A Decomposed Projective Method based on FBT-VMD and t-ELM for Facial Emotion Recognition

Abstract

The performance of affective computing, in particular, of facial emotion recognition, is often limited by the low interpretability of the network structure, parameter redundancy, the high computational complexity with large consumption caused by CNN strategy in extracting features. The decomposed projective method, which integrates dimensional conversion, signal analysis, and dimensionality reduction for facial emotion image processing, is considered in this presentation. The main contributions of this work are the following: 1) a dimensional conversion for a facial image, which transforming the 2-D image matrix into a 1-D signal using Fan Beam Transform (FBT); 2) a hierarchical analysis method for the signal with mode decomposition (VMD) for feature selection; and 3) a method combining the t-SNE algorithm and Extreme Learning Machine (ELM) is introduced to reduce the dimension of feature vectors and predict classification. Experimental results present a satisfactory performance with 80.17% accuracy.