Research On Facial Expression Recognition In The Wild

Emotion plays a very important role in the process of transmitting information in human social communication.Facial expression is one of the most powerful,spontaneous and universal signals of human state and intention.The establishment of an automatic face emotion recognition system is a hot topic in recent years.At present,the deep neural network is frequently applied to learn the automatic facial expression recognition tasks.Due to the complex scene in the wild,it is difficult to collect the facial expression data set in the wild.The existing deep learning algorithms are easily overfitting due to the lack of sufficient training data on the data set in the wild.It takes a lot of training data to drive the deep learning model to capture the subtle distortion associated with expression.In addition,with limited training samples,complex environmental factors in wild environment,such as illumination change,occlusion,large attitude change and other influencing factors,which are the main factors leading to low emotion recognition rate.In view of the above two problems,this thesis proposes facial expression recognition methods in wild environment based on few-shot learning and disentangled representation learning.The main contents and innovations of the thesis are as follows(1)Facial expression recognition method in the wild based few-shot learning This method introduces the idea of few-shot learning to construct a FER deep learning framework--Convolutional Relation Network(CRN)in the wild.The framework learns by comparing the feature similarity among the sufficient samples,which is applied to recognize new classes with few samples.Specifically,the method aims to learn a metric space in which the classification can be done by calculating distances,and the predictive power of the network can be generalized by employing the strong discrimination ability of depth expression features.In order to realize this,the features are constrained to maximize the distance between the features of different classes and find the commonality of the same class.Extensive experiments on three public datasets in challenging wild environments(RAF-DB,FER2013,and SFEW)show that the proposed method is significantly superior to the most advanced methods(2)Facial expression recognition in the wild based on disentangled representation learning.This method introduces the idea of disentangled representation learning to learn the significant emotional features of human face in the wild.Through the encoder to get the higher-level features,then face is obtained by Gaussian sampling to generate factors associated,finally,from the factor is obtained by weighting the emotional related factor,emotional characteristic extraction,and to classify emotions.This method reduces the influence of non-affective correlation factors on the recognition task by dissolving the affective correlation factors from the given image,so as to reduce the influence of complex background factors in the wild on facial expression recognition.The experimental results show that the method of extracting emotional salient features based on disentangled representation learning can further improve the performance of emotion classification in the mainstream wild environment data set,and can tolerate environmental changes to a certain extent(3)Design and implement the prototype system of facial expression recognition in the wild.This system uses the framework of MATLAB,Python programming language and deep learning algorithm to design and implement the prototype system The system includes face detection,model loading and emotion recognition.Among them,the face expression recognition method based on few-shot learning and disentangled representation learning are both embodied in the prototype system.The effectiveness and practicability of the proposed method are verified by the prototype system.