Facial Expression Recognition Based On Deep Learning

Facial expression,as one of the ways to express human emotion information,has important research value.The excellent automatic facial expression recognition technology has a wide range of applications.For example,to provide a friendly man-machine interaction experience;It also plays an important role in maintaining public security,through the timely and accurate analysis of human facial expressions in public areas,prevent terrorist violence and provide aid.In the medical field,the patient can also be monitored in real time to reduce the accidental death rate.With the development of modern technology,artificial intelligence and machine learning have gradually invaded people's daily lives.In the fields of image and computer vision,the research of facial expression recognition technology has become a hot topic.In recent years,deep learning has developed well.It has shown unique advantages in text analysis,natural language processing,image processing and so on.This paper combines the convolutional neural network with facial expression recognition,and gives a further study on it.The main work is as follows:One is the improvement of traditional Inception structure.Inspired by the traditional Inception structure,this paper proposes a deep convolutional neural network for low pixel facial expression recognition.By using different convolution kernels to extract features from the input information,and try to maintain the consistency of the local receptive field of the convolution block.Therefore,the consistency of local input information is maintained in the feature maps.Compared with the simple splicing method of feature fusion in the traditional Inception structure,the more obvious features are retained by taking the maximum value of the feature element,it also reduced the number of parameters,Simplified the complexity of the model.In order to solve the problem of gradient disappearance and over fitting due to the increase of network depth,we learned from the DenseNet network.On the basis of the original network structure,we added the simplified Dense structure and the pool structure of the traditional Inception structure,so as to improve the performance of the network model.The other is that the rotation invariance of general convolutional neural network is poor.Therefore,a new convolution method is proposed in this paper.That is,in the case of keeping the general network structure,add the rotation angle information to the convolution process for the large rotation angle.The convolution method adapted to the pictures of different rotation angles is used,improved the rotation invariance of the convolutional neural network.Experiments show that this algorithm makes up for the decline of accuracy due to the introduction of rotating pictures.Not only kept the accuracy about the upright face images in FER2013,but also improved the recognition rate with rotation picture rate.The validity of this method is proved.