CAPTCHAs, those visual jumbles of numbers and letters, have been used to separate the humans from the bots because software couldn't read visuals all that well. Now, Google has created software for its Street View cars to read street numbers in its images -- and may have overturned the value of CAPTCHAs in the process.
In a post Wednesday on the Google Online Security Blog, the company's reCAPTCHA Product Manager Vinay Shet noted that the technology finds and reads street numbers in Street View imagery, and then correlates those numbers with existing addresses so that they can be shown on Google Maps. The software has been described in a scientific paper presented at the International Conference on Learning Representations 2014. It is able to detect and read difficult numbers with a 90 percent accuracy, which can reach 96 percent in some cases.
Street View obviously needs such rigorous software if the millions of recorded street addresses are going to be available through Google Maps without manual tagging. Weather conditions, varying quality in street numbers and signs, and lighting conditions also add to the difficulty of the problem.
99 Percent Accuracy
To accomplish this task, Google brings together the three main components of such visual interpretation -- localization, segmentation and recognition -- through the use of a neural network that is optimized for image recognition.
CAPTCHAs have been used for longer than a decade to prevent automated software from conducting transactions on Web sites, but Shet says the Street View technology can decipher the "hardest distorted text puzzles" with more than 99 percent accuracy.
In fact, according to various anecdotal reports on the Web, this accuracy rate is now far higher than what many humans can achieve in trying to figure out the distorted and twisted text. And, since accurately reading street numbers is a frustrating task for any carbon-based lifeform, Google's technology is likely outperforming humandom there as well.
ReCAPTCHA is a free service from Google to keep automated software agents out of Web sites. The solved text or deciphered images have also been used to help digitize hard-to-solve text, to annotate images or to build machine learning datasets, as humans correctly deciphered parts of old texts that standard optical character recognition software could not.
'Advanced Risk Analysis'
Google's reCAPTCHA department apparently knew where the Street View department was heading, so last year the company announced reCAPTCHAs would reduce their reliance on text distortions to tell humans from bots.
In October, Google said that reCAPTCHA was updated to use "advanced risk analysis techniques, actively considering the user's entire engagement with the CAPTCHA -- before, during and after they interact with it."
To accomplish that, different classes of CAPTCHAs were released for different kinds of users. More details were promised in the next few months. At the moment, however, how Google tells which of us are real and which of us are algorithms is a dark secret, probably kept in the same box as its search engine ranking routine.