Research On Facial Activity Recognition Based On Restricted Boltzmann Machine

Facial expression is the most important way for human to express their emotion.Only if computer analyze human’s facial activity accurately, can computer or robotunderstand human’s emotion, and do Human-Computer Interaction or ArtificialIntelligence achieve truly. However, the technology of facial activity recognitionhas been taken lots of attention in decades, and has accumulated a large number ofachievements, it is still far from application. A great number of difficulties andchallenges, such as the uncertainty of images caused by illumination andocclusion, different faces between different subjects, the rich expressions. Thispaper will have a deep research on these issues. Facial feature points tracking a ndfacial action unit recognition are included. The main contents of the dissertation areas follows:This paper analyze the drawbacks of linear model on modeling the complexexpression, and improve the process of tracking facial feature points. General priormodels are difficult to capture the sharp changes of expression. To solve thisproblem, this paper propose a shape prior model based on Gaussian RestrictedBoltzmann Machine. This method includes two important steps. First, model theshape relationship of facial feature points by GRBM. Then, get the location of thepoints by Gibbs Sampling. In this way, the measurements of facial feature pointscan be restricted by prior model, and their accuracy will be improved.This paper also analyze the problems in independent recognition of facialaction units, and clearly the importance of semantic relationships between AU.Current models can only capture the local relationships between pairs of actionunits. This paper propose a model based on Restricted Boltzmann Machine toresolve this problem. The model takes full advantage of specialities of RBM, suchas whole connections between layers, unsupervised learning. This model canefficiently and effectively capture not only local, but also global dependenciesamong AUs. Finally, the relationships are used to correct measurements of facialaction units.This paper design reasonable experiments to verify the effectiveness andversatility of above models.