Differentiating Between Spontaneous and Posed Facial Expression using Inception V4

Abstract

This thesis proposes a way to simplify and make solutions for spontaneous and posed facial expression analysis more efficient. Traditional approaches have been using hand-crafted features and two image frames to be able to differentiate between spontaneous and posed facial expressions. The solution aims to be as flexible as possible and introduces two models to differentiate between posed and spontaneous facial expression. We introduce Inception V4 as an algorithm to solve this task. The results indicate that Inception V4 may be too deep and unable to differentiate between spontaneous and posed facial expression accurately. A shallow CNN model is also introduced. The shallow CNN model performs better than the Inception V4 model. None of the two come close to the state-of-the-art results. This may indicate that to differentiate between spontaneous and posed facial expressions the difference between the onset and apex frame of an expression is needed as input. This thesis, also suggests an alternative algorithm based on our findings. For further work, an algorithm which is not as deep as Inception V4 is needed. However, by using parts of the Inception V4 architecture, we may be able to capture facial features better. The task of differentiating between spontaneous emotion and posed emotion has also been investigated; however, the results do not show great promise. The task does not have any state-of-the-art results to compare our approach with. Our models, although lacking in performance, does seem able to capture relevant facial features from the dataset.