Facial Expression Recognition Research Based On SIFT Feature

Facial expression recognition which has been widely applied to human-computer interaction,medical treatment and cartoon industry is an important research direction in fields like computervision and pattern recognition. Developing a facial expression recognition algorithm will have abroad application prospect.This paper uses Dense SIFT(Dense Scale Invariant Feature Transform) algorithm to extractfacial features, analysis the data using the k-means clustering algorithm and uses spatial pyramidalgorithm to improve the clustered data and finally uses the support vector machine to make theclassification. The main research work are as follows:(1)Take a detailed study on the feature extraction process of SIFT algorithm and study asimplified algorithm of SIFT, called Dense SIFT. Experiments show that Dense SIFT algorithm issuitable for facial expression recognition.(2)Study the way to generate a dictionary using the bag of words model. This paper studied theeffect on the final recognition rate when using extraction interval as2,4,6,8between pixels anddictionary number as100,150,200,250,300,350,400,450,500,550,600respectively.(3)The data get by k-means clustering analysis is locality feature and have lost the locationinformation. Combined with spatial pyramid method, get the holistic features of every word in thedictionary. Study the effect on the facial expression recognition rate before and after the use ofspatial pyramid algorithm.(4)Combined with SVM(support vector machine), make the classification on the seven facialexpressions(anger, disgust, fear, happiness, neutral, sadness and surprise).Use a kind of self-definedkernel function, that’s HIK(histogram intersection kernel). Use different kinds of kernel functionslike RBF kernel and histogram intersection kernel to find the best recognition rate. The result showsthat the histogram intersection kernel is more suitable for facial expression recognition. The resultshows that the algorithms used in this paper are effective on JAFFE, the highest average recognitionrate is93.333%.