LCNN: A Lightweight Convolutional Neural Network with Improved Activation Function for EEG-based Emotion Recognition^*

Abstract

A pivotal realm within artificial intelligence, affective computing, has garnered substantial attention from researchers. The core idea behind affective computing is to empower machines with emotional capabilities. Its technological advancements have resulted in groundbreaking opportunities within the domain of human-computer interaction. This paper introduces a lightweight convolutional neural network (LCNN) designed for emotion recognition based on EEG, incorporating an enhanced activation function to augment the network’s nonlinearity and improve its classification performance. To capture time-frequency domain information, we compute differential entropy (DE) features utilizing the Short Time Fourier Transform (STFT). Furthermore, brain topographic maps are employed to transform these features into matrices, enabling the inclusion of spatial position information from electrodes. Ultimately, the LCNN is utilized for emotion recognition tasks. To gauge the efficacy of our approach, we conducted experimental validation on the SEED dataset, achieving satisfactory results with a success rate of 90.00% in subject-dependent experiments and 76.21% in subject-independent experiments. This breakthrough opens up the potential for emotional interaction between humans and computers, offering avenues for meaningful emotional support in the future. This is particularly relevant in the context of assisting robots in delivering effective emotional care.