Research And Implementation Of Facial Expression Recognition System

In recent years, with the rapid development of computer and information technology, facialexpression recognition technology draws more and more attention. How to effectively andaccurately extract expression features and classification have also become hotspots.In this thesis, some key issues which appeared in feature extraction and classification arestudied. Several improved algorithms and methods for these tasks are proposed and validated byexperiment results. The main work is as follows:1. It is MLBP(multiple local binary patterns) that is proposed. Based on preserving theadvantages of original LBP(local binary patterns), the method is reasonable to consider the roleof the center pixel and the relation of the gray value of neighborhood pixels by adding a binarycode, which successfully overcomes the original LBP’s shortcomings that ignoring the role ofcentral pixel point and the relation of the gray value of neighborhood pixels and improves textureidentification. The validity and advantages of the improved algorithm have been verified bymeans of a series of experiments which are performed on standard expression databases.2. It is MLBP combined with wavelet decomposition expression feature extraction methodthat is proposed. First, four sub-images with different frequency information are obtained by thesource expression image using wavelet decomposition. The facial expression feature ischaracterized by the texture information and deformation of facial expression, but which is notpresented in the diagonal high-frequency image. So the low-frequency image, horizontalhigh-frequency image, the vertical high-frequency image are selected to produce expressionfeatures. These features are put in series to form expression feature vector. Compared withfeature extraction without using wavelet decomposition, in the proposed method, subsequentdata which is produced by expression feature extraction and identify is reduced by1/4. It may be appropriate to increase the rate of feature extraction and recognition. The experimental resultsshow that this method can effectively improve the accuracy of classification of expression.3. It is Fuzzy Laplacian SVM that is proposed. The method which is based on the LaplacianSVM which is obtained by adding manifold regularization to C-SVM method imports fuzzyvectors to debug the penalty value depending on the different data. It can overcome theshortcomings of traditional SVM. For example, large consumption of space and time resultedfrom the need of solving quadratic programming, large-scale data; inadequate training resultedfrom insufficient samples in practice. The improved algorithm not only takes advantage of thelabeled and unlabeled samples, but also can ignore data of a very small impact on theclassification results, which improves greatly the performance of SVM.4. It is the real-time facial expression recognition system that is designed and implementedby using object-oriented methods, the running interface of which has been provided. The systemhas achieved relatively better recognition results.