Perl programs are, by definition, text. So "paint splatters are valid Perl" implies that there's a mapping from paint splatters to text.
Do you have a suggested mapping of paint splatters to text that would be more "accurate" than OCR? And do you really think it would result in fewer valid Perl programs?
I agree that a symbolic representation of the splatters would probably be more interesting. The whole point is that random character sequences are often valid Perl, though, so changing the generation method wouldn't change that aspect.