Research On Partial Occlusion Facial Expression Recognition Technology Based On Improved CNN

Facial expression recognition(FER)refers to the calculation and analysis of changes in a person's facial muscles,morphology,and key features through a computer to determine their internal emotions.In the field of pattern recognition,FER is a very active and challenging field,which is widely used to make human-computer interaction environment intelligent,distance education and psychopathological analysis etc.However,despite deep learning and convolutional neural networks(CNN)have been successfully applied to FER in recent research,partial occlusion of FER without constraints is still a challenge and has become a major obstacle in its practical applications.On the one hand,in the case of partial occlusion,FER is usually not well studied and discussed in the deep convolutional neural network(DCNN)model due to the lack of data sets.And the traditional Softmax Loss does not have the ability to distinguish features,and can not play a better role in the expression classification problem.In order to solve the above problems,this paper reconstructs partially occluded facial expression data sets based on JAFFE and KDEF data sets and performs a series of necessary preprocessing operations.A deep convolutional neural network model based on Convolutional Neural Network and residual block mixing is proposed to complete the partial occlusion facial expression recognition,which can solve the degradation problem of DCNN.In addition,the Large Margin Cosine Loss function is used to improve the distinguishing ability of the loss function.The partial occlusion facial expression recognition based on the deep neural network model is completed and the better classification results can be achieved.Secondly,for DCNN-based FER,the partial occlusion of facial expression data is not large enough,which will cause DCNN to perform poorly in cross-database verification.Moreover,the network with deeper depth has problems such as the disappearance of the gradient and the greater oscillating of the optimization process,which reduces the training efficiency and generalization of the network.In order to solve the above problems,this paper adopts the strategy of mixed features and combines the traditional image method Gabor filter with Adaboost classifier,and processes it in parallel with CNN network.Through parallel processing,the classification results obtained by each are combined according to certain weights,and finally the classification vector is output.In addition,we replaced the traditional Stochastic Gradient Descent optimization algorithm with a stochastic gradient descent optimization algorithm with momentum,and added batch standardization.Therefore,the deep network model can improve the cross-database verification ability when the data set is not large enough,and the oscillation problem caused by SGD.