Research On Facial Expression Recognition Based On Multiscale Image Local Feature Descriptors

The progress of information technology has laid the foundation for the development of artificial intelligence and pattern recognition. As one of the hot research topics,facial expression recognition is an indispensable key technology in the field of human computer interaction. However, the difference of facial expression brings great difficulty to face recognition. In order to solve this problem, this paper focuses on the correct classification of different facial expressions in face recognition.For the diversity and complexity between facial expressions, this thesis mainly carries out the study of four parts:the face image preprocessing, feature extraction, classifier design and experimental verification. The specific implementation procedures of the proposed algorithm are given as follows:1. Facial expression image preprocessing. The pretreatment of this thesis consists of two aspects:Geometric preprocessing and gray level preprocessing. The geometric preprocessing, which includes face images, tilt correction and geometric normalization, is used to obtain the final expression image. The gray preprocessing is to perform image histogram equalization and to reduce the influence of illumination on recognition results2. Facial expression feature extraction. In the feature extraction process, this thesis proposes a center multiple binary patterns (CMBP) operator the operator is of multiple second value spread mode (MBP) operator and improve the ability of feature classification. Gabor wavelet and CMBP operator combined feature extraction algorithm applied to facial expression recognition, improve the recognition class. The proposed operator has the following advantages:(1) The CMBP operator uses the difference between the "nearest neighbor" for the improvement of the identification accuracy. (2) The CMBP operator considers the center pixel point, and assigns the maximum weight to the center pixel, which improves the accuracy of the measurement and reduces the dimension of the feature.3. Classifier design. This thesis uses the support vector machine to design the classifier. We analyzed four key problems, which are the selection of cross validation method, the selection of kernel function, the determination of parameters and the construction of multiple classifiers.4. Experimental verification. In this thesis, we use Gabor wavelet, LBP, MBP, CMBP, and the united method of Gabor and LBP were separately used for feature extraction, and the united method of Gabor and MBP were also used for feature extractionand, and also compared with the proposed algorithm based on Gabor wavelet and CMBP feature extraction methods.The experiment was implemented to verify the superiority of the proposed algorithm.