Face Expression Generation Based On Multi-Domain Mapping Generative Adversarial Networks

In recent years,the development of deep convolution neural network has made human life more intelligent.Its applications include automatic driving,intelligent retail,intelligent financial services,intelligent security and other industries.From the initial classification network(Alex NET,Google NET,etc.)gradually developed into a variety of deep neural networks with their unique functions,among which the classical classification network includes Res NET,VGG16,Se NET and so on.The internal structure of classification network,as the basis of feature extraction of deep neural network,is still developing.On the basis of classification network,a complex target detection network and a countermeasure generation network are also derived.Target detection networks such as SSD,YOLO,Faster RCNN and Mask RCNN have been widely used in industry.In 2014,Goodfellow first proposed a self-learning network that allows two convolutional neural networks to confront each other and gain feedback from each other's losses.Its function is to enable the generator network to generate realistic pictures and give the machine imagination.In this paper,the existing algorithms of Generative Adversarial Networks are studied,their inherent mathematical principles are analyzed,and the popular multi-mapping GAN structure is improved in facial expression generation.By adding the key point information of face to optimize the results of the algorithm,the natural conversion of the foreground without changing the background of the picture and the fusion of interpolation are realized.All right.The essence of GAN is a parameter estimation of high-dimensional probability distribution.However,because the training data is in the form of discrete points in the distribution and the training process can only be expected by sampling,it is difficult to guarantee diversity and the fidelity of the generated pictures on the premise of satisfying the lifelikeness of the generated pictures.Therefore,in the traditional GAN loss function,network structure and training mode,there are often problems such as training convergence,parameteradjustment and mode collapse.In this paper,the current algorithms to effectively reduce these problems are analyzed,and the multi-mapping Generative Adversarial Networks is effectively improved,and the experimental results are analyzed.