Masking and Compression Techniques for Efficient Action Unit Detection of Children for Social Robots

This paper proposes masking and compression techniques for action unit detection of children to be used on a social robot. The aim of the study is to develop a real-time action unit-based emotion recognition module based on limited processing power and lower memory comsumption on the robot. For this purpose, the effect of various preprocessing techniques and image compression methods on action unit recognition of children are observed. The results show that preprocessing steps that will constrain the robot's response time, can be excluded in our real-time applications.