Research On Facial Expression Recognition Based On Deep Learning

In human social life,human face is the first "pass",the change of facial expression is an important way to convey information,and the instantaneous reaction of facial expression(micro-expression)is far more real and reliable than the information presented by sound,language,behavior and so on.At present,computer is mainly used to extract and classify facial expression feature areas in images to complete the task of facial expression recognition.With the continuous progress of the society,the research upsurge of facial expression recognition technology has been set off,and it has been widely used in a variety of fields.However,due to the complex and changeable natural environment and human behavior,the sample images taken in non-laboratory areas are affected by a series of factors such as natural light stimulation and shooting methods,which lead to the interference of facial expression related feature information in the images.Although the image normalization processing can eliminate most of the interference information,it cannot solve the problem that part of the image information is blocked.Moreover,the features of facial expression are complex and changeable,and the traditional expression recognition technology is not sufficient to extract the feature information.The convolutional neural network in deep learning can automatically extract deeper essential features from a large number of data.Reduce the influence of human factors and external environment,improve the accuracy of expression recognition.This paper explored and improved the facial expression recognition algorithms based on deep learning,and proposed a facial expression recognition algorithm based on deep learning:(1)Aiming at the problem that weak feature extraction ability in expression recognition,an expression recognition method based on the attention mechanism of convolutional network was proposed.Based on the convolutional network,two kinds of attention mechanisms,channel and space,were connected in parallel to achieve weight allocation in different dimensions and positions of the facial regions extracted by the network,and focused on the nuance feature information in the key points of facial expressions.Then,an improved high-order residual module was introduced.The depth-separable convolution in the residual module was used to change the number of channels and reduce the number of redundant parameters.The Squeeze-and-Excitation module was added to weight the input channels to improve the importance of some channels.The residual mechanism was used to establish a direct correlation channel between general features and advanced features.Through the detailed module,the multi-scale features of different depths were extracted to further extract the depth facial information.The joint loss function was introduced at the end of the network to expand the distance between expressions and reduce the differences between the similar expressions.FER2013 and CK+ datasets were respectively verified.Experiments showed that the expression recognition accuracy rate of the network model with improved features was higher than the expression recognition rate of the traditional convolutional network model,which proved the feasibility of the improved network model in the field of expression recognition.(2)In order to solve the problem of low recognition accuracy and large number of parameters in the experiments of traditional convolutional neural networks on face occlusion images,a facial expression recognition method based on the attention mechanism of residual network was proposed.In this method,the residual network was used as the basic network,and the cutout was added to cover any area and any size of the image to improve the robustness of the network.Ghost convolution of packet convolution was replaced by depth-separable convolution in Ghost module to reduce the number of network parameters,and a squeeze excitation module was introduced to reduce the noise interference of feature extraction.Channel attention mechanism was used to assign different weights to channels of different importance to increase the ability to capture key information.Multi-scale features were combined with spatial attention mechanism to extract information from different feeling wild to improve the ability of network extraction.The joint loss function was used to increase the out-of-class distance and reduce the in-class distance.The network was applied to FER2013 and CK+ datasets,and the experimental results showed that the recognition rates were 64.81% and 96.86%,respectively.The number of parameters was 5.21 M.Compared with Res Net network,it was found that the model not only reduced the number of redundant parameters,but also improved the recognition accuracy.The feasible of the improved model in face recognition was proved.