It's ironic that these hard problems, such as character recognition, have been known to be hard for a long, long time, and yet almost as soon as people make crypto things out of them, they get solved. Actually it's not quite the way you think because what's happened is that for the examples that get automatically generated, there are special techniques which work just because they've been created deliberately. It's not trivial to produce things that are really hard to solve, and some of the ideas for distorting characters have been quickly broken, but I believe there are some around which are quite robust.
展开▼
