modern tech should let us beat drum scanners for cheap

[oysteroid]

Drum scanners are old technology. And film scanners like the Nikon Coolscans have old CCDs in them. And most the flatbeds are pretty old tech.

You'd think that with all the amazing advances in digital sensors, it wouldn't be too hard to make something using that technology to scan the film that would surpass what a flatbed scanner can do, and perhaps even a drum scanner. All you really need is good, strong, calibrated, even backlighting, a fairly common, good quality sensor like those they put in something like a D5200, an excellent macro lens that will put 5000ppi of info onto the sensor, and a robotic stage that holds the film and moves it to capture a precise grid of images, maybe taking multiple exposures if necessary to get a large Dmax, with all of those images stitched together automatically. Surely, something like that shouldn't cost more than $1500. Hell, forget even the D5200's sensor. Use a cell phone camera sized sensor and a rather small but good macro lens and just take a whole bunch of little shots and stitch them. It should take less time by far than my flatbed or a drum scanner.

Or what about a technique like that used in those Hasselblad DSLRs:

"Hasselblad has announced a camera capable of producing 200MP images based on shifting its 50MP sensor. The H4D-200MS uses an extension of the company's sensor-shift, multi-shot (MS) technology to create a 200 megapixel file from six images taken at slight offsets. It can also use the four-shot mode used by the H4D-50MS that shifts the sensor by one pixel in each direction to capture all colors at each position."

Since your film isn't moving, you don't have to worry about it requiring multiple shots. Why not do that on a cheaper level with a cheap sensor? Then you wouldn't need a robotic stage to move the film for multiple shots.

I saw something on TV about a new camera system for spy drones that was developed in which they just used a bunch of cell phone camera chips, and it stitches all the images together on the fly, yielding video with 1.8 gigapixels of resolution! Those cell phone chips are cheap! Interesting. I think this is it here:

http://www.extremetech.com/extreme/110873-new-spy-drone-has-1-8-gigapixel-camera

But I honestly don't know why an entry-level DSLR with a good macro lens and a proper stage couldn't be used to digitize film with quality equivalent to a drum scanner. The more I think about it, the more I think a flatbed scanner is just a bad design to begin with, given that things are moving during the exposure, allowing vibrations to limit resolution.

[C. D. Keth]

The build quality and optics to be that precise is expensive. It's that simple. What you mention is possible but not enough people would pay the cost to make it worth developing, manufacturing, and marketing.

[scm]

http://www.largeformatphotography.info/forum/showthread.php?84769-Making-a-scanner-with-a-DSLR&highlight=dslr

[oysteroid]

Here is a video about the ARGUS sensor on the drone. It is pretty astonishing.

https://www.youtube.com/watch?v=0p4BQ1XzwDg

This made me think that maybe small sensors are the way to go for something like scanning. They are cheap and readily available. Small sensors use smaller lenses. And macro is easier to do with them. And the small pixel pitch means you don't have to achieve magnifications quite as high. Why not take a number of stitched shots using a small camera? Have you ever put a small convex lens over your cell phone camera lens? You can get amazing magnifications easily, and the DOF is suprisingly high.

And the limitations of less than perfect optics and sensors can be overcome with numerous samples. We just need to think outside the box here. What is difficult is recording all the information on a sheet of film in one go. That requires amazing optics and sensors. Sampling a bunch of times doesn't if your method is intelligent. Some ideas could be gleaned from the way astronomers increase resolution when seeing conditions and equipment are less than ideal. There are all sorts of other ways to approach these problems.

[oysteroid]

scm,

Thanks for the link. I'd kind of like to start a new thread though. As I'd like to explore the idea of trying something perhaps a little different than the usual DSLR idea.

Anyway, I grabbed an old 8mp compact Canon ELPH that I have laying around. I stuck a little stack of some small convex lens elements I had on the end of the lens, holding them on with my fingers, handholding the camera, and took a shot of my screen as close as possible. Here is a 100% crop:

88149

The total width of the picture amounts to half an inch of my screen, and the image is 3264 pixels wide. This yields roughly 6500ppi. The optics are far less than ideal. And yes, the image is noisy. But with the right little lens coupled with a more modern, higher resolution small sensor, and some multi-sampling, what might be possible? I think something surprisingly good could be made surprisingly cheap with some know-how and some can-do.

[oysteroid]

Okay, I was reading through the thread that scm linked to. It seems it does address a lot of what I am getting at. Maybe I should read that. If a moderator wants, my posts in this thread could be moved over there and this thread deleted, if it bothers anyone.

The more I read from that thread, the more ignorant I feel! :) That's a good thing. I am learning.

[richardman]

Actually, the sensors in the LS-9000 and probably the drums aren't so bad since they scan line by line, so they don't have that Bayer filter. i.e. the colors are accurate without interpolation.

[jslabovitz]

You've pretty much described the Imacon (now Hasselblad) Flextight "virtual drum" film scanners that were first released over a decade ago. They use a non-Bayer linear CCD array about 8000 pixels wide with excellent dynamic range, strong backlighting, a good Rodenstock macro lens, and a precise mechanical structure with which to move the lens (or maybe it's the sensor) closer or farther to the film. Additionally, the film is mounted in a magnetic holder which is bent slightly as it's pulled through, to keep the film parallel to the sensor without fluid mounting.

Although these machines originally cost $10-20,000, you can easily find them on eBay for about $2,000. I bought my Flextight Precision II for $1,500 (sans film holders), just about the price you're talking about. Being older machines with SCSI interfaces (later ones are Firewire, but are more expensive), they do require a bit of patience to get going, but once set up, they're wonderful beasts.

--John

[mortensen]

... and perfectly good drum scanners can also be had in the $1500 range, assuming you are patient, thorough and bit lucky.
For all the further reading, testing, discussions, thoughts etc on the DSLR scanner project, see these threads:
http://www.largeformatphotography.info/forum/search.php?searchid=1311362

... all the practical (and VERY real) obstacles rise to the surface. Apart from all the technicalities in building the stage and making it work automated, tethered etc, what I find to be the most troublesome obstacles (from reading, that is - haven't been part of the prototype projects) is the stitching (a slight vignetting from the lens, for instance) and aaaallll the manual dust removal afterwards - from Peter's test it seems worse than the Imacons (and that is bad, believe me).

[Nathan Potter]

oysteroid, I've been wanting to set up a DSLR scanning system for 35 mm and 4X5 films. Have so far just dabbled the past few days using a resolution target with a D800E and an older Vivitar Series 1 macro lens - the f/2.5 version. That lens is pretty low contrast but reportedly quite high resolution. I'm using a steel vertical optical bench for stability and tungsten illumination through a GG diffuser. Initial results seem encouraging.

It is not easy obtaining equivalent MTF values from the key chain of functional elements involved.
These include the light source, lens, and sensor so my initial value for MTF includes that chain. None of the elements can be conveniently separated from one another.

So here is a first shot using the Koren approach. Lens set at f/2.5, shutter 1/2500 sec., 3200K diffuse lighting, D800E mirror lockup mode, Manual focus on a wedged glass resolution plate. I selected the region of best focus which was very close to the D800E spot focus point. The resolution plate was tilted slightly in reference to the sensor array to reduce interference anomalies.

Image moved to PS for contrast evaluation. White and black points adjusted to histogram end points. The K values were then translated to the 0 to 256 8bit scale.

I picked 14 µm line pairs to scan with the eye dropper which yielded an average of 235 for clear and 165 for opaque. Employing the difference divided by the sum yields an uncorrected .175 (17.5%) contrast for 14 µm lines well above the Rayleigh limit which can be taken at about 9%.

Adding a correction factor for the low frequency contrast (measuring a large clear area and a large dark area) I get 250 for clear area and 60 for the opaque area. Difference divided by the sum yields .613.

Now normalizing using 100 X .175/.613 we have .285 X 100 = 28.5% at 14 µm linewidth. That is 28.5% at 36 lp/mm. or equivalently 28.5% 1816 spi.

I was hoping for better but this is considerably better than my results on the V750 Epson scanner.

I see a number of things that can improve this though.

1. Need to shoot at the lens critical aperture.
2. Need to reduce flare from the light source.
3. I think a collimated light source will improve the contrast - maybe drastically.

Clearly the fact that a Bayer pattern is used reduces the resolution and contrast expected because of the demosaicing required; but this is a black hole that cannot be accessed.

Ah, this should be in the scanner section.

Nate Potter, Austin TX.

[oysteroid]

Nathan,


So here is a first shot using the Koren approach.

Did you mean to attach or link to an image here? If so, I don't see anything.

Why would a collimated light source make a difference? I genuinely would be interested. I know that the rays of collimated light are parallel. That's about it.

I am surprised you are shooting at such a wide aperture. Your optical performance would be far less than ideal at f/2.5 and you wouldn't have enough DOF to handle some curvature in the film.

Also, as far as lights go, I once did a bit of testing like this and I just taped a piece of white paper to the wall and shined a number of lights onto it, trying to get the illumination as even as possible. I just put the film a distance away from that. It seemed to work fairly well in principle, though it didn't make a neat package, and it isn't the most energy efficient way, but you can give it as much light as you want and can use any type of bulb. But for experimenting at least, if someone doesn't have a diffuser, this will work. And there seems to me to be some advantage in holding the film vertical rather than horizontal, as it doesn't collect as much dust. And there is no need for anything like a copy stand. Just mount the camera on a board with something to hold the film, along with a shield around it, perpendicular to the board. If you want an easy way to get things aligned properly, just get a piece of mirror glass and clamp it tight to your film holder and draw a little crosshair in the center or something. Then look through the camera and point it such that the center of the view is aligned with the crosshairs. If you do that, at least the pitch and yaw must be right on. Aligning the roll is easy.

Also, it is critical with this kind of thing to put a dark shield around the film. You don't want any light bleeding around it that might cause some flare and loss of contrast. This is true of scanning as well. I am surprised at how many people tell how to do fluid mount scans and they don't suggest a light shield.


mortensen,


.. and perfectly good drum scanners can also be had in the $1500 range

Yes, but there are a lot problems there. And how long will it be possible to use those drum scanners? It just seems to me that all this new sensor tech should be useful for digitizing film. And it might be possible to use devices we already have. It surprises me that most still think the best devices for the job are so old. Digital imaging technology has come a long way since drum scanners became the standard way of digitizing images. And I don't see why such a big, bulky, power-hungry, slow, messy, technically challenging device along with a Mac museum should be the only way to get a quality digitization of a film image. And as far as something like DMax goes, I don't see the trouble at all. If you can't see into the denser areas of the film, use more light or a longer exposure! If you can't get it all in one exposure because of lacking dynamic range, take multiple shots and fuse them with Photomatix or something.

I did some stitching of DSLR shots in the past and used Hugin or PTGui and I had no problem with even slight vignetting. It made seemless stitches if set up correctly. It has features to match brightness and contrast and whatnot between the shots. As for dust, well, yes. But I have that problem anyway with my V750. The dust removal algorithms (ICE or Silverfast's own algorithm) all seem to do major damage to the image if they are sensitive enough to make a difference in the dust. So I just do it manually. It is indeed a nightmare to remove it all by hand. The image has to be worth it. But I am getting to pretty fast and accurate with my Wacom tablet and the cloning tool.

Also, as far as the bayer pattern on digital sensors goes, it seems to me that it is a simple matter of getting enough magnification so that you can downsample the digital image and still get to the resolution you want, just throwing away the bayer crap. If I remember someone's analysis correctly, downsampling a bayer image by a factor of about .7 gets you to something like pixel-level resolution. So if you want 3000ppi, just get enough magnification for 4300ppi, stitch some of those samples together, and downsize to 3000ppi.

With some programming and electronics know-how, I don't know why all this couldn't be automated and done for a LOT less than that Gigapan setup costs. I don't see why precision robotics or anything is needed. Some simple gears and some basic stepping motors should really be all that is needed. The stitching software can make up for some error. Some overlap is needed for the stitch anyway. There is no need to be pixel-precise or anything. We would be downsampling a bit anyway.

I am tempted to learn some scripting and electronics just for this purpose. If I did it, I would even open source it all. I don't know why good scans have to be so inaccessible. Some seem to be making this problem more complex than it needs to be. But maybe I am just naive, as I haven't tried to put it all together.

[mortensen]

Oysteroid, your enthusiasm is great. By all means, keep up your research and keep us posted - it will be greatly appreciated in here!
Initially I was very tempted by the DSLR-scanner idea, but I ended up buying a 20 year old drum scanner. The good things about drum scanners are:
They are usually built like tanks for demanding production - mine is in really good shape, despite its age.
They don't have to be that big and heavy. Look at the smaller Howteks, the Screen and Scanview scanners, most om them are tabletop.
Wet mounting might be a bit time consuming (haven't reached that yet...), but once it is done you just leave the scanner humming... and you will have remarkably less dust spotting to do afterwards.
Mac museum... well, I got a 10 year old powerbook for $100 - all in all my drum scanner and old mac take up approximately the same space as a DSLR-scanner rig.

I don't think it takes longer to wet mount and scan 4 4x5's on a drum scanner @ 4000ppi than it does to photograph, stitch and dust spot the same 4 pieces of film using a DSLR-rig... but please prove me wrong! Both time- and quality wise

[mortensen]

... a short extra note on the drum scanners (if you should ever fancy one): Choose a brand that is serviced locally ;) For me that meant one option, Scanview... so I got a Scanview

[Jody_S]

With some programming and electronics know-how, I don't know why all this couldn't be automated and done for a LOT less than that Gigapan setup costs. I don't see why precision robotics or anything is needed. Some simple gears and some basic stepping motors should really be all that is needed. The stitching software can make up for some error. Some overlap is needed for the stitch anyway. There is no need to be pixel-precise or anything. We would be downsampling a bit anyway.

I am tempted to learn some scripting and electronics just for this purpose. If I did it, I would even open source it all. I don't know why good scans have to be so inaccessible. Some seem to be making this problem more complex than it needs to be. But maybe I am just naive, as I haven't tried to put it all together.

The mechanics of doing this are almost identical to building a 2-axis CNC router using a Dremel tool as a cutter, and instructions for doing that are all over the 'net. It's not even very expensive, if you use threaded rods from the hardware store instead of proper lead screws (if you don't need 1/1,000" accuracy, why not). The software is pretty much just a freeware CNC controller again available on the 'net, with a relay for a cable release instead of a relay to turn on your cutter. Your 'pattern' for the CNC software will be simply a pattern of holes drilled into a flat 8x10" piece of stock, with each 'hole' calculated to be the center of one image. The software won't know you're snapping a photo instead of drilling a hole.

I don't personally use stitching software, I don't know if that part can be automated with scripts. But the 'scanning' certainly can.

[Nathan Potter]

oysteroid, no image attached, and not intended. I'm just fooling around with a primitive setup so far to sort of look at feasibility of obtaining some high level of quality.

I ran all f/stops from f/2.5 to f/22 using the wedged glass resolution target so I could extract DOF information as well as resolution and contrast. But see only a slight improvement in contrast at the critical aperture of f/5.6 or f/8. Clearly need to work on stray light baffling as you suggest.

I used a vertical setup because that is the hardware I had available that was ultra stabile. If DOF is a problem then the film can be sandwiched between glass after driving off moisture from glass with heat.
The DOF I'm seeing at a reasonable aperture (f/5.6 or f/8) may not require film flattening for 1:1 magnification or higher.

The notion of using a collimated light source is simply to reduce light scattering by the emulsion. In principle this can be carried to extremes where no intergrain modulation occurs between silver clumps.
That would be roughly analogous to drum scanning where a single, focused, small spot is all that is seen by the detector. Initially it is awkward for me to set up such a collimated source but I'm working on that.

My interest in this is partly intellectual and partly practical. If quality can exceed the V750 (seems likely) then I'd scan some 4 X 5 films where the image quality warrants. I might try 35mm if I could exceed the quality of the Nikon LS 5000 (doubtful).

I should take a picture of the setup just for documentation purposes and will do soon.

Nate Potter, Austin TX.

[Peter De Smidt]

Daniel and a few other friendly and knowledgeable people and I have been working on automating the scanning process for awhile now. We are using an Arduino Uno, 2 EasyDrivers, and stepper motors. Currently, we have the xy programing done, and we're working on menus and such things. It shouldn't be too long before we're ready to unveil fully working prototypes.

Ludvig has had a working version for awhile now, and he's talking about using an infrared light source along wiht a modified camera to use automatic dust cleaning for dye-based films. Since my film work is 99.9% BW, I haven't worked on that at all. Note that I did no special cleaning of my test scans, nor did I try wet-mounting the film.

Imo, it shouldn't be too hard to better the quality given by a Coolscan 5000, especially with grainier film. (I used to have a Coolscan V, and it did really poorly with HIE, for instance.) Basically, you can get whatever quality you need. If you want, you can go to higher magnifications using, say, a Mitutoyo 5x APO machine inspection lens on a 100mm tube lens, which would give 2.5x magnification and very high quality. Going much higher than 1:1 would probably require focus-stacking, but that's not a big deal, although it would at least double the amount of exposures needed. The Gigapixel project routinely scans opaque objects at 5x magnification, and they use telecentric optics and focus stacking.

[Jody_S]

....using, say, a Mitutoyo 5x APO machine inspection lens on a 100mm tube lens,....

I have a few (3?) OGP optical comparator lenses lying around somewhere, I would consider donating one for this project if postage could be arranged (they're quite heavy). They're probably much too large and heavy though for what you're trying to do, I think 15-20lbs each.

[Peter De Smidt]

Hi Jody,

Thank you for the offer, and maybe I'll take you up on it down-the-road, but lenses really are a small issue at the moment. Even an old 55mm Nikkor F3.5 macro lens at 1:1 gives better results than a consumer flatbed. It even gives my Screen Cezanne a run for it's money. I also have a bunch of enlarging lenses on hand, the Rodenstock lens from an Imacon, a 2x Mitutoyo APO, a 4x microscope objective..... Really, this type of stuff amounts to refinement. Just fyi, the best lens at 1:1 that I've seen tested is a Printing-Nikkor 90mm (or maybe it was a 105mm), a very rare and expensive lens.

This highlights a really great aspect of the project. Namely, once the basic structure and system is developed, people can easily modify it to suit their needs. Need a fast rig for duplicating slides at 1:1 with your full-frame dslr? No problem! No stitching necessary, and it's much faster than any traditional scanner. Need super high res? Knock yourself out with the highest quality and magnification optic you'd like. Want tremendous dynamic range? Use multiple exposures and exposure blending. What higher contrast? Try a collimated light source. Want really good color negative scans? Develop an RGB Led system where the light source balances out the orange mask.

[mortensen]

This highlights a really great aspect of the project. Namely, once the basic structure and system is developed, people can easily modify it to suit their needs. Need a fast rig for duplicating slides at 1:1 with your full-frame dslr? No problem! No stitching necessary, and it's much faster than any traditional scanner. Need super high res? Knock yourself out with the highest quality and magnification optic you'd like. Want tremendous dynamic range? Use multiple exposures and exposure blending. What higher contrast? Try a collimated light source. Want really good color negative scans? Develop an RGB Led system where the light source balances out the orange mask.

Daym... that was a bunch of really, really cool features you mentioned!!! Looking forward to your further experimentation and results!