Research On Face Completion And Replacement Based On Coordination Generative Adversarial Network

Generative adversarial network based on game theory was first proposed by a researcher named Ian Goodfellow from Google Brain in 2014 and it has got the attention of many researchers in the field of machine learning.In the field of computer vision,generative adversarial network has a broad application prospect for image completion.By analyzing the existing research results based on the generative adversarial network and consulting the traditional image implementation and face replacement methods at the same time,in this paper,we build the improved generative adversarial network,the generator and the discriminator of which are constructed by convolution neural network.By using the related image processing tools,this paper has successfully completed two important tasks which are face completion and face replacement.Through a large number of scientific and systematic experiment at the same time,the good performanc of the network in these tasks by the objective evaluation index and the typical investigation method has been confirmed.Inspired by deep convolution generative adversarial network,this paper has first proposed the improved deep convolutional generative adversarial network and the improved generative adversarial network based on Wasserstein distance and do the experiment of face completion respectively,comparing the peoformance of the two models under different training iterations.At the same time,aiming at the problem that the completion effect is not good enough,the coordination generative adversarial network is first proposed in this paper by using two generators to aim at the whole contour and the local area to generate real samples.Through a large number of logical experiments,this paper has proved that the network performance in face completion in performance than the previous two kinds of improved model promoted by at least 30%.In addition,this paper has creatively proposed the combination of loss function based on cross entropy and pixel mean square error,making the coordination generative adversarial network have good performance when dealing with large area face completion,which has improved at 50% in proformance.Finally,this paper also creatively puts forward using coordination generative adversarial network to realize face replacement through the weight sharing between the convolution layers and confirmed the validity and rationality of the model in face replacement.