Research On Convolutional Neural Networks For Expression Recognition

With the continuous development of human-computer interaction,people's lives have been greatly facilitated.People have gradually put forward new requirements for human-computer interaction.People hope that computers can recognize people's emotional state and provide better service according to the they emotional state.As the most intuitive expression of human emotions,expressions are buried with a lot of important information.Making full use of this information can further improve people's living standards,and thus has a huge application scenario.Facial expression recognition is the task that hope computer can accurately recognize people's expressions,which is the basis of human-computer emotional interaction.Therefore,how to make the computer accurately recognize facial expression has become the focus of researchers.As an important algorithm in deep learning,convolutional neural networks have achieved good results in many computer vision tasks.Therefore,facial expression recognition based on convolutional neural networks has attracted the attention of researchers.The paper focus on three problems when using convolutional neural network to facial expression recognition: convolutional neural network is not sensitive enough to the subtle changes of expression images,convolutional neural network does not have rotation invariance,and when fussing the low-level features and high levels features to facial expression recognition,the number of parameters will growth rapidly.Proposing corresponding solutions.The main contributions of this paper are as follows:(1)To solve the problem that facial expression recognition based on convolutional neural network is not sensitive enough to the change of detail of facial expression,the paper propose a facial expression recognition method of convolution neural network fusing ROI(Region of Interest)probability distribution.Firstly,the image is divided into a series of ROI images;then,the ROI images is input to the convolutional neural network for training,making convolutional neural network learn the global information and the local information of images;finally,fusing and normalizing the ROI category probability distribution that output by the Softmax classifier during the testing,obtain the category probability distribution of the test image,selecting the maximum value of the category probability distribution corresponding to the category as the category of the test image.Experiments show that the proposed method improves the perceptual ability of convolutional neural networks to express expression changes,and significantly improved the accuracy of facial expression recognition.(2)To solve the problem that the convolutional neural network can't learn the spatial position information from the training image,the STN(Spatial Transformer Network)is introduced into the expression recognition task.The STN network is placed in the first layer of the convolutional neural network,so it can adjust its own affine transformation parameters with the loss value of the convolutional neural network,thereby convolutional neural network has rotation invariance.Experiments show that the introduction of STN network can improve the expression recognition accuracy.(3)To solve the problem that when fusing low-level feature and high-level features learned by fusion convolutional neural networks to recognize facial expression,the number of parameters will growth rapidly,proposing a double-channel convolution convolutional neural network.The feature maps which output by pooling layer is divided into two channel for convoluting,the upper channel is convolved with 1×1 convolution kernel,getting low-level feature values,the lower channel input into next convolution layer to learn high-level features.Then fuse low-level features and high-level features and input into classifier.Experiments on two datasets show that the proposed method can effectively reduce the amount of parameters.