For many years I am thinking about building a film scanning device that enables a similar approach like in professional scanners such as fuji frontier or noritsu lab scanners. But I was only thinking, reading articles and patents and so on, but never started to actually build something. Now I discovered GitHub - jackw01/scanlight.
This light source for scanning has almost everything I need for my approach, but lacks an infrared channel. Now I think about adding one, but I wonder what wavelength to choose: Close to visible light might not give enough SNR between scratches/dust and the image itself, but being too far off might cause too much out-of-focus blur and cause issues with the IR filter of usual cameras. So, does somebody have any insight into what wavelength fuji or noritsu are using? In this article: Input color profile to use for negatives - #13 by NateWeatherly, @NateWeatherly seems to have some insight, but maybe others know as well.
maybe I can supply a little bit of information from the PIE-scanners (reflecta DigitDia6000).
The IR-LED is centered at a wavelength of about 880nm with a width of about 60nm. The detector used for the IR is the G-band CCD. According to the CCD’s data sheet its filter opens up at longer wavelengths so it is sensitive enough at that wavelength.
Close to visible light might not give enough SNR between scratches/dust and the image itself, but being too far off might cause too much out-of-focus blur and cause issues with the IR filter of usual cameras.
Long ago, I used to fix watches and I bought an UV lamp and, based on Airy Disk size, removed the IR/UV blocking filter from my Sigma/Foveon DSLR. It captured dust and scratches very well, even without an UV lens. Method based on diffraction which is much less at say 350 nm compared to say 880 nm. Method obviously not applicable in this case when only talking about scanners.
In the general case, it remains that: the lower the illuminant wavelength, the higher the MTF per Rayleigh et al.
Because ideally, for a sufficiently long wavelength, the film substrate and the chemicals on it are transparent. Then, you only see the the scattered radiation from dust/scratches, and not the image itself, giving a nice mask.
This is of course not true with visible light, you get scattering but also record the image, so it’s difficult to untangle both. In the UV, for a sufficiently short wavelength, it might work if light is absorbed homogeneously on the substrate surface without reaching the silver particles. But it might be difficult to get a UV led with a short enough wavelength, and then the transmission of the scattered radiation through the lens will be horrendously low (I suspect you need something well below the 370nm cut-off of normal glass).
This seems to be close enough to the visible band such that optics and sensor may still work good enough to get some picture. I do own a reflecta scanner myself, so I can confirm that dust detection works reasonably well.
Thanks for this information, but now I wonder: How were you able to get it? Typically, LEDs are not marked with type names …
If @Jossie is correct, then it seems that for a decent separation, wavelengths close to the visible spectrum can be used. Looking at the infrared picture from my reflecta scanner, I typically get a ghost image of the actual picture, but I can clearly draw a separation threshold between the ghost image and the dust/scratch data, which is typically much more opaque, see an example image here: Scanned image scratch removal with “ICE” - #77 by chris. Furthermore, the dust image does not need to be perfect, a small amount of false positives will not affect the overall result, given a good inpainting algorithm is used. Here’s a comparison of different inpainting methods: Scanned image scratch removal with “ICE” - #21 by chris with data from my reflecta scanner.
As said, the goal is not a perfect IR image for the mask, and I suspect a little bit of lens blur is even beneficial for a good mask image. 880 nm are not too far off from the visible light anyway. I’ve written a little g’mic script to semi-automate the mask processing and application which you can also find information for in the thread linked above, and what this script does is to increase the detected mask regions a bit anyway such that no nasty edge effects happen.
Typical IR RCs for TVs are operating also between 840 and 940 nm, and typical cameras are able to “see” the light, such that also the IR filter of the cameras may not be an issue.
The spectrum of the IR-LED was provided by an engineer at PIE, obtained via PIE-US. The sensitivity curves of the CCDs+filter is from the data sheet from Toshiba.