Facial Expression Recognition by Estimating Movement of Facial Muscles from Stream Data

Abstract

In order to analyze and recognize various types of facial expressions, discriminating slight movements of facial muscles is essential. However, it is difficult because of complex structures of facial muscles. In many traditional methods of facial expression analyses, discriminative movements of facial muscles are manually defined in advance and facial expressions are discriminated by analyzing only the predefined movements. This approach is problematic because finding and defining appropriate movements is quite troublesome and time-consuming. To solve this problem, we propose an effective method to automatically find useful movements. We accurately estimate the movements of facial muscles and represent them as time-series data called multistream. Since multistream generally contains several redundant streams, we remove useless streams by evaluating the usefulness of each stream based on AMSS (Angular Metrics for Shape Similarity) which efficiently measures the similarity between streams. We verify the effectiveness of the proposed method through several facial expression recognition experiments.