A Hybrid Framework Combining CNN, LSTM, and Transfer Learning for Emotion Recognition

Deep learning has substantially enhanced facial emotion recognition, an essential element of hu-man–computer interaction. This study evaluates the performance of multiple architectures, including a custom CNN, VGG-16, ResNet-50, and a hybrid CNN-LSTM framework, across FER2013 and CK+ datasets. Preprocessing steps involved grayscale conversion, image resizing, and pixel normaliza-tion. Experimental results show that ResNet-50 achieved the highest accuracy on FER2013 (76.85%), while the hybrid CNN-LSTM model attained superior performance on CK+ (92.30%). Per-formance metrics such as precision, recall, and F1-score were used for evaluation. Findings high-light the trade-off between computational efficiency and recognition accuracy, offering insights for developing robust, real-time emotion recognition systems.