Optimal Attention Deep Learning-Based Multi-Class Emotion Detection and Classification Model on fMRI Images

Abstract

This study proposes an optimal attention deep learning (DL)-based multi-class emotion detection (ED) and classification (OADL-MCEDC) model for recognizing human emotions from the images obtained from functional magnetic resonance imaging (fMRI). The approach integrates comprehensive fMRI preprocessing, hybrid capsule networks for robust feature extraction, and an attention-based bidirectional LSTM for emotion classification. Emotions such as angry, happy, sad, neutral, blank, and scrambled are identified. The Nadam optimizer enhances training stability and performance. Experimental results demonstrate that the proposed method outperforms baseline and state-of-the-art models, including support vector machine (SVM), random forest (RF), convolutional neural network (CNN), deep neural network (DNN), and Extreme Gradient Boosting (XGBoost), across multiple evaluation metrics.