Facial Expression Recognition Based On Convolutional Neural Network

Facial expression recognition is an important research topic in the field of computer vision.It plays a key role in non-verbal communication and can be applied to human-computer interaction,social robotics,video games and other fields.Facial expression recognition is affected by many factors,such as age,gender,race,culture and so on.Even if the same person's expression image,its illumination,head posture may be different.The traditional expression recognition algorithm needs complex manual feature extraction,which takes a long time,and the accuracy of expression recognition in complex scenes is not high.Convolutional neural network(CNN)has a strong ability of autonomous learning and can automatically extract expression features.Therefore,this paper mainly studies facial expression recognition based on convolution neural network.Aiming at the problems of deep neural network in expression recognition,this paper mainly does the following work:(1)Aiming at the problems of large parameters and resource consumption in traditional VGG network for facial expression recognition,a face expression recognition method based on Modified VGG13(M-VGG13)is proposed.Firstly,a batch normalization(BN)layer is added after each convolution layer of traditional VGG13 to speed up the convergence of the network,and the Global Average Pooling(GAP)layer is used to replace the full connection layer,which greatly reduces the amount of network parameters.Secondly,the size of the input image is reduced to reduce the complexity of the model,and the group convolution is adopted to further reduce the amount of parameters.Finally,the suitable optimization algorithm and image enhancement strategy are selected in the training.The effectiveness of the improved algorithm is proved on real scene data sets RAF-DB and FER2013 Plus.(2)Focus on the problem that M-VGG13 still has many redundant features,two model compression strategies are designed.In the first strategy,deep separable convolution is introduced into M-VGG13,part of the convolution layer is replaced by deep separable convolution layer,and three lightweight M-VGG13 models are designed.The second strategy uses the channel pruning method of network slimming,takes the ?factor in BN layer as the index to measure the importance of the channel,and prunes the channels with smaller ? factors according to the preset threshold to realize the simplification of M-VGG13.The effectiveness of the two compression strategies is proved on RAF-DB data set and FER2013 Plus data set.