removing the AA filter

[ErikKaffehr]

Hi,

It's actually quite possible to regain information if the Point Spread Function (PSF) is known.

Not that I have expertise in the area, but you may check "Digital Image Processing" by Gonzales and Woods or google on "PSF deconvolution Richardson Lucy" and you ought to find examples.

An example is here: http://www.bialith.com/Research/BARclockblur.htm

Best regards
Erik

NO, you cannot regain lost sharpening, you will only increase contrast, and preceived sharpness, no matter the channel, it is applying the same effect.
Some result better than others, but even with some samples posted here, you will see that you CANNOT regain the information that native sharpness resolved.

[Phil Indeblanc]

Hi,

It's actually quite possible to regain information if the Point Spread Function (PSF) is known.

Not that I have expertise in the area, but you may check "Digital Image Processing" by Gonzales and Woods or google on "PSF deconvolution Richardson Lucy" and you ought to find examples.

An example is here: http://www.bialith.com/Research/BARclockblur.htm

Best regards
Erik

Erik, it would be great to see the revits reappear from the 50D image crop you posted.  The SD14 shows them nice and seprate.  Perhaps a visual would be most convincing.  I have invested too much time with all this and so far ZERO visual support[size="2"][/size].  Also, there are I am sure examples that might be able to do this as calculations are obvious ....but the tiny revits is a good example to try and REVIVE!

[ErikKaffehr]

Phil,

I don't understand the question. If you mean the rivets I don't really know which rivets you mean, If you could encircle that part on one of the original photographs I could see what you mean.

Best regards
Erik

Erik, it would be great to see the revits reappear from the 50D image crop you posted.  The SD14 shows them nice and seprate.  Perhaps a visual would be most convincing.  I have invested too much time with all this and so far ZERO visual support[size="2"][/size].  Also, there are I am sure examples that might be able to do this as calculations are obvious ....but the tiny revits is a good example to try and REVIVE!

[Phil Indeblanc]

Phil,

I don't understand the question. If you mean the rivets I don't really know which rivets you mean, If you could encircle that part on one of the original photographs I could see what you mean.

Best regards
Erik

Yes, there is a recatangle shape area that has rivets in it.  the difference is really not much at all as this again is 4.6 vs 15.1mp.

But I am taking a few images now with Non AA camera, and will post in the next 30min?

[Phil Indeblanc]

hmmmm, I dont post images online, and not having much luck.  Anyone have a way I can upload. I dont have any image hosting places.

[Bart_van_der_Wolf]

NO, you cannot regain lost sharpening, you will only increase contrast, and preceived sharpness, no matter the channel, it is applying the same effect.
Some result better than others, but even with some samples posted here, you will see that you CANNOT regain the information that native sharpness resolved.

Phil,

As Erik already said, YES it is possible to restore blur, to a significant degree anyway. There is a lot of new research that shows even better results than from a regular Richardson Lucy (RL) restoration. You can convince yourself that RL works with a free implementation that apparently is implemented in the RawTherapee converter.

Cheers,
Bart

[bjanes]

Phil,

As Erik already said, YES it is possible to restore blur, to a significant degree anyway. There is a lot of new research that shows even better results than from a regular Richardson Lucy (RL) restoration. You can convince yourself that RL works with a free implementation that apparently is implemented in the RawTherapee converter.
Bart

Bart,
Roger Clark has also demonstrated use of RL in lieu of or in addition to traditional unsharp masking. However, he did not explain how he derived the point spread function parameters (PSP) used for that restoration. Just guessing? I understand that deconvolution can also be used for defocusing error and and diffraction correction. Some photographers report good results using deconvolution (FocusMagic) to offset the effect of a blur filter in capture sharpening. Unfortunately, it is difficult to derive the PSP function to be used in the deconvolution. Do you have any pointers in this area?

[Phil Indeblanc]

Phil,

As Erik already said, YES it is possible to restore blur, to a significant degree anyway. There is a lot of new research that shows even better results than from a regular Richardson Lucy (RL) restoration. You can convince yourself that RL works with a free implementation that apparently is implemented in the RawTherapee converter.

Cheers,
Bart

I am open to it, I will try it out....is this method also available to do using PhotoShop, or is this something Focus Magic specific.  I will have to check out the links.

It is a bit hard for me to believe it.  I almost took the chance today and purchase a 50D as it is the closest to my 14mp, so I will see within the week, as I was thinking of waiting about 3-4months and purchase the 5dMk2.....But I don't need either as I have enough bodies, and they are secondary to my MFDB, so the 50D is still something I might get...ok, enough rant.

How can I post my raw files?

[Bart_van_der_Wolf]

Bart,
Roger Clark has also demonstrated use of RL in lieu of or in addition to traditional unsharp masking. However, he did not explain how he derived the point spread function parameters (PSP) used for that restoration. Just guessing?

Hi Bill,

Yes, it is basically guessing (although with visual feedback from a small image area that guess can be pretty good!), unless one has a good method of determining the characteristics of the optical system. I use Imatest to calculate the PSF of my optical system (lens/aperture/AA-filter/microlens/sensel aperture), and I am able to derive a pretty good PSF that way. The guessing part also works fine though     .

I understand that deconvolution can also be used for defocusing error and and diffraction correction. Some photographers report good results using deconvolution (FocusMagic) to offset the effect of a blur filter in capture sharpening. Unfortunately, it is difficult to derive the PSP function to be used in the deconvolution. Do you have any pointers in this area?

The interesting thing is that an application like FocusMagic can already do an impressive job based on a single (radius) parameter. It is capable of tackling defocus/diffraction/AA-filtration, even motion blur, which all have distinctly different PSF shapes, so I assume its PSF assumption may be a bit more elaborate (or an average of different shapes) than a simple defocus model of a PSF, or the algorithm is somewhat adaptive. Anyway, it proves that with limited input one can already achieve impressive improvement results.

As an example, a C1 conversion of an EOS-1Ds3 image at f/8 without sharpening, and the FocusMagic version with a small PS layer blending tweak:
[attachment=19347:OPF_Debl...rop01_1_.png] [attachment=19348:OPF_Debl..._125__1_.jpg]

IMHO those who seek for a solution by removing an AA-filter will add more problems than the benefits of an AA-filtering + proper sharpening can deliver ...

Cheers,
Bart

P.S. I've kept the unsharpened image in the losslessly compressed PNG file format to allow comparisons with other sharpening utilities.

[Jonathan Wienke]

NO, you cannot regain lost sharpening, you will only increase contrast, and preceived sharpness, no matter the channel, it is applying the same effect.
Some result better than others, but even with some samples posted here, you will see that you CANNOT regain the information that native sharpness resolved.

As others have mentioned, sharpness can be restored if the point spread function (think of it as a lens blur profile) is known. I'm currently working on a software program that does this exact thing. All in all, my impression so far is that it is better to use an AA filter to filter out aliasing artifacts, then use deconvolution to reverse the blurring of frequencies below Nyquist.

[Jonathan Wienke]

Several new and exciting technologies are going to hit the imaging world and I think it is about time if we find aliasing to be too objectionable to just use a camera which has an AA filter and move on.

Care to share what those might be?

[joofa]

Care to share what those might be?

Yes, for one thing: Hyper-/multi-imaging. I have talked about it a little here on this forum before. We have worked a little on that and are blown away at the prospects of it - ranging all the way from food inspection, to digital special effects especially for filmmaking, to medical imaging, to early-cancer detection, to user-specific camera curves resulting in more control than traditional 3-color RGB, to digital pathology, to ....

However, I'm not sure if Adobe/Apple/etc. are yet ready for it. New algorithms might need to be developed for such data.

Roger Clark has also demonstrated use of RL in lieu of or in addition to traditional unsharp masking. However, he did not explain how he derived the point spread function parameters (PSP) used for that restoration. Just guessing? I understand that deconvolution can also be used for defocusing error and and diffraction correction.

Those looking for a general solution might google "blind image deconvolution" for which you may not need to know the psf. However, in theory that is fine, but in practise many constraints are imposed such as the Gaussian assumption that helps in closed form solutions for methods such as maximum-likelihood. Richardson-Lucy algorithm can be shown to be a special case of these general frameworks.

[Bart_van_der_Wolf]

As others have mentioned, sharpness can be restored if the point spread function (think of it as a lens blur profile) is known. I'm currently working on a software program that does this exact thing. All in all, my impression so far is that it is better to use an AA filter to filter out aliasing artifacts, then use deconvolution to reverse the blurring of frequencies below Nyquist.

Indeed, I do as well, and it's my impression that the vast majority of the Digital Signal Processing (DSP) community is of the same opinion. When over-sampling is impractical, add a low-pass filter before sampling and then restore.

Cheers,
Bart

[Bart_van_der_Wolf]

Those looking for a general solution might google "blind image deconvolution" for which you may not need to know the psf. However, in theory that is fine, but in practise many constraints are imposed such as the Gaussian assumption that helps in closed form solutions for methods such as maximum-likelihood.

The main issue with blind deconvolution methods is noise amplification.

Richardson-Lucy algorithm can be shown to be a special case of these general frameworks.

In a way, yes because RL is based on a Bayesian approach to Poisson noise (just like Photon shot noise), but with improvements like additional insensitivity to random noise.

Cheers,
Bart

[Phil Indeblanc]

As others have mentioned, sharpness can be restored if the point spread function (think of it as a lens blur profile) is known. I'm currently working on a software program that does this exact thing. All in all, my impression so far is that it is better to use an AA filter to filter out aliasing artifacts, then use deconvolution to reverse the blurring of frequencies below Nyquist.

Maybe I can email the file to someone who can post it.

[ErikKaffehr]

Hi,

Many of us have used:

http://www.yousendit.com/

For sending pictures.

You can mail to yourself. Yousendit sends a short mail with a download link. You can just post the download link on this forum.

Maybe I can email the file to someone who can post it.

[ErikKaffehr]

Hi,

RawDeveloper from Iridient has also Richardson Lucy, in part probably because RL is part of the image processing libraries on Apple. I presume that it's based on a guess regarding PSF.

My guess is that it may be possible to guestimate PSF in some cases. The AA-filter acts a beam splitter so it probable deflects some of the light in four adjacent cells.

http://books.google.com/books?vid=ISBN0240...2yfoNY#PPA30,M1

Best regards
Erik

Bart,
Roger Clark has also demonstrated use of RL in lieu of or in addition to traditional unsharp masking. However, he did not explain how he derived the point spread function parameters (PSP) used for that restoration. Just guessing? I understand that deconvolution can also be used for defocusing error and and diffraction correction. Some photographers report good results using deconvolution (FocusMagic) to offset the effect of a blur filter in capture sharpening. Unfortunately, it is difficult to derive the PSP function to be used in the deconvolution. Do you have any pointers in this area?

[Phil Indeblanc]

Hi,

Many of us have used:

http://www.yousendit.com/

http://www.yousendit.com/download/VGlmbUpRNDRHa04zZUE9PQ

[Phil Indeblanc]

http://www.yousendit.com/download/VGlmbUpRNDRHa04zZUE9PQ

This is the longest pause this post has had...anyone DL the image and look it over?

[Phil Indeblanc]

http://www.yousendit.com/download/VGlmbUpRNDRHa04zZUE9PQ

I just opened this image up to expose it a bit, and I lowered the color noise some also, and I hardly added some sharpening, and it is rather tack without any sharpening.