Research On Facial Expression Classification Technology Of Static Images Based On Convolutional Neural Network

Facial expressions contain rich emotional information and occupy an important position in people's daily communication.With the rapid development of computer technology,people urgently hope that computers can accurately simulate human behavior patterns,thereby enhancing the existing intelligent life experience.Facial expression recognition technology has broad application prospects as an important means for computers to distinguish human facial expression states.It is widely used in personalized recommendation,distance education,medical assistance,driving assistance,smart city,human-computer interaction and other fields.In recent years,with the development of artificial intelligence,facial expression recognition has become a research hotspot in the field of computer vision,attracting the attention of a large number of researchers.Using convolutional neural networks in deep learning to identify facial expressions in static images is the main research content of this article.The specific work is as follows:1.In traditional machine learning methods,excessive preprocessing of the original image,superabundant image enhancement,and insufficient data samples have an impact on facial expression recognition.In response to this,the original information of the FER2013 dataset and the CK + dataset is retained in the article.Facial expression images are cropped at the four corners and the center,and then flipped horizontally to achieve the purpose of expanding data samples.2.In view of the high complexity of manually extracting features in traditional machine learning and the inadequate extraction of facial expression features by shallow convolutional neural networks,VGGNet-19 GP and ResNet are selected to exclude the suspicion that too deep networks are likely to cause waste of computing resources and excessive feature extraction.VGGNet-19 GP is originated from the improvement of VGGNet-19.The maximum and average global pooling are performed on the depth descriptors which are learned by the last convolution layer,and then L2 normalization is applied on them,respectively,and the two obtained features are cascaded through a fully connected layer toachieve the purpose of classification.During the experiment,both networks use ReLU activation function and the stochastic gradient descent training algorithm,and the dropout mechanism is added to prevent over-fitting.Finally,the VGGNet-19 GP and ResNet-18 networks achieve an average accuracy of 71.848% and 72.271% on the FER2013 dataset,respectively,and achieve an average accuracy of 91.107% and 92.845% on the CK+ dataset.3.Aiming at the problem that the single feature extraction during the expression recognition process of single convolutional neural network will result in the low accuracy of final expression classification,and drawing on the idea of integrated learning,a new facial expression recognition method based on integrated convolutional neural network is proposed.The method is to integrate the VGGNet-19 GP model and the ResNet-18 model,and build an ensemble network which is the EnsembleNet model.The model uses the outputs of the two networks,reserves the feature vectors of the 7 types of expression energy values after the dual-network FC layer,and cascades into new feature vectors.By obtaining the maximum value in the new vector,its index is returned as the prediction of the expression to complete the facial expression classification.Through experimental research,it is found that the strategy of weighted average integration is the best when the weights of 0.47 and 0.53 are set for VGGNet-19 GP and ResNet-18,respectively.The average accuracy rates of 73.854% and97.611% are obtained on the FER2013 and CK+ datasets,respectively,which achieves the research purpose of high classification accuracy and strong generalization ability.