While not AI based, Silverfast has a plugin for Photoshop called SRDx - Smart removal of defects. It is not perfect (does not work well in high-frequency detail areas) but it works better than the Photoshop filter IMHO.[/url]
Yes, I tried it a while back and I agree: it's better than the Photoshop filter, but its primary benefit is in open areas—especially those that can be masked off—where spots and scratches are particularly apparent. It's of limited value in areas with high-frequency detail. It has been a while since I've used it, however, and probably I should check the current trial version of the plug-in. Thanks.
My feeling is that if some company has been willing to go to the effort to create a good training set of images, reducing if not completely eliminating spots and scratches should be a tractable problem with current machine-learning techniques. Of course there is a threshold economic issue of making the investment to create the training set, which no doubt is labor-intensive: you would need to collect enough original scans and their competently retouched counterparts for the neural network to "learn" how to differentiate between the source and destination images. I know Microsoft has done some research in this area and even offers a product (MS-
Windows only, alas), but I'm hoping someone knowledgeable with no stake in the outcome has reviewed the various alternatives for a post-processing (i.e., post-scan) workflow.