16MP FHD Camera vs Canon HRT2i

[ray_parkhurst]

A while back, there was an extended discussion involving forum member Justwalking and the use of sensors with small pixels to improve depth of field. I was interested in his results, and the concept of shooting at lower magnification and higher pixel density, seeing the potential for using the method for lens evaluation. I supposed shooting with higher pixel density might be useful to measure lens quality, in a similar way as doing SR, or pixel shifting, or shooting with a teleconverter or objective to magnify the image.

Over the years I've tried several USB C-mount cameras, but all gave very poor results. In another forum devoted to coin photography, Justwalking posted a very nice looking image of a coin taken at an angle, with good IQ and DOF. I asked him about the camera he was using, and we were able to determine that one of the "Full HD 16MP Industry" cameras for sale on eBay was similar to the one he used. I purchased one of these, and it took over a month to get to me from a Chinese seller. I had low expectations, and indeed the workflow for these cameras is distasteful to me in that the only way to get the 16MP image is to view over HDMI (with USB unplugged), snap the shot with the handheld wireless controller, and save it to a MicroSD card, but in the interest of science I was willing to hold my nose to this workflow.

The other major limitation I've found with all USB C-mount cameras is the very poor tethering software. None that I've used have decent WB, exposure, saturation, or sharpness setting capability. This camera also has poor capability in software control over USB, and limitation to 1080P output file size. But unplugging from the USB, and using the remote control, is a completely different experience. The on-screen menus are much more capable than the software, giving good control over WB, exposure, sharpness, and saturation parameters. Snapping the shot with the controller resulted in a real 16MP (4608x3456) image, unlike other high-MP C-mounts I've tested, which gave "high resolution" through digital upscaling.

My first test images were simply to compare the output image quality of the 16MP FHD camera with my 18MP HRT2i. My subject is of course a Lincoln Cent, with the date and MM centered in both sensors. Magnification is ~1:1, which over-fills the HRT2i sensor with the 19mm dia Cent. Lens is the 85MV at f4.5. For the FHD camera, which has a much smaller sensor (6.2mm vs 22.3mm), the image "appears" to have much higher magnification, but it's really only looking at a small crop area at the center. The smaller pixel size of 1.34um vs 4.3um makes the output image size about the same even with the small crop area, so the result is a significantly increased sample density for the FHD.

Here is a comparison of the overall images from the FHD and HRT2i. Ignore the WB differences, as I failed to properly MWB the HRT2i:

HRT2i


16MP FHD


Cropping just around the date and mintmark on both, and then upscaling the HRT2i and downscaling the FHD to match, gives the following two images to compare:

HRT2i


16MP FHD


Honestly, I was shocked with the comparison of these two images. At f4.5, the lens is operating at f9 effective, which is beyond the DLA of the HRT2i (f6.7) and far beyond the DLA of the 16MP FHD (f2.1). I expected the 16MP FHD image to be quite blurry, but I see a very sharp and natural looking image with the FHD, and a very unflattering image from the HRT2i!!

Here is a 100% crop from the FHD image:



It's abundantly clear that in this case the 85MV, even at f4.5, has far more resolution than can be captured by an 18MP sensor with 4.3um pixels.

I think what I have done is a perfectly valid comparison and explanation, but do let me know if I've made any mistakes. If not, then I have a lot to think about regarding my future evaluations and imaging, and my thoughts of future camera purchases.

[rjlittlefield]

It's abundantly clear that in this case the 85MV, even at f4.5, has far more resolution than can be captured by an 18MP sensor with 4.3um pixels.

I agree. Let's run the numbers to see why.

Calculating for 550 nm green light and effective f/9, the cutoff frequency nu_0 is 1/(0.55*9) = 0.202 cycles/micron, or 4.95 microns per cycle. Minimum Nyquist sampling is then 2.475 microns per pixel. So, yep, even without worrying about Bayer, your 4.3 micron pixels are not nearly fine enough.

Meanwhile the FHD sensor at 1.34 microns per pixel, again ignoring Bayer, is giving about 3.7 pixels per cycle. That's enough to give quite good sampling.

In fact it should be almost double the minimum sampling necessary to capture all the detail that's present in the optical image. We can crosscheck that assertion by shrinking the 100% crop FHD image to 50%, then doubling it to restore its original size, and flashing between the half/doubled version and the original to see how much was lost in the halving. Here's the result of that test, using Photoshop Bicubic Spline interpolation for each step, and flashing 1 second for each version.



To my eye this matches expectation in that the halved and doubled image seems quite similar to the original 100% crop. In fact it seems so similar that I opted to put one little white pixel in the lower right corner of one image so that I could tell for sure the animation was working.

So, what you really have here is roughly 4 MP of image content, captured with a 16 MP sensor.

As for why the HRT2i image looks so bad, that's a good question. But the thing that really puzzles me is why the specular highlights appear in such different places in the two images. It's like you had completely different lighting setups, not just two different cameras. I understand that you didn't match color balance and exposure, but that's not going to move the location of the highlights around. So, what's going on with the lighting?

--Rik

[ray_parkhurst]

I agree. Let's run the numbers to see why.

Calculating for 550 nm green light and effective f/9, the cutoff frequency nu_0 is 1/(0.55*9) = 0.202 cycles/micron, or 4.95 microns per cycle. Minimum Nyquist sampling is then 2.475 microns per pixel. So, yep, even without worrying about Bayer, your 4.3 micron pixels are not nearly fine enough.

Meanwhile the FHD sensor at 1.34 microns per pixel, again ignoring Bayer, is giving about 3.7 pixels per cycle. That's enough to give quite good sampling.

In fact it should be almost double the minimum sampling necessary to capture all the detail that's present in the optical image. We can crosscheck that assertion by shrinking the 100% crop FHD image to 50%, then doubling it to restore its original size, and flashing between the half/doubled version and the original to see how much was lost in the halving. Here's the result of that test, using Photoshop Bicubic Spline interpolation for each step, and flashing 1 second for each version.
....
To my eye this matches expectation in that the halved and doubled image seems quite similar to the original 100% crop. In fact it seems so similar that I opted to put one little white pixel in the lower right corner of one image so that I could tell for sure the animation was working.

So, what you really have here is roughly 4 MP of image content, captured with a 16 MP sensor.

As for why the HRT2i image looks so bad, that's a good question. But the thing that really puzzles me is why the specular highlights appear in such different places in the two images. It's like you had completely different lighting setups, not just two different cameras. I understand that you didn't match color balance and exposure, but that's not going to move the location of the highlights around. So, what's going on with the lighting

I apologize for the uncontrolled nature of those images. The cameras were rotated 180-deg vs one another, hence the lighting weirdness.

I re-shot the comparisons in exact same orientation and lighting conditions, see below. These are again at 1:1, f4.5, and are properly white balanced, though there are still big color differences.

With the consistent lighting, the comparison is much more clear.

HRT2i at 1:1


16MP FHD at 1:1


I also shot at 2:1. Here are the images:

HRT2i at 2:1


16MP FHD at 1:1


I see less difference between the cameras at the higher magnification, I assume because both are farther into diffraction limitation. Going the other way, at lower magnification, I'd expect to see greater differences, and I do plan to test this.

[Pau]

Hi Ray,
At your second set of images the final magnification is matched between both cameras so the actual magnification on sensor is not just because the difference in sensor size as you pointed at your first post, could you clarify it?

[ray_parkhurst]

Hi Ray,
At your second set of images the final magnification is matched between both cameras so the actual magnification on sensor is not just because the difference in sensor size as you pointed at your first post, could you clarify it?

Optical magnification is the same between the images being compared. All I did is to swap cameras, keeping the lens in the same position, and adjusted the camera position to re-focus.

I did re-size both images in each comparison pair. The most recent ones were initial crops of 1600x1600 for the 16MP FHD, and 500x500 for the HRT2i. This corresponds well to the ratio of sensor dimensions. I resized both pairs of images to 900x900 for publishing at same final size.

Does this answer your question Pau?

Edited to add: you questioned "magnification on sensor is not just because the difference in sensor sizes". What I meant was that both the sensor and pixel sizes are different, and both of these affect how the images appear in published form. It turns out that in the case of these two cameras, the two almost cancel each other, with the final images both being in the 16-18MP range, so full-size published images are about the same. Furthermore the heights of both sensors is 3456 pixels, so I ended up emphasizing the similarity by cropping the 5184 pixel wide HRT2i shot to match the 4608 pixel wide 16MP FHD shot. This made the final images identical in pixel dimensions for convenience.

[Pau]

Optical magnification is the same between the images being compared. All I did is to swap cameras, keeping the lens in the same position, and adjusted the camera position to re-focus.

I did re-size both images in each comparison pair. The most recent ones were initial crops of 1600x1600 for the 16MP FHD, and 500x500 for the HRT2i. This corresponds well to the ratio of sensor dimensions. I resized both pairs of images to 900x900 for publishing at same final size.

Does this answer your question Pau?

Yes, now it's clear for me, thank you
I missed the nature of the test, used as I am to just compare actual 100% pixel crops. Downsizing one and upsizing the other to compare them was not in my mind

[rjlittlefield]

With the consistent lighting, the comparison is much more clear.

I agree that this comparison is much more clear.

What it shows, to my eye, is that throwing 10 times as many pixels at the same image area does a much better job of capturing all the detail that is present in the optical image. I don't find that surprising, given the calculations that I summarized earlier.

What I do find surprising is your comment in the first post:

I expected the 16MP FHD image to be quite blurry, but I see a very sharp and natural looking image with the FHD

The reason I find this surprising is that previously you have expressed great frustration that you could never get your APS-C images to look sharp at 100%. Now, I remember running the halving/doubling test on those APS-C images. They clearly lost a lot of detail when halved and doubled, but you complained they were not sharp. Now we have this FHD image, which loses hardly anything when halved and doubled, and you write "very sharp".

At this moment, I can only assume that "sharp" means something different to you than it does to me. But I have no idea what definition you're using.

Can you clear up my confusion?

--Rik

[Perl]

Hi

Do you have and Link to 16MP FHD ?

So many diffrent models on ebay

Best Regards
Pär

[ray_parkhurst]

Hi

Do you have and Link to 16MP FHD ?

So many diffrent models on ebay

Yes, quite a number available. The one I bought was:

https://www.ebay.com/itm/16MP-1080P-60F ... 3147285682

[Perl]

Hi

Many thanks !
Best Regards¨
Pär

[ray_parkhurst]

With the consistent lighting, the comparison is much more clear.

I agree that this comparison is much more clear.

What it shows, to my eye, is that throwing 10 times as many pixels at the same image area does a much better job of capturing all the detail that is present in the optical image. I don't find that surprising, given the calculations that I summarized earlier.

What I do find surprising is your comment in the first post:

I expected the 16MP FHD image to be quite blurry, but I see a very sharp and natural looking image with the FHD

The reason I find this surprising is that previously you have expressed great frustration that you could never get your APS-C images to look sharp at 100%. Now, I remember running the halving/doubling test on those APS-C images. They clearly lost a lot of detail when halved and doubled, but you complained they were not sharp. Now we have this FHD image, which loses hardly anything when halved and doubled, and you write "very sharp".

At this moment, I can only assume that "sharp" means something different to you than it does to me. But I have no idea what definition you're using.

Can you clear up my confusion?

Sorry for the confusion on this. I suppose rather than "very sharp" I should have said "relatively sharp", ie in comparison with the HRT2i image. But in absolute terms it is also sharper than I expected given that it's operating so deep into diffraction limitation. I suppose I need to do the experiment to see the image quality from the HRT2i at similar level of diffraction blurring relative to pixel size. From equivalent image perspective it should be the same, so I guess my surprise is that the FHD image was so "relatively sharp" vs my expectation given operation at f9 with a DLA of f2.

Edited to add: some of the improved sharpness could be related to demosaicing. For the HRT2i, the image is up-scaled to match the size of the down-scaled FHD image.

[ray_parkhurst]

Something of potential interest is the difference in geometry between the two cameras. The images are scaled to nearly exactly match in the vertical FOV, yet the horizontal is significantly different. This causes some distortion betwen these two sensors. The horizontal FOV is larger for the FHD, so I assume the pixels are slightly narrower than they are tall. Or maybe the Canon pixels are slightly taller than they are wide? Is there a physical standard that manufacturers use? This is something like "sensor distortion".

A reminder...these images are identical, taken with different sensors, and then processed to match scale.

[rjlittlefield]

Sorry for the confusion on this. I suppose rather than "very sharp" I should have said "relatively sharp", ie in comparison with the HRT2i image. But in absolute terms it is also sharper than I expected given that it's operating so deep into diffraction limitation. I suppose I need to do the experiment to see the image quality from the HRT2i at similar level of diffraction blurring relative to pixel size. From equivalent image perspective it should be the same, so I guess my surprise is that the FHD image was so "relatively sharp" vs my expectation given operation at f9 with a DLA of f2.

Thanks for the clarification. Yeah, it's hard to get the right expectations by reading other people's stuff. I still have vivid memories of struggling to understand, back in 2012, how it could be that cambridgeincolour calculates DLA as f7.1-f/10.7 for 15 MP APS-C, while my experiments were clearly showing that even an f/16 optical image contained at least two more USAF elements than were captured by my Canon T1i (15 MP APS-C). The answer, as best I ever figured out, is that cambridgeincolour is interested in what aperture you need, given the sensor you have, while we're interested in what sensor you need, given the aperture you set. The questions are different, and the answers are surprisingly different also.

--Rik

[rjlittlefield]

Something of potential interest is the difference in geometry between the two cameras. The images are scaled to nearly exactly match in the vertical FOV, yet the horizontal is significantly different. This causes some distortion betwen these two sensors. The horizontal FOV is larger for the FHD, so I assume the pixels are slightly narrower than they are tall. Or maybe the Canon pixels are slightly taller than they are wide? Is there a physical standard that manufacturers use? This is something like "sensor distortion".

That's a very interesting observation. The eBay ad that you point to is explicit that "Pixel size: 1.335x1.335um". Meanwhile Canon quotes the T2i as "Image Format 22.3mm x 14.9mm (APS-C size)" and "5184 x 3456 pixels" which calculates out to be 0.004302 by 0.004311. I've always assumed that Canon's pixels are actually square and that discrepancy in the third digit was due to rounding error in the sensor size. But maybe that's not the case. Seems like the question could be clearly resolved by shooting a stage micrometer in landscape and portrait alignments and seeing where scale differences appear.

--Rik

[ray_parkhurst]

Something of potential interest is the difference in geometry between the two cameras. The images are scaled to nearly exactly match in the vertical FOV, yet the horizontal is significantly different. This causes some distortion betwen these two sensors. The horizontal FOV is larger for the FHD, so I assume the pixels are slightly narrower than they are tall. Or maybe the Canon pixels are slightly taller than they are wide? Is there a physical standard that manufacturers use? This is something like "sensor distortion".

That's a very interesting observation. The eBay ad that you point to is explicit that "Pixel size: 1.335x1.335um". Meanwhile Canon quotes the T2i as "Image Format 22.3mm x 14.9mm (APS-C size)" and "5184 x 3456 pixels" which calculates out to be 0.004302 by 0.004311. I've always assumed that Canon's pixels are actually square and that discrepancy in the third digit was due to rounding error in the sensor size. But maybe that's not the case. Seems like the question could be clearly resolved by shooting a stage micrometer in landscape and portrait alignments and seeing where scale differences appear.

--Rik

Going with the Canon numbers, you would expect the FOV of the two images to be the same by cropping 2 horizontal pixels from the Canon image. Experimentally I can use the images already taken and crop them to match FOV. Going back to the original, rather than making a 500x500 crop, I did a 498x501 crop, then upscaled it to 900x900. Other than a sub-pixel shift, these now match FOV pretty well. Here's the re-cropped and scaled image:



So instead of the expected 0.9979 scale factor per Canon spec, the required scaling is 0.9940, almost 3x. It could be that the pixels on the FHD are not really 1.335 x 1.335. For all the numbers we are trusting the mfr's data (sensor size on Canon, and pixel size on the FHD) and both of them could be slightly incorrect. The combined effects are that the two images are ~0.6% different, which is fairly big.