Real-Time Emotion Detection using Hybrid CNN-BiLSTM Deep Learning Model

Abstract

Emotion recognition from facial expressions is a critical task in affective computing and human-computer interaction, with applications spanning healthcare, education, surveillance, and entertainment. Traditional convolutional neural networks (CNNs) have shown promising results in extracting spatial features from facial images, but often lack the temporal sensitivity needed to capture nuanced emotional patterns. In this, the hybrid deep learning model is proposed that integrates CNN and Bidirectional Long Short-Term Memory (BiLSTM) layers for real-time facial emotion recognition. The CNN layers effectively extract hierarchical spatial features from the FER-2013 dataset—a benchmark dataset consisting of 48x48 grayscale facial images categorized into seven basic emotion classes. These extracted features are reshaped and passed to BiLSTM units that model temporal dependencies and contextual relevance within spatial encodings. Our method results in notable enhancements in performance over baseline CNN models, with enhanced recognition accuracy, particularly in distinguishing subtle emotions like fear and sadness. Experimental evaluation using confusion matrices and classification reports confirms the robustness of the hybrid architecture. The results suggest that the integration of BiLSTM with CNN offers a more expressive and context-aware solution for real-time emotion detection, paving the way for more adaptive and emotionally intelligent systems.