MIT scientists have learned to more accurately recognize images

A team of scientists from the Laboratory of Computing and Artificial Intelligence of the Massachusetts Institute of Technology (MIT) published an article with intermediate results of its work in the institute edition of TechTalk. Specialists under the leadership of Antonio Torralba (Antonio Torralba) have developed a new technology of image indexing, which will help to quickly and more accurately recognize objects in digital photos and look for similar ones, for example, on the Internet.



The essence of the new method is a more sophisticated, in comparison with modern algorithms, technique of creating an image hash, thanks to which its size does not exceed 1 Kb. For clarity, in their experiments, scientists worked with a database describing almost 13 million photographs, the size of which was only 600 MB.



Despite the fact that the work is still far from being published, Torralba mentioned that a large role in pattern recognition is assigned to the definition of the object's shape and its comparison with the base of the samples. In any case, apparently, the rapid production of such a miniature hash requires much more computing power than modern desktop PCs can provide.