Facial Expression Recognition Method Based On Depth Convolution Neural Network

People's emotional information is mainly expressed through a rich facial expression,With the advent of the information age,intelligent computing technology is becoming more and more popular in people's lives,and the recognition of facial expression features has become more and more obvious in home entertainment,intelligent terminal human-computer interaction,security field and even medical health.However,in the study of machine application,the expression is a more difficult to extract the characteristics of the traditional artificial selection of the characteristics of the method,it is difficult to obtain efficient interpretation of facial expressions,and for the traditional machine learning classification algorithm,the artificial The choice of facial features trained by the shallow network classifier with the generalization ability is still at a large level of limitations.Since 2006,the depth of learning to the rapid development of the situation,and demonstrated better than the traditional machine learning algorithm for the powerful performance of various applications for the field has brought new breakthroughs and development prospects.Different from the traditional method of robot learning,the depth learning algorithm is more in line with the law of natural biological self-learning,and it can be self-learning of the sample,like the natural creatures,under the condition of supervision or unsupervised,So as to get the abstract features of the sample,and have excellent network generalization ability.In this paper,according to the particularity of 2D facial expression images,the depth convolution neural network is applied to facial expression recognition to study the feature.In order to solve the problem of face image in addition to the partial organ posture region,the other regions are relatively redundant In this paper,the original facial expression images are preprocessed,and several facial features regions are detected as input images,and the following three kinds of network structures are designed for feature learning:1.In order to expand the facial features as much as possible in the process of facial expression recognition and to improve the separability of the extracted facial features,a fusion of improved LBP expression features and deep convolution(LBP-DCNN),the LBP operator is improved by the operation rules matching the expression feature region,and the convolution neural network with optimal performance is obtained by contrast experiment.Then,the expression vector is extracted The improved LBP feature is then trained in an 8-layer depth convolution neural network to extract the expression depth of abstraction,abstract the abstract feature and improve the LBP feature to fuse,train a seven-class output layer,and finally get the recognition result of the expression.2.A deep convolution neural network(AL-DCNN)based on auxiliary task estimation is designed to take into account the special characterization ability of facial expression in deep swept neural network.Firstly,a master task is trained to learn the DCNN network to get the shared feature weight matrix,and then the eyebrow gesture and the mouth posture local map of the expression sensitive area are merged as a supplementary task to estimate the branch task to share the feature weight matrix mapping.The task of estimating the classification results of the task,and finally optimizing the classification performance of the main task learning and improving the generalization ability of the deep convolution network in the expression recognition.3.In order to avoid the fact that the pooling layer in the general convolutional neural network,due to the simple downsampling operation,will lose some of the characteristics of the layer of the convolution layer and the full connection layer output contains only abstract information and the lack of a lot of shallow local features(WP-DCNN),which combines the multi-scale wavelet transform,which ensures that the characteristics of the convolution layer can effectively perform the complete feature transfer in the pool layer,But also in the full connection layer to expand the shallow learning to get the local characteristics of the expression,so that the entire network structure of the expression features can be better described and improve the recognition results.From the experimental results of this paper,it can be seen that the design of three different improved depth convolution neural networks is more conducive to the improvement of recognition performance in the field of expression recognition than the simple deep convolution neural network.Open database JAFFE expression database and CK + expression database were compared and analyzed.It can be seen from the experimental results that the three improved depth convolution neural network expression recognition methods can achieve high classification accuracy.