Lafodis160: a DIY Large-Format Digital Camera

What an interesting project! It's hard to believe that the guts, minus lens, only cost $50.00.

As long as there no screws the size of baby ants I’d be interested

bill winslow hansen wrote:

As long as there no screws the size of baby ants I’d be interested

No such thing.

I've been delayed getting a more DIYer-friendly version out, but it's still coming. I've slowly realized Lafodis160 is a bit too big a step for most DIYers, so first I'll be releasing KISS-E. It's a lot smaller and simpler, but shares most of the same tech. It's a roughly $20 E-mount camera using the same sensor: KISS-E is Kentucky's Interchangeable-lens Small Sensor E-mount camera:

The photo above is the second KISS-E prototype, which I was originally going to release, but I discovered that some of the parts varied more than I expected depending on vendor (especially the $3 rear OLED display), so I'm in the middle of the third version.

Graham Gibson wrote:

Very cool project!

Thanks. It was sort-of "done" when the paper went to Electronic Imaging, but it seems a lot more broadly useful, so that's how it's evolving now.

Did you have issues with the micro lens array in the sensor being optimized to the angle of incidence, therefore causing problems trying to stitch the images together?

Lafodis160 is still being tweaked to be easier for others to build and use, and I'm actually going to have two undergrad RAs working on cleaning-up the software for it this Summer. There are a bunch of tweaks; the latest one is I found it was worthwhile to buy actual NIR-blocking filters... that cost less than $1.

There is a problem getting raws from the camera. Fortunately, it can deliver JPEGs with a "raw" tone mapping, but there is some darkening off-axis which hurts DR. I don't think it's from microlenses, because it has the same issue using the built-in lens, but maybe? We'll just model the lower SNR as lower confidence in the values for stitching.

I read about another photographer attempting a similar project with the Raspberry Pi HQ cam in a 4x5 back. Building a camera like this into a large-format back would also help in composing the image.

Not happening. The first prototype was a more traditional 4x5 X,Y, and it's just way sloppier and harder to build. I'm pretty well sold on the angle, radius scan control.

I have considered having a second ESP32-CAM mounted with the default lens as an approximate EVF, but that will probably be an optional add-on. There also will probably be a fast framing scan mode.

I have seriously considered moving a bigger sensor around, even an EOS-M programmed using Magic Lantern, but that's just not as easy nor effective as one might hope.

I've thought about something like this. But the question remains- why not multiple sensors? If they are that low cost, then I see little reason not to explore this.

Here's the thing- yeah, you now have to move two sensors accurately, which you could claim would be bulkier overall, riskier and demand more advanced tech.

But in return you get an idea of how this project could be expanded to be faster. Aside from buying larger and bigger sensors, using multiple is the only other route I could forsee.

How about this. Try exploring placing sensors along an Archimedes spiral (in whatever form you choose) that rotates. Could be a nightmare for stitching and resolution loss, but why not?

Graham Gibson wrote:

Here's a link to the thread discussing the 4x5 back that I mentioned if you're interested:

https://www.fredmiranda.com/forum/topic/1665827/1#15365498

I thought your approach would run into the same challenge with the sensor if it was angle-optimized in the same way.

Very interesting!

It does not seem to be optimized that way for the included lens, but it does have the generic problem that the edges are always a bit darker -- including with the supplied lens. The camera hardware does have a "lens compensation" option which helps no matter what lens is used. Fundamentally, I simply think of the pixel values having decreasing confidence further off axis... but it could be that the microlenses aren't positioned correctly for the bundled lens. It's just had to understand why Omnivision would do that...?

BTW, now that I've added an NIR blocking filter, I see color tinting like you saw, which seems to come from the edges of the NIR filter.

I would love to try building one of these as a project. I have a 3D printer and a Nikkor 90/8 that I use for shooting film (6x17).

I'm still trying to simplify and improve the design, and have two students (who just started yesterday) working on clean-up of the software this Summer... so it's coming.

BTW, I actually don't have a problem with using a better sensor... it's just that things get more complex because the ESP32-CAM can't do it all then.

eggplanted wrote:

I've thought about something like this. But the question remains- why not multiple sensors? If they are that low cost, then I see little reason not to explore this.

That's always been on the list... however, it's not as big a win as you'd hope because we are trying to optimize the scan order dynamically, and that isn't as effective with multiple sensors in fixed relative orientation.

Here's the thing- yeah, you now have to move two sensors accurately, which you could claim would be bulkier overall, riskier and demand more advanced tech.

It's  actually trivial to mount multiple ESP32-CAMs on the shuttle that's driven along the radius.

But in return you get an idea of how this project could be expanded to be faster. Aside from buying larger and bigger sensors, using multiple is the only other route I could forsee.

How about this. Try exploring placing sensors along an Archimedes spiral (in whatever form you choose) that rotates. Could be a nightmare for stitching and resolution loss, but why not?

I think you'd consider our current stitching scheme an unbearable nightmare because it handles per-pixel-value certainty computations for merging and can dynamically change the order. A spiral scan is one of the simplest reasonable fixed patterns. We had built gigapixel stitching hardware more than a decade ago using a computer-controlled telescope mount -- even that used a modified Hilbert curve scan order.

As my interest is occasionally peaked by the prospect of 60mm width line scan cameras for a medium format camera, I thought to return as this seems pretty exciting. Reading the thread again I realise some of my questions were answered before my post.

ProfHankD wrote:

eggplanted wrote:

I've thought about something like this. But the question remains- why not multiple sensors? If they are that low cost, then I see little reason not to explore this.

That's always been on the list... however, it's not as big a win as you'd hope because we are trying to optimize the scan order dynamically, and that isn't as effective with multiple sensors in fixed relative orientation.

I think you'd consider our current stitching scheme an unbearable nightmare because it handles per-pixel-value certainty computations for merging and can dynamically change the order. A spiral scan is one of the simplest reasonable fixed patterns. We had built gigapixel stitching hardware more than a decade ago using a computer-controlled telescope mount -- even that used a modified Hilbert curve scan order.

I'm not familiar with scan ordering techniques so am trying to get up to speed as much as possible on "angle, radius scan control".

By having multiple sensors in a fixed relative orientation that might rotate in a sweep, it does prevent the dynamic movement this camera offers, but what are trying to do when you are optimizing the scan order dynamically?

I can't quite understand this in terms of the sensor's behaviour at the time of exposure. Simply put, how does it move during exposure?

By "dynamically changing the order" of the scan during exposure based on a certain value, I personally would envision the scan order in a very slow camera like this prioritizing movement in an image-

that is, the sensor moves to scan the next part of the image based on where the most motion blur/movement in the merged image was detected.

.

Once I understand the need to optimize the scan order dynamically, I'll have an answer to why using multiple sensors isn't effective, because from where I'm coming from even if it's difficult, its going to drastically speed up exposure times.

In order to call this a success we have to agree that imaging at 1MP scan speed for a 500MP image (at the least) is an acceptable rate, and I struggle to agree, even if your scan method is vastly superior to the linked Pi camera. I'm sure you're right about larger sensors being more difficult, but I'd be curious for an elaboration.

A line scan camera should always have the benefit of speed, and its just a matter of waiting the market out to acquire resolution and sensor size. To be done affordably I envision using multiple of the same model, as single high resolution & large size cameras are not climbing down in price, and top out about 12-16k anyway.

GFX100 100MP has 8736 pixels image height. You could get close with 4x2k line scan cameras. The 400MP mode is 17,472 pixels high. You could get close with 4x4k line scan cameras.

No, its not the 500MP-2.6GP that yours offers, but if you wanted to give up some speed you could maybe use them with a scan order aswell.

Eggplantt wrote:

As my interest is occasionally peaked by the prospect of 60mm width line scan cameras for a medium format camera, I thought to return as this seems pretty exciting. Reading the thread again I realise some of my questions were answered before my post.

ProfHankD wrote:

eggplanted wrote:

I've thought about something like this. But the question remains- why not multiple sensors? If they are that low cost, then I see little reason not to explore this.

That's always been on the list... however, it's not as big a win as you'd hope because we are trying to optimize the scan order dynamically, and that isn't as effective with multiple sensors in fixed relative orientation.

I think you'd consider our current stitching scheme an unbearable nightmare because it handles per-pixel-value certainty computations for merging and can dynamically change the order. A spiral scan is one of the simplest reasonable fixed patterns. We had built gigapixel stitching hardware more than a decade ago using a computer-controlled telescope mount -- even that used a modified Hilbert curve scan order.

I'm not familiar with scan ordering techniques so am trying to get up to speed as much as possible on "angle, radius scan control".

Classical scanning was done in a pre-determined, fixed, order. For a line scanner, it's just a simple sweep. For a scanner using a point or rectangular sensor, it's a raster scan much like that used in a TV set -- a control loop like:

for (y=0; y<Y; y=y+1) for (x=0; x<X; x=x+1) sample(x,y);

The good news is that order has sample(5,8) taken just one time unit after sample(4,8) and one before sample(6,8). However, that order also means sample(5,7) and sample(5,9) are temporally separated from sample(5,8) by X units of time. With a potentially changing scene, that means there is a high probability that scene content has changed and you'll get stitch errors. In fact, this same scan pattern is what causes the distortions associated with focal plane shutters and rolling electronic shutter. Thus, when I did my first gigapixel scanners in the 2000s, I used more complex scan orders that typically have much smaller temporal differences between when neighboring locations are sampled. For example, the Hilbert curve:

A Hilbert curve, from Wikipedia CC BY-SA 4.0 by TimSauder

For lafodis160, the actual drive system runs in polar coordinates: you have 360 degrees of rotation and 80mm of radial movement (moving 0 to 80 mm from the center). Despite that, you can still drive it in a raster scan or in a scan like the Hilbert curve above, but the sensor will be changing it's angle throughout the scan -- which makes stitching harder, but helps de-correlate any sensor defects in adjacent sample positions. There are also more efficient scan orders possible using the angle, radius drive, such as spirals and Hilbert-like curves mapped to polar coordinates.

By having multiple sensors in a fixed relative orientation that might rotate in a sweep, it does prevent the dynamic movement this camera offers, but what are trying to do when you are optimizing the scan order dynamically?

Adjacent samples actually overlap a bit to facilitate alignment. Thus, we can tell when the scene content for a previously-sampled portion of the scene has changed in a way that would cause a glitch in the stitching. We also know when neighboring samples were captured, hence can predict which ones are likely to be affected by the scene change. Thus, some already-sampled locations should be resampled... and that's the main reason the algorithm should dynamically change the scan order. However, it also can detect where portions of the scene are essentially content-free -- such as an evenly blue sky -- and could scan those areas faster with lower resolution and quality, just sufficient to confirm that there is "nothing to see here."

I can't quite understand this in terms of the sensor's behaviour at the time of exposure. Simply put, how does it move during exposure?

By "dynamically changing the order" of the scan during exposure based on a certain value, I personally would envision the scan order in a very slow camera like this prioritizing movement in an image-

The priority is internal consistency of scene details in the stitched resulting image.

that is, the sensor moves to scan the next part of the image based on where the most motion blur/movement in the merged image was detected.

We might have it detect motion blur, but that's not a priority. The current goal is simply making sure that adjacent samples will integrate seamlessly.

.Once I understand the need to optimize the scan order dynamically, I'll have an answer to why using multiple sensors isn't effective, because from where I'm coming from even if it's difficult, its going to drastically speed up exposure times.

The easy way to use multiple sensors is to move them together, and the polar coordinate system means that the utility of movements will vary with radius. If multiple sensors are lined-up at different radius offsets, the inner ones will be moving less than the outer ones -- so, we could use superresolution techniques to get higher resolution and improve SNR near the center of the stitched image, but that's not really a huge help. Placing multiple sensors on an arc so that they all have the same radius would work better, but would mean that all sensors would have to "go along for the ride" if any one needed to re-sample a position in the dynamic ordering... so there is a lot more complexity to defining the optimal scan motions.

Also imagine using Lafodis160 as a security camera. You can use a second, low-res, camera to detect scene change and then drive Lafodis160 to sample what's interesting at full resolution... perhaps even tracking subject motion, but with no external camera motion....

In order to call this a success we have to agree that imaging at 1MP scan speed for a 500MP image (at the least) is an acceptable rate, and I struggle to agree, even if your scan method is vastly superior to the linked Pi camera. I'm sure you're right about larger sensors being more difficult, but I'd be curious for an elaboration.

Well, if budget wasn't a big concern, a Sigma fp Lwould make a great scan sensor.

That would reduce the maximum "native" (non-superresolution) resolution to about 1.5GP, but scan speed and IQ would undoubtedly be way higher.

A line scan camera should always have the benefit of speed, and its just a matter of waiting the market out to acquire resolution and sensor size. To be done affordably I envision using multiple of the same model, as single high resolution & large size cameras are not climbing down in price, and top out about 12-16k anyway.

Line scan has good temporal properties, but large-format line scan sensors are not commodity products. It would actually be easier to build a full-size large-format sensor. The LargeSense folks hit a surprisingly good price point, although their cameras have HUGE pixels and hence low resolution... which isn't unreasonable if you think about how images are really used.

GFX100 100MP has 8736 pixels image height. You could get close with 4x2k line scan cameras. The 400MP mode is 17,472 pixels high. You could get close with 4x4k line scan cameras.

No, its not the 500MP-2.6GP that yours offers, but if you wanted to give up some speed you could maybe use them with a scan order aswell.

Well, that's only 44x33mm, which is even easier to do as a simple 3D-printed manual shifter for a FF camera -- which I've built and will soon post too. The upcoming "Budgie" allows up to 48x36mm capture using a FF E (FE) body and a lens that can be adapted to Leica M mount. Most FF lenses don't quite cover 48x36mm, and those that do still tend to have lousy corners, but here's one of my first test shots using Budgie on my A7RII:

Budgie on an A7RII, 3 OOC JPEGs stitched using Hugin (25% to minimize posted file size)

The above JPEG is full resolution, but I compressed it using 25% quality to keep upload bandwidth reasonable... any blocky artifacts in the OOF regions are from that, not from the stitching. Also keep in mind that any IBIS-based multi-shot high-res modes will still work with this, so on an A7RIII or A7RIV....

As far as I'm concerned, this is just another low-hanging fruit on the way to a new generation of scanning technology... sort-of like APSC2 (APS-C Squared) Rotate-and-Stitch Adapter . Hopefully, people will get comfortable making and using these simpler devices and thus become comfortable with things like Lafodis160, which is currently using really cheap parts, but is really at least a full generation ahead of any previous scanning tech... and Lafodis160 is still coming as a fully open source DIY design. I have two undergraduate students working this Summer on making Lafodis160's software better and more user-friendly....

Hopefully, people will get comfortable making and using these simpler devices and thus become comfortable with things like Lafodis160, which is currently using really cheap parts, but is really at least a full generation ahead of any previous scanning tech... and Lafodis160 is still coming as a fully open source DIY design.

I hope so. Many thanks for the response- does clear things up.

ProfHankD wrote:

Line scan has good temporal properties, but large-format line scan sensors are not commodity products. It would actually be easier to build a full-size large-format sensor. The LargeSense folks hit a surprisingly good price point, although their cameras have HUGE pixels and hence low resolution... which isn't unreasonable if you think about how images are really used.

Actually, if you go searching for Dalsa in the Business, Office and Industrial section, you would be very surprised. There's hundreds of results.
I've found multiple 8-12k monochrome line scan cameras of 56mm diameter for anywhere between £84-230 with free shipping, ready to make offers on. Two of those together would give you 4x5 diameters, and 2x12k = 24k should give you enough resolution to play with.

At that price, all I'll say is watch this space...

Eggplantt wrote:

ProfHankD wrote:

Line scan has good temporal properties, but large-format line scan sensors are not commodity products. It would actually be easier to build a full-size large-format sensor. The LargeSense folks hit a surprisingly good price point, although their cameras have HUGE pixels and hence low resolution... which isn't unreasonable if you think about how images are really used.

Actually, if you go searching for Dalsa in the Business, Office and Industrial section, you would be very surprised. There's hundreds of results.
I've found multiple 8-12k monochrome line scan cameras of 56mm diameter for anywhere between £84-230 with free shipping, ready to make offers on. Two of those together would give you 4x5 diameters, and 2x12k = 24k should give you enough resolution to play with.

At that price, all I'll say is watch this space...

Not for me.

Scanning an image with a line scan camera requires very precise and repeatable positioning, which I don't think 3D-printed parts can deliver. In contrast, minor sensor-positioning misalignments with rectangular cameras are trivially recognized and repaired by analysis of overlapping features. You know line scan cameras don't capture images -- they capture one line -- it's assumed the subject is moving past the camera at a precisely controlled rate so timing line sampling gives relative position. There's also to interface issue; most line scan cameras don't  use an interface that is compatible with cheap controllers.

ProfHankD wrote:

Not for me.

Scanning an image with a line scan camera requires very precise and repeatable positioning, which I don't think 3D-printed parts can deliver.

No, but the need for precise, repeatable positioning is found in a wide variety of mature, mass-manufactured devices and components; if the thrust of this project is low cost, then we should be taking advantage of this.

I would find it strange to suggest there's no off-the-shelf parts that come close to the convenience/availability/dependability of 3D printing filament. There's more to mention about what this attitude risks, but I'll save it for now.

Pneumatic linear actuators of various sizes and weights can be had under £70. All in one printers are being chucked in landfill and can have their scanning components swapped out for line scan camera. In even simpler terms, solid aluminium tubing, metric T Slots... the science would be in getting a consistent velocity to match timing. But as I say below, documentation...

I understand however it's difficult for me to come up with something equal to what 3D printing proposes- having one box which is supplied with one material, which doesn't require specialist tools or environment to produce your components.

In contrast, minor sensor-positioning misalignments with rectangular cameras are trivially recognized and repaired by analysis of overlapping features. You know line scan cameras don't capture images -- they capture one line -- it's assumed the subject is moving past the camera at a precisely controlled rate so timing line sampling gives relative position.

If using 3D printed parts, then yes, but once using some of the components I talk about I'm curious to see what the risk here is. I also wonder in your example how much repairing would have to be done the more you speed things up, given the lack of precision in 3D printed parts.

A benefit of a one-line camera could be the ability to travel across a lens' field curvature incredibly accurately, about as accurately as you can imagine before you use a curved sensor. This would require more advanced movement control, and maybe not worth it, but it may not have crossed your mind when thinking about what precise control of one line offers...

There's also to interface issue; most line scan cameras don't use an interface that is compatible with cheap controllers.

No, but the manuals provide you all the information required to do something about that - pinouts and how to change the settings. There's little else I can imagine you'd need, neither might it require C++ knowledge/a skillset to adjust, and there's a large body of knowledge often in layman's terms for these cameras online that isn't irrelevant to my use case.

P.S. No hard feelings- this helps me clarify my options, as long as you don't feel it's going off topic.

Eggplantt wrote:

ProfHankD wrote:

Not for me.

Scanning an image with a line scan camera requires very precise and repeatable positioning, which I don't think 3D-printed parts can deliver.

No, but the need for precise, repeatable positioning is found in a wide variety of mature, mass-manufactured devices and components; if the thrust of this project is low cost, then we should be taking advantage of this.

I am. The whole Lafodis160 is under $50 to make using low-quantity retail part costs. You're talking about costs at least an order of magnitude higher. I certainly have considered using 3D-printer parts (as opposed to parts made on a 3D printer), but so far I haven't seen that as a win; the systems are bulkier, heavier, and more power hungry.

... I understand it's difficult for me to come up with something equal to what 3D printing proposes- having one box which is supplied with one material, which doesn't require specialist tools or environment to produce your components.

I've had more than a few research projects become open source hardware+software, and I'll politely say that most people grossly overestimate what a person replicating someone else's design can deal with. I'm honestly even a bit worried people will have trouble wire-wrapping a dozen connections between components in Lafodis160... might make a simple custom board for it.

In contrast, minor sensor-positioning misalignments with rectangular cameras are trivially recognized and repaired by analysis of overlapping features. You know line scan cameras don't capture images -- they capture one line -- it's assumed the subject is moving past the camera at a precisely controlled rate so timing line sampling gives relative position.

If using 3D printed parts, then yes, but once using some of the components I talk about I'm curious to see what the risk here is. I also wonder in your example how much repairing would have to be done the more you speed things up, given the lack of precision in 3D printed parts.

No parts are perfect. Keep in mind that precision of parts does NOT imply precision of the device. For example, Lafodis160's radius drive screw was printed using 0.25mm-tall extrusion, but it can accurately move that axis in increments that are more than 100X smaller (moving in one direction; there is some slop when switching directions).

A benefit of a one-line camera could be the ability to travel across a lens' field curvature incredibly accurately, about as accurately as you can imagine before you use a curved sensor.

You have this oddly backwards. Line scanners don't bend, and the long, narrow shape makes them have more error in following a curved focus plane.

This would require more advanced movement control, and maybe not worth it, but it may not have crossed your mind when thinking about what precise control of one line offers...

I have considered 3-axis motion for Lafodis so it could essentially autofocus. The 3rd axis doesn't need much motion, but would need very fine control. Keep in mind that the sensor is small enough area so that focus plane curvature is never significant across it, so it would be easy to do a field flatness calibration for a lens and then apply the correction at each position in the field to give perfect correction for curvature of the focus plane.

There's also to interface issue; most line scan cameras don't use an interface that is compatible with cheap controllers.

No, but the manuals provide you all the information required to do something about that - pinouts and how to change the settings. There's little else I can imagine you'd need, neither might it require C++ knowledge/a skillset to adjust, and there's a large body of knowledge often in layman's terms for these cameras online that isn't irrelevant to my use case.

...and hard-to-get and/or expensive connectors, additioning chips needed for interfacing, etc. still make it a problem. Honestly, more of a problem for people to replicate than for me and my research lab to build and use one of. 

ProfHankD wrote:

I am. The whole Lafodis160 is under $50 to make using low-quantity retail part costs. You're talking about costs at least an order of magnitude higher. I certainly have considered using 3D-printer parts (as opposed to parts made on a 3D printer), but so far I haven't seen that as a win; the systems are bulkier, heavier, and more power hungry.

If it wasn't for this thread, then I probably wouldn't've done a more detailed search on line scan cameras- so you have succeeded in your aforementioned goal of this being "just another low-hanging fruit on the way to a new generation of scanning technology".

I believe I'm getting a magnitude higher camera mainly because of speed. I felt encouraged to respond because I feel it meets the terms you set out, but in a different way.

I believe this is so much of a win that I'm prepared to explore and document how I minimize the weight and the bulk of everything required, because the on paper benefits are a 24k 4x5" camera that all in costs <£600, compared to the £65,000+ of the aforementioned 8x10, where the drawbacks are atleast known far in advance, and are not particularly unique; weight, bulk, strength and precision.

The (blithe) confidence there comes from what I've garnered from the world of amateur telescope making, which has a consistent need for precision in optics and mounting that's extremely well detailed across many threads, books and the like. There should be far more crossover than currently exists with the photo world.

For example, lightweight reinforcing materials for telescope tubes, or truss-only designs, how to treat cardboard to give it extra strength. All stuff I could translate into reducing the weight of the proposed camera, because it's been done for far narrower gains in more difficult circumstances.

You may note I haven't included time as a cost, because from my position and what needs to be done, they are an equal time sink, differing in content.

I've had more than a few research projects become open source hardware+software, and I'll politely say that most people grossly overestimate what a person replicating someone else's design can deal with. I'm honestly even a bit worried people will have trouble wire-wrapping a dozen connections between components in Lafodis160... might make a simple custom board for it.

I don't disagree, but I don't think the aforementioned limitations of bulk, weight and power requirements are going to be solved with tricky/hacky solutions- lightweight rigid parts come from advanced materials, weight reduction comes from getting rid of the housing in a frame grabber, etc.

It is very obvious if something will work or not. That doesn't mean there isn't work to do, and learning to be done, but when we're doing a theoretical proposal I'm not sure I'm seeing roadblocks that completely write it off. If we're talking of the next generation of scanner cameras, then I'm able to include this as part of it. So, the next big and new thing would be an effort to document a number of common line scan camera modules on the surplus market, and how they could be used.

No parts are perfect. Keep in mind that precision of parts does NOT imply precision of the device. For example, Lafodis160's radius drive screw was printed using 0.25mm-tall extrusion, but it can accurately move that axis in increments that are more than 100X smaller (moving in one direction; there is some slop when switching directions).

You have this oddly backwards. Line scanners don't bend, and the long, narrow shape makes them have more error in following a curved focus plane.

That's correct- in my implementation the lens would be fixed in place, and the single line sensor would move along the path of the lens' field curvature.

That is, the camera wouldn't just move straight across a flat image plane- if coming from the left, it would start infront of the image plane, and as it moved closer to the centre it would move backwards. Across one, back one, record. There wouldn't be any rotation of the camera around the curve, it would be assumed that the sensor narrowness would be good enough.

On the contrary, to best use a square (area) sensor it would have to rotate around the field curvature.

This all assumes the field curvature of your lens is that bad that you would need this, though- as you describe below, you don't find it significant enough in your test case.

I have considered 3-axis motion for Lafodis so it could essentially autofocus. The 3rd axis doesn't need much motion, but would need very fine control. Keep in mind that the sensor is small enough area so that focus plane curvature is never significant across it, so it would be easy to do a field flatness calibration for a lens and then apply the correction at each position in the field to give perfect correction for curvature of the focus plane.

...and hard-to-get and/or expensive connectors, additioning chips needed for interfacing, etc. still make it a problem. Honestly, more of a problem for people to replicate than for me and my research lab to build and use one of.

Hardly much of an equivalent problem to what I described, and one which is very liable to research beyond the proposal stage.

For example, a 6 pin Hirose power cable is pretty cheap and the MDR26 used on the linked Piranha2 can be had for £15. It is inconsistently priced and requires further research. Some modern line scan cameras I have seen use USB 3.0. As part of the industrial surplus you are wading into, you might have a bundle of relevant parts you can make an offer on.

Eggplantt wrote:

ProfHankD wrote:

I am. The whole Lafodis160 is under $50 to make using low-quantity retail part costs. You're talking about costs at least an order of magnitude higher. I certainly have considered using 3D-printer parts (as opposed to parts made on a 3D printer), but so far I haven't seen that as a win; the systems are bulkier, heavier, and more power hungry.

If it wasn't for this thread, then I probably wouldn't've done a more detailed search on line scan cameras- so you have succeeded in your aforementioned goal of this being "just another low-hanging fruit on the way to a new generation of scanning technology".

I believe I'm getting a magnitude higher camera mainly because of speed. I felt encouraged to respond because I feel it meets the terms you set out, but in a different way.

I believe this is so much of a win that I'm prepared to explore and document how I minimize the weight and the bulk of everything required, because the on paper benefits are a 24k 4x5" camera that all in costs <£600, compared to the £65,000+ of the aforementioned 8x10, where the drawbacks are atleast known far in advance, and are not particularly unique; weight, bulk, strength and precision.

A very different price point....

The (blithe) confidence there comes from what I've garnered from the world of amateur telescope making, which has a consistent need for precision in optics and mounting that's extremely well detailed across many threads, books and the like. There should be far more crossover than currently exists with the photo world.

For example, lightweight reinforcing materials for telescope tubes, or truss-only designs, how to treat cardboard to give it extra strength. All stuff I could translate into reducing the weight of the proposed camera, because it's been done for far narrower gains in more difficult circumstances.

Telescope are certainly closely related in a variety of ways....

You may note I haven't included time as a cost, because from my position and what needs to be done, they are an equal time sink, differing in content.

Yeah, you don't ever want to count dev time.

I've had more than a few research projects become open source hardware+software, and I'll politely say that most people grossly overestimate what a person replicating someone else's design can deal with. I'm honestly even a bit worried people will have trouble wire-wrapping a dozen connections between components in Lafodis160... might make a simple custom board for it.

I don't disagree, but I don't think the aforementioned limitations of bulk, weight and power requirements are going to be solved with tricky/hacky solutions- lightweight rigid parts come from advanced materials, weight reduction comes from getting rid of the housing in a frame grabber, etc.

The 3D-printed parts of Lafodis160 are already pretty sneaky and advanced materials in that no part is less than 75% air (max 25% internal structure)... but there are other methods that are competitive. Just as a reminder, Lafodis160 is only 877g including lens! Put another way, the whole thing is about 40% the weight of the Nikon Z Noct lens alone.

It is very obvious if something will work or not. That doesn't mean there isn't work to do, and learning to be done, but when we're doing a theoretical proposal I'm not sure I'm seeing roadblocks that completely write it off. If we're talking of the next generation of scanner cameras, then I'm able to include this as part of it. So, the next big and new thing would be an effort to document a number of common line scan camera modules on the surplus market, and how they could be used.

No parts are perfect. Keep in mind that precision of parts does NOT imply precision of the device. For example, Lafodis160's radius drive screw was printed using 0.25mm-tall extrusion, but it can accurately move that axis in increments that are more than 100X smaller (moving in one direction; there is some slop when switching directions).

You have this oddly backwards. Line scanners don't bend, and the long, narrow shape makes them have more error in following a curved focus plane.

That's correct- in my implementation the lens would be fixed in place, and the single line sensor would move along the path of the lens' field curvature.

But the lens fields curve in two dimensions simultaneously -- there is no straight path.

That is, the camera wouldn't just move straight across a flat image plane- if coming from the left, it would start infront of the image plane, and as it moved closer to the centre it would move backwards. Across one, back one, record. There wouldn't be any rotation of the camera around the curve, it would be assumed that the sensor narrowness would be good enough.

On the contrary, to best use a square (area) sensor it would have to rotate around the field curvature.

A 1/4" sensor doesn't span a large enough area to have any significant curvature within a single capture -- you don't even need to tilt it to align well. The long dimension of the OV2640 is 3590um! Even tilting a line sensor, you'll have much more error along it's length.

This all assumes the field curvature of your lens is that bad that you would need this, though- as you describe below, you don't find it significant enough in your test case.

Don't think telescopes. A little curvature of field is part of the charm of large-format photography... so the real issue is if correcting is improving the image or making it look more like what you'd get from stitching with a smaller camera re-aimed. Continuity of bokeh is one of the best reasons for large format, and small degrees of curvature cause subtle, continuous, deformation of bokeh over a large region of the coverage.

I have considered 3-axis motion for Lafodis so it could essentially autofocus. The 3rd axis doesn't need much motion, but would need very fine control. Keep in mind that the sensor is small enough area so that focus plane curvature is never significant across it, so it would be easy to do a field flatness calibration for a lens and then apply the correction at each position in the field to give perfect correction for curvature of the focus plane.

...and hard-to-get and/or expensive connectors, additioning chips needed for interfacing, etc. still make it a problem. Honestly, more of a problem for people to replicate than for me and my research lab to build and use one of.

Hardly much of an equivalent problem to what I described, and one which is very liable to research beyond the proposal stage.

For example, a 6 pin Hirose power cable is pretty cheap and the MDR26 used on the linked Piranha2 can be had for £15. It is inconsistently priced and requires further research. Some modern line scan cameras I have seen use USB 3.0. As part of the industrial surplus you are wading into, you might have a bundle of relevant parts you can make an offer on.

Power supplies, cabling, etc. all adds up. USB 3 isn't a cheap microcontroller interface... but I suppose if we're assuming you can plug the camera into a high-end laptop....

ProfHankD wrote:

A very different price point....

I was overestimating the cost of mine based on what I could acquire to get this working even with my limited current knowledge, in order to show how it comes up exponentionally cheaper (and then some) than the next best thing. You are right that this figure is very different (cough, 10x!!) to yours, but it is a demonstration of what magnitudes more of a scanning camera would look like- simple movement, semi standardised connectors, fast shutter speeds, big pixels but still big resolution.

I would more realistically target £350, with the two cameras not really going below about £80-120 and then all the other bits being where money is saved.

Still alot more money than yours, and any decent speed improvements to yours would really hurt this option bad, seeing as the high readout speed of these cameras will always be wasted to some extent by what sensor velocities are practical/feasible.

The kinetic energy figures get pretty hairy pretty fast moving seemingly small objects to moderate shutter speeds done only by cloth or rubber. Amending this would assume you'd be comfortable stripping parts off the camera. Some models seem to have big heatsinks, yet others are single board computers with no apparent extra cooling (check the last three images).

.

.

The attraction was to have 1-2 sensors moving in one simple direction, but I would like to target something closer to your price point. To do this I would instead focus on the high throughput of line sensor cameras, and use just one to scan across an image plane larger than its diameter, in some pattern.

However, as you might imagine the discussion reverts to what you were already talking about, which I need more knowledge of.

You've already elaborated on many benefits your existing 3D printed area sensor design already has, and thought about the speed penalty, but from what I can see your main issue with line cameras these was their cost. If we assume for a minute that cheap options do exist (and remembering we're getting a big speed boost for maybe just a little more cost), would you pursue them, and how would it change the design? Would the potential for error be too much for your tastes?

The 3D-printed parts of Lafodis160 are already pretty sneaky and advanced materials in that no part is less than 75% air (max 25% internal structure)... but there are other methods that are competitive. Just as a reminder, Lafodis160 is only 877g including lens! Put another way, the whole thing is about 40% the weight of the Nikon Z Noct lens alone.

Useful numbers- I shall keep these in mind. I would probably begin with an aim to equal the weight of a Noct, or a typical 6x7 SLR with prism + grip + back, mainly because instead of limited TTL metering, loud shutter, film flatness issues and rareness, you're getting 4x5, big resolution, quick exposure images. Yes, the weight penalty is what alot of modern 4x5 camera designs are trying to minimise, but I'm saying that I should be able to start from what people already put up with for smaller negatives, and give them new benefits that might make it worth it.

Also, if its possible to make the sensor cooled with more lightweight components suitable for physical movement, or maybe cool it before/after exposure, then I feel a little more at ease. Static computer or power components are less of an unknown than the practical limit on how fast the sensor can move, governed by what cooling it requires and when that might be required.

You have this oddly backwards. Line scanners don't bend, and the long, narrow shape makes them have more error in following a curved focus plane.

That's correct- in my implementation the lens would be fixed in place, and the single line sensor would move along the path of the lens' field curvature.

But the lens fields curve in two dimensions simultaneously -- there is no straight path.

D'oh! Someone didn't have their Weetabix. You're right. Oh well..

A 1/4" sensor doesn't span a large enough area to have any significant curvature within a single capture -- you don't even need to tilt it to align well. The long dimension of the OV2640 is 3590um! Even tilting a line sensor, you'll have much more error along it's length.

Hah, that's very interesting. And you're right, sorry I didn't understand the first time around. Yes, one long line sensor whos length = image diameter risks tilt on certain axis.

Don't think telescopes. A little curvature of field is part of the charm of large-format photography... so the real issue is if correcting is improving the image or making it look more like what you'd get from stitching with a smaller camera re-aimed. Continuity of bokeh is one of the best reasons for large format, and small degrees of curvature cause subtle, continuous, deformation of bokeh over a large region of the coverage.

I shall endeavour to remember this. One of the main attractions for me when it comes to digital large format is the big pixels, low noise. I really want to be blown away by this, and line scan cameras do above all else offer some good options here. For example, the 12k Dalsa Piranha 2 across its 56mm diameter has a pixel pitch/size just smaller than the GFX 50S, at 5.2um. So already starting from a pretty good place. Some might go for the 8k models at the same diameter for 7um. Two together would still give handsome resolution but killer low noise.

(There are more reasons than just pixel pitch for the GFX's low noise abilities, but the Dalsa isn't any less of a camera..)

Power supplies, cabling, etc. all adds up. USB 3 isn't a cheap microcontroller interface... but I suppose if we're assuming you can plug the camera into a high-end laptop....

Alot of the off the shelf framegrabbers I'm seeing from cursory research are as expensive as the camera I'm using them for, and PCIe. This is not my expertise, but is one part I'd like to bring the cost down.

I would be interested to explore if there's any room for using less capable/powerful frame grabbers, because I'm not sure I'd ever reach the maximium throughput that these devices are capable of (nor could physically move them fast enough across the frame). Therefore, could you cap it at a much lower level, and ditch the need for a high end laptop or portable computer. I'm not even sure that might work with the data connectors used, but you can see where I'm going.

ProfHankD wrote:

Graham Gibson wrote:

Very cool project!

Thanks. It was sort-of "done" when the paper went to Electronic Imaging, but it seems a lot more broadly useful, so that's how it's evolving now.

Did you have issues with the micro lens array in the sensor being optimized to the angle of incidence, therefore causing problems trying to stitch the images together?

Lafodis160 is still being tweaked to be easier for others to build and use, and I'm actually going to have two undergrad RAs working on cleaning-up the software for it this Summer. There are a bunch of tweaks; the latest one is I found it was worthwhile to buy actual NIR-blocking filters... that cost less than $1.

There is a problem getting raws from the camera. Fortunately, it can deliver JPEGs with a "raw" tone mapping, but there is some darkening off-axis which hurts DR. I don't think it's from microlenses, because it has the same issue using the built-in lens, but maybe? We'll just model the lower SNR as lower confidence in the values for stitching.

I read about another photographer attempting a similar project with the Raspberry Pi HQ cam in a 4x5 back. Building a camera like this into a large-format back would also help in composing the image.

Not happening. The first prototype was a more traditional 4x5 X,Y, and it's just way sloppier and harder to build. I'm pretty well sold on the angle, radius scan control.

I have considered having a second ESP32-CAM mounted with the default lens as an approximate EVF, but that will probably be an optional add-on. There also will probably be a fast framing scan mode.

I have seriously considered moving a bigger sensor around, even an EOS-M programmed using Magic Lantern, but that's just not as easy nor effective as one might hope.

If you attached the tripod to the body of the camera and could remove the back of the camera, then you could replace the back of the camera with a screen that you could use to compose. A 4x6 piece of glass for a photo frame that was scratched with very fine grit sandpaper might work for a cheap ground glass composition screen.