Re:Google's Street View computer vision can beat reCAPTCHA w...
by Theodores on Thursday April 17, 2014 @06:40AM
(#7602838)
> The street address/book scan approach that Google uses is interesting in that the exact solution is not known, so they presumably have to be somewhat forgiving in accepting answers (as their machine learning might have gotten it wrong).
In the original captcha algorithm there are two images shown, one where the answer is known and one where the answer is not known. The unknown - the street sign or that extract from an old newspaper - is guessed at by those filling in the captcha.
If 90 people say it looks like '11' and 5 people say it looks like '71' and another 5 say it looks like '77' then you can be 90% certain that it is '11'. The machine has therefore 'learned' what that number is, it is '11'. This can continue being used in the captcha however, now that it is known, people will have to type '11' for it.
That is how it fundamentally works and why the two snippets of information 'to get right'. You can try it for yourself, get the first one of the street sign deliberately wrong and see if it lets you through.
In the original captcha algorithm there are two images shown, one where the answer is known and one where the answer is not known. The unknown - the street sign or that extract from an old newspaper - is guessed at by those filling in the captcha.
If 90 people say it looks like '11' and 5 people say it looks like '71' and another 5 say it looks like '77' then you can be 90% certain that it is '11'. The machine has therefore 'learned' what that number is, it is '11'. This can continue being used in the captcha however, now that it is known, people will have to type '11' for it.
That is how it fundamentally works and why the two snippets of information 'to get right'. You can try it for yourself, get the first one of the street sign deliberately wrong and see if it lets you through.