Hello Mark and Bill. I believe that's because deconvolution cannot restore out of focus images: defocus and dof cannot be modeled accurately as convolution with a blurring kernel.
Hi Jack,
Those are two different things and I can add some other detractors too. The point is that mathematically, convolution can be exactly reversed by a deconvolution, provided we can use high enough precision calculations. It's like dividing and multiplying, e.g. 1/3*3 = 1 , although some calculators insist on a different outcome due to lack of precision. Convolution is not much more than multiple multiplications and additions, and deconvolution does the inverse order of operations.
The real issue is finding the exactly correct Point Spread function (PSF) that caused the defocus , and then we also need to deal with noise. In practice we do not know the exact convolution PSF, so this then becomes a process of (user assisted) blind deconvolution with noise regularization. And even then, we will find that precision is not enough to fully restore all detail, although we can come a long way.
Attached I've added an example of a synthetic image, a noisy image would be harder to find a model for:
1. A sinusoidal star image,
2. An image convolved with a defocus kind of PSF,
3. The PSF,
4. The result after deconvolution.
I stopped at 10000 iterations of the special Deconvolution algorithm, because this small image already took some 47 minutes to complete, a full size image would take days.
t's not a perfect restoration, but it does achieve a significant restoration of what seemed to be lost. I'm just showing that there is more possible that acutance enhancement.
In images captured with excellent technique there are very few quantifiable sources of blur left (diffraction, AA, pixels, various lens aberrations). Restoration by deconvolution works best when those are known (in type, if not quantity).
Yes, but then there will still be other factors, like the demosaicing of the sensor signal and of the noise.
Also the a6300 does not appear to have an AA filter, its green channel hitting Nyquist around 0.25 cycles/pixel with an FE55mm at f/5.6.
The presence of aliasing artifacts will again make it harder to restore the original image signal.
Cheers,
Bart