Facial Expression Recognition Based On WMCBP-WWEF Feature Fusion Using Random Forest

Facial expression recognition is a hot research topic of computer vision andpattern recognition, causes widespread concern of an increasing number ofdomestic and foreign researchers. The goal of facial expression recognition is tomake artificial intelligence can automatically identify person’s expression, and thenanalyze their inner emotions. It can enhance the friendliness and intelligence ofhuman-computer interaction, so it will has a broad application prospects andpotential market value.The expression recognition system mainly consists of the following parts: facedetection, image preprocessing, expression feature extraction and expressionclassification. This dissertation researches on some key issues of the static imageexpression recognition, improves some algorithms, and verifies by experiments.The main work and innovations are as follows:(1)Expression image preprocessing. Due to a variety of image size, angle, andthe influence of the illumination condition, the image is preprocessed by usingmean-variance normalization and geometric normalization. Successful pretreatmenthas laid a good foundation for improving the expression recognition rate.(2) Feature extraction. To deal with the deficiency of the existing multi-scalecentralized binary patterns, a feature extraction method called wavelet multi-scalecentralized binary pattern (WMCBP) is proposed based on wavelet transformation.This method can not only get more accurate multi-scale information, but alsogreatly reduce the computational complexity. Weighted wavelet energy feature(WWEF) is also introduced to get the details of the sub-picture information in orderto further enhance the facial expression recognition accuracy. These two types offeatures are complementary to some extent, and the fusion of them can enhance theperformance of WMCBP in facial expression recognition without increasing thecomputation obviously.(3) Expression classification. As a famous ensemble learning classifier,random forest algorithm (RF) is one of the classic classification methods in patternrecognition. RF can effectively deal with large data sets, give the better variableimportance estimation in the classification process and can’t produce overfitting.So RF is selected as facial expression recognition classifier. Some parameters of the RF model, such as the number of decision tree, the number of candidateattribute, node splitting strategies are discussed.Experimental results show that the feature extraction and expressionclassification methods used in this dissertation are feasible in facial expressionrecognition and can obtain good recognition results.