Multiple Facial Action Unit Recognition Enhanced by Facial Expressions.

Jiajia Yang,Shan Wu,Shangfei Wang,Qiang Ji
DOI: https://doi.org/10.1109/icpr.2016.7900274
2016-01-01
Abstract:Facial expressions and facial action units (AU) respectively describe facial behavior globally and locally. Therefore, the dependencies between expressions and AUs carry crucial information for facial action unit recognition, yet have not been thoroughly exploited. In this paper, we propose a novel facial action unit recognition method enhanced by facial expressions, which are only required during training. Specifically, we propose a three-layer restricted Boltzmann machine (RBM) to capture the probabilistic dependencies among expressions and AUs. The parameters of the RBM model are learned by maximizing the log conditional likelihood with gradient ascent. After that, the learned RBM model combines AU measurements with the AU-expression relations it captures to perform multiple AU recognition through probabilistic inference. Experimental results on three benchmark databases, i.e. the CK+ database, the ISL database and the BP4D database, demonstrate the effectiveness of our method on capturing the joint relations among AUs and expression to improve AU recognition.
What problem does this paper attempt to address?