I'm making my own film scanner and will attempt to sell it

[brbo]

If the backlight spectral power distribution multiplied by the masking spectral transmittance and sensor spectral sensitivity per channel achieved a trichromatic response which neutralises the orange mask (produces the same code values per channel) without the need for any post white balancing, and then the overall exposure were adjusted such that D-min transmittance sits at a code value of (2^bit_depth-1) / 2 per channel. That would allow for the cleanest capturing from D-min to D-max.

Yes, although gains are pretty small. Even with 12bit sensors. This is a link to a post where I compare "straight" and "neutralized" scan.

Thank you for clarifying this. I wonder if all digital cameras use this approach. I wish the manufacturers tell us more about what our cameras are actually doing with our images.

If you think about it, it makes no sense to do it otherwise. The "width" and "depth" of the pixel wells are predetermined, you can only capture a certain amount of photons. Even if you could adjust individual R, G, B pixels' voltage gain (different "ISO" for R, G, B) it would make absolutely no difference in 99.99+% pictures taken with digital cameras. Especially with modern sensors which are basically ISO invariant in a range that far surpasses the needs of a scanner.

 

[koraks]

Especially with modern sensors which are basically ISO invariant

That's a valid point, yes. In the end, it won't matter much if it's digital or analog gain being applied beyond a certain point. You'd end up introducing either some noise, or some mild posterization. The net quality difference will end up being quite minimal.

 

[gswdh]

Some images of the prototype almost complete. Please bear in mind this is a prototype!

The electronics here uses a BW only sensor at 2000 pixels which gives 2000dpi. This is the electronics I developed almost 8 years ago but it's still working strong.

I'm waiting for some LEDs to arrive to finish the back light then I'll be able to test the optics.

The film transport is working excellently. It needed a little fettling for 10mins to smooth off one of the corners but is great now.

The rubber band does look a bit amateur but works very well at applying pressure onto the film to grip the rollers. Overall I'm very happy and optimistic it will work to some extent. How well? I'm not sure!

 

[brbo]

One could easily swap out the nifty fifty for a proper lens. I think the sensor distance adjustment, even at the current design, is sufficient. Make three R, G, B, IR pulses, combine the BW sensor readout and you're in business...

 

[gswdh]

One could easily swap out the nifty fifty for a proper lens. I think the sensor distance adjustment, even at the current design, is sufficient. Make three R, G, B, IR pulses, combine the BW sensor readout and you're in business...

Absolutely, this setup was designed for easy changes like that. I'm in the process of sourcing at lens at the moment, I'd like to be able to test that with the production sensor on this setup.

Ha, yes! It could be done with RGB but the scan would be very slow as I don't think I could switch the LEDs that fast, worth a try tho maybe.

 

[koraks]

I don't think I could switch the LEDs that fast

How fast would be needed? There's a good chance you could, depending on the driver. The LEDs themselves are super fast.

 

[Bronson Dugnutt]

Nikon's recently announced digital microscope bears a striking resemblance to some of their old CoolScan units. It's easy to imagine a similar product tailored for film scanning but so is imagining pigs that fly.

That prototype is looking good - best of luck with your project!

 

[Steven Lee]

How fast would be needed? There's a good chance you could, depending on the driver. The LEDs themselves are super fast.

He uses a line sensor, so you'll have to do 3 exposures per pixel. Even assuming instant LED response, you'll need at least 1/125s of exposure times 3. That's no more than 3-4 lines per second, i.e. over 10 minutes per frame.

 

[brbo]

He uses a line sensor, so you'll have to do 3 exposures per pixel. Even assuming instant LED response, you'll need at least 1/125s of exposure times 3. That's no more than 3-4 pixels per line, i.e. over 10 minutes per frame.

How did you come up with 1/125s minimum time?


Btw, there were also ccd line sensors with 3 lines without R, G, B filters - so they still enabled monochromatic capture but at 3x the speed. Are those still available?

 

[koraks]

Even assuming instant LED response, you'll need at least 1/125s of exposure

Why would it have to be so monumentally slow? I was thinking of maybe a microsecond per exposure or so. Maybe even less; let's say 1 to 10 MHz. Depends mostly in the dynamics of the sensor. The timings can be gleaned from the datasheet.

 

[brbo]

Well, you will be limited by sensor sensitivity, illumination power and lens speed so I don't think you will be sampling film at a MHz, but 1/125s does seems awfully slow.

 

[Steven Lee]

Why would it have to be so monumentally slow? I was thinking of maybe a microsecond per exposure or so. Maybe even less; let's say 1 to 10 MHz. Depends mostly in the dynamics of the sensor. The timings can be gleaned from the datasheet.

Because of the distance between the light source and the sensor as shown in the photo. A camera scanning rig of comparable geometry would need around 1/30s for a LED-based light source without active cooling. Subtract the light loss from a color array and you're looking at 1/60 to 1/125s for ISO 100 @f/5.6.

The OP uses a line sensor here, and I haven't looked up their specs, but can we expect them to have vastly superior sensitivity to modern BSI area sensors from Sony?

 

[brbo]

I'll let the OP to chime in (he did state that currently he can sample a roll in 2 minutes, multiplied by 3 this is still as fast as anything currently out there), but I think you are forgetting that you only need a very narrow beam of light for a line sensor. That narrow beam can be much brighter (per Watt used) than a much bigger surface that you need to illuminate for an area sensor.

 

[Steven Lee]

@brbo You're referring to an image circle of a scanner lens. A bunch of them got tested on https://www.closeuphotography.com/ blog over time. They are not that different from full frame macros in that regard: comparable image circle, f/5.6 aperture. But you are right, let's wait for the OP to chime in.

 

[Adrian Bacon]

t's not so easy to make a signal path that actually yields the additional data and not just noise when going from 14 bit to 16 bit

True, the analog part before the actual ADC may still have a noise floor that's only 13-14 bits down from full ADC saturation, but even then, I'll still take a 16 bit ADC over a 14 bit one any day of the week simply because (and especially because of negative films), that gives me 4 times more discrete tone values for the same amount of noise if I put the film base plus fog right under the ADC clipping point. While where the noise floor is important, actually having a lot of discrete tone values significantly cuts down on posterization artifacts as you twist and pull the samples around to get your positive image.

 

[Adrian Bacon]

Does anyone know how white balance is actually implemented in digital cameras? Adjusting the gain of individual channels before digitization seems more logical than applying the corrections in post-processing.

Yes. It's a combination of digital gain and analog gain. For example, many Canon cameras have a native white balance of approximately 4600K (plus or minus a couple hundred K). This is where the digital multiplier between the red and blue channel is the same, or very close to the same. The green channel is typically about a stop more sensitive, so those digital multipliers are applied to the red and blue channels to get them to the same code value as the green channel if the light hitting the sensor was the same Kelvin as what the camera was set to. Those multipliers change as you walk up and down the Kelvin scale, and for the most part, it's digital gain via those multipliers until you start to get to the extreme ends of the Kelvin scale. As soon as your multiplier for the red or blue channel exceeds 2.0, Canon will actually change the analog gain for that color channel by a full stop, then reduce the multiplier down to close to 1.0, effectively having the green and blue channel at ISO 100 and the red channel at ISO 200 for example. Then as you continue to walk up to Kelvin scale, the multiplier for that channel will keep increasing to keep the white balance correct until it exceeds 2.0 again, then wash rinse, repeat. This is standard issue behavior for pretty much every digital camera. The bad ones never change the color channel analog gain and white balance is purely a digital gain function, and the good cameras will change it as needed, usually in full stop increments, so when measuring a camera for dynamic range and noise levels, the white balance you do it at, and the color of the light you're using is going to effect the outcome quite a bit, which is why you should take the measurements that various sites give with a grain of salt.

 

[koraks]

A camera scanning rig

...is a totally different animal. Pretty much every assumption you made about the parameters will work out differently for the kind of scanning setup proposed here. Light source, effective aperture, sampling strategy - the works.

As soon as your multiplier for the red or blue channel exceeds 2.0, Canon will actually change the analog gain for that color channel by a full stop

Where is this behavior documented?

 

[koraks]

Well, you will be limited by sensor sensitivity, illumination power and lens speed so I don't think you will be sampling film at a MHz, but 1/125s does seems awfully slow.

The main limitation is really in the sensor readout since this is a serial process.

 

[Adrian Bacon]

...is a totally different animal. Pretty much every assumption you made about the parameters will work out differently for the kind of scanning setup proposed here. Light source, effective aperture, sampling strategy - the works.



Where is this behavior documented?

In the raw file. You look at the white balance multiplier and the color channel bias. On recent canon cameras, the per color channel bias for ISOs 50, 100,and 200 is ~511. For all ISOs above 200, it’s 2047.

The per color channel bias will look like 511 511 511 511 (for RGGB) if the color channels are all set to the same ISO, and the RGGB White balance multipliers will be something like 1500 1024 1024 1500 if the camera is at its native white balance (This varies with camera model, but that’s the gist, for Canon cameras, you divide the WB value by 1024 to get the actual multiplier, and don’t apply the multiplier until after you’ve removed the channel bias). If you put the camera in kelvin mode, and take pictures as you walk up and down the Kelvin scale, and inspect the bias and multipliers at each Kelvin setting, you’ll see it change depending on what the camera is doing. Where on the scale it does it varies from camera model to camera model, but typically, it’ll do it when analog gain would be cleaner than the digital gain.

 

[brbo]

@brbo You're referring to an image circle of a scanner lens.

No, I think I was quite clear that I was referring to illumination. Theoretically, you only need a small fraction of power to illuminate a sliver of film that is being scanned at certain moment by a line sensor compared to an area sensor. Which also means that you can have brighter narrow strip of illumination with the same power you use for your camera scans.

Furthermore, staying with BW line sensor might not bring as much time penalty as one might think. Every tri-line colour filtered sensor will be less efficient than non-filtered BW sensor since, well, it's not filtered.

The main limitation is really in the sensor readout since this is a serial process.

Maybe that's the limitation that you really can't do anything about. But in every single scanner I've had it was the light source that was the real limitation to faster scanning.

If you have a scanner where you can set an analogue gain (effectively longer exposure) you will see that scan times will more or less follow the exposure. With analogue gain set to 2, my scan time would roughly go up by 2. In both cases there is only one sensor readout. Now, lets set multisampling to 2x. Scan times roughly double again. At 2x multisampling scanner is doing two exposures and two sensor readouts. My take from that is that sensor readout is not a dominant limiting factor of scanner speed. Again, I can imagine that if you had extremely fast lens and extremely bright light source sensor readout could become the limiting factor.

 

[gswdh]

I don't have a lot to go by calculating the exposure times etc. Currently the pixels are clocked out at 1MHz, the total number of pixels including the vertical transfer over head for this sensor is around 2100. So I'll be getting around 476 lines per second or 1/500s exposure time. I don't know if this is enough yet.

 

[brbo]

If you put the camera in kelvin mode, and take pictures as you walk up and down the Kelvin scale, and inspect the bias and multipliers at each Kelvin setting, you’ll see it change depending on what the camera is doing.

That alone doesn't mean that the camera changes individual photosites analogue gain, though.

I'm not saying it doesn't happen, I just came up empty searching for anything written on this feature...

 

[Steven Lee]

No, I think I was quite clear that I was referring to illumination. Theoretically, you only need a small fraction of power to illuminate a sliver of film that is being scanned at certain moment by a line sensor compared to an area sensor. Which also means that you can have brighter narrow strip of illumination with the same power you use for your camera scans.

Good point, I haven't thought of that.

 

[brbo]

Besides, I think you (and consequently me too) got confused by "1/125s exposure time enabling you to scan 3-4 lines per second". That's obviously not true. 1/125s would enable you to scan 125 lines in BW or 40 lines in RGB per second. Not great, but also not terrible. That 3-4 lines made me think you were saying that exposure time will be 0.125s which seemed like an order of magnitude too slow. 1/125s wasn't that crazy far off, 3-4 lines/s was.

Anyway, what I wanted to say or ask @gswdh was what does he think he will be doing differently than other consumer scanners to get significantly higher speeds?

1. I can see that getting much faster lens with adequate optical properties for scanning application will definitely be beyond his reach.
2. I really don't know how much development ccd line sensors got in the last decade, can you really run them significantly faster (higher sensibility) than two decades ago with same or lower S/N ratio?
3. I keep hearing that LEDs are getting better (brighter), but don't know nothing about them. Is this the area where the difference can be made?
4. Scanning entire rolls with perfect flatness enabling you to get away with fixed focus is a great time-saver, but will that alone be enough?

 

[Adrian Bacon]

That alone doesn't mean that the camera changes individual photosites analogue gain, though.

I'm not saying it doesn't happen, I just came up empty searching for anything written on this feature...

That’s because at the end of the day, it doesn’t change how you process it or change the white balance on a raw file. Even if the analog again was changed on a single color channel to get a cleaner signal, you can still do exactly the same thing with the digital gain to change the white balance after the fact. The only difference is the actual multiplier used will be different, but even then, (at least for Canon cameras) the multipliers you should use for the WB scale for that image are also documented in that raw file, so if you want to change the WB, you just look up what multipliers you need to use in the file and change it.