New image recognition algorithm from Google is able to recognize CAPTCHA with an accuracy of 99.8%

Today, the Google Street View project team has published interesting news. The fact is that the developers of the service created a new image recognition algorithm in order to more accurately recognize house numbers and street names, photographed by Google cars.


')
And this algorithm turned out to be very effective in another matter - in recognizing CAPTCHA. And Google's captcha from the company's own algorithm recognizes with unprecedented efficiency - 99.8%. This is a much better result than even a man shows. You know yourself how many times you need to type captcha characters before the system recognizes them to be correctly recognized.

The old image recognition algorithm worked on house numbers with an efficiency of 90%, which is quite small, because the volumes of the analyzed information are simply huge, and some errors were laid out on Street View. At the moment, the system has already recognized about 100 million numbers worldwide. For the operation of the new algorithm, neural networks are used (some technical details, in English, here - deep convolutional neural network ). And here is a joint article by Street View and reCAPTCHA teams on the problem of recognizing house and street numbers.

As for the captcha, the Google developers are not worried about this. On the contrary, it has now become clear how you can still protect your own CAPTCHA so that it can be worse recognized by automatic systems of various kinds of intruders / spammers. In general, reCAPTCHA will become even more reliable, and there will be even fewer errors on Street View.

Via techcrunch