| In recent years,facial expression recognition has been widely used in human-computer interaction,security and distance education.However,the traditional facial expression recognition method has many problems,such as complex feature design,difficult recognition,difficult to promote and so on.In order to achieve the purpose of high efficiency,high accuracy and automatic recognition of facial expression,it is necessary to study the method of facial expression recognition based on deep learning.In this paper,convolutional neural network is studied,and a new facial expression recognition method based on deep learning structure is proposed.Fer2013 data set is used in the experiment,and the main results of this paper are as follows:1.Two structures of net are proposed to solve the problem of low recognition rate of alexnet facial expression,a smaller convolution kernel is used to solve the problem of poor nonlinear expression ability of the network.The new network structure is reduced from two channels to a single channel,and half of the hidden cells in the full connection layer are reduced,which makes the network more suitable for solving the problem of less classification.The network is composed of four layers of convolution layer and three layers of full connection layer,and the recognition rate of facial expression is increased by 3%from 65%to 68%compared with the traditional methods(alexnet structure).The performance of the two network structures is better than that of the traditional methods,and the recognition rate of facial expression is improved by 3.38%and 3.07%respectively.Three structures of net are proposed to solve the problem of long training time of vggnet,the structure is simplified to solve the problem of too many network parameters;the training time of layer 13 network is reduced by 40%.The performance of three network structures is better than that of traditional methods,and the recognition rate of facial expression is improved by 4.14%,4.41%and 4.67%respectively.Two structures of net are proposed to solve the problem that the network with more than 20 layers is difficult to train,the residual network(resnet)is used to solve the problem that the training set is difficult to fit.The basic unit is the order of batch standardization activation function weight(bn-relu-conv).Two full connection layers are replaced by global average pooling.Resnet adopts residual learning block,which can build a deeper convolution neural network,and the number of layers can reach 100 or deeper.The convergence speed of the new structure network is faster and easier to train.The performance of the two network structures is better than that of the traditional methods,and the recognition rate of facial expression is improved by 7.08%and 7.44%respectively.2.Aiming at the problem that some pictures in fer2013 dataset are difficult to recognize,a method of facial expression recognition based on adaptive enhanced vggnet is proposed,which solves the problem that all images are equally concerned by the network.Using vggnet as a weak classifier,we train different vggnet by different weighting to the data set,and integrate all the vggnet.Vggnet is composed of five convolution groups and three full connection layers,and the hidden units of three full connection layers are reduced,and the output is set to 128,32 and 7 respectively.Compared with a single vggnet,the recognition rate of facial expression is increased by 6%from 68%to 74%.3.Based on the above research,an integrated facial expression recognition system is designed.The system combines opencv,dlib and other deep learning libraries,uses opencv to detect human face,and uses dlib to extract facial feature points;the system is simple and designed,running smoothly on personal pc,with fast recognition speed,and the recognition rate of facial expression is better than that of similar domestic technologies.The facial expression recognition method proposed in this paper can be used in artificial intelligence,education and other fields. |
PDF Full Text Request