Luminous Landscape Forum
Raw & Post Processing, Printing => Digital Image Processing => Topic started by: Josh-H on October 11, 2011, 12:29:54 am
-
You Tube Video of the Sneak (http://www.youtube.com/watch?v=xxjiQoTp864&feature=youtu.be) - pretty impressive stuff!
-
I saw it this morning and was ask very very impressed...
-
I've been using Focus Magic, although it only works in 32-bit Photoshop, and Topaz Refocus. The Adobe Max 2011 looks like focus Magic on double steroids!
-
I wish the video was clearer.
It also seemed that this method applied to camera shake only. The map which popped up after the analysis showed the distribution of a data point over the time of the exposure. There was nothing about simple missed focus or poor lens quality, etc.
Still, deconvolution is an exciting prospect. I look forward to seeing where this goes.
-
Wow, I like!
-
It's a shame the person videoing it hasn't got access to it.
-
Many of us have been using deconvolution for years. Unfortunately for Mac users most of the software comes from the astrophysics area who seem to be partial to Windows. It will be interesting to see how Adobe is handling it. The calculations and iterations required take up a lot of processing time . Perhaps Adobe have found some solution. Blur has numerous causes and requires different strategies. For example the Richardson -Lucy algorithm is usually most favoured but the Van Cittert can come through a winner in some situations. Adobe is very good at whizz bang demos that sometimes don't quite match product performance.
-
Many of us have been using deconvolution for years. Unfortunately for Mac users most of the software comesfrom the astrophysics area who seem to be partial to Windows. It will be interesting to see how Adobe is handling it. The calculations and iterations required take up a lot of processing time . Perhaps Adobe have found some solution. Blur has numerous causes and requires different strategies. For example the Richardson -Lucy algorithm is usually most favoured but the Van Cittert can come through a winner in some situations. Adobe is very good at whizz bang demos that sometimes don't quite match product performance.
As exampled by the "I'll just select a group of parameters" part of the demo.
-
Exactly!
It is not a "new technology", but for sure is new to Photoshop and very welcome if it works properly.
As one of the presenters pointed, the "content aware fill" caused a similar fuss on the sneak preview, but I think it ended up being useless.
-
content aware fill is anything but useless!
-
content aware fill is anything but useless!
Well...
I guess I should have said "useless to my needs".
To me the bottom line is: It does not make my life easier as I still have to do a lot of manual work to make the image acceptable. I find that it only works perfectly on about 5% of the images, on all the rest I still have to do a lot of work to remove artefacts and to ease transitions, so much work that would be faster to clone and heal manually.
-
Well...
I guess I should have said "useless to my needs".
To me the bottom line is: It does not make my life easier as I still have to do a lot of manual work to make the image acceptable. I find that it only works perfectly on about 5% of the images, on all the rest I still have to do a lot of work to remove artefacts and to ease transitions, so much work that would be faster to clone and heal manually.
Complete agreement. I once used it to successfully extend a body of water, and it did ok at lining up ripples, but lots of artifacts remained to be cleaned up. Adobe's magic software demos never quite hold up in the real world usage.
-
Many of us have been using deconvolution for years. Unfortunately for Mac users most of the software comes from the astrophysics area who seem to be partial to Windows. It will be interesting to see how Adobe is handling it. The calculations and iterations required take up a lot of processing time . Perhaps Adobe have found some solution. Blur has numerous causes and requires different strategies. For example the Richardson -Lucy algorithm is usually most favoured but the Van Cittert can come through a winner in some situations. Adobe is very good at whizz bang demos that sometimes don't quite match product performance.
Indeed. The advantage for Lucy Richardson in astronomical imaging is that you either know the PSF because you have incredibly precise and well calibrated instruments (even if you implement the initial design somewhat wrongly cf Hubble) or you have a reasonable way to reverse engineer it from a supposedly perfect point source, a "just enough but not too much exposed star". Approximation is the key here,as Wikipedia puts it "It has been shown empirically that if this iteration converges, it converges to the maximum likelihood solution for uj." empirically and if being keywords here. It's always been, at least for me, a matter of trial and errors and multiple attempts. And let's not even talk about the role of noise, who should just be the right kind when you RE the PSF.
But there are other techniques and since one of the author of the paper works at Adobe that could be what they use.
http://www.cse.cuhk.edu.hk/~leojia/projects/motion_deblurring/index.html
http://www.cse.cuhk.edu.hk/~leojia/projects/motion_deblurring/deblur_siggraph08.pdf
There's an exe there you can try. I did when I first read the paper, with mixed results.
-
More info here:
http://blogs.adobe.com/photoshopdotcom/2011/10/behind-all-the-buzz-deblur-sneak-peek.html
-
This is already in another thread. I can make a custom PSF, blur an image, then reverse it. I can improve an image with a partial PSF, then improve it more with another partial. I had to do this on a picture where a buffalo started running.
If you want adaptive richardson-lucy now, use images plus. If you want van cittert now, use images plus. If you want custom PSFs now, use images plus. If you want adaptive smoothing that will preserve fine detail ... you guessed it.
Let me know when it can figure out the psf on the fly. Then i will buy photoshop.
-
Hows about focus stacking in macro, an area where software help would be greatly appreciated.
Here I stopped on the leg as it got artifacts. You can keep on going to focus the body.
-
Custom deconvolution now
http://www.youtube.com/watch?v=9D6-BTRivG8&feature=player_detailpage (http://www.youtube.com/watch?v=9D6-BTRivG8&feature=player_detailpage)
Before
(http://farm7.static.flickr.com/6094/6259414538_4ccf27d35e_b.jpg)
After
(http://farm7.static.flickr.com/6111/6259414544_78cb9ce3c5_b.jpg)
-
Fine_Art, that's some nice work there, but I don't see how it compares with the improvements shown in the links provided by Eric Chan and others.
The image of the antique cash register looks to be just a lens induced diffraction type clarity issue which you can easily fix in Adobe Camera Raw and Lightroom. I've done it on 100's of images in ACR.
-
I guess you havent been following the Adobe Max story, They faked one of the pictures. Its on DPR, Adobe's own blog, TOP, etc. They used a known PSF to blur it then used the same PSF to deblur it perfectly. You can never get that in a real test. They presented it as "This is what we are working on, this is what you can expect in a future product with your pictures". It is a lie. It is impossible to perfectly know the PSF from an image.
What you don't understand about that image (probably from the low res video) is it was hand held in a museum. The mild shake was real. I selected a small spot of dust in the image for the program to determine a PSF. It then took the smear reversing it back to (close to) a point source. This is not clarity, which is just local contrast enhancement, it is like a corrective lens (glasses).
-
Hi,
It is not impossible to know the PSF. For instance, the Hubble space telescope had some problem with it's mirror and it could be corrected because the PSF was known. Sometimes PSF can be estimated from the image.
Best regards
Erik
I guess you havent been following the Adobe Max story, They faked one of the pictures. Its on DPR, Adobe's own blog, TOP, etc. They used a known PSF to blur it then used the same PSF to deblur it perfectly. You can never get that in a real test. They presented it as "This is what we are working on, this is what you can expect in a future product with your pictures". It is a lie. It is impossible to perfectly know the PSF from an image.
What you don't understand about that image (probably from the low res video) is it was hand held in a museum. The mild shake was real. I selected a small spot of dust in the image for the program to determine a PSF. It then took the smear reversing it back to (close to) a point source. This is not clarity, which is just local contrast enhancement, it is like a corrective lens (glasses).
-
The PSF in the hubble was estimated based on point source stars. It was not good enough. If you remember they put together another mission at high expense to install corrective hardware. If you remember the video of the astronauts doing a spacewalk installation of the "glasses for the telescope" was on the news.
Would they have done that if they could just do it in software? No.
-
This article put the cost of the mission at 1.5 billion $. 1.3 was just the space shuttle.
http://www.company7.com/c7news/nasa_sts61.html (http://www.company7.com/c7news/nasa_sts61.html)
A supercomputer would have been cheaper if it could work. If all the brainiacs at NASA couldn't do it for a billion $, what are the chances it will be in the next photoshop upgrade for $300? Its obscene snake oil salesmanship.
-
To be more precise on "clarity", it is a color contrast stretch. A normal contrast enhancement stretches the luminance channel. Clarity stretches the color channel. There is nothing special about it.
-
Hi,
Yes, but if you have missed your picture the best alternative is probably to use deconvolution. NASA can afford to modify the telescope in space and reshoot the pictures, but for a photographer the exact subject is normally gone.
Best regards
Erik
This article put the cost of the mission at 1.5 billion $. 1.3 was just the space shuttle.
http://www.company7.com/c7news/nasa_sts61.html (http://www.company7.com/c7news/nasa_sts61.html)
A supercomputer would have been cheaper if it could work. If all the brainiacs at NASA couldn't do it for a billion $, what are the chances it will be in the next photoshop upgrade for $300? Its obscene snake oil salesmanship.
-
Hi,
Yes, but if you have missed your picture the best alternative is probably to use deconvolution. NASA can afford to modify the telescope in space and reshoot the pictures, but for a photographer the exact subject is normally gone.
Best regards
Erik
You say that like you think a billion has no value to them. If it could be done perfectly in software they would have done it. Adobe reversing a known PSF and presenting it as an upcoming product is fraud. They could have taken a mild shake then done an approximate fix. They knew the crowd would not be roaring with delight. They chose to present a fraud. A real product will never be able to do that.
-
I don't know to stop all this misrepresentation. Software floggers, politicians, priests, bankers- on it goes.It remains to be seen if street marching does anything but replace one lot of crooks with another. All I can think of is try and do without, if not possible buy another product and lock up or redistribute immoral funds gained by anyone proved to be lying, manipulating, molesting, slaving, invading, stealing,bullying, ruining cultures/ environment/ health/ childhood/intelligence.
Start at the top , not with minions, couriers, petty salesmen, bank clerks, local politicians, foot soldiers etc. I know, I know . They have the money, they pay the lawyers and own the press. Back to photography. For a start don't buy any software or hardware unless it improves your work and does what it claims to. Investigate alternatives.Don't assume big must be best. If rigour proves without reasonable doubt the product or it's hype is a con , tell everyone. Got to start somewhere.
-
To be more precise on "clarity", it is a color contrast stretch. A normal contrast enhancement stretches the luminance channel. Clarity stretches the color channel. There is nothing special about it.
Where am I seeing improvements between your two cash register shots, Fine_Art? They look almost identical. None look blurred. What was your point in posting those two images?
The link below to one of my images in my "Jpeg VS Raw" folder gallery will show a more pronounced improvement in what I'm talking about similar to what you've demonstrated in your images. I agree there is a difference between Clarity and Blurry but not how you describe it.
Blurry is either movement induced or narrow "Bokeh" DOF induced or just out of focus. It's all about bringing back the "Mach Band" (explained in link below) effect our human optics use to judge clarity and sharpness by. It's just contrast/clarity starts at the micro fine edge and radiates out to broader tonal variances as shown in the flower image. Blur broadens the tonal roll off between contrast variances.
I don't know where or how you correlate luminance and color channels from in determining Clarity. Maybe you're referring to how algorithms handle it. Not sure.
http://photo.net/photodb/photo?photo_id=12334632
http://www.cambridgeincolour.com/tutorials/unsharp-mask.htm ...scroll down to gradient for Mach Band.
-
1024 is the largest size flickr will let me post. If you look carefully you will see detail enhanced. Ill put crops in as attachments. Sorry, i should have checked to see how well the difference shows.
-
So from those you should be able to see a clear improvement from selecting in the picture. There was a clear shake from the button press. It is not a perfect deconvolution by any means. It also tends to create some artifacts that become enhanced from spots of noise. With this adaptive RL version you can select the noise cutoff. I usually go with 5 standard deviations on the slider. The artifacts aren't really noticeable unless you are pixel peeping. On the overall picture you should see clear differences on the vertical faces facing forward. Look at the bottom. A the top. Look at the zigzags in the thistles.
A USM will find an edge then increase the contrast. An edge that is not sharp really needs the width of the edge fixed not the amplitude. A deconvolution shrinks the diameter of edges. If you run too many iterations it creates dark ring craters around edges. It starts to look like USM but it is different. It is like taking a pile of sand and moving it into a taller thinner pile. Too much makes a crater around it.
I opened the original again, here is the speck of dust I selected with the custom PSF shown. As you can see it is a blur point. Using this software you can pick these things. Look at how the numbers are not symmetrical around the peak. Using a standard Gaussian would be inferior.
-
Again, I don't see a big difference compared to what PierreVandevenne's deconvolution links demonstrate including what I did with the white flower (did you even look at the white flower-you seem to be talking past me).
The kind of improvements you're espousing requires you DO have to pixel peep to see it and it isn't anything better than what I can accomplish with a combination of Curves, Clarity, Sharpening, a custom camera profile, HSL, and Contrast tweaks in Adobe Camera Raw.
Here's a 100% crop ACR preview of what I'm talking about.
-
I did look at the white flower. It looks completely artificial. What is your point? It has nothing to do with debluring an image. It doesnt even look like a rose, it looks like a 3D studio max creation.
As posted on DPR the Plaza shot was also rigged. The blur on the left is supposed long exposure motion blur. Look at the legs on the improved image. There is no motion in the walking legs.
http://blogs.adobe.com/photoshopdotcom/files/2011/10/Plaza.png (http://blogs.adobe.com/photoshopdotcom/files/2011/10/Plaza.png)
-
I'm done with you.
Pinning you down on your logic is like playing wack-a-mole.
-
You clearly have no comprehension on what the demo was supposed to be demonstrating. Your grainy coins do not look sharp, they look contrast enhanced.
-
This is what a rose is supposed to look like. Your looks like it could be a plaster sculpture.
(http://farm7.static.flickr.com/6030/5927130275_bca1b5638f_b.jpg)
-
I did a quick conversion of a white rose to compare to your method.
My Rose
(http://farm7.static.flickr.com/6119/6265068432_bf79ebb280_b.jpg)
Your Rose
(http://gallery.photo.net/photo/12334632-lg.jpg)
Which looks more realistic?
-
Which looks more realistic?
Mine.
Yours looks soft, murky and hazy looking and I'm viewing on an X-rite calibrated display.
You weren't in that garden when I shot that white flower. The clarity, brilliance and intensity from a crispy 10 AM morning sunlight was overwhelming and that is what I tried to convey in the final rendering based on my memory and emotional response of that scene.
-
Did you rub it with carbon paper? What is the black stuff?
-
The black stuff is a result of reducing saturation to -50 for both orange and yellow channels in ACR's HSL panel. ACR's interpretation of my camera's Auto WB tends to be on the warm side and I do remember the flowers having an off white appearance but not yellow/orange as in the unedited jpeg. If I were to reduce saturation to make the flower more neutral by adjusting WB then the entire image would be too blue. That's the neat thing about that panel in that it can do some unexpected things to the images color table just by playing around.
I'll admit I took liberties in rendering the flower and went overboard with local contrast sharpening in Photoshop which over emphasized the Mach Band effect (similar to over cranked tonemapping in HDR). My eyes tend to adapt to the appearance of clarity where I tend to not see this as I'm editing.
It was mainly to see just how much clarity I could bring out in such a soft image.
It was shot at f/16 with a cheap kit lens that induces diffraction at that aperture.
Below is the non-Photoshop sharpened version straight out of ACR, but showing what would've happened if I used WB adjust instead of reduce orange and yellow saturation to make the rose more neutral.
-
I think there are arguments for both sides: at the time the Hubble fix was launched, the fastest super computer was the Connexion Machine 5 - at about 1.5 millions per node, some of the fast machines having 512 nodes and delivering 60 GFlops. That's about what you get from a current mid to high level i7, and less than what you'd get from GPUs. For about $30000, knowledgeable users can build stuff that runs in the hundreds of teraflops range (most notably this http://web.hpu4science.org/ - and yes, one of the goal was deconvolution). So the computional abilities have increased tremendously and that has to be taken into account. Algorithms have made some progress as well, if not a a fundamental level - there is data loss after all and I don't think any way to RE a multiple factors PSF has been proven - simply because some of the paths are now become practical.
I think it is reasonable to say that Adobe's demo is a bit misleading, but then they don't claim it is ready and universal. And it is also fair to say that a generic, automatic, one size fits all sources of loss, deconvolution algorithm leading to images satisfying photographers is a very very distant perspective.
BTW, I've always found the "content aware fill" buzz a bit ironic because in essence it is the same phenomenon that the demoes exploit: give us a couple of well chosen test cases and our minds automatically fill the vast void between them with the expectation that it will work for all images...
And any way, I would venture to say that, as soon as there are three photographers in one room, there isn't a single issue that is consensual :)
-
Adobe reversing a known PSF and presenting it as an upcoming product is fraud. They could have taken a mild shake then done an approximate fix. They knew the crowd would not be roaring with delight. They chose to present a fraud. A real product will never be able to do that.
This is a bigger mis-representation of the situation than Adobe's "cheat". According to Adobe, the first two images were real camera shake. The third image was simulated camera shake, but the reconstruction program was not given the simulation PSF.
-
This is a bigger mis-representation of the situation than Adobe's "cheat". According to Adobe, the first two images were real camera shake. The third image was simulated camera shake, but the reconstruction program was not given the simulation PSF.
And you believe everything you are told. Look at the link to the plaza shot above, it is from Adobe's own blog site. Look at the legs of the people walking around. They are sharp. There was no slow shutter speed causing shake. Even looking at the shake it looks unnatural. When a photographer takes the shot it is usually close to stable, then the button press causes a dip plus rotation clockwise or the photographer decides to move to track something mid shot. There is an acceleration. The blur shown looks like it was generated by a constant continuous pattern. There is no acceleration. No rotation.
If you refuse to believe the actual picture from the demo it is because you want to believe something like this is real.
A coverup is not in Adobe's interests. It is better to blame the presenter for pulling the wool over everyone's eyes with no authorization. Can him, then buy one of the small companies making this specialty software.
-
In a seemingly crafty (but nevertheless transparent) corporate side-step, Adobe Inc. has evidently decided to try to "deflect the heat" generated by the controversies surrounding their recent posted video featuring public claims made by employee Jue Wang while in the employ of Adobe at:
http://tv.adobe.com/watch/max-2011-sneak-peeks/max-2011-sneak-peek-image-deblurring/
and at the Adobe Inc. blog web-page authored by Adobe employee Cari Gushiken, published at:
http://blogs.adobe.com/photoshopdotcom/2011/10/behind-all-the-buzz-deblur-sneak-peek.html
... by instead referring interested parties to a web-site operated and maintained by Jue Wang himself (while he remains in the employ of Adobe Inc, and is acting on Adobe's behalf, anyway).
.
Jue Wang's Adobe Employee Profile can be found at:
http://www.adobe.com/technology/people/seattle/wang.html?PID=2159997
Wang's web-site maintained for profit on behalf of Adobe Incorporated is here:
http://juew.org/deblurFamousPhoto.html
... where one will find the publishing of an entirely different set of images represented by Adobe Inc. to be processed by a prototype "de-blur" algorithm. Suddenly, Adobe's story has changed ...
The resulting images posted are marginal, noisy, not useful, and not representative of results that are even remotely close to what the (alleged by Adobe to be) legitimate "*Plaza.png*" depicts:
http://blogs.adobe.com/photoshopdotcom/files/2011/10/Plaza.png
.
Despite the fact that the "*sarah-dog.png*" image published by Adobe at:
http://www.adobe.com/technology/people/seattle/wang.html?PID=2159997
... with the direct link to the image being here:
http://blogs.adobe.com/photoshopdotcom/files/2011/10/sarah-dog.png
... shows absolutely no visible signs of there existing any difference whatsoever between what has been claimed by Adobe to be a "before/after" representation of the original Android cell-phone recorded image, a new and clearly entirely different image showing some visible results of some sort of processing, and containing visible image-noise/artifacts (without any explanations or admissions to the obvious differences by Wang himself) is presented here on behalf of Adobe, Inc:
http://juew.org/blur/2010-12-22_16-43-14_520_out.jpg
... published Oct 23 2011, and appearing to have actually been processed in some manner relative to:
http://juew.org/blur/2010-12-22_16-43-14_520.jpg
In well-managed and ethical companies, the engineering department is not forced by the marketing department to make known to be false public representations regarding the the status and operational efficacy of in-development prototype products. My condolences to Mr. Jue Wang.
After evidently experiencing pressure by his employer to turn their desire to "sell a sizzle from a yet uncooked steak" to the public, Mr. Wang has been reduced to jeopardizing his personal professional reputation as an evident condition of employment with Adobe, Inc., and (furthermore) evidently been pressured into maintaining a web-site in his personal name that ends up publishing a (much more accurate and realistic) "whimper" following Adobe's "big bang", stating:
"We agree that the algorithm does not always work well, and this is the case it fails right now."
-
Talk about a meaningless story totally blown out of proportion. Really, who cares whether an Adobe prototype works or not?
I'd have more problems with claims related to an actual product being sold, but even then we all know that marketing tends to show a product under its most favorable angle.
Then it is our responsibility as buyers to check the claims of software vendors. Trials are typically available for that very purpose.
Cheers,
Bernard
-
Wonder why all the links "Detail Man" posted have been severely changed with some getting the "404 File not found" associated with Jue Wang's blog images.
Well, I guess there is no such thing as bad press.
I agree with Bernard. It's much ado about nothing from a controversial aspect.
Do you think someone as smart and educated as Jue Wang would jeopardize his career over this? Come on!
Frankly I'm quite impressed with Jue Wang's results with the images he did fix with his algorithms he posted in his blog. He clearly shows the possibilities and potential behind his work. It's not a total misrepresentation.
Need more Dwight, please!
-
Wonder why all the links "Detail Man" posted have been severely changed with some getting the "404 File not found" associated with Jue Wang's blog images.
The links are working, just remove the semi-colon at the end and they will work just fine. That is a quite common mistake induced by the "autolink" feature on some forums, everytime a link does not work, check for that.
Regards.
-
Say, is there any way to set up this LL Forum viewing system so that (in addition to seeing bombastic posts, which is inevitable), I don't have to stare at ["disturbing imagery"] in the "avatars", as well? TIA!
PS - The links to the URLs in my original post are (now) all working fine. Sorry about that! If one's video does not seem to be working, it just might be time to visit those "hallowed halls of Adobe" in order to get the latest Plug-ins for one's browser ...
-
Say, is there any way to set up this LL Forum viewing system so that (in addition to seeing bombastic posts, which is inevitable), I don't have to stare at "ugly mugs" in the "avatars", as well? TIA!
Why don't you grow some manners, shitbird?
-
Geraldo,
The links are working now.
But still the ones with the AdobeTV video are just showing a white box.
-
... that is, when applied at the level of "raw" image-data prior to de-mosaicing and prior to NR (which helps to attenuate some of the DD artifacts that inevitably result). DxO Optics Pro 6.x "Lens Softness" corrections can be impressive, indeed. Note that DD is not the only component of the DxO "Lens Softness" corrections:
http://www.dxo.com/us/photo/dxo_optics_pro/optics_geometry_corrections/lens_softness
I have created some comparison examples for viewing. Best to download them and line them up to compare them in an image-viewer (as the DPR display-system downwardly re-samples these images, thus distorting the results).
.
A comparison between two approaches to the business of "sharpening" in order to enhance "image acuity": DxO Lens Softness corrections followed by mild Unsharp Masking; OR Lightroom 3.4 Sharpening (which uses USM and Deconvolution Deblurring):
(1) DxO 6.6 (without Lens Softness corrections) > LR 3.5 Sharpening (Equal Parts USM + Deconvolution Deblurring: Radius=0.5; Strength=100%; Detail=50; Masking=3) > XnView 1.982 (Loss-less JPEG); or
(2) DxO 6.6 (with Lens Softness corrections) > Sagelight 3.1 Sharpening [16-bit USM: Radius=0.5; Strength=100%, (non-linear) Threshold=10] > XnView 1.982 (Loss-less JPEG).
Notes:
In the process of sharpening the 16-bit TIF output image-file resulting from DxO RW2 processing, Lightroom 3.5 was used only as a (USM + DD) sharpening tool. The LR 3.5 "Masking" (linear threshold control) was set to 3/100 (around 5 EV, or "stops", below maximum level). All other LR 3.5 functions were either set to Zero (0), or disabled completely.
In the process of sharpening the 16-bit TIF output image-file resulting from DxO RW2 processing, Sagelight 3.1 was used only as a (16-bit USM) sharpening tool. The SL 3.1 "Threshold" (an exponentially scaled threshold control) was set to 10/100 (estimated to be around 5 EV, or "stops", below maximum level). No other SL 3.1 functions were used.
XnView 1.982 was used in both case to convert the 16-bit TIF results to an 8-bit loss-less JPG (using no Chroma Sub-sampling, and with Quality Factor = 100%).
Below are working links to example images that show the result of applying those two approaches (at 3272x2454 original crop size):
DxO (without Lens Softness corrections) > Lightroom (USM+DD Sharpening):
http://www.dpreview.com/galleries/4464732135/download/1476868
DxO (with Lens Softness corrections) > Sagelight (USM Sharpening):
http://www.dpreview.com/galleries/4464732135/download/1476869
.
So, do things get any better at a reduced (1600x1200) pixel-size? ...
A comparison between two approaches to the business of "sharpening" in order to enhance "image acuity": DxO Lens Softness corrections followed by mild Unsharp Masking; OR Lightroom 3.4 Sharpening (which uses USM and Deconvolution Deblurring):
(1) DxO 6.6 (without Lens Softness corrections) > Artizen 2.86 (16-bit Lanczos-3 re-sampling to a 1600x1200 pixel-size) > LR 3.5 Sharpening (Equal Parts USM + Deconvolution Deblurring: Radius=0.5; Strength=100%; Detail=50; Masking=3) > XnView 1.982 (Loss-less JPEG); or
(2) DxO 6.6 (with Lens Softness corrections) > Artizen 2.86 (16-bit Lanczos-3 re-sampling to a 1600x1200 pixel-size) > Sagelight 3.1 Sharpening [16-bit USM: Radius=0.5; Strength=100%, (non-linear) Threshold=10] > XnView 1.982 (Loss-less JPEG).
Notes:
In the process of sharpening the 16-bit TIF output image-file resulting from DxO RW2 processing, Lightroom 3.5 was used only as a (USM + DD) sharpening tool. The LR 3.5 "Masking" (linear threshold control) was set to 3/100 (around 5 EV, or "stops", below maximum level). All other LR 3.5 functions were either set to Zero (0), or disabled completely.
In the process of sharpening the 16-bit TIF output image-file resulting from DxO RW2 processing, Sagelight 3.1 was used only as a (16-bit USM) sharpening tool. The SL 3.1 "Threshold" (an exponentially scaled threshold control) was set to 10/100 (estimated to be around 5 EV, or "stops", below maximum level). No other SL 3.1 functions were used.
XnView 1.982 was used in both case to convert the 16-bit TIF results to an 8-bit loss-less JPG (using no Chroma Sub-sampling, and with Quality Factor = 100%).
Below is the result of applying those two approaches (at a 1600x1200 pixel size):
DxO (without Lens Softness corrections) > Lightroom (USM+DD Sharpening):
http://www.dpreview.com/galleries/4464732135/download/1476933
DxO (with Lens Softness corrections) > Sagelight (USM Sharpening):
http://www.dpreview.com/galleries/4464732135/download/1476934
.
(All methods of applying) DD to images suffer from the possibility of significant artifacts (if/when the number of re-circulations in the Deconvolution-processes is increased). (I think) that the idea of being able to apply DD processes after the fact, and "willy nilly" on random images is a dubious pipe-dream at best [whether or not some algorithm has attempted to "guess" the total system Point Spread Function (PSF)]. See:
In practice, finding the true PSF is impossible, and usually an approximation of it is used, theoretically calculated or based on some experimental estimation by using known probes. Real optics may also have different PSFs at different focal and spatial locations, and the PSF may be non-linear.
http://en.wikipedia.org/wiki/Deconvolution#Optics_and_other_imaging
Various people appear to be reading all kinds of fantasy-applications in to this Adobe "toy" still being tinkered-upon by "Santa's Elves". Such thoughts say more about their naivete (probably a big plus for Adobe's marketing department?) than it does about what is really going on, and what is possible. Adobe has proven that all that a marketeer has to accomplish is to "fool some of the people all of the time", and (in the situation of a high enough volume of sales), "Fat City" is not that hard to (for a time, anyway) find.
BTW - Jue Wang is clearly a highly technically competent individual. Please understand that I harbor no ill-will towards Mr. Wang himself. (However), the way that Adobe Inc. has handled this public relations "FUBAR" is unfortunate, indeed. Those who have actually worked within a design/development engineering environment (as have I) will likely understand exactly what I am talking about ... ::)
.
Some Deconvolution Deblurring links for the genuinely curious to follow:
Here is the Richardson-Lucy deconvolution approach arising out of the Hubble Space Telescope:
http://adsabs.harvard.edu/full/1994ASPC...61..296
.
Some (freeware) command-line executable files (the example-images for which look pretty slick):
http://www.cse.cuhk.edu.hk/~leojia/projects/motion_deblurring/index.html
Here's a paper that is co-authored by Jue Wang of Adobe Systems you might like:
http://cg.postech.ac.kr/research/deconv_outliers/
There are a number of things already in existence as (GUI) plug-ins and stand-alone applications that you might dig:
The Refocus (freeware) plug-in for GIMP (Wiener Filtering deconvolution; a real POS).
FocusMagic (not sure if it's Wiener Filtering or Richardson-Lucy; a few people still like using it).
InFocus (looks slicker than FocusMagic; more current from the standpoint of development).
RAW Therapee (freeware) Richardson-Lucy Deconvolution Deblurring (4 controllable parameters). The 4.x releases are much improved from the (still buggy with Windows OSs) Versions 3.x. RT 4.x has what is (in my viewpoint) a better user-interface than DxO, LR3.x /CR 6.x, or Silkypix (SE or Pro 4.x versions).
(And my personal favorite), DxO Optics Pro 6.x "Lens Softness" corrections which incorporates Deconvolution Deblurring along with other elements (for RAW/JPG processing of supported camera/lens combinations). The DxO (RAW-mode) Optical Corrections Modules work the best. The DxO (JPG-mode) Optical Corrections Modules are impressive (but do not benefit from the originally mentioned ability to perform the DD processes at the "raw" image-data level, prior to de-mosaicing, and prior to any NR applied in processing).
-
Hi,
I don't think its considered a bad practice in science to first convolve (blur) an image with an artificial kernel, add some noise and later deconvolve (deblur) it with the known kernel. If the kernel is not known, but estimated from the image, it would be called "blind deconvolution". Much of the research is based on similar experiments.
So what Adobe did was something based on scientific approach, but that may have not been fully appreciated by the public.
My humble view may be, that it is beneficial for the photographic community that Adobe and others develop usable tools for blind deconvolution. Sometimes we fail on something in taking pictures and it's nice if the failed picture can saved/restored. More often than not it is not possible to reshoot.
Adobe is doing marketing, and they of course promise a bit more than what the technology can deliver. They find examples that work and that make impression.
Finally, I wouldn't judge anything based on a "you tube" quality presentation.
Best regards
Erik
Talk about a meaningless story totally blown out of proportion. Really, who cares whether an Adobe prototype works or not?
I'd have more problems with claims related to an actual product being sold, but even then we all know that marketing tends to show a product under its most favorable angle.
Then it is our responsibility as buyers to check the claims of software vendors. Trials are typically available for that very purpose.
Cheers,
Bernard
-
Hi Erik Kaffehr,
I was wondering if you might possibly further explain your thoughts expressed (quoted below in italics). If you don't want to, don't sweat it!
I don't think its considered a bad practice in science to first convolve (blur) an image with an artificial kernel, add some noise and later deconvolve (deblur) it with the known kernel.
What would be the significance and utility of simply convolving and then de-convolving using the same Point Spread Function (PSF)? Other than adding some noise, you seem to be describing performing a convolution process simply followed by then subsequently performing the inverse convolution process. If one wanted to determine the relative noise-immunity of a de-convolution process, then taking such an action does seem to make sense.
However, my own (possibly limited) understanding is that a relevant problem with deconvolution deblurring processes seems to be the artifacts that the processes themselves create. This artifacting does (as I understand it) increase when processing an input with a poor Signal/Noise Ratio (SNR). Is it your understanding that, given an (albeit theoretical only) infinite SNR image-data source, that there would exist no resulting artifacts from a deconvolution deblurring process?
If the kernel is not known, but estimated from the image, it would be called "blind deconvolution". Much of the research is based on similar experiments.
Yes. Cases where the Point Spread Function (PSF) is unknown or unknowable (and various attempts may be made to in some manner estimate the PSF).
So what Adobe did was something based on scientific approach, but that may have not been fully appreciated by the public.
Indeed, blind deconvolution techniques are more common than deconvolution techniques where the PSF is previously known or characterized.
My humble view may be, that it is beneficial for the photographic community that Adobe and others develop usable tools for blind deconvolution. Sometimes we fail on something in taking pictures and it's nice if the failed picture can saved/restored. More often than not it is not possible to reshoot.
It's not clear to me that any person on the planet would argue with your sentiment expressed. What seems more interesting are the practical limitations of the use of deconvolution processes to address optical system de-focusing phenomena and motion-blur. Any (actual) "scientist" speculating about the practical likelihood of the actual successful implementation of such schemes would be subject to "review" (by "peers", as well as by others) in the course of such "scientific approaches" being explored. But Adobe, from their proprietary perch, is not offering anything remotely resembling such a process ...
Adobe is doing marketing, and they of course promise a bit more than what the technology can deliver. They find examples that work and that make impression.
Well, the term "marketing" seems to (one might hope) have a separate meaning than the phrase "deceptive marketing". Courts of Law (who take a "scientific" and "evidence-based" approach) surely do. If these concepts have become fused within our consumer-culture, what does that say about the sensibility of accepting promises by commercial entities that no person actually expects to be truthful and accurate?
Finally, I wouldn't judge anything based on a "you tube" quality presentation.
I take it that you use the term "quality" in a generalized sense above (as it relates to a certain "lack of believability" of anything that appears in that medium)? If so, is it true that corporations that intend to market products for monetary gain may be expected to suspend ethical practices (including an expectation of accuracy and truthfulness) as a result of using such a suspect medium as "YouTube"?
If you use the term "quality" in relation to the actual optical resolution of the presentation, we do not need to rely on hazy glimpses of video shot of presentation-screens published by Adobe on "YouTube". We have Adobe's previously published "Plaza.png" at:
http://blogs.adobe.com/photoshopdotcom/files/2011/10/Plaza.png
When you yourself have a look at the images (now) posted at:
http://juew.org/deblurFamousPhoto.html
... and the general lack of efficacy, as well as the visibly low Signal/Noise Ratio (SNR), of the de-blurring processes demonstrated, have you (as I have) "scratched your head" and wondered exactly why it would be that none of the examples (now) posted at the link directly above appear to come anywhere even close to the seemingly miraculous quality (and relatively high output SNR as compared to all of the other examples published) that is displayed in Adobe's publishing of the "Plaza.png" "before/after" image?
My interest here surrounds the more straightforward and immediate questions relating to a "scientific method", such as "where's the beef?", and "where is the evidence?". Is not that "just the stuff" that the "men in white coats" think about in the hallowed halls that constitute the "cathedrals of science and industry" that seem to have (somehow) earned such an automatic sense of (knee-jerk) deference among members of "the public" these days?
And (from a sociological standpoint) is this modern "techno-tribalism" - where individuals experience an emotional brand-loyalty that is divorced from logic to corporations that they pay money to (and not the other way around) - at all wise and healthy (and a "scientifically sound") trend? Should we elect to "suspend our rational disbelief" merely on the emotional basis that our persona's identification with a corporation (i.e., Apple. Adobe, etc.) may happen to give us a sense of existential meaning, belonging, and a competency imbued from without (on the sole basis of product "consumption" via the liquidation of "disposable income")?
Such questions may not prove to be answered by coherent answers from "the public". But it seems (to me) that the day when nobody either asks such questions, or is capable of answering them, may well be day when the "consumer state" becomes little but the indentured servants of a modern techno-feudalism in the willing servitude of corporate monoliths, where a once "eternal vigilance" has since devolved into a "blissful indifference" ...
Regards, and "top of the mornin' to ya'", kind sir.
PS - I have been perusing some of the links provided at the end of your post, including your personal web-site, and your articles. Some interesting materials, indeed. You clearly are a knowledgeable and thoughtful fellow, and I look forward to the possibility of learning from you and your interesting mind. I can see that the brevity of your post on this thread does not reveal your extensive amount of thought and consideration regarding some of these issues surrounding photography, and I am not in any way underestimating your intellect by raising some questions in my post above about your thoughts on this particular matter of Adobe's recent publicizing of information concerning their developmental work in the area of deconvolution deblurring tools.
-
Hi DetailMan,
First I may need to state that I'm not really involved with image processing, also many thanks for your comments on my writing.
The points I wanted to make are essentially that:
1) I got the impression that it is a very common practice to test algorithms using known PSF to convolve and deconvolve test images in presence of noise or disturbance. I have a book called Digital Image Processing by Gonzales and Woods, and the type of experiment I mention is used throughout the book. On the other hand, the use of this methodology is probably not well known for the public. So when it is discovered that some of the images were artificially degraded the public regards it to be cheating, while for the developers it is "Standard Operating Procedure".
2) In my view deconvolution is not a substitute for exacting work. I see several areas however where it can be quite useful. One such area is of course the rescue of images with technical problems. I don't expect wonders from deconvolution but some times very good results can be achieved. Also, deconvolution may be useful to restore OLP filtered images or enhance images suffering from diffraction. This page of my own testing indicates some possibilities: http://echophoto.dnsalias.net/ekr/index.php/photoarticles/49-dof-in-digital-pictures?start=2
3) Adobe did not give a product presentation, just a sneak peek into their research. But even in a product presentation they would normally demonstrate cases where the method works and avoid cases where it does. I guess that it is Standard Operating Procedure for most industries.
Finally, I'm essentially underhelmed with almost any presentation on U-tube. We work with very high quality media, while digital video at best (full HD) corresponds to two megapixels.
I hope that this explanation is elaborate enough.
Best regards
Erik
Hi Erik Kaffehr,
I was wondering if you might possibly further explain your thoughts expressed (quoted below in italics). If you don't want to, don't sweat it!
I don't think its considered a bad practice in science to first convolve (blur) an image with an artificial kernel, add some noise and later deconvolve (deblur) it with the known kernel.
What would be the significance and utility of simply convolving and then de-convolving using the same Point Spread Function (PSF)? Other than adding some noise, you seem to be describing performing a convolution process simply followed by then subsequently performing the inverse convolution process. If one wanted to determine the relative noise-immunity of a de-convolution process, then taking such an action does seem to make sense.
However, my own (possibly limited) understanding is that a relevant problem with deconvolution deblurring processes seems to be the artifacts that the processes themselves create. This artifacting does (as I understand it) increase when processing an input with a poor Signal/Noise Ratio (SNR). Is it your understanding that, given an (albeit theoretical only) infinite SNR image-data source, that there would exist no resulting artifacts from a deconvolution deblurring process?
If the kernel is not known, but estimated from the image, it would be called "blind deconvolution". Much of the research is based on similar experiments.
Yes. Cases where the Point Spread Function (PSF) is unknown or unknowable (and various attempts may be made to in some manner estimate the PSF).
So what Adobe did was something based on scientific approach, but that may have not been fully appreciated by the public.
Indeed, blind deconvolution techniques are more common than deconvolution techniques where the PSF is previously known or characterized.
My humble view may be, that it is beneficial for the photographic community that Adobe and others develop usable tools for blind deconvolution. Sometimes we fail on something in taking pictures and it's nice if the failed picture can saved/restored. More often than not it is not possible to reshoot.
It's not clear to me that any person on the planet would argue with your sentiment expressed. What seems more interesting are the practical limitations of the use of deconvolution processes to address optical system de-focusing phenomena and motion-blur. Any (actual) "scientist" speculating about the practical likelihood of the actual successful implementation of such schemes would be subject to "review" (by "peers", as well as by others) in the course of such "scientific approaches" being explored. But Adobe, from their proprietary perch, is not offering anything remotely resembling such a process ...
Adobe is doing marketing, and they of course promise a bit more than what the technology can deliver. They find examples that work and that make impression.
Well, the term "marketing" seems to (one might hope) have a separate meaning than the phrase "deceptive marketing". Courts of Law (who take a "scientific" and "evidence-based" approach) surely do. If these concepts have become fused within our consumer-culture, what does that say about the sensibility of accepting promises by commercial entities that no person actually expects to be truthful and accurate?
Finally, I wouldn't judge anything based on a "you tube" quality presentation.
I take it that you use the term "quality" in a generalized sense above (as it relates to a certain "lack of believability" of anything that appears in that medium)? If so, is it true that corporations that intend to market products for monetary gain may be expected to suspend ethical practices (including an expectation of accuracy and truthfulness) as a result of using such a suspect medium as "YouTube"?
If you use the term "quality" in relation to the actual optical resolution of the presentation, we do not need to rely on hazy glimpses of video shot of presentation-screens published by Adobe on "YouTube". We have Adobe's previously published "Plaza.png" at:
http://blogs.adobe.com/photoshopdotcom/files/2011/10/Plaza.png
When you yourself have a look at the images (now) posted at:
http://juew.org/deblurFamousPhoto.html
... and the general lack of efficacy, as well as the visibly low Signal/Noise Ratio (SNR), of the de-blurring processes demonstrated, have you (as I have) "scratched your head" and wondered exactly why it would be that none of the examples (now) posted at the link directly above appear to come anywhere even close to the seemingly miraculous quality (and relatively high output SNR as compared to all of the other examples published) that is displayed in Adobe's publishing of the "Plaza.png" "before/after" image?
My interest here surrounds the more straightforward and immediate questions relating to a "scientific method", such as "where's the beef?", and "where is the evidence?". Is not that "just the stuff" that the "men in white coats" think about in the hallowed halls that constitute the "cathedrals of science and industry" that seem to have (somehow) earned such an automatic sense of (knee-jerk) deference among members of "the public" these days?
And (from a sociological standpoint) is this modern "techno-tribalism" - where individuals experience an emotional brand-loyalty that is divorced from logic to corporations that they pay money to (and not the other way around) - at all wise and healthy (and a "scientifically sound") trend? Should we elect to "suspend our rational disbelief" merely on the emotional basis that our persona's identification with a corporation (i.e., Apple. Adobe, etc.) may happen to give us a sense of existential meaning, belonging, and a competency imbued from without (on the sole basis of product "consumption" via the liquidation of "disposable income")?
Such questions may not prove to be answered by coherent answers from "the public". But it seems (to me) that the day when nobody either asks such questions, or is capable of answering them, may well be day when the "consumer state" becomes little but the indentured servants of a modern techno-feudalism in the willing servitude of corporate monoliths, where a once "eternal vigilance" has since devolved into a "blissful indifference" ...
Regards, and "top of the mornin' to ya'", kind sir.
PS - I have been perusing some of the links provided at the end of your post, including your personal web-site, and your articles. Some interesting materials, indeed. You clearly are a knowledgeable and thoughtful fellow, and I look forward to the possibility of learning from you and your interesting mind. I can see that the brevity of your post on this thread does not reveal your extensive amount of thought and consideration regarding some of these issues surrounding photography, and I am not in any way underestimating your intellect by raising some questions in my post above about your thoughts on this particular matter of Adobe's recent publicizing of information concerning their developmental work in the area of deconvolution deblurring tools.