Low Resolution Expression Recognition Based On AMMS-DCNN

As the main way of expressing human emotions,facial expressions play an increasingly important role in life.Nowadays,with the rapid development of artificial intelligence technology,understanding human emotions is an inevitable direction for the future development of human-computer interaction.In addition,as an important means of emotion recognition,facial expression recognition has broad application scenarios in entertainment,education,medical care,automatic driving and other fields.In the research process of facial expression recognition,convolutional neural network(CNN)has become a popular method for expression recognition because of its powerful feature extraction capabilities.In recent years,many neural network models for expression recognition have been proposed,all of which have achieved good results in the laboratory environment.However,in real application scenarios,due to various conditions,the resolution of the real face images collected is often low,but the common neural network for expression recognition has a higher resolution of the input image.Requirements,do not meet the real application scenarios.Secondly,most of the common recognition data sets used for expression recognition are collected by web crawlers or from film and television materials,which do not conform to real expression recognition scenarios.In response to the above problems,this paper conducts research on data sets and neural network network structure optimization.The main work and contributions are as follows:(1)In view of the fact that the existing public emotion recognition data sets have more or less defects and do not fully conform to the real expression recognition scene,this paper designs and collects its own data set by referring to the design theory of the existing public data sets.The Fer2013 data set was used as an extended sample to improve and upgrade,and finally formed a special low-resolution expression data set(Low Resolution Fer,LRF).The data set has a total of 7 kinds of expressions and 43184 samples.The improvement and upgrading of the data set provides data support for subsequent research.(2)Aiming at the problem of low resolution of face images in real scenes,this paper proposes an expression recognition model based on dilated convolutional neural network(DCNN).According to different dilated convolution structures and positions in the model,this paper proposes Front＿1＿2＿3,Front＿2＿2＿2,Front＿3＿3＿3,Back＿1＿2＿3,Back＿2＿2＿2,and Back＿3＿3＿3 have a total of six hollow convolutional network models,and the tenfold cross-validation method is used for comparison and verification on the LRF dataset,and the recognition accuracy rates are 66.7%,66.5%,66.4%,and 66.3%,respectively.%,65.4%,65.1%,which prove that the performance of the Front＿1＿2＿3 model is better,which is 1.5% higher than the 65.2% accuracy of Base Net.(3)Aiming at the problem of how to further improve the recognition accuracy of DCNN at low resolution,this paper proposes a multi-scale feature fusion based hole convolution expression recognition network model(MSF-DCNN).In this paper,by comparing and analyzing the principles of the Concatenate and Add algorithms,the fusion method is reasonably designed,and the shallow features extracted by the shallow network and the deep features extracted by the deep network are effectively fused with multi-scale features,and finally the network can extract more The feature information further improves the model performance.(4)Although more feature information can be extracted based on the multi-scale feature fusion structure in the MSF-DCNN expression model,it inevitably brings about the problem of increasing redundant feature information.The more important feature information for facial expression recognition has become the key to further improving the accuracy of model recognition.Therefore,this paper proposes an attention mechanismbased multi-scale feature fusion hole convolution(AMMS-DCNN)expression recognition model,so that the model can adaptively focus on the feature information with higher importance,so as to achieve More accurate feature information extraction,and the effectiveness of the improved algorithm has been proved by experiments.