Deconvolution sharpening

[TonyW]

In another thread, Dave Ellis mentioned deconvolution sharpening. It is clear from basic mathematical principles that this cannot exactly negate the effects of diffraction blur and other types of blur but nevertheless could be effective.

My question really is what software is available to implement it. Is it in Photoshop, for example? I have not seen it there, at least not by that name.

[dje]

Tony it difficult to get hard and fast information about Adobe software (for obvious commercial reasons) but from what I've read (including comments from people "close" to Adobe), deconvolution sharpening is used in ACR/Lightroom and PS.

In ACR on the sharpening and NR tab, there is a slider called "Detail". I believe if this is positioned on the RHS, de-convolution sharpening is employed whereas if it is positioned on the LHS, something closer to Unsharp Mask is employed. In between there is a blend of methods, whatever that means ! I believe it is also employed in Smart Sharpen in PS but I'm not sure about that.

De-convolution sharpening is also available in RawTherapee and is used in FocusMagic I think. It's probably used in some of the plug-ins like Topaz and Nik too but it's not specifically identified as such.

Dave

[NorthernFocus]

It is the "claim to fame" for FocusMagic. In spite of the name, magic it ain't. Granted I haven't used it extensively but I don't see a dicernable difference with USM when applied to the same image. But my information is dated and I didn't pursure it extensively because after a small sampling of tests it did not appear to have any obvious advantage.

[Richard Lundberg]

Focus Magic has improved in its latest incarnations. There are several other apps, just google it. I have all of them except the expensive one. None of them were able satisfactorily to sharpen an image I scanned of a modem label that was about 1/16 of an inch above the glass. It is true that LR Details all the way to the right is deconvolution.

[TonyW]

Thanks Dave, Dan and Richard.

Since I have Photoshop and have some idea about how to use it, out of laziness I am disinclined to look at alternatives unless there is a clear advantage.

I see the Detail slider in ACR and will experiment with it. So far it hasn't done wonders. I often use Smart Sharpen in Photoshop, especially when there is a likelihood that noise will occur, and I find that it is effective to a certain extent. I don't know what method it uses.

[Richard Lundberg]

not sure Detail in ACR is the same as LR but it should be.

[tclune]

I'm sure there are others, but FocusMagic and TopazLabs' InFocus both employ deconvolution methods. I have both applications and use FocusMagic routinely. I find the UI of InFocus hard to understand, while FocusMagic is very simple to use. I use it as a plugin. As with any post-processing routine, it works best when it is needed least. Generally, the suggested sharpening is a bit too much with the application. There is a motion blur "undo" option, but I've never found it to produce usable results (not surprisingly, I guess.)

Deconvolution appears to be unrelated to unsharp masking, in that I have tried pushing USM as far as it will go and deconvolution as far as it will go on an image. You can combine the two without getting artifacts (I'm not recommending that you do this -- I just wondered whether you'd see the Gibbs effect sooner with one after having done the other. From what I could tell, you don't.) FWIW

[xpatUSA]

RL deconvolution is available in Raw Therapee as a specific alternative to unsharp mask and can vary radius, amount, damping and no. of iterations. It uses the Gaussian model but it works well and is my preferred method for slightly blurred shots.

Eric Chan (Adobe scientist involved with ACR development) has confirmed elsewhere that the ACR detail slider begins to apply it above a certain setting. Sorry, no link for that. Search LL.

[ajohnw]

I don't think that deconvolution is a correct name for it. If you google that this is what comes up.

In mathematics, deconvolution is an algorithm-based process used to reverse the effects of convolution on recorded data. The concept of deconvolution is widely used in the techniques of signal processing and image processing.

If you look at the wiki on the subject you will find a link to convolution and some where down that it will mention that an out of focus image is a convolute of a sharp image. It's a very very intensely mathematical area. So deconvolution usual means either focusing an out of focus image or if they are known removing the optical errors that are in a lens. This is probably what the military gave to the Hubble project before it obtained it's spectacles. Initially refocus routines where called blind deconvolution and there were a few around on the web pointing out that lacking info on what the problem is can be problematic but they did improve things. I'd strongly guess this is now being called deconvolution sharpening.

But there is also convolution. That can do all sort of things, blur or sharpen. It's used by people who are into image restoration - astronomy, microscopes etc. I fail to see how they decide on what to put in the matrix that is used but having seen a load of tat changed to a half decent image they can. There is a facility for this in the GIMP and a video that explains what goes on and gives a couple of examples.

https://www.youtube.com/watch?v=DogFik9p6jM

The whole area is a very good way of getting a very extreme headache.

John
-

[Richard Lundberg]

Deconvolution is exactly that. Blind deconvolution works because there are so many convolutions in the image capture process that a Gaussian function works very well (central limit theorem).

[ajohnw]

Only within limits Richard. Past some point artefacts appear and detail is lost. In other words too high a degree of convolutes. They can be very apparent for instance past some out of focus detail level. Strong contrasty detail may be restored but if for instance some ones face is in the image at some point it will loose it's "shape" and that is likely to be distorted at relatively low out of focus levels. This is why I wonder if deconvolution sharpening actual is but there are are degrees of how much is applied. I did have some examples but a silly lost all of my processed shots recently. I deleted a desktop folder by mistake and also messed up recovery. Silly me. I have now replaced the desktop entry with a link to the real folder. Fortunately I can restore the base photo's. It's about time I had a clear out anyway.

The GIMP uses a convolute matrix for sharpening with a brush. I can't help wondering if that is what this type of sharpening is. A question on the G'MIC site would probably answer that as it's supposed to have it but all I can see is sharpen - deblur. Probably AKA blind deconvolution.

John
-

[revi]

Only within limits Richard. Past some point artefacts appear and detail is lost. In other words too high a degree of convolutes. They can be very apparent for instance past some out of focus detail level. (...)

John
-

The artefacts are typical of appying a Fourier Transform to a sharp edge (ringing).
A gaussian curve is basically infinite (never reaches zero), so in practice you have to cut it after a certain distance. That creates a sharp edge in your function (even a discontinuity), and thus the ringing artifacts. If you use a moderate intensity, the ringing will be undetectable by eye, but if you are applying a too 'intense' deconvolution, the artifacts will be visible (and they are ugly).

Increasing the area used in the calculations could diminish the visibility of the ringing, but that increases processing time quite a bit ( O(N²) iirc).

Note that the gaussians here are already present in the image: they are the density distribution around each of the points making up the image, and deconvolution tries to reduce them to their optimal size: a pixel. So there is no way you can avoid the sharp cut-off in the processing.
And of course we have to use 'blind' deconvolution: there's no easy way to get the specific set of deconvolution functions needed for each point of the image (and the required functions can vary over an image, e.g. for focus correction).

[TonyW]

This is an interesting discussion. Any convolution which represents blurring is usually something like a discrete approximation to Gaussian blur, which is analogous to the diffusion of heat in a solid. It is well known that a solution to the heat equation, which is a form of convolution, exists for any starting conditions but if you try to solve the equation backwards, which would be a form of deconvolution, a solution does not exist in general and that the process is unstable. Using deconvolution to remove blur has the same problems but nevertheless it should be possible at least in some cases to reduce the blur to a limited extent before instabilities become too great.

[ajohnw]

What I found Revi is that sharp detail, grass, brick wall, may be restored rather well on a fuzzy image but faces in the image loose their true "shape". The features became angular, the nose depth is distorted and other features flatten and become angular. A striped tee shirt was restored and kept it's "shape". This suggests that in situations where the sharper features in an image are close to some degree it can work well. It falls down on a smooth face because there aren't any other than the nose and the eyes. The eyes come up well, the nose gets pointed and the facial features flatten and tend to become angular plains.

The Hubble aspect is a little different. It seems that there are algorithms around that can reduce certain optical aberrations. On the other hand image restoration by experts in that area seems to be a little prescriptive. In other words the person doing it knows what should be there. There was an example in one of the UK astronomy magazines some years ago that restored a poor image of the moon. I became interested in convolution filters. The best option seems to be to search for convolution kernels on the web. A rather expensive book didn't really help at all.

In terms of this thread I still question if deconvolution sharpening is a form of refocus and not really a form of convolution sharpening. There are a few kernels on this page - one takes a blurred image and cleans it up.

http://pippin.gimp.org/image_processing/chap_area.html

The results of unsharp on this one is interesting and it links to an unsharp mask section

http://en.wikipedia.org/wiki/Kernel_(image_processing)

Me, well I look around the web at kernels from time to time but so far don't use them.

John
-

PS There is an interesting pdf on sharpening on this page - With a Stiletto.

http://en.wikipedia.org/wiki/Unsharp_masking

-

[xpatUSA]

It has been mentioned that "blind" deconvolution can be less than satisfactory. For example, the radius of deconvolution required is not necessarily known but is set by twiddling the slider.

Deconvolution sharpening has been discussed at greater depths than in this thread over on Luminous Landscape (use their search). In particular, Bart van der Wolf discusses the technique and introduces a method to determine a radius here:

http://www.luminous-landscape.com/fo...?topic=68089.0

An example from there:



As to the original post, ImageJ and ImageMagick are mentioned there as being able to apply deconvolution sharpening.

[dje]

Yes Ted I have a lot of respect for Bart's knowledge in such matters. His web based tool works quite well.

Dave

[TonyW]

It has been mentioned that "blind" deconvolution can be less than satisfactory. For example, the radius of deconvolution required is not necessarily known but is set by twiddling the slider.

Deconvolution sharpening has been discussed at greater depths than in this thread over on Luminous Landscape (use their search). In particular, Bart van der Wolf discusses the technique and introduces a method to determine a radius here:

http://www.luminous-landscape.com/fo...?topic=68089.0

. . .

As to the original post, ImageJ and ImageMagick are mentioned there as being able to apply deconvolution sharpening.

Thanks for the link and info about ImageJ and ImageMagick. I must look them up

[Richard Lundberg]

ImageJ is free and not really an editor but a processing tool. Its FFT is excellent. ImageMagick is also free but extremely geeky.
Thanks for the concise explanation of FFT ringing, revi, and to TonyW for the note on deconvolution instability. I have been able to get good results on most images, others have ringing before success. Motion Blur reduction in Focus Magic also has worked well in some instances where it's possible to determine the direction.

[TonyW]

I tried to use the custom filter in Photoshop to experiment with deconvolution sharpening but unfortunately it does not allow non-integer entries in its filter matrix. Pity!