Research On Key Technology Of Facial Expression Recognition Based On Multi-level Salient Region

Facial Expression Recognition(FER)aims to recognize the class,for example,anger and surprise of expression based on a facial image.It has been widely used in various fields such as artificial transportation,human-machine interaction and healthcare.It has also attracted increasingly research interest in the fields of computer vision and deep learning(DL).Despite the fact that many researchers have proposed a lot of methods regarding facial expression recognition and achieved remarkable improvement on accuracy and reliability of the recognition results,facial expression recognition is still a challenging research topic.One of the most significant components of a FER system is feature extraction,which aims to obtain discriminative and robust image representation for recognition.This thesis is concerned of feature learning and salient region detection for FER,which targets to improve the accuracy of the method.The main content and contributions of this thesis are summarized as follows:(1)this thesis studies the basic convolution neural network(CNN)feature extraction method for facial expression recognition(FER).Traditional methods directly use convolution layer output feature map as image feature vector,which has high dimension and poor generalization,and is not conducive to practical application.This paper studies several methods to construct feature vectors based on pooling,multi-scale local pooling and aggregation.The advantages and disadvantages of several methods are compared and analyzed through experiments.The advantages of CNN image features of R-MAC compared with traditional image features in face expression recognition are mainly analyzed,which provides a reference for facial expression recognition system to select convolution neural network(CNN).(2)this thesis studies a method based on multi-level salient region detection for facial expression recognition(FER).This thesis analyzes the visual attention model and the salient regions in face images,and designs a coding network for processing multi-level feature images.The designed neural network consists of two parts: image feature extraction network and multilevel salient region detection network,which are used to extract different levels of feature map and detect salient region respectively.This method can combine the shallow features and deep semantic features of the image to predict the salient region of the feature map.Compared with the salient region detection method based on single level,this method has higher robustness.Through experimental comparison and result analysis,it is found that the recognition accuracy of 73.06% and 81.86% is achieved on fer 2013 and RAB-DB data sets.(3)this thesis studies a new image feature learning method based on triplet network structure to face expression recognition.Because the traditional convolutional neural network(CNN)based facial expression recognition(FER)method can not effectively learn the intra group and inter group differences of images,this paper studies the construction of triplet data set suitable for facial expression recognition,and trains the convolutional neural network(CNN)based on triplet loss function.The obtained model can minimize the distance of the same class images in the feature space,and maximize the different class images at the same time Like the distance in the feature space.This method provides a solution to solve the imbalance problem of intra group and inter group differences in facial expression recognition(FER).Experiments on fer 2013 and CK + face recognition data sets show that the accuracy is 75.61%% and 97.85%.The results show that the proposed method can effectively improve the classification accuracy of facial expression images.