Facial Expression Intensity Estimation And Recognition Based On Rankboost

Facial expressions provide a wealth of information that can help us understand a person's emotions and attitudes better.And the intensity of facial expression is very important for detecting and tracking the change of expression.It has a very broad application scene in real life.In daily life,the change of facial expression is a process of dynamic changes over time.So the problem of estimating the intensity of facial expression can be converted into the sequencing problem of expression.RankBoost algorithm is a very good sorting algorithm.In addition,facial expression estimation and recognition are inseparable.In this paper,we present a framework based on RankBoost algorithm to estimate the intensity of facial expression and recognize expression.The details are as follows:(1)I research the problem of estimating the intensity of expression based on RankBoost and make some changes to the original algorithm.RankBoost is a kind of integrated learning algorithm,which is a series of weak learners obtain a strong learner by linear combination.Every weak learner plays a crucial role.Therefore,this paper enhances the ranking function of the weak hypothesis and makes it more distinguishable.Moreover,in order to get a more correct and practically result,we add more prior information into the loss function.(2)I apply the RankBoost to recognize expression.Through the algorithm learning,we can get the function score.Build the distribution of scores histogram to find the appropriate threshold for facial expression recognition.Indeed,a large number of experiments on Cohn-Kanade+and MMI database show that the algorithm presented in this paper has better performance than the previous ones.(3)Traditional machine learning features are diverse and the features are extracted from different aspects of the images.Different features including LBP features,Haar features and Hog features are studied in this paper.The advantages and disadvantages of each feature are analyzed and we combine them through the co-training using test data to improve learning ability.At the same time,deep features based on CNN have strong generalization ability.We learn the features from CNN directly from the original image as input,and compare it with the traditional image features.