Real World Facial Expression Recognition

Recent years,facial expressions recognition has remained a challenging and interesting problem,especially for faces in the real world.Most traditional static facial expressions recognition methods are based on Action Units(AU)or low level image features.These methods have achieved very impressive approaches on lab-controlled facial expression databases,such as CK/CK+,JAFFE etc.However,because of the complex background,illumination and different personal habits,real world expressions recongnition is more challeng-ing than recongizing lab-controlled expressions,Thus these traditional methods always gain bad results in real world.According to researches,facial expressions are the result of faical musles movements.This paper apply a method that combines facial local region with attributes learning.Using an automatic framework to extract "Mid-Leve" fea-tures from a number of class-pairwise local facial regions.Compared with tradi-tional texture features,the "Mid-Level" features are more discriminative.After Mid-Level features extraction,a feature selection method based on Adaboost is applied on these large number of Mid-level features to get the top discriminative features.This step can not only reduce the dimension of the features and pre-vent over-fitting problem,but also promot the preformance of real-world facial expression recognition.In this paper,we firstly analyze the facial expressions difference between traditional and real world.Then we apply an automatic framework to extract and select the discriminative Mid-Level features for expression recognition.In experiments,four facial expression benchmarks(CK+,SFEW,RAF-BASIC,RAFCOMPOUND)are evaluated.And our method achieves state-of-the-art performance compared with recent approaches.