Emotion recognition and classification using Inception EfficientNet based on electroencephalography signals

Abstract

Emotions are intricate psychological phenomena arising from the interaction of internal cognitive states and external environmental inputs. The manual extraction of electroencephalography (EEG) signals results in less optimal performance of learning models. To overcome this, a novel EEG-based emotion recognition and classification (EEG-EMRE) model has been proposed for the detection and classification of emotions. Initially, the input EEG-Signals are pre-processed using quantum signal processing (QSP) to enhance the quality by removing the noise from the signal. The enhanced signals are fed into an improved Inception EfficientNet for extracting the relevant features. The Penta types of emotions, such as happy, sad, anger, scared, and anxiety, are classified using a bidirectional-k nearest neighbors (KNN) classification network. The performance of the proposed EEG-EMRE approach is evaluated using the F1-Score, recall, specificity, accuracy, and precision. The proposed Inception EfficientNet for feature extraction network improves the overall accuracy by 0.41%, 1.52%, 0.63%, 1.55% better than ResNet, AlexNet, GoogleNet, and DenseNet. The proposed EEG-EMRE method achieves an overall accuracy by 0.68%, 1.77%, and 0.52% better than the linear formulation of differential entropy (LF-DfE), extreme learning machine wavelet auto encoder (ELM-W-AE), and attention-based convolutional transformer neural network (ACTNN), respectively.