EFFICIENCY OF APPLICATION OF LSTM NEURAL NETWORK LAYERS FOR THE PROBLEM OF CLASSIFICATION OF EMOTIONS ON HUMAN FACE

Abstract

The article thoroughly examines the effectiveness of long short-term memory (LSTM) layers for solving the task of emotion classification on human facial images. Particular attention is paid to comparing the performance of two models: a standard convolutional neural network (CNN) and a hybrid model that includes LSTM layers. The results of the study show that although CNN slightly outperforms LSTM on clean data in terms of quality metrics, LSTM demonstrates significantly better performance under noisy conditions. This is especially important for tasks related to emotion recognition in real-world scenarios, where noise is a common factor that can affect the system's accuracy. Additionally, LSTM offers advantages not only in noise resistance but also in data processing speed and a smaller number of parameters, making it much more efficient in terms of computational resource usage. The article also explores the potential of LSTM for predicting changes in emotional states, which opens new perspectives for the analysis of emotional dynamics. This capability is crucial for building future emotion-oriented systems, where not only recognition accuracy but also the ability to anticipate changes in users' emotions is essential. The models were trained on the FER2013 dataset, which includes images of seven different emotion categories, providing a comprehensive approach to model evaluation. The study emphasizes the practical potential of LSTM for use in real-world emotion recognition systems, where factors such as noise impact, processing speed, and computational efficiency are critical for system reliability and performance. LSTM's application in such systems enables the development of more adaptive solutions capable of high-accuracy recognition and prediction of users' emotional states.