Title:
Estimation of Human Emotion by Analyzing Facial Expression Transition of Image Sequences

Speaker:
SIRITANAWAN, Prarinya (JAIST)

Abstract:
Face is a tool to deliver our emotional message to those around us, and we can read them in the otherwise. It is one of the most visible part in the non-verbal communication, and has a greatly influence in our way of interacting with others. Conventional emotion research stated that the expressing and understanding of facial expression are universal characteristics across races and cultures. There is a particular set of facial muscles that always occurs under the same emotional contexts. For example, people reveal a smile in the happiness moment, or show a sign of distress by the wrinkle around your forehead and eyebrows. Among the early studies of facial expression's universality, they defined the emotion into categories by a few English words, which are obviously insufficient to explain the whole spectrum of emotions. These categories were inherited to the modern studies of emotions in computer vision and machine learning, in which scientists attempt to build an intelligent machine that can understand human’s emotional messages by classification techniques. By limiting the number of class, which inaccurately describe human’s emotional states, results in an impractical outputs. With several considerable evidences opposed to the universality statement, we have been convinced that the expressions of emotions are the by-product of both biological traits and the cultural influences. In this research, we created the personalized emotional recognition framework from a spatiotemporal feature aiming for the human-robot interaction. We proposed the cumulative change of features frameworks to measure the facial muscles activities across space and time. Our robust temporal feature overcomes the limitations of typical facial expression recognition systems, which usually used a set of still images to learn the features and cannot recognize the subtle expressions. In contrast to the previous studies, which they reduced the face features into the geometrical points or distances, our proposed features can be visualized in facial activities in spatial resolution, and can detect the subtle actions of facial muscles which are usually difficult to distinguish by the geometrical features. In order to describe more complex facial expression in the similar fashion to psychological studies, we used our robust temporal feature and discriminative subspace method to automatically learn Action Units (AUs) according to Facial Action Coding System (FACS) standard.