COMPARATIVE ANALYSIS OF METHODS FOR IMBALANCE ELIMINATION OF EMOTION CLASSES IN VIDEO DATA OF FACIAL EXPRESSIONS

Subject of Research. The imbalance of classes in datasets has a negative impact on machine classification systems used in applications of artificial intelligence, such as: medical diagnostics, fraud detection and risk management. This problem in facial expression datasets also degrades the performance of classification algorithms. Method. The paper discusses the main approaches for the class imbalance reduction: resampling methods and setting the weight of classes depending on the number of samples observed for an each class. A histogram of oriented gradients is used for the face area localization in the frame stream, then an active shape model is applied, which detects the coordinates of 68 key facial landmarks. Using the coordinates of key landmarks, informative features are extracted that characterize the dynamics of facial expressions. Main Results. The results of the study have shown that the proposed approach to the extraction of visual features exceeds the accuracy of human emotion recognition by facial expressions. The considered methods of the class imbalance reduction in the set of facial expressions have provided the improvement of machine classifier performance and showed that the existing class imbalance in a training set has a significant effect on the accuracy. Practical Relevance. The proposed approach to the extraction of visual features can be used in automatic systems for human emotion recognition by facial expressions, and result analysis of applying methods that reduce class imbalance can be useful for researchers in the field of machine learning.