FACIAL EMOTION RECOGNITION SYSTEM WITH LOCALIZATION IN PORTUGUESE, ENGLISH, RUSSIAN AND UMBUNDU

Abstract

The article proposes a new emotion recognition system based on facial expressions and their descriptions in four languages – Russian, English, Portuguese and Umbundu. Emotions play a significant role in determining the true intentions of the implementation of a communicative act in various types of practical activities, along with traditional verbal communication. There are a large number of emotion recognition systems that are able to recognize human emotions by class and use them to control a communicative act. The known systems have low validation accuracy, large size and long execution time, and also do not have localization of emotions for a specific country or region speaking Portuguese, Russian and Umbundu. For emotion recognition, a classifier using the machine learning method implemented in the Keras library has been built, and a new deep convolutional neural network (DCNN) classifier architecture has been created. The training of the classifier was carried out on a sample of FER 2013 data. Two trained models for emotion recognition were compared: the basic model and the model developed by us. According to experimental data, the model we developed achieved the best results in validation accuracy (70.41%), execution time (0.00043 s), and size (12.7 MB). The most adequate translation correspondences of these units have been established in order for the emotion recognition system developed by us to provide information about the emotional state of customers in public service systems according to the language situation of this country or region. The system provides translation of emotions into Portuguese, English, Russian and the traditional Umbundu language in order to obtain information about customers and minimize conflicts in public service systems.