Study Of Facial Expression Recognition Method Based On Weber Local Descriptor

Facial expression recognition has been received extensive attention by domestic and foreign researchers in the field of intelligent information processing, digital image processing and so on. In this dissertation, the thesis firstly makes further research of Weber Local Descriptor (WLD). After that, against its inadequate on the aspects of local minutiae characterize and partial occlusion in facial expression recognition, related improved algorithm and fusion methods are proposed and verified by a lot of experiments. The main work and innovations are as follows:(1) In order to solve the insufficient that facial expression cannot be well characterized by single feature, a facial expression recognition method based on weighted Weber Local Descriptor (WLD) and Histograms of Oriented Gradients (HOG) is proposed. First of all, the images are divided into blocks and weighted according to their contribution to facial expression recognition and extract WLD features of these sub-blocks. Then HOG features of the sub-blocks with large contribution are extracted to achieve the local significant features. Under these conditions, the final histograms are generated by cascading of WLD and HOG. At last, the chi-square distance and nearest neighbor method are used for classification. Compared with other related algorithms, the results verify the effectiveness of this method.(2) Against the limitations of the representation of original Weber Local Descriptor, a new method named Pyramid Weber Local Descriptor (PWLD) is proposed. Different layers are selected in this method based on the size of the images and richness of information. The width and height of the blocks in each layer is the half of those in upper layer. By adjusting the parameters in WLD, the thesis extracts the WLD features of each block and merges them to form the final PWLD features with a certain weight. Experimental results show that PWLD features have better representation of local information than WLD features.(3) Aiming at the problem which is partial occlusion in facial expression recognition, a method that fuses the global and local features is proposed in this dissertation. In global aspects, the thesis firstly calculates the differential excitation of the images and utilizes the information entropy, Otsu method, opening operation and closing operation in morphological to detect the occluded area based on different richness of information in occluded portion and non-occluded portion. Then replace the occluded portion with the images reconstructed by Principal Component Analysis (PCA) and extract PWLD features of the processed image as an overall description. After that, the outputs of SVM are fitted to the probabilities of the target category by using S-type function. In local aspects, through partitioning the images into overlapping blocks, the thesis extracts the WLD features of these blocks weighted adaptively by information entropy. Finally, chi-square distance and face similar block summation methods are used to obtain the probabilities belonging to each expression. In the phase of classification, Dempster-Shafer theory of evidence is utilized to fuse the features from different aspects so that a more reliable and accurate result can be achieved. Experiments on simulated occluded facial expression database verify the validity and fault tolerance of this method.