Facial Expression Recognition Method Based On Monogenic Wave And Sparse Representation

The research on facial expression recognition is regarded as an important part in the field of visual system and man-machine interaction, it attracts extensive attention from researchers all around the world. This thesis specifically studies the feature extraction and classification of facial expression. Researches on application of monogenic signal and sparse representation classification have been explored, and based on these researches, the corresponding solutions of facial expression recognition towards facial occluded conditions are proposed in the thesis. The main research contents and the creative views of the thesis could be divided into several following parts:(1)Because of the limitation of monogenic waves' extraction from expression texture, an adaptive facial expression feature fusion method is proposed based on monogenic binary pattern (MBP) and histogram of monogenic orientated gradients (HMOG). To begin with this method, we get information of orientation, phase and magnitude component from filtered images by monogenic waves, besides, extract and fuse texture information from orientation and magnitude component to recognize facial expression. However single texture feature have some inadequacies for facial expression recognition. First of all, this thesis describes a new kind of HMOG feature inspired by thoughts of histograms of oriented gradients (HOG) to indicate shape information of facial image. Secondly, the best adaptive weights of expression block are acquired through information entropy. Finally, we classify the expression image using nearest neighbor method to get the final result of proposed method. In order to verify the efficiency of the methods above, we do a large number of experiments and it turns out our methods have great ability to explain expression feature.(2)In order to make up for ignorance of monogenic phase information in second chapter, and enhance robustness of sparse dictionaries towards individual identity difference, this thesis proposes a novel facial expression recognition method based on the monogenic multiple features and sparse representation fusion. First of all, the preprocessed images of facial expression are filtered by the monogenic signal with the purpose of acquiring the information of monogenic magnitude, orientation and phase. Secondly, the final facial expression features are fused by the information of monogenic magnitude, orientation and phase using the histograms of monogenic oriented gradients, enhanced monogenic phase and monogenic binary pattern method individually. Finally, the l1-regularized least squares method is used to optimize the value of the three classifiers and achieve decision-making fusion of facial expression recognition.(3) In order to solve the facial expression recognition problem with partial occlusion existing in real life. The thesis proposes a double-deck model. First of all, we adopt sparse representation classification based on global expression state feature in global layer, then compute the confidence level of classified results using sparsity concentration index (SCI), after that, we put test samples with low confidence value into local layer constructed by local monogenic feature. Finally, several results of subarea sparse representation classification are voted and weighted to get the final results.