 1640SU reigns over 8 megs.
- And then there were those models that are supported by SANE's
snapscanbackend – Perfection 660, 1270, 1670, 2480, 2580, 3490, 3590. Interestingly enough, the original Agfa Snapscans never exactly enjoyed a reputation of being speed demons, and the 660 apparently still wasn't, but later ones seem to have become reasonably, err, snappy (USB 2.0? Use of Epson ASICs?).
- Traditionally, the professional models would digitize the image data as close to the CCD as possible (on the carriage if need be), while consumer models would send it over the ribbon cable in analog form. Thus it is not surprising that consumer models would be a lot more prone to banding (which is picking up analog interference), a common complaint back in the days of the Perfection 1200. 1st-tier consumer models had upgraded to on-carriage ADC by the time the Perfection 3200 was out – actually it looks like the 1640SU was the first to do so.
- AFAICT, the same basic optical setup (though possibly not lens) is used on the Perfection 636 (GT-7000), 610, 1200 and 1240U. These have 4 mirrors. Afterwards, another (5 mirrors) appears to have been used in 1640SU through 3200, with a new one appearing in the 4870.
Perfection 636(U) / GT-7000(S/U)
- Enjoyed a reputation of being fast at the time, but quite definitely cannot keep up with better / newer models. An 8.5x11" page (close enough to A4) was timed to take 39 seconds to scan in color at 300 dpi at the time, regardless of whether a SCSI or USB model was used, which should tell you that throughput is well below the limits even of USB 1.1. That being said, computers were much slower at the time, and Ken Rockwell's hot-rod dual G4 Mac yielded much slower results for the '1640SU back in 2000 compared to what I got with more modern hardware.
- Measured scan times (USB on a Core 2 Duo E6550 machine, Epson Scan 3.04 in
Professional Mode, "hacked" Expression 1600 driver, unsharp masking off, from
push of "Scan" button to image appearing):
Target dpi bpp Time Prescan - - 8 s A4 200 24 17 s A4 200 8 13 s A4 200 1 7 s A4 300 24 31 s A4 300 8 16 s A4 300 1 10 s A4 400 24 58 s A4 400 8 23 s A4 400 1 12 s A4 600 24 2:06 min A4 600 8 56 s A4 600 1 21 s A4 800 24 3:30 min 210x29.5 mm
1200 24 1:00 min 210x29.5 mm
1600 24 1:39 min 5x7" (13x18)
300 24 15 s 5x7" (13x18)
300 24 16 s Scan start @ 300 24 t0 + 3.5 s Scan start @ 600 24 t0 + 5 s Scan start @ 1200 24 t0 + 7 s Scan start @ 1600 24 t0 + 7 s
- The 5x7" (13x18 cm) scan would be noticeably quicker in landscape mode if things weren't limited by the interface – carriage return is at 13 seconds. Lower scanning carriage travel (fewer lines to scan) always wins.
- Scan performance still seems to be somewhat quicker than the 1650 (as the specs would suggest), which was tested to take 16 seconds for both a 200 dpi A4 greyscale scan and a 5x7. It cannot keep up with the speedy 1660, however, and would not count as anything out of the ordinary these days. (That being said, some modern photo scanners may be surprisingly slow in such mundane tasks.)
- After obtaining a matching EU-33 transparency unit, I did some performance
testing with this as well (newfangled Sandy Bridge i3 machine, film type =
slide, 24 bits per pixel, gamma 2.2, unsharp masking on, 36x24 selection in
upper slot, with reflective scans in about the same position for comparison):
Target dpi Time Prescan TPU - 20 s 36 x 24 mm TPU 600 27 s 36 x 24 mm TPU 1600 1:04 min 36 x 24 mm refl. 600 11 s 36 x 24 mm refl. 1600 25 s Scan start @ Prescan TPU t0 + 7 s Scan start @ 600 dpi TPU t0 + 9 s Scan start @ 1600 dpi TPU t0 + 10 s Scan start @ 600 dpi refl. t0 + 6 s Scan start @ 1600 dpi refl. t0 + 6 s
- Exposure in TPU mode appears to be well-calibrated, with the lit background yielding brightness of about 250 (out of 255). I think the scanner contains some sort of exposure tables, though I wouldn't discount the possibility of automatic exposure based on white balance information.
- Dark areas show some noise with banding, worst below 400 dpi (I presume sensor exposure is reduced to speed up scanning there). This is only of concern for transparencies with high dynamic range though, as all of this is happening way below the blackest blacks of any reflective target. (According to Epson Scan, all of this stuff is going on below 19, and I cannot remember ever having to go below 30 for the black point, 45 being more likely. That being said, the 1640XL gets to around 6 when the room is not too bright. I am guessing that the 1640SU uses the less expensive AD9822 CCD frontend like the 1240U does, rather than the appreciably less noisy AD9814 in the 1640XL.) I did a scan with the lid open in order to look at this! Alternatively, you can also use a TPU while covering the image area but not the white balance slot up top.
- It looks like the glass has sagged slightly near home position (double-stick tape sag?). Will need to look at that. Focus seems to be bang on the glass surface though, and 800 dpi and below in particular turn in some beautifully sharp scans. Update: I tried again with positions further down the glass, and found focus a bit more than one 1 Eurocent coin up quite consistently (1/2/5 cent coins are 1.67 mm thick). Nothing you'd even see at 300 dpi, but at 1600 it's obvious. This is about as much as needed for the film holders coming with the TPU.
- The ADF on my Office model came lacking the paper holder and won't load properly without one (separator not too fresh any more?). I can only assume that this caused people a fair amount of grief, since yellowing indicates that it was never even installed. The adapter piece and holder will cost more than the entire scanner, but hey. Meanwhile the ADF works beautifully with a makeshift paper holder.
- Hint: If you ship one of these with an ADF installed, always unplug that. The plug on mine acquired an interesting oval shape, but thankfully still gives good enough contact. My unit likes to emit a high-pitched squeal (inverter / coil noise) after running for a while though, which does not appear with the ADF unplugged (it apparently has to do with whether you've run a prescan through the ADF, and tends to disappear after a scanner reset), so maybe contact isn't that great after all. Or maybe it's just old age. I found some slightly colored lines appearing as well, sometimes right in the middle of the scan, which I suspect may be interference making its way into the image sensor via not-entirely-clean supply voltages, so possibly some electrolytic capacitors in there have seen better days.
- Measured power consumption: 13-15 W idle (lamp on), 21 W max in flatbed operation, 26 W max with ADF. Power factor 0.61-0.64. Off consumption below limits of power meter (<1 W) but known to be non-zero (about 0.4 W has been stated – there is little more than the rectifier and filter capacitor left when the unit is turned off).
- Could scan about twice as fast as its USB interface will allow. Seems to have pretty generous buffer memory.
- Scan times over USB: A4 (208.7 x 293.7 mm) at 300 dpi in color, 32 seconds (scan starts at 5 seconds, carriage return at 21 seconds). Greyscale 16 seconds, halftone 13 seconds. 600 dpi 1:52 minutes (net scan time minus pauses 1:09 minutes), greyscale 1:09 minutes, halftone 22 seconds.
- Network scanning slightly improves throughput. The 300 dpi A4 scan now clocks in at 29 seconds, so it seems I'm getting about 1.07 MB/s instead of 0.95 MB/s over USB. It's not quite the speed boost I'd hoped for (the displayed network throughput of about 1.2 MB/s is so close to 10 MBit/s that I double checked whether the interface connects at 100 MBit, and it does), but it's a bit at least, and network scanning ability is quite handy. The Firewire card is extremely elusive, and USB2 to SCSI adapters seem to have gone the way of the dodo. My network card came with a 2008 date stamp, which I assume means that it originally graced something like a GT-15000 or GT-20000.
- "Black with the lid open" comes in at about 5-6/255, and that still is limited by ceiling reflections and some streaks on the document glass. That's what I call pro-level dynamic range. Obviously things do get kinda noisy at this point. Still, it's miles ahead of its consumer-level colleague at the time, the Perfection 1640SU. AD's then top of the line CCD front-end / ADC finds use here, the AD9814 (whose analog noise level actually still is a tiny bit lower than in its 16-bit successor, AD9826). I presume the 1640SU used the budget alternative also found in the 1240U, the AD9822, which is almost 3 times as noisy – with pricing being just about inversely proportional (both of these parts are still listed as being active, at $3.42 vs. $9.69 a pop if you buy 1000, though with no RoHS-compliant version of the AD9814 being available I suspect they're merely sitting on a pile of old stock of that one).
- Mine seems to have slight skew, will have to find out where that comes from. Unique mechanics don't make troubleshooting any easier, and service instructions are quite intimidating.
- Its USB 2.0 interface seems to be a bit of a diva, often giving 1.1 only. Maybe that has something to do with the supplied cable, which was noted to not be USB 2.0 ready at the time. Stick with Firewire if you can, though the expected speed benefit isn't that big.
- Focus on this scanner inevitably is a compromise between reflective and TPU operation, never mind individual calibration issues.
- Was much faster than its predecessor at high resolutions – but also produced a good deal more image noise in slides. This generation seems to have been focused on speed.
- Effective resolution limited by optics and not any better than on 2450.
- This one was found to be quite speedy. 12 s per A4 page in 48 bit color is nothing to sneeze at.
- Featured all-new optics with ICE for both film and prints and was noted to be sharper than its predecessors. Seems to have focus adjustment for infrared (film ICE) only, so focus for transparencies vs. reflective documents would still seem to be a compromise.
- Not as blazingly fast as the 3170, but still fairly speedy at 20 seconds per 300 dpi A4 color page. I guess they backed off a bit on speed in order to get noise back down.
- This modern photo scanner is surprisingly slow in mundane tasks. 35 seconds for an A4 color page in 300 dpi? That's barely faster than ye olde Perfection 636 / GT-7000. And my 1640SU can complete a 600 dpi scan of the same in significantly less than 2:26 min, while being bottlenecked by USB 1.1. (For comparison, the V700 needs 45 seconds, and a 3170 could do it in 40.) I can only guess that the 6400 dpi sensor must need a lot of light. This model seems to be strictly targeted towards the film scanning crowd.
The most common issue by far appears to be haze on the inside of the scanner glass – usually after several years, but sometimes right from the factory (oops). A scanner inevitably has to be just about airtight in order to keep dust out (unless it's a Plustek OpticPro UT24, that is…), and outgassing from lubricants and plastics eventually produces a film on the glass thick enough to become disturbing. Sure enough, this seems to be more common on "plastic fantastic" models, and virtually standard in e.g. the 2400, 1660 or 1240U. At least the outgassing rate would be expected to be decaying exponentially over time, so with a bit of luck the cleaning would not have to be repeated any time soon.
Lighter vertical lines are another reasonably common problem, it seems. This happens when dust gets onto the white strip in the calibration area that is hidden from view. Again, cleaning fixes this.
I have been able to find service manuals for the following models on the interwebs: Perfection 4870, 3200, 1640SU, 1240U, 1200, 610, 636 (or GT-7000, rather), plus Expression 1640XL, GT-12000 (Ex. 836XL), GT-10000, GT-9600 (Ex. 800), GT-9500 (Ex. 636), GT-5500, and some oldies like GT-8500, GT-5000 (Action Scanner II) and GT-300.
The following opening procedure seems to work for many 1st/2nd-tier consumer models:
- Work in a clean environment with no air movement. Wait for dust to settle. Make sure you have ample space around the scanner, as you may need to put the cover on the side. Keep all the required cleaning utensils handy. Get someone to help if feasible.
- Remove document cover.
- This should expose two screws at the hinges. Some newer models like the V500 have two more lurking on the underside of the unit near the front. (1640SU: 3 at the back, plus two under the front for the front cover that needs removing first.)
- Lift the back of the top case and push it towards the front to release some hinges in the front. (Newer models may have a bunch of nasty clips that are all kinds of fun upon reassembly.)
- Now you should be able to lift off the top case. There'll be a cable of very finite length going to the buttons, so carefully examine which side the top case could be flipped over to without damage. You may have or want to unplug said cable.
- Now you ought to be able to clean the inside of the scanner glass. Do so thoroughly but quickly in order to avoid getting too much dust in. You can try working with dishwashing detergent first and window cleaner (or alcohol + water) second if streaks are stubborn.
- Reassemble in reverse.
Driver Fun with Old Scanners
Epson's support policy is pretty standard for big companies like that – driver support runs for a given time, typically 5 years after discontinuation in this case, and after that they won't move a finger no matter what (though they'll happily sell you a new device, of course). Technical feasibility does not figure in at all, so it's pure bureaucratic randomness. (You could have had even more fun by buying an HP though. They thoroughly wrecked their reputation by shipping lousy scanner software in the mid-2000s, and then there were cases of the warranty having run out before the scanner was even bought. All of which was a bit of a pity since the hardware actually was quite good and really fast in the higher-end models, even if the duplexer in their ADFs had a tendency to jam and make the whole scanner hiccup. Another "fun" company would have to be Microtek and their non-support – I thought they went bust a few years ago, but they still seem to be around.)
So what do you do with an older, well-built, perfectly functional device that just doesn't happen to have any drivers for a more modern operating system that you are using?
- You could buy VueScan. It supports just about a gazillion scanners, runs on Windows, OS X and Linux, has pretty fair licensing terms, and even the Pro version is quite moderately priced for what it does (and certainly cheaper than a good new scanner, especially when we're talking higher-end or A3 size). Unfortunately they take payments by credit card only in order to minimize overhead (it's only a 2-man operation after all), which I bet has cost 'em some sales in credit-card-averse areas like my native country.
- You could buy Silverfast. That's more geared towards graphics types in terms of interface, tends to be more expensive, and you have to buy a license per scanner model.
- You could set up a virtual machine with an OS that does support your scanner. Maybe Windows XP, or some Linux distro.
- Or, of course, you might exercise some creativity or at least some Google skills…
Let's focus on the last item. Like I said, support policy is all bureaucracy with no technical background. Now thankfully Epson's command language (at least for the higher-end models) was pretty well-established by about 2000. If you have the common problem of no Windows 7 drivers, it is often possible to coax the driver for a related newer model Into accepting yours by modifying the USB device ID in the respective INF file(s). (The SANE epson2 list of supported devices tends to be helpful here.) For example, people have used Expression 1600 drivers for the Perfection 1640SU, CX3200 for ye olde Perfection 610, Perfection 4990 drivers for the 2450 and Perfection 2400 drivers for 1200U and 1640SU (though the Expression 1600 tends to be a better match for the latter, save for color profile maybe – the lamps are different; if in doubt you could always buy an IT8.7 test target and get LProf and roll your own ICM). Actually 2400 or 3200 would seem closest to 2450 technically, but to each their own.
I have found that when using an Expression 1600 driver for the Perfection 1640SU, it is best to run setup and install the regular Expression 1600 driver first before pointing Device Manager to the modified INF file. Otherwise you'll get the English-language interface only (instead of whatever other language you'd prefer).
Dmax, Scanning Speed and Pink Sky
You may be wondering what that Dmax spec is good for. Maybe you've heard that a higher value gives better results on slides. We'll try to, err, illuminate this subject a bit.
A scanner like the ones here uses a CCD sensor. Like any optical sensor, this has some limits:
- It only accepts a maximum number of photons before saturating. Each time a photon gets through, an electron is being collected in an internal "bucket". Once that is full, output voltage cannot increase any further.
- If light intensity gets very low, each cell (subpixel) may only be collecting a very limited number of photons. As you may know from statistics, the exact number of photons becomes very random. So if it's an average of 2.5 photons per cell, one cell might collect 5, another 2, another 1, another 4, and another none at all. You can try that by throwing pebbles into a bunch of buckets. In an image, such randomness is nothing but noise. In addition, the electronics used to read out the CCD and perform analog to digital conversion also contribute noise, and back in the day it would not have been unusual to see it dominate the overall noise level entirely.
Now the scanner manufacturer has to couple this given dynamic range to the amount of light that the sensor receives in a sensible way. Now that depends on:
- The light source. Note that a scanner has two different light sources if it is equipped with a transparency unit.
- How much light the scanned object either reflects (when using the lamp in
the carriage) or transmits (when using the transparency unit). As in many other
fields, a logarithmic measure has proven practical – here it's called the
optical density D.
D = -log10 (T) with T being the linear transmissivity (so we're looking at the transparency unit case).
D = 1means that only 1/10 of the light gets through, with
D = 2it's only 1/100, and so on. A typical reflective target might show a minimum D of 0.3 and a maximum D of 2, while a slide might cover almost 0 to 3.0 or even 4.0.
Converting D to photographics stops (or effective number of ADC bits) is easy, by the way:
#stops = 10/3 * D
D = 3.6amounts to 12 stops or bits, which means that a real-life 12-bit converter is likely to achieve rather less. (If you look at scanner models with 12 bits per channel, you'll notice that none of them are specified with a Dmax over 3.3 = 11 stops.)
- Exposure time. A scanner has no variable aperture like a camera does, but exposure time can be controlled electronically.
You'll probably notice that there is some degree of freedom here. You
obviously cannot do much about the kind of things to be scanned, but you do
have some choice when it comes to what lamp to install (and what sort of power
to run it with), and you can vary exposure time widely, though you probably
don't want to go over a few milliseconds, or else it becomes the limiting
factor in scanning speed. You're going to want to coordinate these two factors
in such a way that the minimum D ever to be encountered just about aligns with
the sensor's saturation limit. Otherwise you're either wasting dynamic range
and getting more noise in the shadows than necessary, or clipping the
highlights, which is even worse. I assume that Epson's Dmax specification
assumes this condition, in which case dynamic range = Dmax. Otherwise the
dynamic range is given by the difference of Dmax and Dmin.
0 can, unfortunately, not be taken for granted, so "cheating" is easily
Have you ever wondered why some scanners produce pink sky on slides? What
happens is that the scanner manufacturer apparently messed up sensor dynamic
range alignment in transparency mode (or optimized for much less-transparent
negatives, or felt like playing the Dmax numbers game), yielding badly
overexposed highlights with red, geen and blue channels all being clipped and
maxed out. And then automatic white balance comes along and pulls the green
channel down (a bit of a green cast is not unusual in flourescent lamps). The
result: Pink sky. Yuck.
You can try rectifying this issue by stacking an ND 0.3 (ND2) neutral density filter on top of the slide though; the loss of a polarizer may also do the job. (Noise in the shadows will obviously increase, but unlike blown-out sky it may be inconspicuous or could be addressed with noise reduction software.)
Now how do we get a massive D range equalling 12 bits or more to display on conventional graphics cards and monitors with 8 bits per channel? Simple, nonlinear quantization. IOW, the possible 8-bit values are not equidistant. Ever heard of a "gamma curve"? That defines the nonlinear input/output relation. This makes sense because our eyes do not register light in a linear fashion either. Straight linear 8 bit would give more resolution than needed in bright areas and not enough in dark ones. The nonlinearity permits getting a lot more mileage out of these bits.
Note that it is quite possible to apply gamma within the scanner and output 8 bits per channel only, giving a speed advantage if interface throughput is the bottleneck. (This, however, means that for best results you must set approximate black and white levels and gamma in the scanning software, and cannot rely on doing it all afterwards.) Obviously for this to work properly, the scanner's (linear) ADC must provide a fair bit more resolution, typically at least 12 to 14 bits depending on CCD.
We were talking about speed earlier. What are the limiting factors here?
- Exposure time. High-resolution sensors at full resolution are usually light-starved. Can you guess why the Expression models, Perfection 4990 and V7x0 use a tracking light source in the transparency unit now? An A3-sized scanner would hold an advantage here, by the way, since it can scan an A4 document sideways (reducing the number of lines to scan). Reducing exposure time at lower resolutions seems to be a common tradeoff – the line of thought being that you'd only be using these for reflective documents anyway, so some more (but still irrelevant) noise wouldn't matter, while speed would.
- Image processing speed. ADC sample rate and processor power tend to be rather finite.
- External interface throughput. My Expression 1640XL, for example, could churn out data almost twice as fast as its measly USB 1.1 interface will allow when using, say, 300 dpi. (The Perfection 1640SU still is about 25% faster than USB 1.1.) And if you do the math on the monstrous GT-30000's real-life 8.57 seconds for a 600 dpi A3 scan, you'll know why that one came with an UW SCSI interface… (Hint: That's 199 MiB per scan.)
- Maximum stepper motor speed. That's usually the limit for very low resolutions, like in prescan.
Focusing and Sharpness
Most scanners use fixed-focus optics, with a focal point just slightly above the scanner glass (maybe 1 mm). That's perfectly fine at normal resolutions like 300 or 600 dpi, but once you are getting into 1600 dpi and higher territory and you need that kind of resolution because you want to scan negatives or slides, the range of optimum sharpness shrinks and shrinks. Now the negative holders you got with your scanner may or may not position the material at an optimum distance, as calibration in consumer scanners does not tend to be that precise. Some experimentation is definitely worth a shot. (I like the coin method. Australian 5 cent coins appear to be 1.3 mm thick, so this 4870's focus in transparency mode must be about 5 mm above the glass, rather higher than it should be.) Flatness of the material also becomes increasingly critical, of course.
Epson's "dual focus" models have two selectable focal points, one at about 1 mm above the scanner glass for reflective targets and the other at about 3 mm for transparencies. This works by inserting a piece of glass in the optical path (and thereby lengthening the latter) for one setting. Having continually adjustable focus (like the Expression models and other high-res pro scanners do) is even better, of course, sparing you from the potentially tedious process of tweaking height.
What manufacturers don't tell you is that basic depth of field (DoF) not only is much greater in CCD-based scanners when compared to their CIS cousins (all the scanners listed here use CCD sensors, which admittedly have other drawbacks like potential geometry issues and limited maximum sharpness introduced by their optics), there also is some variation among CCD units. Many of them seem to range around 3 to 7 mm at moderate resolutions, but I have also seen values of several cm quoted.
You may remember that DoF in cameras is tied to focal length and aperture, and scanners are no different. Now for a given focal length, aperture also determines how much light gets to the sensor, which brings us to the dynamic range vs. speed tradeoffs that I mentioned in the previous section. If you want to make a speedy high resolution scanner with high dynamic range, you'll need enough light for that, so you'd want a larger aperture, but that in turn reduces DoF and potentially the resolution of the optics as well (see typical effect of stopping down in lenses), so your chances of actually being able to use the high sensor resolution drop. Has anyone ever thought of producing variable-aperture scanner optics?
If you need to make super sharp scans with absolutely no geometric distortion, CIS actually is your best bet – but only if your original can be quite literally pressed against the scanner glass.
Did you realize that a 1200 dpi scan of a 35mm slide (36x24 mm) only gives a 2 megapixel image? Now granted, those would count about as much as 4 area sensor (camera) megapixels, but still, we're a bit past that sort of resolution already. (Some people apparently scan their slides at 600 dpi and are happy with that!) Good film on a sharp shot may give about 100 lp/mm, that's around 5000 dpi. Since flatbeds have a really hard time getting anywhere near their nominal resolution even with focus being bang on (a Perfection V700 might reach the resolution equivalent of around 2500 "real" ppi), this should explain why film scanners stil hold a resolution advantage. Flatbeds are handy for medium format though.