Scientists have recently created a new artificial intelligence (AI) system that can identify individuals on a video by detecting age and gender more rapidly and with high accuracy. According to the researchers at Russia’s Higher School of Economics, the technology to detect age and gender has already been the basis for offline detection systems in Android mobile applications.
Modern neural networks could detect gender of people on videos with highest accuracy of 90%. However, the conditions for predicting the age has always been much more complicated. Existing neural networks look for discrete values of age; in each video frame, the network evaluates the probability of the individual in the image being of a certain age.
For instance, if in 30% of the video frames, the network’s top prediction is the person’s age as 21 years and in 10% as 60 years, it concludes that the person is 21 with a probability of 30% and 60 years with a probability of 10%. Due to varying conditions of observations and even slight rotation of head, an individual’s age prediction in different video frames differs in the range of 5 years, more or less.
In a new research study, the researchers discovered a way to optimize the operations of neural networks. They implemented a novel technique to assemble confidence levels generated by neural network for each video frame. Software systems integrated for facial recognition analysis usually have multiple separate neural networks; while one identifies a person, another determines the gender, and so on.
The researchers developed a novel neural network with numerous outputs which conducts several tasks simultaneously such as predicting age and gender, producing a set of 1000 numbers that particularly characterize each individual, and allowing them be distinguishable from other people.
The researchers further explained that the software can also be operated with low-performance smartphones.