Facial Expression Recognition Based On Image And Deep Learning

In the 21st century,the society is not only the communication between people,but also the interaction between people and machines.This is becoming more and more frequent.The intelligent machine is gradually changing people's life style.Therefore,it is particularly important for machines to learn to analyze human emotions.In the past,researchers used various mathematical methods to construct image features.These are combined with traditional machine learning methods to recognize facial expressions in images,and some results are achieved.However,there are some problems in the facial expression,such as insufficient samples and unbalanced data;the information loss of artificially constructed features and the low accuracy and generalization ability of traditional machine learning methods for facial expression recognition;the individual facial expression data sets have the problems of tag anomaly,image anomaly,and the height difference of face geometry structure and facial appearance.Facial expression recognition still faces great challenges.In the past ten years,deep learning method has become a hot topic in the field of machine learning and attracted more and more attention in the field of machine learning.Therefore,more and more people used various neural network models to study facial expression recognition,and obtained better model performance.In view of the above difficulties in facial expression recognition,the main research work of this paper is as follows:(1)In this paper,a feature selection network(FSN)is proposed.A feature selection mechanism is embedded in Alexnet to automatically extract and filter face features.The feature selection mechanism designed in this paper can effectively filter irrelevant features,and emphasize relevant features according to the learned feature map.Experimental results on several databases show that FSN is superior to Alexnet.In addition,the generalization ability of FSN is better than Alexnet in cross validation experiments of different data sets.(2)Aiming at the problems of tag error,emotion ambiguity and picture error in fer2013 data set,this paper proposes a sample weight allocation algorithm to distinguish the difficulty of sample identification in training set.On the other hand,it makes the learning process of network from simple to complex.By comparing the expression recognition on weighted samples and non weighted samples,it is verified that the algorithm can improve the accuracy of expression recognition and accelerate the convergence speed of the network.(3)This paper proposes a regularization method based on feature sparseness,which can learn deep features with better generalization ability.The regularization constraints are integrated into the loss function and optimized by deep learning.Compared with neuron sparseness and weight matrix sparseness,the algorithm has excellent performance.And in the cross dataset experiment,it also shows the better generalization ability and better model generalization ability.