Scanning with digital camera and stitching

[alanrockwood]

OK, it's not a new idea, but I have been putting together a simple system to scan film using a microscope objective to capture small pieces of the film image and then stitching the results together.

I am doing this on an ultra-low budget. The most expensive part is probably the 4X microscope objective that I bought on ebay for $14.99 and some extension tubes ($11 for two sets of tubes, including shipping) and some adapters to adapt my camera and microscope objective to the extension tubs (I don't recall the price, but probably about $15), also bought on ebay. Most of the rest of the parts were things I had laying around the house, like scraps of wood, an old relatively small light table, a film holder from a scanner, nuts and bolts, epoxy glue, a short piece of abs tubing (bought actually, but I don't remember the price).

So far I have used photoshop's automatic stitching method for the stitching operation, but I might use hugin once I master its use.

I have only done this on a very small number of negatives so far. I was going to post one of them (a flower shot from a fraction of the frame) but I could not compress the jpeg enough. I kept getting "file too big" errors, even when I set the compression to 5%, and by that time (5%) the quality was really suffering. If I can find a way around this I will post a photo later. Anyway, I think the photo turned out pretty good.

A technical note: The microscope objective I am using is a plan-achro type that is nominally 4X. I am using at at 3.75X magnification, and with my 1.6X crop sensor Canon XTi the effective magnification is 6X. That means that, considering overlap required between shots, I will need to take about 64 shots to completely capture full a 35mm frame. I haven't tried that yet. So far I have only captured a fraction of a full frame. I am using an old 10Megapixel digital slr to do the capture. With this setup, to do a full frame capture I will end up with producing a photo with about 360 Megapixels (over a gigabyte of disc space). Even the partial frame photo I took of the flower ended up producing an image with 129 megapixels.

If anyone is interested I can give more details about my setup. It might appeal to all of you cheapskates out there, but I don't want to make the present post too long.

 

[shutterfinger]

Images posted on this site can vary in size. The limit seems to be length/width and resolution. A 5x7 image at 150 dpi will upload but an 8x10 at 150 dpi will not but an 8x10 at 96 dpi will. Make a copy of the image then resize it in editing software for posting here.
Use an image hosting site or your google drive (if you have one) to upload the full resolution image to then link to it in your post here.

 

[ssssjjjj]

Have you tried saved as tiff or PSB? I used similar method to shoot a 4x5 film (1:1)m and then merged it to a 700mp file. I had to save it as PSB.

 

[alanrockwood]

Images posted on this site can vary in size. The limit seems to be length/width and resolution. A 5x7 image at 150 dpi will upload but an 8x10 at 150 dpi will not but an 8x10 at 96 dpi will. Make a copy of the image then resize it in editing software for posting here.
Use an image hosting site or your google drive (if you have one) to upload the full resolution image to then link to it in your post here.

Thanks. That gave me some ideas to try, and here it is as a smaller file.

And here is a crop of the center part.

Keep in mind that even the main image was already a crop from the original negative, as described in an earlier post.

 

[MattKing]

If you resize a colour image down to 800 pixels on the long size, it will upload.
If you resize a black and white image down to 1200 pixels on the long size, it will upload.
JPegs in both cases.

 

[alanrockwood]

I am attaching some photos of my setup. Please excuse the clutter in the photos.

I think there are probably enough views to understand how it goes together. It is basically a set of rails and a camera holder made from mostly scrap pieces of wood. The ABS pipe is there to keep the camera from falling over. It's not strictly necessary but seems like a good idea. Focusing is accomplished by the 1/4 X 20 screws. I epoxied some 1/4 X 20 nuts to the camera holder, and the screws are threaded into them. Positioning of the rails and the camera holder is done by scraps of popsicle sticks glued to the rails and camera holder. Some of these are necessary to prevent movement when operation the screws. The screws are there to provide focusing by raising or lowering the camera holder. Not shown in the photos is a cable release for the camera. I find that if I set the iso to 100 then a 1/10 second exposure is the right amount of exposure.

This setup is very basic. Some improvements to come later: 1) a focusing magnifier to help focus on the details of the film (on its way from ebay, about $28), and 2) some spacers to help align the film holder. The spacers will go on top of the light table to help align and position the left/right positions. I will probably make them with my 3D printer, or maybe I will cut them from strips of wood. I am also working out how to position it in the other direction. It might just amount to using a ruler. Precise positioning is not important, as long as there is enough overlap between shots. For the time being I am just eyeballing the positioning. 3) I can implement fluid scanning by using a piece of glass or acrylic sheet in place of the film holder.

I calculated that with the canon XTi and the 4X microscope objective I will get the equivalent of approximately 8000 dpi resolution. after the shots are stitched together.

Here's how I figure the resolution. The effective magnification is 6X because I have the extension tube length such as to give 3.75X, and with the 1.6X crop factor it is 6X. The XTi is a 10 Megapixel camera, so the stitched image is about 360 megapixels. On a Bayer sensor (before demosaicing) 80 megapixels of those are red and 80 megapixels are blue and 160 megapixels are green. That works out to be about 8200 dpi of effective resolution, assuming that the anti-aliasing filter of the camera. is not limiting. With the anti-aliasing filter it's probably a little under 8000 dpi effective resolution. I figured this for the red (or blue) sensor(s) because they are the limiting sensors as far as resolution is concerned. If It were just up to the green sensor, and if it weren't for the anti-aliasing filter, the resolution would be almost 12,000 dpi. It seems to me that this may rival a drum scanner, though there may be some other issues (such as noise issues) where a drum scanner might be better.

8200 dpi translates into about 160 line pairs per mm of resolution at the Nyquist limit. It probably won't quite achieve that, but my guess is that it might achieve something like 130 line pairs per mm. TMAX film supposedly resolves 200 line pairs per mm or better, so it won't quite reach a theoretical maximum that a fine grain film is capable of, but with the lens/film combination most photos probably don't reach much more than 100 lines pairs per mm, even with pretty good technique, so this level of performance is probably a good compromise. Velvia is listed as capable of 160 line pairs per mm, but as with TMAX, by the time you add a lens and things like focusing error, camera shake, non-optimal aperture, etc. few photos are likely to get much more than 100 line pairs per mm, even with pretty good technique.

An additional advantage of these extreme dpi counts is that grain aliasing is likely to be negligible.

The biggest issue in this scheme is the stitching. It would require stitching about 64 shots to make a 35mm scan (taking overlap into account). Stitching that number of images might not be trivial. So far I have only stitched part of a 35mm frame. I gave one example in an earlier post.

 

[Chan Tran]

Why would you need that high dpi?

 

[alanrockwood]

Why would you need that high dpi?

I just edited my last post to address the issue of why extreme dpi specifications are desirable.

 

[Chan Tran]

I just edited my last post to address the issue of why extreme dpi specifications are desirable.

If you scan large format it may make sense but for 35mm a typical 24MP camera with resolution of 4233 dpi is more than sufficient.

 

[alanrockwood]

If you scan large format it may make sense but for 35mm a typical 24MP camera with resolution of 4233 dpi is more than sufficient.

I partly agree... more than sufficient for some purposes for 35mm, but a 24mp digital camera not enough to extract all of the information available in some cases. In fact, a full frame 24mp camera with a Bayer sensor is effectively a camera with 3 interleaved sensors, a 12mp green sensor, a 6mp red sensor and a 6mp blue sensor. It therefore cannot even prove true and accurate RGB color and intensity for every one of the 24mp sensors, so it is not a true 24mp sensor. It must use some kind of mathematical algorithm to try to fill in the missing information during the demosaicing process. The best a 24mp full frame Bayer sensor camera can guarantee is 42 line pairs per mm of artifact-free RGB imaging. It can sort of do better, but only in the sense that some errors may be present such as color mosaicing or other artifacts in some pictures.

Even if a "24mp camera" used a true 24mp sensor (i.e. 24million red sensors, 24 million green sensors, and 24 million blue sensors) it would not be able to extract all of the information available in a tmax negative or a velvia slide in cases where the information content of the image is limited by the film and not the lens or other factors such as focusing accuracy or camera motion. To get all of the information from a film like TMAX that can resolve up to 200 line pairs per mm implies at least 138 million pixels. That number is set by the Nyquist theorem and cannot be gotten around by any kind of interpolation scheme.

One of the issues with film is that there is a gradual rolloff in the modulation transfer function, so there is no hard cutoff in terms of line pairs per mm that it can resolve. This has all kinds of interesting implications, but there isn't space or time to discuss all of them here, but one consequence is that as one looks at a film image more and more closely there is a gradual fall-off of image quality, unlike a typical digital sensor which has a flat modulation transfer function up to some limit, where the modulation transfer function then falls rapidly to zero. Consequently a picture obtained by digital capture that looks real sharp up until the point where it fairly suddenly looks terrible.

 

[Billy Axeman]

The sensor of the Rebel XTi gives you an image 3888 x 2592 px (as RAW), which can be translated directly to the desired number of pixels and the number of parts to stitch for a given negative size.

For example, when the desired resolution is 4000 dpi, you need 5668 px horizontally and 3780 px vertically for a 36x24mm negative:

Sensor Canon EOS Rebel XTi: 10.5 Mp, 3888 x 2592 px (RAW)
Negative 24x36 mm
1" = 25.4 mm
36 mm = 36/25.4 = 1.417"
24 mm = 24/25.4 = 0.945"

Needed number of pixels horizontally and vertically:
4000 dpi * 1.417" = 5668 px horizontally
4000 dpi * 0.945" = 3780 px vertically

With this sensor you can do that by stitching 2 rows and 2 columns (including some overlap).

 

[Helge]

I partly agree... more than sufficient for some purposes for 35mm, but a 24mp digital camera not enough to extract all of the information available in some cases. In fact, a full frame 24mp camera with a Bayer sensor is effectively a camera with 3 interleaved sensors, a 12mp green sensor, a 6mp red sensor and a 6mp blue sensor. It therefore cannot even prove true and accurate RGB color and intensity for every one of the 24mp sensors, so it is not a true 24mp sensor. It must use some kind of mathematical algorithm to try to fill in the missing information during the demosaicing process. The best a 24mp full frame Bayer sensor camera can guarantee is 42 line pairs per mm of artifact-free RGB imaging. It can sort of do better, but only in the sense that some errors may be present such as color mosaicing or other artifacts in some pictures.

Even if a "24mp camera" used a true 24mp sensor (i.e. 24million red sensors, 24 million green sensors, and 24 million blue sensors) it would not be able to extract all of the information available in a tmax negative or a velvia slide in cases where the information content of the image is limited by the film and not the lens or other factors such as focusing accuracy or camera motion. To get all of the information from a film like TMAX that can resolve up to 200 line pairs per mm implies at least 138 million pixels. That number is set by the Nyquist theorem and cannot be gotten around by any kind of interpolation scheme.

One of the issues with film is that there is a gradual rolloff in the modulation transfer function, so there is no hard cutoff in terms of line pairs per mm that it can resolve. This has all kinds of interesting implications, but there isn't space or time to discuss all of them here, but one consequence is that as one looks at a film image more and more closely there is a gradual fall-off of image quality, unlike a typical digital sensor which has a flat modulation transfer function up to some limit, where the modulation transfer function then falls rapidly to zero. Consequently a picture obtained by digital capture that looks real sharp up until the point where it fairly suddenly looks terrible.

Bravo! You just saved me typing all that up (again. .

One thing I’d add, is that it is actually very possible to extract or raise contrast of much of that low contrast detail.
It’s always been a part of the enlarging process to some degree, both with B&W and C41.
But of course usually with digital manipulation, the problem was that grain got pulled up and “enhanced” as well.
That is changing right now, with far more intelligent sharpening, based on deep learning networks, doing exactly what humans have done for decades.
That is, look through the grain and abstract from it, and even enjoy it as a separate emotional element (the last one is probably still exclusive to humans .

Of course all that matters little when you actually use the resolution for what it’s meant for:
Printing big or projecting. When you inspect a big print the low contrast detail is perfectly visible and actually looks perfectly natural combined with the grain.
Everybody can experience this magic by buying a decent projector and a roll of slide.

 

[Les Sarile]

The sensor of the Rebel XTi gives you an image 3888 x 2592 px (as RAW), which can be translated directly to the desired number of pixels and the number of parts to stitch for a given negative size.

For example, when the desired resolution is 4000 dpi, you need 5668 px horizontally and 3780 px vertically for a 36x24mm negative:

Sensor Canon EOS Rebel XTi: 10.5 Mp, 3888 x 2592 px (RAW)
Negative 24x36 mm
1" = 25.4 mm
36 mm = 36/25.4 = 1.417"
24 mm = 24/25.4 = 0.945"

Needed number of pixels horizontally and vertically:
4000 dpi * 1.417" = 5668 px horizontally
4000 dpi * 0.945" = 3780 px vertically

With this sensor you can do that by stitching 2 rows and 2 columns (including some overlap).

Mathematically speaking . . .

However, consider the following results I get from 36MP Nikon D800 which has 7360 x 4912 pixels applied to the same frame of 35mm compared to the Coolscan 4000dpi.

Full res -> http://www.fototime.com/8372250EA44CB06/orig.jpg

You can see that the D800 scan applies more pixels to the same area compared to the Coolscan but yet the Coolscan compares very favorably. Of course the film itself has considerable more detail that was not resolved when you look at the 4.5X optically magnified crop on the right. So the film itself was not the limiting factor.

 

[alanrockwood]

Here are two cropped images from the same negative. One was generated by the method described earlier using a digital camera and stitching. The other was from a scan from my Canon 4000 dpi fs4000us scanner. The one from the canon was upsized to the same number of points (using bicubic interpolation).

I don't think there is much to choose from in terms of resolution. It is probably limited by the negative itself. I think the one from the digital camera/stitching looks less grainy.

I have a resolution target that I will try to get around to scanning before too long. This may tell us if there is much difference in actual resolution.

 

[Adrian Bacon]

I partly agree... more than sufficient for some purposes for 35mm, but a 24mp digital camera not enough to extract all of the information available in some cases. In fact, a full frame 24mp camera with a Bayer sensor is effectively a camera with 3 interleaved sensors, a 12mp green sensor, a 6mp red sensor and a 6mp blue sensor. It therefore cannot even prove true and accurate RGB color and intensity for every one of the 24mp sensors, so it is not a true 24mp sensor. It must use some kind of mathematical algorithm to try to fill in the missing information during the demosaicing process. The best a 24mp full frame Bayer sensor camera can guarantee is 42 line pairs per mm of artifact-free RGB imaging. It can sort of do better, but only in the sense that some errors may be present such as color mosaicing or other artifacts in some pictures.

Even if a "24mp camera" used a true 24mp sensor (i.e. 24million red sensors, 24 million green sensors, and 24 million blue sensors) it would not be able to extract all of the information available in a tmax negative or a velvia slide in cases where the information content of the image is limited by the film and not the lens or other factors such as focusing accuracy or camera motion. To get all of the information from a film like TMAX that can resolve up to 200 line pairs per mm implies at least 138 million pixels. That number is set by the Nyquist theorem and cannot be gotten around by any kind of interpolation scheme.

One of the issues with film is that there is a gradual rolloff in the modulation transfer function, so there is no hard cutoff in terms of line pairs per mm that it can resolve. This has all kinds of interesting implications, but there isn't space or time to discuss all of them here, but one consequence is that as one looks at a film image more and more closely there is a gradual fall-off of image quality, unlike a typical digital sensor which has a flat modulation transfer function up to some limit, where the modulation transfer function then falls rapidly to zero. Consequently a picture obtained by digital capture that looks real sharp up until the point where it fairly suddenly looks terrible.

there is a really big difference between extracting all of the information and extracting all the pictorially useful information that we can actually see with our eyeballs. Let’s not forget that a bayer sensor very closely mimics the color resolution of our own eyes. This whole “we have to have full RGB color for every capture sample” in my opinion is a false premise simply because our own eyes don’t see that way. Yes, cutting color resolution out is generally a bad thing, but if you’re still at or above the color resolution of our eyes, it’s hard to tell the difference. Case in point with video color sampling: 4:2:0 color sampling looks fine, 4:2:2 color sampling looks better, but 4:4:4 color sampling doesn’t look appreciably better than 4:2:2, except in very, very specific scenarios like a scene was shot under red light, or a very strong color cast.

the same goes for the whole bayer interpolation artifacts argument. Yes, those artifacts can happen when you’re taking a picture with a digital camera out in the world, however in my experience (something I have a lot of experience with as I scan a lot of film with a bayer sensor) those issues are significantly mitigated when using a bayer sensor to scan data. Because of the random nature of film grain and the fact that film rolls contrast off as you add finer detail, a lot of the issues with bayer interpolation just either go away or are significantly reduced to the point that it’s just not really so much of a problem that I see. I’d rather get my film scanned at a totally usable resolution and color fidelity and get on with life than hork around trying to overcome perceived issues and limitations.

if there’s an issue or limitation that totally kills it, show an actual film scan with a bayer sensor that shows it. I’ve seen a lot of film scanned with a bayer array and not seen any show stoppers that makes me want to change what I’m doing or try to improve upon it.

I recently upgraded my scanning system from 24MP to 32.5MP. Guess what? The files are bigger, but at any reasonable viewing size or print size it does not look appreciably better, both in terms of sharpness, or color fidelity, despite the pretty significant jump up in resolution. If I zoom in to a ridiculous “photographer sickness” level, yeah, the grain is sharper, but there is not significantly more pictorially useful information present. For 35mm film, a 24MP bayer sensor is approaching the point of diminishing returns.

 

[Adrian Bacon]

The sensor of the Rebel XTi gives you an image 3888 x 2592 px (as RAW), which can be translated directly to the desired number of pixels and the number of parts to stitch for a given negative size.

For example, when the desired resolution is 4000 dpi, you need 5668 px horizontally and 3780 px vertically for a 36x24mm negative:

Sensor Canon EOS Rebel XTi: 10.5 Mp, 3888 x 2592 px (RAW)
Negative 24x36 mm
1" = 25.4 mm
36 mm = 36/25.4 = 1.417"
24 mm = 24/25.4 = 0.945"

Needed number of pixels horizontally and vertically:
4000 dpi * 1.417" = 5668 px horizontally
4000 dpi * 0.945" = 3780 px vertically

With this sensor you can do that by stitching 2 rows and 2 columns (including some overlap).

or you can do something even easier and just get a higher resolution camera. The xti (I have one) is pretty low resolution by today’s standards.

 

[Patrick Robert James]

Based on the experiments I've done I would say that you need a pixel density with a Bayer sensor of at least 11,000 dpi to extract the information off the film. 4000 is just scratching the surface frankly. The thing though is that 11,000 is pretty time consuming and/or expensive to get. Anything less than that though is only giving a false sense of sharpness and color, although that may be good enough and usually is. I don't shoot enough color film to rigorously pursue this, but if I did I would probably buy one of the newfangled high megapixel cameras that have pixel shift. The issue then though is a lens good enough to resolve.

I just keep using my Coolscan which is good enough for what I do with color images, which isn't much.

 

[Billy Axeman]

For a more realistic setup, and still with a lavish resolution, an affordable FF camera with pixel shift is the Pentax K1-II ($1800,-). It can be used to copy a 24x36 frame at 5200 dpi without stitching:

Sensor Pentax K1-II: 36.4 Mp, 7360x4912 px (RAW)
Negative 24x36 mm
1" = 25.4 mm
36 mm = 36/25.4 = 1.417"
24 mm = 24/25.4 = 0.945"

Achieved resolution:
7360 px / 1.417" = 5194 dpi
(4912 px / 0.945" = 5194 dpi)

The perceived resolution is actually better with the pixel shift feature activated.

However, the elephant in the room is that the camera can make better photos than the negatives it is scanning, by using it directly without detour.

 

[Helge]

In response to Adrian and Alan, I'd say, as always it very much depends on the lens and film used, and taking circumstances.
Incidentally we where never told the original posters film setup. IE. lens, film, whether tripod was used, how sure of focus he was, etc?

Exactly the same reservations goes for a digital cameras of course.
Less than optimal lens, taking circumstances and sensor, post processing and so on, takes the result many notches down in absolute resolution. Just like on film.
Film like magnetic tape though, is often more elegant and sympathetic to the way the human senses work, in how gracefully it appears to degrade under such circumstances.

Slow slide, Portra and Ektar is going to give to you a surprising amount of solid, very real resolution, even in 135 format. If you really want it that is.
Far in excess of the usually prescribed upper “practical limit”, of the equivalent of 24MP.
Sure there will be grain, just as there is physical media substrate noise in all other recording methods (tape hiss, groove noise etc.).
With a sensor there is pixelation, due to the attempt at basic abstraction of the image.
B&W film is really not that far ahead in resolution, outside a few very low ISO specialist films, if at all.

No one is contesting that good enough is good enough most of the time.
Many people, including most people making a living of images, would be absolutely fine with iPhone resolution photos ninety percent of the time.
Any additional resolution is at best wasted.
What I’m talking about, is printing poster size and above. Or projecting. Or just wanting to study a photo in detail (pixel peeping), for a variety of possible reasons.

The colour perception models used in compression and sensor arrays is based on loose, dated (as in some of it being from the first half of last century and older) assumptions and rules of thumb about how human vision works. So they only “work” in the broad sense of the word.
The “working” also includes the fact that human senses are very eager to adapt, compensate and understand, while pushing small niggling processing difficulties to the back of the attention hierarki, as much as possible.
A well known example is being able to follow a conversation for hours through strong background noise.
Sure you can get used to it, but it’s like being handed a glass of water in the desert, when the noise floor is removed.
Subject your senses to a problematic “new normal” long enough though, and suddenly anything else seems weird and wrong.

The borderline hand-wavy pseudo philosophicals above aside, the problems with the bayer sensor arrays need for spatial interpolation, artifacting naturally goes away gradually when you go beyond 1:1 macro. Diminishing rapidly in importance when approaching the 1:4 mark.
What does not go away as rapidly however, is grain aliasing, and the simple fact that the filter colours used in the array, have a less well separated and distributed colour response than would probably be useful to get the full gamut stored on the film.

 

[Les Sarile]

However, the elephant in the room is that the camera can make better photos than the negatives it is scanning, by using it directly without detour.

Goes without saying but cameras don't make photos, photographers do and that can be with whatever the photographer chooses to use . . .

 

[Adrian Bacon]

In response to Adrian and Alan, I'd say, as always it very much depends on the lens and film used, and taking circumstances.

exactly. The point I was trying to make is that the physical film itself is one piece of a total system. All the little bits and pieces affect total system resolution and it’s a lot better to focus on improving the total system or start with a reasonably good system and have serviceable images for the vast majority of uses. For 35mm film, a 24MP bayer sensor coupled with a really good macro lens and a solid scanning set up and light source is pretty difficult to top without significantly increasing the amount of time and complexity required to get there, and I would argue that the improvement wouldn’t really be significantly better for why you would go for all that extra resolution, which is print super huge at super high resolution. Why would you do that with 35mm? You can get much better performance going to a larger negative size.

 

[Adrian Bacon]

What does not go away as rapidly however, is grain aliasing, and the simple fact that the filter colours used in the array, have a less well separated and distributed colour response than would probably be useful to get the full gamut stored on the film

I’d love to see some actual evidence of this on film scanned with a bayer sensor. Both the raw scan before processing to get a color positive, and the processed color positive after.

the color filters are not unique to cameras. Scanners also have color filters that do the same thing.

 

[alanrockwood]

"... Why would you do that with 35mm? You can get much better performance going to a larger negative size."

True, larger format is one approach, and yet some people willingly pay good money (e.g. $30 to $100 per image) for a good 8,000 dpi to 11,000 dpi drum scan for 35mm, so it must be worthwhile to some people.

 

[Adrian Bacon]

True, larger format is one approach, and yet some people willingly pay good money (e.g. $30 to $100 per image) for a good 8,000 dpi to 11,000 dpi drum scan for 35mm, so it must be worthwhile to some people.

lol... people do a lot of things that make sense to them and don’t make sense to anybody else. I do the same thing.

 

[wahiba]

I scan and stitch 120 images using the 35mm facility of my Epson V300 and PE7. I also make up panoramas from digital and analogue and unless I have left no overlap, or far too little for the software to work on I have yet to find the software join.