Is that even possible at the moment? Most software I've tried just isn't ready for something like that (granted, I haven't tried any of the commercial products). Tesseract is pretty good, but definitely not perfect.
The general answer to the character recognition problem is negative (currently not possible). However, several important subproblems are solvable. For example, if you know that a particular image came from a printed text, or from handwriting, or a (printed) form, etc, there are some very good solvers. Not perfect (neither is a human, in case of handwriting or a poor fax), but good to very good.
For further googling, see terms ICR, OCR, character recognition.
As a user, I have had very good results with Finereader. Have not tried Tesseract. Parascript was good with online character recognition, but that market is small, and I have not looked at them for a while (disclaimer: I used to work in a previous incarnation of the company).
"Not perfect" is good enough for me. There's a whole class of applications where being able to take a photo of some words and have at least a few of them understood would be useful. I'd like to be able to index against the words that can be understood, and present the original image as the search result.
I think it would be a fantastic idea be able to hold up a cameraphone, take a picture of a sign in a foreign country, send it to the cloud and have the text recognised and Google-translated.
"Not perfect" would be fine for many non-life-or-death signs.
