Facial Expression Recognition Based On Gabor Wavelet Transform And Classification Tree Method

Facial expression recognition is an important part of emotional calculationresearch, which involves many areas, and have important applications on driving, thenetwork teaching, human-computer interaction etc. Accurate recognition of facialexpressions, can promote the development of human-computer interaction intelligentsystem, and make the computer more effectively supervise the auxiliary work.Firstly, this paper explains the background and the meaning of research, itintroduces facial expression of the current situation of domestic and foreign research.Secondly, it introduces the general framework of facial expression recognition, theanalysis of the classical method of facial expression recognition and relatedimprovements, and summarizes the advantages and disadvantages. The task in thispaper focuses on the key issues of facial expression recognition, such as featureextraction and facial expression classification. The main research contents andinnovative work in this paper are shown as follow:(1) In order to obtain a good result of facial expression recognition, the first stepis image preprocessing. In this paper, the position of the pupil and the horizontalcenterline of mouth have been located quickly and accurately, and according to theposition information and the structural characteristics of face, the more efficientexpression sub-regions are cut accurately. At the same time, these sub-regions wouldbe normalized.(2) The method of the feature extraction based on Gabor wavelets are discussedin details. Firstly, optimizing the Gabor filter’s window, not only strengthen the Gaborfilter feature extraction capability, but also accelerate the convolution time. Then, inorder to overcome the high-dimension of Gabor features, a method which codes thedifferent orientation of Gabor’s amplitude at the same scale is proposed. It helps toachieve more effective information while it reduces the dimension of features. Finally,a way which fused the local facial regions is presented to improve the recognition rate.The recognition result which experiments on JAFFE database achieves96%. (3) On this basis, in order to strength universality of method, a classification treemethod to identify non specific facial expression has been presented, which uses thecharacteristics between the expressions effectively. Firstly, according to the differencein expression, classified7kinds of expression from coarse to fine. Then, at each layernode in the classification tree, setting different regions of the feature vector, andextract the features for classification using LDA algorithm. With the results on JAFFEdatabase, the recognition rate is82.38%, which verifies the effectiveness of theproposed algorithm.