Dual-Branch Residual CNN-LSTM Model for Emotion Classification using PPG and GSR

초록

Emotion recognition using bio-signals such as photoplethysmography (PPG) and galvanic skin response (GSR) is a practical and efficient approach compared to other bio-signals like EEG and EMG. In this paper, we propose a novel deep learning model with a dual-branch architecture for emotion classification, specifically designed to extract both spatial and temporal features from PPG and GSR signals. Each branch of the model comprises 1D-CNN-based residual blocks for spatial feature extraction, followed by LSTM layers to capture temporal dependencies. The left branch processes PPG signals, while the right branch processes GSR signals. To enhance feature representation, we concatenate the outputs from both branches and pass them through fully connected layers, leveraging the "Swish" activation function. The self-gating behavior of Swish helps to suppress less relevant neural nodes while emphasizing critical ones, thereby improving model performance. Our model outperforms state-of-the-art methods on two publicly available datasets, MERTI-APP and DEAP, achieving higher accuracy in both valence and arousal classification. Detailed results and analyses demonstrate the effectiveness of our approach for bio-signal-based emotion recognition.