Research On Feature Expression And Recognition For Facial Expression

Facial expression plays an essential role in human communication. With rich information on component motion and gestures, facial expressions bear with a great amount of clues on emotions and feelings, and take on as a form of human intelligence, which may proceed in deep insight into response of physical and psychological conditions on testers.In this thesis, the author focuses on pieces of image sequences about different facial expression. As is significant characteristics, Time mark is highlighted as a new entrance, based on which common motion features are summarized. Accordingly, facial expression recognition is turned into analysis of motion trajectory. Now, NMDS(Non-metric Multidimensional Scaling) is employed to get embeddings of original sequence of image data on 2-dimension space, when facial expression classification is done by comparison of trajectory curves. Experiments show a well-proof result.The new method here to explore facial expression based on two points: all working aim to sequential features on time marks; Multi-dimensional Scaling method is introduced to get low-dimensional trajectoryIsomap with Graph embedding will construct a sub-manifold of facial expression on image sequences in low dimension space. This method can provide vital similarity analysis and trajectory estimation on motion preference according to time series features of video sequence. We consider all facial expressions are independently segmented apart, and group them according to each subject in dataset. As same kind of expression tend to share similar line-shape, we explore comparison of geometric curve fitting to complete classification and recognition on semantic content of facial expression.Finally, experiments show, though none of use before, this method achieves a good result on recognition, especially to some specific categories,'surprise'as an example.