mpmendenhall wrote:
And that is exactly what I've already posted. Compare my deconvolved images to the original: all the fine detail/texture is (nearly) perfectly reconstructed. Look at my blurred picture: I'd call that sufficiently blurred to obliterate visible traces of the detail/texture (look, e.g., at the tree bark or the red tile roof upper left). What's missing?
What's missing right now is a part of your own post (emphasis mine):
mpmendenhall wrote:
If I try to deconvolve a more strongly blurred image, or add just a tiny bit of extra noise to the blurred image, the deconvolution falls apart completely into random noise.
AGeoJO wrote:
I am not sure about batching, Fred. The file size will increase by 3X. I would do a pick and choose method only. The adjustment will be incorporated in the new RAW file and LR or any other can pick it up from there.
LR or other applications cannot read the CR2 files corrected by DPP because the file contains multiple RAW values plus some special instructions and headers. Canon has not indicated whether they will include these changes in their SDK and they don't want others to know what calculations they perform. It is very unlikely that third-party will be given the SDK to read the modified CR2 files, at least at this given time.
mpmendenhall wrote:
I was hoping that my images would count for a thousand words of response.
Indeed a good image speaks a thousand words, I think your example has shown this point nicely. If one understands the math it is trivial anyway and they should not be asking to see proof.
KaaX wrote:
What's missing right now is a part of your own post (emphasis mine):
And, as I've tried to repeatedly point out, the reason for this limit is the "noise floor" from rounding/error in the computations. If I wanted to re-write/re-compile my code to work at quad-double precision instead of float precision, I could de-convolve even more blur. How is this inconsistent with my proof-of-principle that under "ideal" conditions (not limited by noise from sensor or computation, or incomplete knowledge of the blur kernel), deconvolution can correctly recover the original image? For any blur radius you want, if you send me a computer with enough horsepower to carry out the calculations to sufficient accuracy, I can deconvolve an ideal synthesized blur back to the original pixels just fine. As far as relevance to the thread topic, I've already demonstrated "ideal" deconvolution to recover much more detail from much more blur than "real world" examples face (where sensor noise dominates much earlier than computational accuracy).
KaaX wrote:
You're saying that the noise in the image spoils the perfect deconvolution -- isn't exactly the same as saying that the fine detail in the image spoils the perfect deconvolution?
No.
He is not talking about noise in the image. Go back to his first examples and read what he wrote again.
He is talking about noise in the computation and "noise" from other sources, including lens sample variation, which is over-rated as to its prevalence.
speedmaster20d wrote:
LR or other applications cannot read the CR2 files corrected by DPP because the file contains multiple RAW values plus some special instructions and headers. Canon has not indicated whether they will include these changes in their SDK and they don't want others to know what calculations they perform. It is very unlikely that third-party will be given the SDK to read the modified CR2 files, at least at this given time.
Ouch, Canon wants to make sure that DPP is the choice of application, I guess. Thanks for clarifying that.
By any chance, do you know whether the modified RAW file can then be converted to DNG? I know it can be converted to TIF but at triple the size, the converted TIF file would be huge. At least, DNG files can still be processed using LR.
AGeoJO wrote:
Ouch, Canon wants to make sure that DPP is the choice of application, I guess. Thanks for clarifying that.
By any chance, do you know whether the modified RAW file can then be converted to DNG? I know it can be converted to TIF but at triple the size, the converted TIF file would be huge. At least, DNG files can still be processed using LR.
The current version of DPP will only convert a RAW file to either tif or jpeg.
bobsofpa wrote:
The current version of DPP will only convert a RAW file to either tif or jpeg.
I realized that DPP doesn't provide any option of converting to DNG; I was actually referring to Adobe DNG converter and not DPP. Meaning, if those modified RAW files can be converted to DNG using that Adobe application then LR can pick it up from there.
mpmendenhall wrote:
And, as I've tried to repeatedly point out, the reason for this limit is the "noise floor" from rounding/error in the computations.
Yes. But not only that. You're also constrained, for example, by the limited precision of the pixel values which are, after all, integers.
mpmendenhall wrote:
...limited by noise from sensor...
That's what I've been harping on. The sensor noise is not a limit.
Consider it from the math point of view. Convolution and deconvolution operate on a matrix of numbers. They completely don't care where these numbers come from or what do they mean. If the image has a lot of high-frequency data, you can't tell without looking at the image whether that high-frequency data is fine detail or noise. That's basically what sensor noise is, it's the high-frequency component. But you can get the same high-frequency component in other, "legitimate" ways, too, where it would represent desirable fine detail.
Consider, oh, I don't know, say images of fractals. Synthetic images, no "noise", right? But there's a huge high-frequency element there. From a convolution/deconvolution point of view, what's the difference from a photograph full of sensor noise?
Clearly, deconvolution works. But also clearly it's a somewhat fragile procedure. Push it beyond it's envelope and it messily collapses. It's not easy to make it work in photographic reality -- it's no accident that only now we're starting to get first de-blur applications.
He is not talking about noise in the image. Go back to his first examples and read what he wrote again.
He is talking about noise in the computation and "noise" from other sources, including lens sample variation, which is over-rated as to its prevalence.
Sure, let's go re-read:
mpmendenhall wrote:
The real-life usefulness of de-blurring is going to be most strongly limited by the noise quality of images coming from the camera. Low-ISO noise levels previously considered completely irrelevant will become highly important when aggressive deconvolution is applied to images.
KaaX wrote:
Yes. But not only that. You're also constrained, for example, by the limited precision of the pixel values which are, after all, integers.
That's what I've been harping on. The sensor noise is not a limit.
Consider it from the math point of view. Convolution and deconvolution operate on a matrix of numbers. They completely don't care where these numbers come from or what do they mean. If the image has a lot of high-frequency data, you can't tell without looking at the image whether that high-frequency data is fine detail or noise. That's basically what sensor noise is, it's the high-frequency component. But you can get the same high-frequency component in other, "legitimate" ways, too, where it would represent desirable fine detail.
Consider, oh, I don't know, say images of fractals. Synthetic images, no "noise", right? But there's a huge high-frequency element there. From a convolution/deconvolution point of view, what's the difference from a photograph full of sensor noise?
Clearly, deconvolution works. But also clearly it's a somewhat fragile procedure. Push it beyond it's envelope and it messily collapses. It's not easy to make it work in photographic reality -- it's no accident that only now we're starting to get first de-blur applications....Show more →
but the difference is high-freq info is something that we want restored and can be, random sensor noise is something totally different even if it is high frequency too, sure it can mess with fixing up the high-freq detail but that is another matter
earlier it had been posted that the 70-300 IS L was not included yet and that only the non-L is included for now, that is wrong, it is the L that is included
KaaX wrote:
Yes. But not only that. You're also constrained, for example, by the limited precision of the pixel values which are, after all, integers.
Quantization noise (limiting sensor readout to a discrete set of integers) is a form of sensor noise, though usually lower than shot noise and read noise in real-life sensors.
That's what I've been harping on. The sensor noise is not a limit.
...
That's basically what sensor noise is, it's the high-frequency component. But you can get the same high-frequency component in other, "legitimate" ways, too, where it would represent desirable fine detail.
Sensor noise typically has components at all frequencies. The high frequency components are more relevant to sharpening/deconvolution, because they will be magnified more.
I've already shown a working proof-of-principle that recovers "legitimate" high frequency detail from a synthetically blurred sample image --- in the sufficient absence of noise (sensor or computational), the deconvolution process works just fine. Sensor noise is "the limit" in the sense that the presence of this noise is a primary obstacle to real-life deconvolution. Lower noise images could be sharpened/deconvolved more strongly to recover more "legitimate" detail.
One more example, I've doubled my blur radius and boosted the internal working precision of my code to double throughout. More blurred image: http://praetoriusphoto.images.s3.amazonaws.com/fmforums/20120322_deconvolution/convolved_double.jpg
recovered deconvolution: http://praetoriusphoto.images.s3.amazonaws.com/fmforums/20120322_deconvolution/deconvolved_double.jpg
this level of precision isn't quite enough to handle all the extra blur, so the results are inferior to my earlier, less demanding blur examples. Nonetheless, my point is that, with sufficiently beefy computer backing, the deconvolution method can be pushed as far as you want for "ideal" cases. The real world limits hit much earlier than even my previous example due to a variety of factors, among which noise (in many forms) is a serious problem.
looked at a 24 1.4 II f/7.1 shot, 5D2, center frame
DPP sharpness 0, turning on DLO did sharpen it up noticeably, if not as dramatically for the OP's shot, but it as looking good until....
i opened the RAW in ACR and simply used it's dumb ACR sharpening at radius .5, level 25, detail 25 and, if anything, I saw a bit more detail from the ACR version
however, it's hard to know what fair parameters are for comparison, i had dlo at maybe 65%, maybe i could have run it nearer to 100 (it defaulted to 50) and maybe i could have added a trace of regular DPP sharpening first too (although not sure that would be fair)
it's hard to compare, but that perhaps flawed test made the whole thing seem a lot less stunning, but again, it might not have been fair to DLO
and i should mention that while the center, if anything, had a trace LESS detail, once I looked near the very far edges the DLO version did fix CA more perfectly and was a bit crisper there I couldn't get ACR to match it on the far edges or corners, no matter what was adjusted in ACR
it's not surprising that it mostly helped in the areas of more extreme aberration
skibum5 wrote:
but the difference is high-freq info is something that we want restored and can be, random sensor noise is something totally different even if it is high frequency too, sure it can maybe mess with fixing up the high-freq detail but that is another matter
The high frequency sensor noise is not part of the image projected by the, despite protestation by some here that it is part of the digital image, which it is.
The principle (in theory and practice) that lens images can be deconvolved is well established and proven many times over.
The high frequency sensor noise, which is part of the digital image, will be deconvolved with everything else and its dots (pixels) will get turned into inverse point spread functions. Again, this is not surprising to anyone who is acquainted with the field.
If you want the most nearly perfect images, shoot at the best ISO (almost always ISO 100 for Canons) with the best production techniques and the best equipment you can and process (convert) the Raws with the best conversion software available (which is now Canon DPP) and post-process with the best post-production software available which is Adobe Photoshop via 16 bit files or via 16 bit direct export to PS.
If you are forced to shoot under less ideal conditions so that ISO is high or lenses aren't on the list or you had exposure or dynamic range problems that need to be rectified with ACR highlight recovery or other problems, then you won't get the very best. But because it is Canon (or in the more general case without DLO, Nikon and others), it will probably be very good if you have done the best you can with what you have and the conditions.
Jeff Nolten wrote:
One could test for this by correcting an f4 image at the 24 end of the 24-105. See if it corrects for the barrel distortion and the vignetting.
All the imperfections of the lens are encoded in the PSF. If you can model the PSF accurately, you can theoretically undo any imperfection. As I said this is the tough part as the real PSF is immensely complicated and varies considerably over the lens. For example the PSF may look like a Airy pattern say in the centre of the lens and at the extreme periphery look like some mutant butterfly extending over a huge number of pixels. What is usually done is to divide the image plane into small regions over which it is assumed the PSF is uniform to make the problem more tractable.
We've already had software that can undo aberrations like distortion and chromatic aberration, but we also have coma, diffraction, AA filters, spherical aberration that contribute to the PSF.
Look, just stop explaining yourself to KaaX. He obviously doesn't understand signal processing, made an uninformed statement about deconvolution, and is now trying to insist that he was right.
wickerprints wrote:
Look, just stop explaining yourself to KaaX. He obviously doesn't understand signal processing, made an uninformed statement about deconvolution, and is now trying to insist that he was right.
