Mitutoyo M Plan APO HR 5X 0.21 Test Results

[rjlittlefield]

couldn't the aliasing be greatly diminished using pixel-shift mode in cameras that supported that function, although that would admitedly also take considerably more time and memory?

If you're talking about half-pixel shift then yes, greatly diminished aliasing. That sort of shifting does not change the MTF curves as computed here, but it does double the sampling frequency so it doubles the Nyquist limit. It also reduces sampling variability by putting 4 samples per cycle at what used to be the Nyquist limit of 2 samples per cycle.

Here is a graph of sensor MTF and overall system MTF when an FF 24 MPix sensor is paired with a 5X NA 0.21 objective, with and without half-pixel shifting:

If you're talking about full-pixel shifting, then that effect is a lot more subtle. In that case the shifting gets rid of Bayer filter effects by actually sampling RGB per pixel, but it does not change the Nyquist limit of 2 pixels per cycle so you can still get intensity aliasing in the pink zone.

--Rik

[Macro_Cosmos]

This graph incorporates the combined effects of diffraction and finite pixel size. What is shown is system MTF = lens MTF * sensor MTF.

Lens MTF is computed from diffraction only, using the equations quoted at viewtopic.php?p=124831#124831 . Sensor MTF assumes perfect pixels (box filter width = pixel width) and is computed as a sinc function as described at http://mtfmapper.blogspot.com/2012/05/p ... d-mtf.html and shown in the 5th chart there. Lambda = 0.00055 mm.

Bottom line, the NA 0.21 image will look sharper and will reveal finer detail, unless you post-process the images so as to compensate for the differences in MTF.

In this particular case, a strong sharpening filter applied to the NA 0.14 image could bring its overall MTF up to match the NA 0.21. But that would come at the cost of also increasing fine scale noise by almost 4X.

The one downside to the NA 0.21 objective is that it provides a lot more opportunity for aliasing, if the subject contains fine detail that resolves above the Nyquist limit of the sensor. The beauty of Nikon's rule is that it guarantees no aliasing, because it matches the cutoff frequency of the objective with the Nyquist limit of the sensor. Unfortunately, for users who don't understand what's going on, that rule can also result in "recommending" an objective with small NA, which will produce an inferior image.

--Rik

Great links there, I'll have a closer look. Sinc functions bring back all the computational nightmares from university, FT and convolution... what's in my mind? The Fourier Transform of a square wave results in that, it probably does have something to do with it, vaguely from memory. I'll read the articles.

I didn't consider MTF at all and never bothered to incorporate it into my calculator, perhaps I should now. It'll require a lot of work in Excel, a program I don't exactly like working with -- too bad my SAS subscription from the university days is gone.

By "fine scale noise", do you mean noise factors such as read-out noise and photon shot noise? So basically, all the noise factors are amplified. As a standard practice, I will decrease the "radius" slider to minimum and apply maximum sharpening and some noise reduction. Sort of similar to "deconvolution", they do make a huge difference especially when down-sampled.

I would generally recommend the cheaper option considering the price of one 5x HR gets the user an array of objectives. It's equivalent to a 5x, 10x, 20x and perhaps even 50x and 7.5x if one's patient. The cheapest 5x HR Mitutoyo I've seen so far is $2500. Below are some typical, lower bound used prices. The 50x fluctuates a lot, sometimes they go for as high as $1200. Must be patient with this one!
- 5x: $450 (as low as $250)
- 10x: $450 (as low as $400)
- 20x: $600 (as low as $550)
- 50x: $700 (as low as $600)
- Carman Haas clone TTL200-A tube lens: $180
At the cost of the cheapest 5x HR I've ever seen online.

I'm not the one to say "the HR is useless because your 24MP sensor wouldn't resolve it" when clearly my tests and far more precise calculations here prove otherwise. The funny thing is, my D810's 36MP shows a far smaller difference, it's worse in some instances compared to the BSI 24MP sensor on the Z6, at ISO100 and with a lowpass filter (vs. 64 and without). I'll attribute this to generational differences and "raw baking" which is yet two more factors in the optical system.

[rjlittlefield]

By "fine scale noise", do you mean noise factors such as read-out noise and photon shot noise? So basically, all the noise factors are amplified.

Right, all that stuff. Think of the observed pixel values as being noise-free signal plus noise. Amplifying the 3000 cycles/FOV signal component by 4X, to compensate for the difference in MTF, means that you end up multiplying the 3000 cycles/FOV noise component by the same 4X.

--Rik

[dmillard]

If you're talking about half-pixel shift then yes, greatly diminished aliasing. That sort of shifting does not change the MTF curves as computed here, but it does double the sampling frequency so it doubles the Nyquist limit. It also reduces sampling variability by putting 4 samples per cycle at what used to be the Nyquist limit of 2 samples per cycle.

If you're talking about full-pixel shifting, then that effect is a lot more subtle. In that case the shifting gets rid of Bayer filter effects by actually sampling RGB per pixel, but it does not change the Nyquist limit of 2 pixels per cycle so you can still get intensity aliasing in the pink zone.

--Rik

Thanks Rik, this clarifies the effects of pixel-shifting very concisely.

Best regards,
David

[chris_ma]

If you're talking about full-pixel shifting, then that effect is a lot more subtle. In that case the shifting gets rid of Bayer filter effects by actually sampling RGB per pixel, but it does not change the Nyquist limit of 2 pixels per cycle so you can still get intensity aliasing in the pink zone

I would argue that on certain subjects the difference can be a lot more then subtle.
here a comparion of the pentax K1-II on single shot an 4 shot pixelshift, have a look at the false detail of the jackets:
https://www.dpreview.com/reviews/image- ... 1108153941

[rjlittlefield]

I would argue that on certain subjects the difference can be a lot more then subtle.
here a comparion of the pentax K1-II on single shot an 4 shot pixelshift, have a look at the false detail of the jackets:
https://www.dpreview.com/reviews/image- ... 1108153941

Hah! What a wonderful comparison!

Yes, I had forgotten the Fundamental Principle of Data Munging:

"Improvements are limited, degradations are not"

(Google says "No results found", so I'm claiming this for myself.)

In the subject image, the actual thing being imaged has a pattern of lines or dots that line up with pixel positions so as to totally confuse the raw conversion software. Instead of deciding that this area is monochrome with a fine checkerboard pattern of brightness, the conversion software has decided that it must be a set of rich colors organized in waves. See here, at 1600 percent from the DNG's:

It should be clear that this problem can be arbitrarily bad. It would be evil, but not difficult, to construct a checkerboard pattern that would render through the imaging pipeline as a frame full of uniform color, with no intensity variation at all.

So I agree, artifacts from Bayer processing can become arbitrarily bad, and those artifacts are prevented by full-pixel-shift shooting. Going farther, I would speculate that full-pixel-shift shooting probably makes actual performance be more like the model, where traditional Bayer processing is somewhat worse at best.

--Rik

[Duke]

Raw images look particulary bad because PPG demosaicing is used on the site.
PPG algorithm is very fast and gives the sharp image, but very prone to colour moire artifacts.
Because of that I've only used VNG4, but then few years ago I've tried AMAZE and it performs much better in terms of colour artifacts to VNG4, while also gives very sharp images. Using it ever since!
Check this out, processed with Darktable:
PPG

VNG4

AMAZE

[chris_ma]

yeah, amaze looks better, but here is how the true image looks like. check the top right and bottom left corner:

[Macrero]

yeah, amaze looks better, but here is how the true image looks like. check the top right and bottom left corner:

Well, you can't expect from a 36MP FF camera with 4X PixelShift to match the output of a 100MP MF camera (or a 240MP 8X shifted file from a 61MP FF camera), regardless of the demosaicing algorithm you use...

The pixel-shifted K-1 II file can be demosaiced a tad better, extracting some more detail, though some of it false. Example:

https://images2.imgbox.com/80/29/wjktWXeI_o.jpg

But again, that's not a fair comparison.

- Macrero

[chris_ma]

Well, you can't expect from a 36MP FF camera with 4X PixelShift to match the output of a 100MP MF camera (or a 240MP 8X shifted file from a 61MP FF camera), regardless of the demosaicing algorithm you use...

sure, the point I was trying to make was that a 36MP bayer sensor not only falls short of resolving 36MP, but in some scenarios can introduce a huge amount of false detail.
what's more, it will often be considered as real detail unless you have a clean image to compare against (because, how can you tell where real detail ends and false detail starts?).

edit: to illustrate the problem:
in the PPG sample the jacket on the top right seems to have vertial stripes, so the setup seems to just about resolve this detail.
looking at the pixel shift version, we see that they are in fact flat diagonal stripes, so not only did the bayer version not resolve that detail, it gave us false information.
ironically this problem gets more pronounced the higher the resolving power of the lens and the smaller the detail of the object.

[Duke]

sure, the point I was trying to make was that a 36MP bayer sensor not only falls short of resolving 36MP, but in some scenarios can introduce a huge amount of false detail.

That's true, nobody says otherwise - it is possible. Though the problem seems to only appear on the ~1px details arranged in periodic patterns - stripes, checkerboards, etc., and only in certain orientations as I mentioned before (p.3). Still for such artifacts to appear, there must be the most unfortunate combination of super high-resolution lens (such as pentax f/1.8 f=77mm), not-so-dense sensor (>5x5mkm pixel), object with high contrast pattern in certain orientation...chances that you will face it in the real objects photography are probably miniscule. I'm not convinced that even such majestic lens as Mitutoyo PlanApo HR 5x 0.21, as a finest example of IQ that one could reach in macro/micro- photography, would outresolve sensor nearly enough for such artifacts to appear, most of the time most of us are in deep diffraction/optical aberrations lens limited territory.
That said, looking on improvement of demosaic algorithms - it is pretty substantial, in my opinion.
However, there is still a room for further advancement - as these artifacts appear only on certain patterns, so, it seems possible to resolve the details and get rid of artifacts using more sophisticated algorithms based on machine learning and image recognition.
I'm fairly certain, such algorithms will be introduced in not so distant future - it is literally tomorrow of bayer matrix demosaicing.

[chris_ma]

Though the problem seems to only appear on the ~1px details arranged in periodic patterns - stripes, checkerboards, etc., and only in certain orientations as I mentioned before (p.3). Still for such artifacts to appear, there must be the most unfortunate combination of super high-resolution lens (such as pentax f/1.8 f=77mm), not-so-dense sensor (>5x5mkm pixel), object with high contrast pattern in certain orientation...chances that you will face it in the real objects photography are probably miniscule.

I get this problem quite frequently on a 47MP sensor with really good lenses on finest patterns. they don't have to be periodic, it also happens on random patterns but there it's not as obvious because the original random pattern looks similar to the random pattern generated by the debayer artifacts.

I agree though that it's more a problem in scientific uses than in pictoral photography where we have to be unlucky that it becomes a significant problem.

[Macro_Cosmos]

sure, the point I was trying to make was that a 36MP bayer sensor not only falls short of resolving 36MP, but in some scenarios can introduce a huge amount of false detail.

That's true, nobody says otherwise - it is possible. Though the problem seems to only appear on the ~1px details arranged in periodic patterns - stripes, checkerboards, etc., and only in certain orientations as I mentioned before (p.3). Still for such artifacts to appear, there must be the most unfortunate combination of super high-resolution lens (such as pentax f/1.8 f=77mm), not-so-dense sensor (>5x5mkm pixel), object with high contrast pattern in certain orientation...chances that you will face it in the real objects photography are probably miniscule. I'm not convinced that even such majestic lens as Mitutoyo PlanApo HR 5x 0.21, as a finest example of IQ that one could reach in macro/micro- photography, would outresolve sensor nearly enough for such artifacts to appear, most of the time most of us are in deep diffraction/optical aberrations lens limited territory.
That said, looking on improvement of demosaic algorithms - it is pretty substantial, in my opinion.

viewtopic.php?f=25&t=41908
Could this be an example of said phenomenon?
It still puzzles me, I've never found a satisfactory answer.

[Duke]

viewtopic.php?f=25&t=41908
Could this be an example of said phenomenon?
It still puzzles me, I've never found a satisfactory answer.

Looking at these pictures now (on the better monitor) I can see that D810 shots also have the same waviness - of exactly the same period and magnitude, just much less pronounced, but lines of a target do oscillate about half a pixel in the latitude direction. So this is definantly a property of the object and not the issue with the sensor.
FSI D810 sensor has much longer light path, filled with reflecting objects, such as copper wiring, which not only scatter the light, but also act as conductive zones allowing for electrons to bounce around with the grater magnitude, also less NA of microlens - overall all this combined results in much more crosstalk between the pixels of the sensor, what essentially acts as an anti-aliasing (opto-electical low pass) filter - it blurs the image, subsequently "smoothing" the lines. Having more pixels also helps.
This is the moire artifact on the BSI sensor of Z6, yes, but on the other hand it does a great job of showing the imperfection of the grating, what speaks in favor of resolving ability (MTF) of the sensor by itself.

[Macro_Cosmos]

Looking at these pictures now (on the better monitor) I can see that D810 shots also have the same waviness - of exactly the same period and magnitude, just much less pronounced, but lines of a target do oscillate about half a pixel in the latitude direction. So this is definantly a property of the object and not the issue with the sensor.
FSI D810 sensor has much longer light path, filled with reflecting objects, such as copper wiring, which not only scatter the light, but also act as conductive zones allowing for electrons to bounce around with the grater magnitude, also less NA of microlens - overall all this combined results in much more crosstalk between the pixels of the sensor, what essentially acts as an anti-aliasing (opto-electical low pass) filter - it blurs the image, subsequently "smoothing" the lines. Having more pixels also helps.
This is the moire artifact on the BSI sensor of Z6, yes, but on the other hand it does a great job of showing the imperfection of the grating, what speaks in favor of resolving ability (MTF) of the sensor by itself.

I see, crosstalk. That makes sense considering 400lp/mm for a 5x objective is really challenging for the lens and the sensor.
It's also an illustration of how megapixels isn't everything, as 36MP is being beaten by 24MP with a better structure, and of course more modern.
The only downside of BSI I can think of is slightly worse read noise and dark current due to the readout structure and back-thinning typically used for Gen IIII sCMOS. I personally think it's negligible.