Facial Expression Recognition Method Based On Feature Fusion

Facial expression recognition is a very challenging research. Because of facialexpression’s potential applications, It has been widespread concern by manyscholars. With the development of multimedia technology and network video, It isvery important to analysis and understand people’s emotion. Facial expression isdivided into three steps: face detection, feature extraction and expressionclassification. Feature extraction is the most important part of this three-step. Theeffect of feature extraction will directly affect the facial expression recognition rate.This thesis summarizes and analyzes the typical algorithm of expressionrecognition at home and abroad. In this thesis, we introduce the key technologies infacial expression recognition. Facial expression recognition based on multi-featurefusion algorithm is carried on depth-study in this thesis. The main work of thisthesis as follows:1、Facial expression method based on Gabor feature fusion is proposed in thisthesis. First of all, cut the sub-images of eyebrows, eyes, mouth and nose from anface image by using geometry feature of three chambers and five holes, thesesub-images can express facial expression effectively. These sub-images areprocessed by Gabor wavelets. The component feature weights are calculated byusing principal component analysis(PCA) algorithm and fisher linear discriminant(FLD) algorithm. The contours of the face change with the facial expression change,so it is very necessary to add global feature in expression features. The weightedcomponent features fuse with the global feature to get a feature fusion matrix, andwe can get the facial expression’s class from this fusion matrix. Experimentalresults show that the algorithm proposed in thesis has much more accuraterecognition rate compared with the global Gabor feature.2、Facial expression method based on higher order autocorrelation (HLAC)feature fusion is proposed in this thesis for improving the efficiency of facialexpression recognition. The features of HLAC can reflect the inherent edges anddetails’ information in the face image, and these features are not sensitive to thedisplacement. We firstly use25HLAC template to extract the features of gradientand details from the global image and local images. And then calculate the weightsby using fisher linear discriminant algorithm and fuse the local features and global features into a weighted HLAC (WHLAC) feature matrix by using these weights.At last, the FLD algorithm is used to classify the expression. Experimental resultsshow that this algorithm has higher recognition rate than the methodes based onlocal binary pattern (LBP) and based on global HLAC. This algorithm also hasbetter real-time than the method of Gabor feature fusion.