Home // INTELLI 2021, The Tenth International Conference on Intelligent Systems and Applications // View article


A Hybrid Model to Improve Occluded Facial Expressions Prediction in the Wild during Conversational Head Movements

Authors:
Arvind Bansal
Mehdi Ghayoumi

Keywords: Artificial Intelligence; conversation; emotion analysis; facial expression analysis; facial occlusion; facial symmetry; head movement; multimedia

Abstract:
Human emotion prediction is an important aspect of conversational interactions in social robotics. Conversational interactions involve a combination of dialogs, facial expression, speech modulation, pose analysis, head gestures, and hand gestures in varying lighting conditions and noisy environment involving multi-party interaction. Head motions during conversational gestures, multi-agent conversations and varying lighting conditions cause occlusion of the facial feature-points. Popular Convolution Neural Network (CNN) based predictions of facial expressions degrade significantly due to occluded feature-points during extreme head-movements during conversational gestures and multi-agent interaction in real-world scenarios. In this research, facial symmetry is exploited to reduce the loss of discriminatory feature-point information during conversational head rotations. CNN-based model is augmented with a new rotation invariant symmetry-based geometric modeling. The proposed geometric model corresponds to Facial Action Units (FAU) for facial expressions. Experimental data show hybrid model comprising a CNN-based model and the proposed geometric model outperforms the CNN-based model by 8%-20%, depending upon the type of facial-expression, beyond partial head rotations.

Pages: 36 to 42

Copyright: Copyright (c) IARIA, 2021

Publication date: July 18, 2021

Published in: conference

ISSN: 2308-4065

ISBN: 978-1-61208-882-2

Location: Nice, France

Dates: from July 18, 2021 to July 22, 2021