Research On Expression Recognition Based On Dense Connection And Attention Mechanism

An important research direction of computer vision is facial expression recognition.People’s facial expression is an important external expression of their emotions,even the most important external expression.Emotional communication between people is often done through our facial expressions.Therefore,the research on computer recognition of facial expressions can effectively help the machine to understand people’s emotions and promote the development of human-computer interaction.However,due to the vague definition of some facial expressions,and the influence of human face posture and the environment around human face,the robustness of the machine in judging facial expressions will be greatly reduced.In this thesis,through experiments,the advantages and disadvantages of different models in facial expression are studied and some models are improved.The main work of this thesis is as follows:(1)In view of the two problems that the traditional convolutional neural network only pays attention to the depth of the network model and ignores the lightweight,resulting in the pressure of computer performance,or only pays attention to the lightweight and ignores the importance of the model depth to complex feature extraction,inspired by the dense connected convolutional network,the dense connected convolutional network is improved to make the model maintain a deeper network depth at the same time,The model greatly reduces the demand for parameters,has high prediction accuracy,and realizes high utilization of model parameters.In the experiment,the mainstream popular networks such as VGG network,Dense Net network and Res Net network were selected to compare with the model we designed,and Fer2013 data set was used as the training and testing data set of facial expression recognition model.During the experiment,a series of data enhancement operations such as turning over and adjusting illumination were carried out on facial expression data set.At the same time,in training,the learning rate of the model is continuously and dynamically adjusted according to the current training state of the model to ensure that the learning direction of the network can always be in a good state.The final experimental results show that the improved model can ensure high accuracy on fer2013 data set,and the number of parameters required by the improved model is reduced to a certain extent.(2)In order to fit the facial expression data set better and improve the generalization and robustness of the network model,this thesis introduces attention mechanism based on convolutional neural network models.Through the introduction of attention mechanism,we can filtrate some feature maps,that is to amplify and enhance the features with strong expressive ability of a given task,and weaken those features with little influence.Through this enhancement and weakening operation,we can pick the target features and finally improve the performance of the network model.By comparing the experimental results of multiple data sets,the influence of different data sets on the accuracy of facial expression recognition is analyzed.The results show that the introduction of attention mechanism into convolutional neural network can effectively improve the prediction accuracy of the model.At the same time,we also find that the clearer the facial expression in the data set,the less the interference of the background of facial expression,and the higher the prediction accuracy can be obtained;And the more facial expression images in the data set,the model can achieve an ideal fitting state in less rounds.