Re: Linearized computations - was OCR
On Fri, Aug 10, 2001 at 01:28:20PM -0400, Trei, Peter wrote:
I think text is the wrong approach. Put up some pictures, and a question - 'Click on the white bunny' 'Click on the chipmunk' 'Click on the Russian icon'. 'Click on the clowns nose' 'Click on the clown's right hand'
Taking text, which is explicitly designed to be readable, and trying to find a degree of distortion which is human (but not machine) readable is probably a bad approach. Try something a 6 year old can do, but which is still a PhD thesis problem for computers.
I had thoughts along a similar line as Peter's, except to use numbers. Include with <img> tags some .GIFs and ask the person/bot to type in the numbers in the GIFs. Generate the filenames randomly and expire them after five minutes. Should take you 10 minutes to write a prototype CGI script in Perl. -Declan
Declan McCullagh <declan@well.com> wrote:
I had thoughts along a similar line as Peter's, except to use numbers. Include with <img> tags some .GIFs and ask the person/bot to type in the numbers in the GIFs. Generate the filenames randomly and expire them after five minutes. Should take you 10 minutes to write a prototype CGI script in Perl.
This is done somewhere already. I can't quite remember where, but I suspect that it's in the EBay registration process. "For security purposes, retype the numbers to the left." The numbers aren't vertically aligned, have erratic horizontal spacing, and are presented against a grid-pattern background. I'm not sure if this actually helps much against standard OCR, although I'm pretty sure that the SVMs I work with could be easily trained to translate the GIFs to text. -- Riad Wahby rsw@mit.edu MIT VI-2/A 2002
participants (2)
-
Declan McCullagh -
Riad S. Wahby