Research On Facial Expression Recognition Algorithm Based On Deep Learning

With the development of human-computer interaction technology,facial expression recognition has attracted more and more attention.Facial expression recognition technology has been widely used in the fields of human-computer interaction,intelligent medical treatment,online education,intelligent transportation and elderly care services.With the maturity and development of deep learning backbone network.The research on human facial expression classification uses a method from the traditional engineering-based complex manual feature extraction to the end-to-end convolutional network processing.Traditional facial expression recognition tasks only need to classify several facial expression categories.However,with the continuous development of facial expression recognition tasks,the number of categories needed to be processed by facial expression recognition algorithms is increasing.The original facial expression algorithms can no longer meet the requirements of current facial expression classification tasks,and the classification accuracy is low on many data sets with large number of facial expression categories.In order to solve the problem of low classification accuracy when there are a large number of expression categories,an improved algorithm is proposed.In view of the interference of complex background information in the real environment,and the difficulty of classification when the number of expression classification is large,this paper mainly does the following work:(1)Although the residual network can extract abstract semantic information well,it pays the same attention weight to all characteristic information,which results in that information related to real classification cannot be focused on and thus the classification accuracy cannot be further improved.A facial expression recognition algorithm based on Resnet network is proposed.Firstly,attention mechanism is added to each residual module in Resnet50 network,so that the network can extract facial expression features more accurately.Secondly,Sigmoid activation function is used at the full connection layer used in classification to transform the multi-label classification task into the superposition of multiple binary classification tasks,and the training is performed on the real scene data set Emotic,which proves the effectiveness of the improved algorithm.(2)To solve the problem of class imbalance in multi-label classification tasks,starting from the loss function of the model,the Softmax+ cross entropy algorithm is proposed.The Cos Face cosine penalty method is selected from the penalty methods of the loss function,and the Softmax and Cos Face are combined to obtain a loss function which is more suitable for processing multi-label and multi-classification tasks.The effectiveness of the algorithm is proved in the EMOTIC data set.