A neural network developed by Russian scientists learned how to determine age from a video with a high degree of accuracy.
An authoritative scientific publication Journal of Physics published an article that describes the results of work on the creation of a neural network that determines the age and sex of a person.
The developers in question implemented their project at the Higher School of Economics under the leadership of Andrei Savchenko. The team proposed a new method of analyzing human data by video, this method is the basis for the work of the neural network, which belongs to the convolutional class .
The method itself consists of frame-by-frame video analysis, with the release of individual frames on separate frames. Further analysis goes in two directions. The first allows you to determine the average age of a person, the second - his gender. As usual, the neural network was originally trained, the base of the video that served as the core of training includes 1165 videos.
The authors indicate that their neural network can determine a person’s age with an accuracy of about 71%, gender - 88%. The authors plan to use their development to create a mobile application for Android.
The innovation of the development lies in the fact that the neural network was taught to work with video. As for the images, neural networks have long been able to determine the age and sex of a person - and the accuracy of the systems in this case is quite high. But if you need to work with video, then the task becomes more complicated, since it’s not so easy to select a clear frame with a person where you can clearly see his face.
The standard scheme uses an estimate of a person’s age from 0 to 100 years, followed by an analysis of the whole age scale, indicating the probability that the person in the image is exactly that many years. For example, the probability of his belonging to the age group of 25-30 years is 10%, 30-35 - 35%, and, say, 50-55 years - 60%.
The algorithm is implemented based on IDE Pycharm with Python 3.6. Many resources of such a neural network are not needed - tests were conducted on a standard desktop PC with an Intel Core i5-2400 CPU, NVIDIA GeForce GT 440 graphics card and 64-bit Windows 7. In addition, the system was also tested on a mobile device with Android OS (Android version and the characteristics of the mobile device are not listed).
Estimated graphical interface of the Android application.
As for the mobile application, its main element is the window with a video demonstration (capture from the camera). The neural network analyzes individual frames and tries to indicate the age and gender of a person.
According to the developers, the main problem in recognizing various characteristics of a person, including his age and gender, is that training of neural networks that specialize in this task is too limited. Databases of video and images are relatively small, and yet all people are very different, including representatives of the same sex and age category.
Interestingly, one of the databases on the basis of which the neural network was trained from the article is that all videos were cut from Indian films. In total, the database had 322 different videos with 34512 frames. The video contained scenes involving hundreds of Indian actors. For convenience, they were divided into 4 age categories: “Children”, “Youth”, “Middle age”, “Elderly”. As for the timeline, it is 1-12 years, 13-30, 31-50, 50+.
Neural networks can determine the gender and age of a person not only from photographs or videos of a face or body. For example, the neural network created by Google and Verily learned to recognize not only these characteristics, but also the average blood sugar HbA1c, body mass index BMI, arterial systolic pressure SBP, arterial diastolic pressure DBP. And the system indicates whether a person smokes or not. And all this - in the fundus picture.
For learning this neural network, developers used the image database, which contained about 300,000 photos. Information provided by EyePACS and UK Biobank. According to doctors, a new approach to diagnosis can help doctors quickly diagnose. AI is able not only to speed up, but also to increase diagnostic accuracy. Doctors simply need this help, as the human doctor is not always able to work quickly and efficiently, especially at the end of the working day. As a result, the accuracy of diagnosis and the correctness of the prescribed course of treatment suffer.