INTEGRATING TEMPORAL DYNAMICS IN FACIAL EMOTION RECOGNITION USING HYBRID CNN-RNN MODELS FOR ENHANCED HUMAN-COMPUTER INTERACTION

Abstract

Facial Emotion Recognition (FER) is still an important branch in computer vision and artificial intelligence, mainly benefiting Human-Computer Interaction (HCI). Existing FER systems, which are mainlybased on Convolutional Neural Networks (CNNs) for analysis of static images, do not support the dynamic evolution of human emotions over time. To address these issues, this work presents a novel model that incorporates temporal information in FER using a hybrid CNN-RNN (Recurrent Neural Network). The proposed method uses CNNs for spatial emotion feature extraction, and RNNs to model the sequential dynamic information of emotions that enables a better understanding of affects. By evaluating on a benchmark FER2013, we investigate three deep learning strategies: a baseline CNN-RNN, a CNN with an attention module, a CNN-RNN with data-enrichment techniques. Experimental results show that the CNN-RNN with data augmentation outperforms the other approaches with a test accuracy of 89%, precision, recall and F1-scores higher than 88%. These results suggest that temporal dynamics along with the synthetic data can be effective in addressing the challenge of class imbalance and data sparsity. Moreover, attention mechanisms enhanced the interpretability and classification accuracy of the model. However, even though good results have been observed, there still exists real time deployment challenges because of the computational complexity and the model sensitivity under various weather conditions. Conclusion Future directions to pursue are an optimal design of hybrid architectures for real-time inference, extension of cross-cultural generalizability, and privacy-preserving learning strategies. This research provides a scalable and effective FER solution that is suitable for use in emotionally intelligent systems in such areas as healthcare, surveillance, education, and HCI.