Facial Expression Recognition Using Parallel Convolutional Neural Networks

In social language,in addition to natural language and body movements,facial expressions can express many emotions that natural language and body movements cannot express.It not only means that it is an expression,but also a psychological state that cannot be concealed.Based on the above reasons,facial expression recognition has a very important research and application value in human-computer interaction,medical treatment and even public safety.Although current researchers have designed some effective methods for feature extraction,there are still some problems that need to be solved urgently.However,in recent years,deep learning has achieved remarkable results in various fields.In particular,convolutional neural networks have shown unparalleled advantages in the fields of image recognition and detection,which brings hope to real-time facial expression recognition.This paper selects the deep learning which parallel convolutional neural network model to recognize facial expressions,and mainly performs the following tasks:(1)Design a parallel convolutional neural network which consisting of two sub-models.In the training process,the two pictures separately sent to sub-models to ensure that the features learned have positive significance for facial expression recognition by the network.In the testing period of the network,the pictures are input into a single convolutional neural network,and then generated the prediction of the pictures is.(2)Design a model for face expression recognition.The model which is referring to the Inception model,is suitable for face expression recognition.Compared with the traditional convolutional neural network models such as AlexNet,Google Net,etc.,it has a smaller amount of calculation,making real-time recognition possible.(3)The design loss function makes it have a lower loss rate when identifying different but the same expression,and reduces the influence of different individuals’ faces on the expression recognition accuracy.This article is train and test on the Fer-2013 data set.Through the final experimental results and has achieved 69.9% accuracy,which exceeds the recognition rate of other methods in this database.