There are a number of lens distortions that seem like AI technology could fix. For instance coma distortion and astigmatism distortion as well as chromatic aberration. These problems are most easily observed in taking images of stars in the night sky. Coma distortion tends to stretch the image away from the center and produces elongated stars towards the corners. Astigmatism distortion tends to produce stars towards the corners that tend to look like flying seagulls. And of coarse, CA is where the RGB colors don’t converge at the same place.
And where would be the best place to offer such correction? Would it make the most sense to provide a way to perhaps take an image of point sources such as stars and to learn the corrections needed for a given lens and to later, be able to select that correction for that lens to be used on any image?
I can see that the ability to correct lens distortions could allow a mediocre lens to be corrected to the point that it produces results similar to much more expensive lenses.
And of course, to be able to batch this type of correction would be most useful.
It may be possible that these types of corrections with AI may produce results that conventional methods have tried to produce for years and have not succeeded very far?
Hi Jared, I was one of the first to suggest to Canon when videoing with the XLH-1 camcorder which showed signs of colour aberration. I printed out a couple of A4 pages with a black dot matrix on white paper. When I took a single video frame, the red green and blue dot displacement was easily observed. This could by easily ‘fixed’ by applying corrective displacement to the R G and B layers in the Photoshop image. I supplied Canon UK with a document showing how to do it, the factors involved and how much the zoom lens affected. My suggestion was to use digital correction. I never heard back. Nowadays (pioneered by Panasonic I believe) ALL video cameras and some still cameras have this sort of correction built in. Post processing requires a clever ‘diagnosis’ - especially if using a camera that has interchangeable lenses. Having pre-characterised maps for different lenses is an onerous task for something like a Topaz, which is essentially a unique post processor with no specific database of lenses. To do it ‘on the fly’ would be impressive indeed!
