Research On Face Expression Recognition Algorithm Based On Deep Learning

Facial expression recognition is an extremely challenging task of pattern recognition,which has important practical significance in such fields as medical research,traffic safety,public security,criminal investigation and interrogation,film and television entertainment,etc.For the foreseeable future of the world of efficient human-computer interaction,the correct and rapid recognition of user expressions is also an essential technology.In the past decade,the research progress of facial expression recognition has advanced by leaps and bounds.Indoor research with small-scale data has shifted to outdoor research with larger data volume.Research methods have also changed from traditional methods to deep learning methods.The main difficulties of deep expression recognition are as follows: the lack of effective training data,the existence of a large number of redundant information unrelated to expressions,and the objective existence of ambiguous expressions.This paper mainly focuses on the ambiguous expression problem.In order to improve the recognition accuracy of deep expression recognition,the following research works have been done in this paper:(1)In view of the subjectivity of expression classification and the reality of life,this paper proposes a deep residual expression recognition network to strengthen the distinction between classes.Firstly,through prior experiment and threshold algorithm,the relationship between classes is quantitatively analyzed to verify the viewpoints of this paper and obtain the set of strong and weak relationships of various expressions.Then,the relationship between facial expressions is integrated into the network branch design through the form of fixed parameters.Finally,do the training on the overall network with the added branches.Experiments on popular large field databases show that the proposed method has excellent performance compared with the basic method,and has certain competitiveness among a series of advanced methods.(2)To solve the problem of insufficient information of single labels in expression classification,this paper proposes an inverse k-fold deep expression recognition algorithm of label distribution,which converts one-hot label into probability distribution that is more consistent with the current cognition of expression recognition.Firstly,the training data was divided into K pieces,and the training was carried out on one piece of data,and the prediction was made on the remaining K-1 pieces of data.Then,through a cycle of K times,the label distribution of each data marked K-1 times is obtained for the training data as a whole.Finally,these labels are used as the label distribution of the data to make predictions for the test set.The method has been applied to several classic CNN frameworks,and the experiments on RAF-DB and Affect Net show that the this method is obviously effective in the case of large data volume with poor label quality(3)In order to solve the problem of insufficient information of single labels in expression classification,a label distribution expression recognition algorithm based on asymptotic truth value is proposed in this paper.Under the premise of not splitting the database,the original information of the database is fully used to complete the generation and utilization of label distribution.Firstly,in data training,single label learning is used to collect the mean value of the overall distribution of data.Then,on the granularity of data batch,the true value of data label is approached gradually.Finally,the whole network model is retrained using the generated data label distribution.Experimental results show that this method can improve the accuracy of the network model obviously,and has certain competitiveness compared with the advanced algorithms.