Luminous Landscape Forum
Raw & Post Processing, Printing => Digital Image Processing => Topic started by: David Eckels on September 11, 2016, 03:18:56 pm
-
I posted another image of the same general frame from the North Rim of the Grand Canyon in User Critiques. Wmchauncey asked me a question about photo stacking, which I did not understand. Then he posted a topic here in this area on super-resolution (http://forum.luminous-landscape.com/index.php?topic=113378.0;topicseen), which I read and got to thinking. Now this may be dangerous, but I thought I would post it for C&C and pose the question as to whether it could be generally applied to landscape photography.
My reading suggested that I'd inadvertently taken an image that would enable me to test this idea: Nine frames, same image, tripod, same exposures, opened as separate files in PS CC, File>Scripts>Statistics>Mean, and I let it crank away. I should add that the nine DNG files had all been reset in LR before opening in PS. The resulting "Mean" image definitely had less noise as might be expected, but to my eye, there was also more detail when compared to one of the single frames or to an image produced by choosing "Median" rather than the average. The image was saved as a TIFF to be processed in parallel with one from a single frame in LR. For comparison, I include two 500 pixel crops from the finished images, one from the "Mean" (left) image, the other from the single frame (right). Mind you, these crops were taken after I processed them in LR and then opened in PS for creative processing and sharpening. The reason I shot the series of nine frames was because the light was changing so very rapidly that I'd hoped to catch one with just the right illumination of the peak. The creation of the TIFF was not difficult or cumbersome as I had written an action to take care of it. I apologize for my ineptness in cropping and guess that the differences in appearances might be due in part to the rapidly changing light condition; it could also be that I screwed up somehow in LR, but I don't think it changes the conclusions. Finishing in PS was the same for each.
Regarding the comments by wmchauncey and the discussion on "super-resolution" I would be curious to see if this might comprise an easy DIY method for obtaining not only lower noise* but also greater detail in images. To wit, select the files in LR, open in PS, run the "Mean" action, and save as TIFF to create the digital "negative." Perhaps others have already discovered this approach.
*Nikon D800E, 28-300 zoom at 28 mm, 1/8 sec, f/16, ISO 100
-
BTW, the finished product, just as an FYI.
-
BTW, the finished product, just as an FYI.
Great image David.
Regarding the comments by wmchauncey and the discussion on "super-resolution" I would be curious to see if this might comprise an easy DIY method for obtaining not only lower noise* by also greater detail in images. To wit, select the files in LR, open in PS, run the "Mean" action, and save as TIFF to create the digital "negative." Perhaps others have already discovered this approach.
It will work better if you upsample the files (e.g. to 200%), and then align them before averaging (using Mean will reduce noise, Median will remove movement, while both will increase resolution).
The trick with upsampling before aligning, is that it allows to align, but now at half pixel precision (compared to the original image size). You can then downsample to almost the original size and it will be sharper than it was with an original single capture, or you can leave it at the enlarged size to not lose anything to downsampling, although the larger file will not be pixel sharp but it does have more detail.
Cheers,
Bart
-
I gotta thank David for his "real world" depiction of the technique...looks great.
Bart, I now understand the process but, why stop at a 200% upsize...why not 300% or 400% for that matter?
-
Thanks, Bart.
BTW, the images were aligned as part of the Statistics script. Is this sufficient for an upsampled image?...but, why stop at a 200% upsize...why not 300% or 400% for that matter?
Is it asymptotic (diminishing returns)?
-
I gotta thank David for his "real world" depiction of the technique...looks great.
Bart, I now understand the process but, why stop at a 200% upsize...why not 300% or 400% for that matter?
No real reason other than requiring many more shots to reach a stable increase.
At a 200% upsizing, we can increase resolution at a virtual quarter pixel precision when compared to the original size. With perfect half pixel pitch monochrome samples, we'd need 4 pixels for which we needed 1, or 16 if we also want to have full colour coverage with a Bayer CFA. But with semi-random sampling we may indeed need something like 20 or more to accidentally also have half pixel luminances and others in between, and probably more like 60 for full colour, coverage.
For a 300% case the required number of images gets unworkable pretty quickly, especially when the stitching alternative is much faster and more accurate.
What 300% does with a low number of semi-random shots, say up to 20 or so, is slightly (hardly visible) increase precision and resolution over the 200% case, just because the samples are slightly more accurately positioned. The file sizes do get to be prohibitively large for registration/alignment and averaging. Slow and large, with a hardly (if at all) a noticeable improvement.
The procedure would be best suited for Raw data, so that Demosaicing can be done on higher resolution data, but doing it with the already Raw converted data still has benefits.
Cheers,
Bart
-
Thanks, Bart.
BTW, the images were aligned as part of the Statistics script. Is this sufficient for an upsampled image?
If the statistics script first does alignment yes, otherwise it's not helping resolution, but only helps with providing noise averaging. Hint: Denoising can be more efficient.
Is it asymptotic (diminishing returns)?
Diminishing returns and computer memory and storage guzzling. If the randomness of the shot positioning is not random enough then the gain in resolution will also not be very isotropic (uniform in all orientations). So many more random shots are needed to get better uniformity of the improvements.
Cheers,
Bart
-
BTW, I discovered this in my searching...http://www.stark-labs.com/craig/articles/assets/Drizzle_API.pdf
-
David--
Your image capture and post work bring "life" to that wonderful canyon landscape. thanks for posting.
-
Nice picture. But it's the light that does it not the pixels. This is Luminous Landscape, after all.
-
Nice picture. But it's the light that does it not the pixels. This is Luminous Landscape, after all.
agreed, but without the pixels you can’t make a decent size print.
-
I just tried it but i am not impressed; ( maybe i did something wrong)
My lenses already give me a 100% cover on the 36Mp of my sensor and the Mean image over 10 shots is in fact less sharp.
Moiré is still there but a bit less pronounced.
The result is unpredictable. Making it a wast of time and effort.
The only thing i can see is less grain, but i can do that with choosing a lower iso value if things are not moving...
-
The OP's two crops are supposed to demonstrate what? They're both muddy,blurry and they are different colors.
Also, where is the evidence that resolution is improved, in the form of discernible line patterns. I suppose that jitter can overcome some of the estimation involved in demosaicing,but I think that adding images and averaging actually just raises the S/N ratio and over comes the MTF of the lens.
-
The OP's two crops are supposed to demonstrate what? They're both muddy,blurry and they are different colors.
Addressed in the OP. Acknowledged that there are differences in PP, but recognize that they are, in fact, different images with slightly different lighting conditions.where is the evidence that resolution is improved, in the form of discernible line patterns.
ExactlyI think that adding images and averaging actually just raises the S/N ratio
Obviously is raises S/N. But as to your "just" qualifier, you seem to agree that there is an improvement in discernible line patterns. Is the juice worth the squeeze? That depends on personal preference. For me, there's probably not so much an advantage that a tripod and low ISO wouldn't compensate, as kers points out and raises the question of what is good enough? It was "just" and experiment and not intended to advocate for a general approach.
-
Much talk of IQ on this forum tends to get quite esoteric, but for me the bottom line is not in pixel peeping, but in printing large. The transition back to an analogue print from digital data tends smooth out most of the pixel peeping differences. While I agree, shooting at base ISO should be better, I've found higher ISOs look just fine on my D800E, and the differences in a print I consider to be quite minor. While I love the lightweight carbon Fiber tripod I bought 3 years ago to replace my world traveled Tiltall, I feel quite liberated shooting hand held at higher shutter speeds. And putting together one row stitches of the camera held vertically works just fine.
-
Much talk of IQ on this forum tends to get quite esoteric, but for me the bottom line is not in pixel peeping, but in printing large. The transition back to an analogue print from digital data tends smooth out most of the pixel peeping differences. While I agree, shooting at base ISO should be better, I've found higher ISOs look just fine on my D800E, and the differences in a print I consider to be quite minor. While I love the lightweight carbon Fiber tripod I bought 3 years ago to replace my world traveled Tiltall, I feel quite liberated shooting hand held at higher shutter speeds. And putting together one row stitches of the camera held vertically works just fine.
+1
-
I did not agree to anything about line resolution. I would like to SEE actual evidence of improved resolution that is not just better signal to noise.
-
Nice image David
I have not tried the technique but I do think there is enough evidence to suggest an improvement in resolution as well as noise is attainable.
From the little I understand camera on tripod is not the best way to go as there needs to be a slight movement between shots so for many cameras it is suggested to hand hold and take a bucket load of shots
http://photoncollective.com/enhance-practical-superresolution-in-adobe-photoshop
https://aperture64.wordpress.com/2014/03/19/improving-image-quality-with-stacking-and-super-resolution/
Is it worthwhile I guess only the individual can answer
-
From the little I understand camera on tripod is not the best way to go as there needs to be a slight movement between shots so for many cameras it is suggested to hand hold and take a bucket load of shots.
Thanks, Tony. So I have tried this exactly twice and while the OP may arguably show some enhancement of resolution and/or S/N, when I tried it with an 8-shot handheld, I couldn't see any benefit; exactly the opposite of what seems to be reported. Honestly, I don't think the juice is worth the squeeze. It was "just" an experiment ;)
-
Hi Bart
I know you have been into deconvolution for several years and were very active on an old forum of the subject which sort of died away.
I have been using a new deconvolution processor "Franzis Sharpen" on my landscapes. The latest version was just updated today. I have been impressed with what they are accomplishing and would like to get your impressions of it. It still has scaling problems with the control icons on a 4K monitor but the image resolution is vastly improved over CaptureOne or PS, and it works very well as a plugin to DXO.
Regards
Paul
-
Hi Bart
I know you have been into deconvolution for several years and were very active on an old forum of the subject which sort of died away.
I have been using a new deconvolution processor "Franzis Sharpen" on my landscapes. The latest version was just updated today. I have been impressed with what they are accomplishing and would like to get your impressions of it. It still has scaling problems with the control icons on a 4K monitor but the image resolution is vastly improved over CaptureOne or PS, and it works very well as a plugin to DXO.
Hi Paul,
Yes, I also have that (the Sharpen Projects Professional edition) in my tool chest. I missed the update, so now I just updated to the latest version as well. I did not notice any obvious new features, so it probably is just a maintenance release.
The nice thing is that the Blind deconvolution handles so-called non-Isotropic blur (= not symmetric in all directions) automatically. The default region it uses to estimate the Point Spread Function (PSF) is usually a bit large, I usually have to reduce it to a radius of something like 10 to begin with, instead of 50 or 60 or so.
Also the default noise reduction is a bit much. Of course, I also maximize the quality setting, but that will make things a bit slower due to the required additional processing. Luckily the usage of the GPUs helps to keep processing times acceptable.
The automatic Photographic settings make it a bit hard to judge the sharpening settings in isolation, so I switch those off, and later may apply a bit of those (like dehaze) again.
The whole graphic user interface is a bit overwhelming and it looks rather technical, but it covers a lot of aspects, including selective masking. The Raw conversion quality is not good enough for my taste. It would be a good addition for those who cannot use Photoshop with plugins, because it is a stand-alone application.
So, all-in-all it does a good job but requires some tweaking of the parameters for better results than Automatic delivers.
Cheers,
Bart
-
Thanks Bart for your input. I haven't tried to do any tweaking yet as the icons and text on my 4K monitors are so small I have to use a magnifying glass, so for now I've basically been using it as a plugin for DXO 11.3.
Regards
Paul