Facial Expression Image Synthesis Based On Machine Learning

Facial expression synthesis has become an important technique in the field of Human-Computer interaction, as well as an active research topic in many areas, such as computer vision, computer graphics, psychology and multimedia. It has widely potential perspectives such as computer aided instruction, game design and virtual reality.This thesis focuses on example-based synthesis method, discusses how to utilize training examples and statistic methods to synthesize facial expressive image. The main contributions of this work are listed as follows.First, a synthesis method based on factorization model is proposed. Here"human identity"and"expression"are regarded as two influencing factors. Then a kernel based bi-factor factorization model is introduced to learn their interactions. By given a test image, the expressions in the training set can be"translate"to the input image by using factorization model and vice versa.Second, a parameter-controlled synthesis method is proposed. This method is based on expression subspace extracted by factorization model, and it can generate any kind of expression by given an emotion-parameter. To conquer the drawback of losing unforeseen face characteristics, Expression Ratio Image (ERI) is introduced into the algorithm. Moreover, we also propose a new expression recognition algorithm based on the subspaces extracted by factorization model and achieve very satisfying accuracy.Third, an emotional viseme synthesis framework which based on the Data-Driven Visual-Speech Synthesis system and the Parameter-Controlled Expression Synthesis system is presented. Given some emotion and viseme parameters, the expression and viseme can be generated by these two modules first and then combined together with high reality by a fusion algorithm.