Abstract

This paper presents a classifier ensemble for Facial Expression Recognition(FER) based on models derived from transfer learning. The main experimentationwork is conducted for facial action unit detection using feature extraction andfine-tuning convolutional neural networks (CNNs). Several classifiers forextracted CNN codes such as Linear Discriminant Analysis (LDA), Support VectorMachines (SVMs) and Long Short-Term Memory (LSTM) are compared and evaluated.Multi-model ensembles are also used to further improve the performance. We havefound that VGG-Face and ResNet are the relatively optimal pre-trained modelsfor action unit recognition using feature extraction and the ensemble ofVGG-Net variants and ResNet achieves the best result.

 

Quick Read (beta)