Electrodermal Activity for Emotion Recognition Using CNN and Bi-GRU Model

Several signals can be collected from wearable devices containing important physiological and psychological information. Understanding various physiological signals is significant for computers to recognize human emotional states. Electrodermal Activity (EDA), originating from the spontaneous activation of sweat glands in the skin, is closely related to mood, arousal, and attention and is the most widely used measurement in the physiological response system for emotional state detection. However, extracting valuable features from EDA signals and making accurate emotional classification predictions has always been challenging. With the continuous development of models with representation learning capabilities, the use of deep learning models to automatically learn physiological signal features and perform classification learning is promising. In order to improve the shortcomings of traditional emotion recognition methods, which require a deep understanding of physiological signals and artificial extraction of relevant features, this paper proposed a Recurrent Neural Network (RNN) -based method for automatic feature extraction from EDA's spectrograms. A Convolutional Neural Network (CNN) is used to learn the extracted features further and output the determined emotional state. The results show that the classification accuracy for arousal and valence has reached 83.4% and 81.2%, respectively, which is promising in extracting features automatically and tackling the emotional state classification problem.