In order to provide users with accurate directions on their maps, Google developed an algorithm that can recognize he numbers of the houses that appear in their street view images with high level of accuracy. However, the same algorithm could be applied in reading CAPTCHAs very precisely.
For some it is a safety need. For others , it is a pain in the heart of mankind. Indeed, the CAPTCHA has a little of both. Although we can not let the malicious bots sow chaos in the Web, the average user accumulates several defeats against a system that seems to spit a mixture of Chinese and Martian . Many methods have been developed to avoid CAPTCHA challenges, with varying degrees of success, but what we have here today is not even the CAPTCHA as a priority.The story begins with Google Street View . If you need an exact address, Google is doing everything possible to deliver it with maximum precision. Teams of Street View and reCAPTCHA developed a new algorithm that reads the numbers on the houses directly from the images, and from these data, the right direction is calculated.
We are facing an extraordinary example of visual recognition , which involved an advanced neural network . When working with regular Street View images, the algorithm has an accuracy of 90 percent . When applied on the “Street View House Numbers Dataset” from the University of Stanford(used in training the neural network) , the number rises to 96 percent , and to recognize individual digits, is 97.8 percent . Still, Google decided in their study that the algorithm takes on some of the more complex examples generated by reCAPTCHA . The result was a chilling 99.8 percent accuracy.I’m not sure that a human can match such performanc.: Google shared in its blog four of the reCAPTCHA that the algorithm recognized, and being honest, I barely managed to read two of them.
Thanks to this “side effect” , Google and the reCAPTCHA team may reinforce the classical verification system. They found that distort text is no longer sufficient, and if it is possible to achieve this level of accuracy in an automated way. The likelihood that someone will develop the equivalent of the neural network to break the entire Google’s CAPTCHA structure are very low, but there is a possibility that a group of intelligent individuals can build a machine to beat it.
