Hybrid CNN-GRU Architecture for Efficient Emotion Recognition in Immersive Augmented Reality Systems

This paper presents a hybrid deep learning architecture that takes gaze data as input and recognizes emotions efficiently and in real time in resource-constrained augmented reality (AR) environments. Unlike conventional systems, which typically rely on facial expressions or speech signals, the proposed system uses eye tracking and gaze dynamics as the main indicators of emotional state, which offers better privacy, stability, and usability in practical AR applications. The model combines a convolutional neural network (CNN), which extracts spatial representations from eye-region images and gaze heatmaps, with a gated recurrent unit (GRU), which models time-varying gaze behavior. The model was evaluated on a self-curated dataset of 3,500 temporally sequenced samples annotated with six emotion classes. The proposed model achieved an accuracy of 91.2%, outperforming the baseline long short-term memory (LSTM) and standalone CNN models while remaining efficient. The GRU-based architecture was also more sensitive to temporal patterns and had a lower computational cost, with an average inference latency of 14.5 ms per sample, which meets real-time processing requirements.