Automated Pain Estimation based on Facial Action Units from Multi-Databases

Automatic pain intensity estimation is emerging as a crucial requirement in painful analysis and has showed a high value in health care applications. The correctness and computational complexity are the two required keys for a pain assessment application. Inspired by the fact that facial expression of pain can be described by the pain-related Facial Action Unit (AU) combinations, we propose a novel three stages training for pain intensity estimation approach. Based on the Inception Resnet architecture, we firstly learn to predict AUs from the combination of DISFA and UNBC McMaster databases using Heatmap regression. Secondly, we learn to predict pain level as linear regression. Lastly, we use a LSTM model to exploit the temporal relation between video frames. With this approach, we take advantage of two databases for regularization to prevent overfitting, while still maintaining the efficiency in term of complexity of the Inception Resnet architecture. Experiments on the UNBC McMaster database show that our approach yields a promising result compared to the state-of-the-art in automatic pain intensity estimation. The code for testing and the models are available to download from https://github.com/glmanhtu/3stages