Face Analysis In Human-machine Interaction

Vision interaction is one of important aspects of natural human-machine interaction, andthe efficient facial feature extraction is crucial to vision interaction. Great achievements havebeen made in face analysis, but almost all of them are based on supervised learning which needinput covered the entire situation and cannot seek the relations of the data. Yet there is a simplequestion which is related to unsupervised learning cannot be answered, given a database of faceimages, exactly how many people are there?According to the unsupervised learning problem, this paper proposed a novel frameworkwhich consists of two parts, the learning part and the clustering part. In the learning part, thebasic goal is to extract and select a feature subset from the anthropometric feature set thatremains the discriminative features of distinct individuals but discard those features that reflectposes, expressions and illuminations. In the clustering part, the main challenge is toautomatically determine the cluster number. We adopt the nonparametric Bayesian methodwhich is capable of automatic model selection or cluster number selection and calculated withvariational inference.Comprehensive experiments are conducted to measure the performance of the frameworkin terms of a variety of quantitative indices. The experimental results show that the proposedface clustering approach is non-sensitive to the model parameters and could get goodexperimental results under a variety of conditions such as illumination changes, large posevariations, and facial expression changes and so on.