Study Of Geometry Features-based Facial Expression Recognition And Its Applications

Facial expression recognition(FER)is the key of such an urgent need as well as the premise of computers' understanding about human emotion.It is an effective way to make the service more intelligent.FER is a classic research topic in the field of computer vision referring to the task of identifying the emotion of a human face from static images,image sequences,or videos and so on.The geometry-based FER is mainly based on the following theory:the change of facial expression must be accompanied by the change of the shape of the face.For example,the emotion of“surprise”is usually expressed by an open mouth,wide eyes and so on.However,some limits exist in this kind of methods based on landmarks:advanced features obtained by landmarks limited in two dimensions may have more inherent characteristics with better performance than itself even if landmarks have outstanding characteristics;besides,not everyone in advanced features has contribution to recognition which means irrelevant and redundant features may be included.Some methods can be used to test and verify if less features can achieve better classification accuracy.Feature extraction is the most critical part of facial expression recognition.A good feature extraction and selection method largely determines the classification accuracy of facial expression recognition.Aiming at the problems in the non-ideal environment and the low accuracy and generalization of facial expression recognition,this paper deeply analyzes the existing facial expression feature extraction methods and proposes new facial expressions based on geometric features The recognition algorithm realizes the facial expression recognition of the face image and designs the primary facial expression recognition system based on this.The main contributions of this article are as follows:1.This paper proposes an IM high-dimensional manifold feature extraction algorithm based on LDDMM-curve.This algorithm is based on two models,NP and PE,to simulate two changes in the shape of the face when the subject undergoes emotional changes,and then extracts the initial momentum vector(IM)as the representative feature of the facial expression.Feature sets are used for feature selection,feature subsets with optimal classification performance are selected for feature fusion,and the fusion feature classification performance is verified.Finally,experimental verification and comparative analysis of the proposed algorithm on two data sets of CK + and Face ++ show that the expression recognition rate and generalization ability of this method are better than other methods.2.This paper proposes a dimensionality reduction feature set extraction algorithm based on FNT-SVM-RFE-CBR.This algorithm implements feature extraction from the original high-dimensional feature set for the first time.Considering that the correlation between features will interfere with the determination of feature importance,for each feature set,we use ten-fold cross-validation to find the optimal SVM model parameters.And correlation parameters,and use the LOO method to obtain the classification performance of the feature subset after feature selection,select the feature subset with the best classification performance for feature fusion,and verify the fusion feature classification performance.While reducing the feature dimension,the classification performance is improved.Finally,experimental verification and comparative analysis of the proposed algorithm on two data sets of CK + and Face ++ show that the expression recognition effect of this algorithm is better than other algorithms.The performance of the proposed algorithm on CK+and Face++datasets was evaluated.Firstly,the proposed algorithm was compared with some existing algorithms to verify the superiority of the algorithm.After that,the proposed algorithm was assessed with the classification accuracy of original IM feature set and subset obtained from SVMRFE,the higher classification accuracy and lower feature dimension proves the effectiveness and superiority of the proposed algorithm(taking the feature subset of NP method of Face++as an example,classification accuracy of original IM feature set,feature subset by SVM-RFE and feature subset by FNT-SVM-RFE are 88.89%,89.91%and 90.12%respectively).Secondly,the fusion feature is much better than any IM feature subset of NP or PE,especially when the fusing rate is 9:1 between NP and PE.Lastly,performance of CK+is better than Face++whether the feature subset of NP or PE or fused feature set(NP:95.37%>90.12%;PE : 94.14%>88.27%;fusion feature : 96.60%>91.67%).In summary,the proposed algorithm in this thesis can extract more inherent feature and select much more important ones to improve the classification accuracy which has a great practical value.