Research Of Facial Feature Localization And Facial Expression Recognition Based On Cascaded Regression

Facial expressions can represent the state of a person,including his mood and cognition.It plays a very important role in the intelligent human-computer interaction system.Many researchers have made efforts in expression recognition and achieved huge success in the field.However,it brings challenges to expression recognition on account of the face appearance differences,as well as makeups which make the change of expression become both subtle and complex.As for the automatic facial expression recognition,how to extract the effective features is one of the key factors to achieve a high recognition rate.This thesis studies the algorithm of facial features localization and face recognition.There are many questions such as overfitting when using traditional cascaded regression model,high dimensions of extracted Gabor feature,low robustness of gesture factor over traditional geometric feature,so we studies the solutions of these problems.Details are as follows:Firstly,describing the cascaded regression model in detail and introducing SIFT feature descriptor which is necessary to cascaded regression model.Overfitting occurs when using ordinary least squares in every stage of cascaded regression,so this thesis adopts Support Vector Regression which constructs linear decision functions in high-dimensional space to solve the problem of the nonlinear decision function in original space,thus enhancing the generalization of model.The feasibility and robustness have been verified on LFPW data sets.Secondly,the multi-scale and multi-directional Gabor filter can express the face features effectively,but it results time-consuming because of the large dimension of feature.To solve this problem,we use LLE to reduce the dimensionality of feature,which not only reduce the dimension but also make the new feature more effective.And the experimental result shows that the LLE is better than the traditional PCA.Thirdly,Traditional facial expression geometric feature are obtained by ASM,AAM etc.However,these methods are not robust enough on account of the factors of illumination,pose,expression etc which lowers the performance of facial expression recognition based on geometric feature.To solve this problem,we use cascaded regression method to locate feature localization so that the geometric feature we acquired can be more feasible and robust.Finally,it's easy to miss the details when using single feature,so we propose a method based on the fusion of texture and geometry to improve the performance of facial expression recognition system.Experiments with images from the JAFFE facial expression database have validated the proposed features,and achieved good recognition performance with a Support Vector Machine classification method.