The Styudy Of Multi-kernel Learning In Facial Expression Anaylsis Based On MultiClass-SVM

Recently,facial expression recognition has drawn great attention in the field of social network and human-machine interaction.And a series of achievement has been witnessed.The most facial expressions in the existing database are with front faces,high resolution and well-controlled lighting conditions.In contrast,the real-world facial expressions are much more polytropical,and thus the existing algorithms are too simple to meet the rea-world tasks’need.To test the performance of the existing algorithms,we studied some factors that might infect the detection of the smile faces in real-world occasions,including illumination equalization,align methods,size of images,features and the selection of SVM kernels.Base the experiment results,we found the illumination equalization played a limited role,but the align methods meant a lot.To solve the multi classification problem,we established a new database:Real-world Affective Face Database(RAF-DB),which contains about 30,000 facial images in great diversity.One point which deserves mentioning is that each image is labelled by about 40 independent volunteers.We did cross education experiment on our database and CK database,and in the same circumstances,the result are much better on our database,which indicates the good performance of our database.We found the that facial expressions classification is a typical imbalanced multi-label classification problem.Based on our new database,for the imbalanced data,we use some strategy such as up-sampling to reconstruct the human face data,which showed higher accuracy rate in the experiment.In terms of the feature representation,we take advantages of the DCNN learned feature besides the common used ones(such as HOG,Gabor,LBP).Based on different database and recognition tasks,different SVM kernels may perform differently.So,we selected the multi-kernel SVM.After abundant experiment,we achieved a remarkable result.