Research On Facial Expression Recognition Methods Based On Feature Learning

Facial expression recognition(FER)has important research value.To achieve effective FER,extracting robust facial expression features is of great importance.However,facial expression images are always affected by facial identity,gender,age,race,or environmental illumination,facial pose and other factors,which seriously interfere with the extraction of facial expression features.In addition,there are also great inter-class similarities between different expressions.The above problems make the FER task a great challenge.In order to solve the above problems,this thesis proposes two FER methods:First,an FER method based on Disturbance-Disentangle Learning(DDL)is developed.In order to achieve effective FER,it is very important to alleviate the influence of various disturbing factors in the facial images.However,most facial expression databases only pro vide the labels of expression,but lack the information for other disturbing factors.Therefore,a new FER method based on disturbance-disentangle learning is proposed.It is capable of simultaneously and explicitly disentangling multiple disturbing factors by taking advantage of multi-task le arning and adversarial transfer learning.Experimental results on five popular databases demonstrate the superiority of the proposed method.Second,an FER method based on Feature Decomposition and Reconstruction Learning(FDRL)is developed.There are high inter-class similarities between different expressions,which limit the performance of the FER model.The expression information can be viewed as the combination of the expression-shared information and the expressionspecific information.Based on this observation,an FER method based on feature decomposition and reconstruction learning is proposed to learn the expression-shared information and the expression-specific information in expression images.Specifically,the proposed method mainly includes two key modules:a feature decomposition learning module and a feature reconstruction learning module.Experimental results on five popular databases demonstrate the superiority of the proposed method.