Research On Several Key Technologies Of Facial Expression Recognition

In recent years, with the rapid development of computer and informationtechnology, facial expression recognition technology draws more and more attention.How to effectively and accurately extract expression features and classification havealso become hotspots.In this thesis, some key issues which appeared in feature extraction andclassification are studied. Several improved algorithms and methods for these tasksare proposed and validated by experiment results. The main work is as follows:This paper presents a new feature extraction method, namely LBP curveletdomain expression recognition. The method was presented applied curveletcoefficients to form features for representing the entire face. In order to classify thefacial expressions, the local facial information needs to be stored. To obtain the localdescription of the expressions, local binary patterns (LBPs) are computed usingselected sub-bands of image pre-processed by curvelet transform. The approach isnon-cross-validated and has been compared with LBP and Gabor wavlet basedmethods. The experimental results show that this method can effectively improve theaccuracy of classification of expression. This method requires relatively low on animage, overcomes the traditional limitations of the approach to noise, light, faceinclination.Semi-BSVMs of face recognition is made in the paper. It introduces a Bayesiansemi-supervised support vector machine (Semi-BSVM) model for binaryclassification. Then classification by constructing Semi-BSVMs, we can achievedhigh recognition rate. Our semi-supervised learning has a distinct advantage oversupervised or inductive learning since by design it reduces the problem of overfitting.While a traditional Support Vector Machine (SVM) has the widest margin based on the labeled data only, our semi-supervised form of SVM attempts to find the widestmargin in both the labeled and unlabeled data space. This enables us to use someinformation from the unlabeled data and improve the overall prediction performance.The likelihood is constructed using a special type of hinge loss function which alsoinvolves the unlabeled data. A penalty term is added for the likelihood partconstructed from the unlabeled data. The parameters and penalties are controlledthrough nearly diffuse priors for objectivity of the analysis. The rate of learning fromthe unlabeled data is reflected through the posterior distribution of the penaltyparameter from the unlabeled data. This formulation provides us with a control onhow much information should be extracted from the unlabeled data without hurtingthe overall performance of our model. Experiments show that this classifier for therecognition rate has been greatly improved.