Camera System for High-Res Scientific Photography

<p>In addition to being a hack hobbyist photographer, I also happen to be a computer vision (image analysis) researcher. I'm looking at setting up a new research project that has some funny parameters, and I could use some advice on equipment purchasing. I am debating between a Nikon D800E system or a digital MF systems. Budget is around $20k, but less is always better! I could also rent the equipment.<br>

<br />Basically, the goal is to resolve objects of 0.25mm in size from as far away as possible using a modest telephoto lens (ie I am not shooting macro or super-telephoto). I can control the lighting, although strobe work is possible. I'd prefer a system with some field portability, but most work would be done in the lab. Dynamic range is a complete non-issue.</p>

<p>What I don't have a good feel for is the quality of the lenses available in MF systems. Lens performance is going to play a huge role in the results, so high quality, highly corrected lenses are essential. <br>

Any guidance on suitable systems would be appreciated!</p>

<p>Hi, a couple of questions for you: 1) is the distance to the objects consistent (both in a plane and not moving), and 2) are you going to try to measure distances between objects?</p>

<p>Questions aside, here are a few thoughts. The D800E seems to have a pixel pitch of about 0.0048 mm, so with a 1:1 magnification, your 0.25mm (diameter) object would cover about 50 pixels (diameter). Obviously this is way more than you need. If you go for about 1/10 of that magnification, the object should still span a diameter of about 5 pixels, which still seems to be plenty. I don't think you'd want to go too much smaller than that, so I think 1/10 magnification would a good starting point.</p>

<p>Using a mag ~ 1/10 would seem that there is no need for a true macro lens, and the limiting spot size due to diffraction will probably not be an issue at any aperture you might want. So I would not expect to see any problems with any competent lens, assuming you can get the working distance you want. If you are planning to do critical measurements, slight focus errors can change your measurements. In this case, you might be better off to look for a more specialized setup using a telecentric lens.</p>

<p>These are along the line of back-of-the-envelope calculations, but might be a good starting point.</p>

 

Regarding pixel density: remember that 2x2 'pixels' are pooled to produce a pixel, so that 'pixel' pitch figures have to be doubled, resolution halved.

<p>Objects will be planar and not moving. There is some likelihood of trying to measure distances between objects, as well as the size of the objects themselves. <br>

<br />I hadn't thought about it from a pixel pitch - lens magnification point of view, although that is very helpful. I am well aware of focus errors, and that is my biggest concern with the experiments. I am considering using calibration targets as a focus aid.</p>

<blockquote>

<p>I am well aware of focus errors, and that is my biggest concern with the experiments.</p>

</blockquote>

<p>You're probably aware of the possibility of using magnified "live view" for focusing, but it needs to be said, just in case. This is probably about the best focus you can get, and with the objects all in a fixed plane, no need for a telecentric lens.</p>

<p>If you go for the ~1/10 magnification, your working distance, lens to object, should be roughly 11 focal lengths, ie, 250mm fl should give roughly 9 feet in front of the lens (better double check my math, though). If it were me, I'd do a quick verification with whatever gear you already have, just to make sure things are in the ballpark. If you don't have any "test objects," you could just photograph a ruled scale and see if some arbitrary dimension covers the right number of pixels.</p>

<p>Hope the project works out.</p>

<p><em>[R]emember that 2x2 'pixels' are pooled to produce a pixel, so that 'pixel' pitch figures have to be doubled, resolution halved.</em></p>

<p>Q. G., I know you know a lot of stuff, so I figure I'm misunderstanding your statement, or else profoundly ignorant of how medium format digital backs work--but at first blush, to me that statement seems quite wrong. In a typical sensor, a Bayer pattern of red, green, and blue filters cover the pixels / photosites. So if the sensor is, say, 40 MP, there are 10 MP of photosites covered with red filters, 20 MP photosites covered with green filters (because green contributes more to our perceived fine detail), and 10 MP photosites covered with blue filters. The raw file contains (basically) the result of the analog-to-digital converter figures for each photosite. The raw conversion software (or camera's firmware, for in-camera JPEGs) performs a "de-mosaic" algorithm on the 10 MP of red pixels, 20 MP of green pixels, and 10 MP of blue pixels, knowing which filters cover which pixels, to produce a full-color image of 40 MP--not 10 MP. So any given pixel might have, e.g., a red filter, but the de-mosaic algorithm uses information from adjacent pixels with green and blue filters to estimate the green and blue values for that pixel. The result is that the <em>color</em> resolution of the "40 MP" image is <em>way</em> less than 40 MP, <em>but</em> the black-and-white resolution should be somewhere approaching 40 MP (subject to Nyquist issues) where the sensor does not have a resolution-lowering anti-alias filter. (Almost all medium format digital backs, and an increasing number of smaller-sensor cameras like the Nikon D800E and the Sony A7r, omit the anti-alias filters generally found on common digital cameras.)</p>

<p>So it would seem that where the ability to differentiate / resolve in black and white would suffice, a sensor without an anti-alias filter (if given a good-enough lens, technique, etc.) ought to be able to resolve something approaching its nominal pixel count (although if the need is to differentiate, e.g., red from blue, the effective resolution will be considerably less).</p>

<p>No? What am I missing? Thanks.</p>

 

No time for a long explanation now, but consider image detail, black and white lines, just small enough and the right frequency for the lines to fall in adjacent wells. Then the image resolution of the 'unmosaiced' sensor is exactly the same as the 'pixel' pitch. But as long as two by two wells are combined to produce a pixel (whether in non-overlapping adjacent arrays or overlapping arrays shifted by just one well) you always get pixels that combine a black and a white line. You can imagine what that does to resolving power.

I'm back at the computer again, so on it goes. When the lines are black and white, you could just ignore the colour mosaic, do not demosaic the image but just get the value from each individual well to create a pixel. But what if a line is red, blue or green? The filter over the well will determines whether it is seen or not. So unless you remove the mosaic (as in a back like PhaseOne's IQ260 Achromatic) there are issues.<br>And if you do, and though the output measured in megapixels may then be still the same as that of backs that construct pixels out of 2x2 arrays, resolving power is not.

Just because the lines are, e.g., red and blue does not mean they aren't resolved. The result the Bayer sensor gives you

is almost like, for illustrative purposes, old black and white photo prints where color was added by watercoloring over the

print. The level of overall detail is pretty high, but color information is comparatively low resolution. In any event, Bayer

sensors don't normally combine four RGBG photo sites to make one color pixel, and even their color resolutions are

considerably better than they would be if that were what they did. I suggest anyone with serious interest in this issue look

at the tests of various Sigma cameras with Foveon sensors (which of course don't user the Bayer mosaic) that DPReview

and </I>Popular Photography</I> magazine (USA) have published--they provide useful information and examples of how

well Bayer sensors do and don't handle fine color detail.

A red line hitting a blue filter will do not much inside the well hidden below the filter. A blue line hitting a red filter too will remain invisible. Etc. 'In between colours' do a little better, but still are problematic.<br>Bayer mosaic equiped sensors do combine arrays of wells to gather the information that is put out as a colour pixel, which indeed means the resolution of the colour image is lower than the 'pixel density'. Foveon's three layer sensor promised to do better, collecting all colour information in on site. I don't know how they do today, but know that that came with a fresh set of problems, the most pressing being the well depth and the angle of incident light and they need a micro lens array. But they (still) sound promising.

<p>Dave, I don't quite agree with the way Q.G. stated things either [edit - prior to the last post], but I wouldn't quarrel that in this case, the objects to be detected should cover at least a 2x2 area. At least in the case of a Bayer-filter sensor.</p>

<p>The possibility exists that an unknown object to be imaged has a color that can only be detected by one of the photosites in the 2x2 grid. So the possibility exists that the object will be missed unless its image is large enough to cover the entire 2x2 grid. There are other ways to argue this, too, but I think this single possible situation is enough to make the case.</p>

<p>I suggested using an even larger image-spot size partially because of the additional possibility of "bad" pixels (photosites) which have been "mapped out" in the camera. One would not be able to see this in a pictorial scene because interpolated data would be used to fill in the "bad" area. It's sort of like the "blind spot" in the human eye - our eye/brain "fills in" the picture so well that most people, on their own, would probably not be able to find the spot.</p>

<p>At any rate, I imagine that the original poster, as a computer-vision researcher, probably has a decent handle on this sort of thing.</p>

Bill, my math works out the same

by working back from a FOV

calculation.

 

Would I gain measurable resolving

power with a medium format setup?

Presumably less sensor noise and

fewer lens issues...

<blockquote>

<p>Would I gain measurable resolving power with a medium format setup? Presumably less sensor noise and fewer lens issues...</p>

</blockquote>

<p>Hi, I'm not the best one to try to answer this. In principle, it seems like larger sensing area is going to be better, but one would expect the smaller 35mm-film-sized cameras to have much larger sales volume, and thus better-developed technology. But I'm just guessing, and really don't know the true state of development.</p>

<p>Something else you might consider, if you don't need color, is a monochrome sensor. Again, I don't know the current market, but I think there may be a Leica digital that has it, perhaps some digital backs, and probably a variety of machine-vision cameras (these are sort of unknown to the general photography world).</p>

<p>I'd probably do my initial camera research on dpreview or similar. Again, best wishes on the project.</p>

 

<p>Dave, just get a D800E. If you can't do the job with that, then I doubt there's anything else out there that will do it. Ignore the previous toing-and-froing about Bayer patterns (stupidest idea in Christendom - but that's just my opinion), because all commercially available cameras use it. That's unless you want a purely B+W image, in which case there are unfiltered options. Nikon would probably de-filter a D800 for you at a price.</p>

<p>Anyway, the D800E has one of the best sensors out there, and also the Nikon body and F mount gives you a very wide range of lens options. Although I strongly suspect that most of that $20K will need to be spent on specialist optics, and maybe on building a concrete plinth to hold the camera absolutely steady. The D800E is genuinely capable of resolving 100 lppmm. But when you say "resolving objects of 0.25mm in size", I take it you don't mean resolving 0.25mm wide black and white lines.</p>

<p>FWIW, I've been able to spot flies and bees resting on a chimney pot at about 30 metres distance when doing some lens tests on my D800. Couldn't see them at all through the viewfinder until the images were viewed 100% on the computer.</p>

<p>As for the lens you need, it'll obviously need to be an Apo spec lens, but the focal length will need to be worked out according to the magnification you need, and probably custom built.</p>

<p>Edit: The high(er) pixel numbers of medium format backs are purely a function of their greater area. If you look at pixel density, which is what fundamentally dictates the resolution of a sensor, then the D800(E) is about as good as it gets. In fact their are DX sensor cameras that have just as high a pixel-density. So it just depends how wide your field-of-view needs to be. Astrophotographers do great thing with tiny little VGA and XVGA sensors stuck on the end of their telescopes. Or maybe you could butcher a Nokia phone for the stupidly high pixel-density sensor that's wasted inside it behind a crappy lens.</p>

<p><em>Bayer patterns ... all commercially available cameras use it.</em></p>

<p>No. Fuji has several cameras with X-Trans-filtered sensors. Sigma has several cameras that use Foveon sensors. And of course there are a few monochrome-only models (Leica, others?). Some new Android devices may be appearing with RGBC (the C being clear) sensors. In the past, Fuji sold cameras with "EXR" filter arrays, and Sony has sold cameras with RGBE (E = emerald, but closer to cyan) pattern filter arrays on its sensors (remember the F828?).</p>

<p>You can debate the merits of different color filter arrays, or color filter arrays versus Foveon-type sensors, but if you want color, you have to accept one of them, and all of them have their downsides.</p>

<p>I realize that <em>medium format</em> color means a Bayer filter array. But insofar as a D800E is being considered and compared, I see no real reason to exclude these other cameras.</p>

 

<p>Dave L., getting clean images of 0.25mm-size objects from a distance for a scientific project requires... scientifically high-quality lenses that only Leica or Carl Zeiss can provide. Second-hand Leica S and Hasselblad V System+digital back are marginally within your budget.</p>

<p>However, if you are to do serious image processing (BTW, what software do you use, MATLAB-level tools or go deeper with your own algorithms?), you are fully aware you will need quite an area of surrounding pixels per pixel on you image for differentiation and other filtering. So, you may be facing with a conflict between many neighboring pixels and pixel size. This, apparently, indicates medium-format systems.</p>

<p>Paul</p>

<blockquote>

<p>scientifically high-quality lenses that only Leica or Carl Zeiss can provide</p>

</blockquote>

<p>Dear me, no, not "only" them. Any Mamiya APO lens (645 or RB67/RZ67) will be in the same class. For example: check out the MTF diagram for the 350/5.6 APO for the RB/RZ.</p>

<p>The new state-of-the-art high resolution Zeiss Tuit 50mm lens only covers the 36x24mm full-frame format and is only available in a limited range of DSLR mounts. And while Leitz and Zeiss lenses might have cut it with film, today's digital sensors are capable of capturing a higher image quality. The shorter register distance of a DSLR also allows a far wider selection of lenses to be used or adapted than does a medium format body.</p>

<p>You also have to take account of the fact that lens aberrations scale with focal length. Meaning that the longer focal length required to cover a wider image circle (for the same field-of-view) brings with it an automatic decline in resolution, or requires a more complex optical formulation to maintain resolution. You only have to look at the phenomenal resolution figures required of a lens to work with the average digital compact camera (mostly diffraction-limited by around f/5.6) to realise that this is the case, and that shorter focal length lenses of high resolution are more economic to produce than those for larger formats.</p>

<p>" So, you may be facing with a conflict between many neighboring pixels and pixel size. This, apparently, indicates medium-format systems." - How so? That's a totally non-sequiter argument and takes no account of pixel density or the field-of-view needed. Almost any camera capable of RAW capture could provide sufficient data for the amount of image processing needed.</p>

<p>Dave, Sigma's Foveon sensor has a pathetic pixel count and density by comparison with almost any "conventional" sensor. That surely puts it out of the running for any high resolution work? Remember Sigma "cheat" on their pixel count by multiplying the number of sensor sites by 3. By comparison the single sensor-site step interpolation of a Bayer pattern gives a higher resolution in the real world. IMHO, 'twould be better if we used a geometry that didn't waste a quarter of the light gathered in excess green sensitivity, but hey-ho, stupid ideas somehow get entrenched as "the" way to do things.</p>

<p>Anyway, since pictures speak louder than words; below is a sample of what a D800 (not D800E) can do when fitted with a relatively inexpensive lens. In this case it was a Tamron 90mm f/2.8 SP Macro @ f/8. I found this old dusty doll and thought it looked a bit like an Egyptian mummy, so was just playing around with background and lighting. The small inset shows the full frame, while the 100% crop shows the <em>dust</em> fibres and button on the doll. Note, the threads of the cloth are about 0.25mm wide. The tiny dust hairs are far thinner - maybe around the 10 to 20 micron diameter mark.</p><div></div>

<p><em>The ... Zeiss Tuit 50mm lens only covers the 36x24mm full-frame format and is only available in a limited range of DSLR mounts.</em></p>

<p>The Zeiss Touit lenses are only available for mirrorless cameras, not DSLRs, and only with "APS-C" (15.6 x 23.5 mm) sensors, not "full-frame" (24 x 36 mm). So if you want to use these lenses, your camera has to be either a Sony E-mount mirrorless or a Fuji X-mount mirrorless. Per the B&H website, you can't even use these on the Sony FE-mount full-frame mirrorless cameras.</p>

<p>As for Sigma and the Foveon, my point was, if you want color, there are alternatives--although all of the alternatives have issues. Nevertheless, I agree that the Sigma marketing is deceptive. The newest Sigma models have what are actually 15 MP Foveon sensors, although the Foveon's design doesn't really benefit from an anti-alias filter, so its effective B&W resolution is more like a conventional (Bayer-sensor with anti-alias filter) camera of about 24 MP, and its color resolution is probably class-leading.</p>

 

<blockquote>

<p>Dear me, no, not "only" them. Any Mamiya APO lens (645 or RB67/RZ67) will be in the same class. For example: check out the MTF diagram for the 350/5.6 APO for the RB/RZ.</p>

</blockquote>

<p>Ray,</p>

<p>Mamiya lenses are excellent pieces, as are lenses of other known brands. I was trying, however, to respond to the original poster's remark that the system was for scientific use. As far as I can recall from my concern when doing my lab-based research, apparatuses should operate within strict tolerance limits--the fight was for just an extra decimal point, to put it that way. Both Leica and Zeiss have an extensive experience on scientific instrument manufacture.</p>

<p>And, speaking of Mamiya, it reminds me the comments flowing in the Internet for a long time on comparing Mamiya 43mm f/4.5 to SWC's Biogon. The comparisons never mention cleaning-room, assembly and testing tolerances. It's not only MTF, is it?</p>

<p>While I will gladly read your comments on this, please allow me not to further respond, due respecting the original post's content and goal.</p>

<p>Paul</p>