Research Of Facial Expression Recognition Algorithm

Facial Expression Recognition (FER) which aims to recognize the facial expression in the non-contact way by the computer and analyze the affection and emotion of the human is an important research topic in the intelligent human and machine interface field. This topic has attracted wide attentions of all the researchers in this field for its important research significance and huge practical value.As a whole recognition system, Facial Expression Recognition includes the facial detection, the character collection and the facial recognition. The main researches of this paper on the three aspects are as follows.(1)In the facial detection aspect, this paper introduces an improved maximum between-threshold which selects the center pixel and neighborhood pixel gray value mean as the reference threshold when selecting the threshold, calculates the image gray value variance,select the reference threshold which maximizes the variance as the last threshold. This paper also compares the result of this method and the result of other research method.(2)In the character collection aspect, this paper collects the texture character with the LBP algorithm and uses the LBP histogram selected by the statistical method to identify the characteristic data.(3)In the facial recognition aspect, based on deeply researches on SVM and the Adaboost algorithm, this paper introduce an improved combined algorithm which not only has good properties in dealing with nonlinear problems, linear non-separable problems and small samples problems as the SVM algorithm but also has the good performance in training weak classifier as the Adaboost algorithm and to receive an idea classifier.(4) Based on the characters of the DAG SVM algorithm, the experimental results and the overall performance, this paper introduces an two classifiers organization structure to recognize the facial expression when classing by the combined algorithm and gets good results.(5)This paper selects the JAFFE expression database as the training sample, and verifies recognition result of this combined algorithm. This algorithm introduced by this paper can get good recognition rate which is proved by the experimental data.