Research And Implementation Of Facial Expression Generation Algorithm Based On Flow Model

Nowadays,as an important branch of face recognition,facial emotion recognition has made certain progress and has been widely used in many scenes.However,in order to truly understand how computers deal with facial emotions,in addition to the study of discriminant models,the study of generative models should also not be ignored.In addition,in recent years,due to the parallel development of three aspects of deep learning theory,computer computing power and data volume,based on the generation model of deep neural network,this model have more methods to build and have better performance than traditional models.However,these methods still have their limitations,such as the difficulty of model training and the unsatisfactory training results.The flow-based generative model is a deep generative model proposed in recent years.Its construction idea is to obtain the ability to generate data by explicitly learning the data distribution.In theory,the ability to restore data is stronger than other generative models.However,its implementation has many limitations,such as strong restrictions on model design,and excessive model parameters and calculations.This thesis constructs a facial emotion image generation model based on parallel linear flow.The model is mainly composed of several coupling layers superimposed to form a multi-scale structure.Each coupling layer contains 1*1 reversible convolution and linear operation modules.Aiming at the expressive ability of the model and the convergence speed during training,the model has the following improvements compared with other flow-based generative models:1.Replace the identity transformation in the coupling layer structure with linear operations without losing the reversibility of the model,and enhance the expression of the model;2.In the training process,the principal component analysis method is introduced in the dimensionality reduction operation before the data enters the linear operation module,the essence of which is the compression of the training data by the model.In view of the above two improvements,this thesis chooses two different convolution kernel models for comparison,and designs an ablation experiment to verify the effectiveness of the model.In order to further explore the influence of the network model on the generative model based on parallel linear flow,the purpose is to reduce the parameters of model training while ensuring the expressive ability of the model as much as possible.This article leads to two different Inception structures for comparative experiments.In the research process,based on the public facial emotion data set,this thesis made a data set Emotion Set suitable for deep generative model training and tested it in the model.The experimental results on the public data set and Emotion Set show that under the traditional convolutional neural network,the 3-layer 3*3 convolution kernel is more conducive to extracting the features of the face image;the introduction of principal component decomposition can improve the convergence of the model Speed: After introducing the Inception structure,the parameters of model training can be reduced on the basis of ensuring model expression ability and training speed.At the same time,the sample image generated by the improved model also shows that the facial emotion image is more natural and smooth than that,which reflects the superiority of the facial emotion generation model based on parallel linear flow in supervised learning.