Scientists from Sberbank's artificial intelligence laboratory have developed an innovative system that can accurately recognize people's emotions even in poor conditions. The technology works in low light, noise, and movement. It can be used on smartphones.
The system is based on the analysis of video, audio, and text data. The development used lightweight neural networks, such as EmotiEffNet and MobileFaceNet, which were trained on open sets of images with various facial expressions. The neural network from Sber allows processing information in real time without overloading devices.
The algorithm filters frames and selects only those where emotions can be recognized with high accuracy. A "late fusion" method is also used, which combines data from different sources. The model is better adapted to new tasks.
At the international ABAW competition, this technology took first place in facial expression recognition, and also entered the top three in other categories. The accuracy of the system has increased by 1.5–15% compared to traditional methods.
Read more on the topic:
Sberbank Launches New Version of GigaChat
Sber to Develop Neural Network for Import Substitution of Production Equipment
First Cyber Threat Model for AI Systems in Russia Presented by Sberbank
