Research On Facial Expression Recognition Based On HCBP And Feature Selection AdaBoost

Facial expression recognition is a hot topic of computer science. As a subject of interdiscipline, the progress of facial expression recognition will have a direct impact on the development of psychology, artificial intelligence, data mining, faculty sciences and other subjects. With the progress of science, its role in research, medical, business and other fields will come to a huge impact on human life. Presently, research on facial expression recognition is still in the theoretical stage, especially the feature extraction and facial expression classification, there are many areas for improvement in order to put it into practical.This thesis based on the feature extraction and classification, summarized the relevant current methods. As to the feature extraction and classification, this thesis put its emphasis respectively on texture feature and classification integration. It put forward some new methods to improve the shortage of LBP (Local Binary Pattern) and AdaBoost. The main content of this thesis is shown as follow:(1) As to the feature extraction, a method called facial expression recognition based on histogram weighted HCBP (Haar-like Centralized Binary Pattern) was proposed. CBP (Centralized Binary Pattern) is an improved algorithm, which overcomes the shortages of LBP in terms of dimension and characterization description. But in the same with LBP, CBP also uses direct manipulation on pixels, it can’t reflect the change of pixels efficiently as well. Haar-like is some kind of feature that operating the pixels regions, it can effectively reflect the change of pixels. This thesis combined Haar-like features with CBP, and divided images into blocks, used information entropy to get weights to differ the contributions of every region of face to the expressions, enhanced the ability to characterize features.(2) As for the feature classification, an algorithm based on feature selection Adaboost ensemble classifier is proposed. It made some improvements to the AdaBoost. Traditional AdaBoost is used to deal with dichotomous questions, this thesis modified the way of weight selection, so that it would suit the requirements of multi-classification; made use of PCA (Primary Component Analysis) in each iteration of AdaBoost for feature selection and sorting, generated different random subspace, this method overcome the disadvantages which the difference between base classifications would reduce while the number of iterations increased, and the dimension of features had been made lower; besides, according to the base classifiers’ effect on every kind of expression, modified the way of ballot to weight the base classifications dynamically, which enhanced the effect of AdaBoost base classifier ensembling.The experimental results show that the algorithm has achieved good results both on the recognition rate and time efficiency.