Research On Image Synthesis Algorithm Based On Generative Adversarial Network

Generative Adversarial Network(GAN)has been widely studied by researchers since it was proposed in 2014.Image synthesis is one of its most successful application fields.Although image synthesis based on generative adversarial networks is more effective and efficient than traditional methods,it also has some disadvantages,such as unstable training and blurred synthetic images.To overcome these shortcomings,a large number of researchers have proposed methods to improve the performance of GAN in the field of image synthesis.This thesis proposes some solutions to the problem that image synthesis model based on GAN is too large and the synthetic image is not real.1.Latent Variable Loss GAN(LVLGAN)is proposed.The LVLGAN is constructed by using the hidden variable loss instead of the dual generator-discriminator network structure based on the dual learning idea.The dual generator-discriminator network structure preserves the image features well and in consequence it brings the problems that the model is too large.We train a separate generator into an approximate identity map and use the output features of its internal encoder as a reference,to correct internal encoder of another generator which is trained against the discriminator.As a result,we enhance the image feature extraction capability of model and reduce a discriminator network.2.It proposes Bipartite GAN.The encoder parts of the two generators are shared and designed to update the encoder only when the training target is an identity mapping generator,and the encoder module is fixed when training the participating generators.This method further simplifies the network architecture of the entire model.3.Apply the proposed methods to image synthesis tasks.Experiment results on multiple datasets show that the two proposed algorithms work well.Compared with multiple benchmark algorithms,the synthesized picture is more real and clear.4.Quantitatively evaluate the effectiveness of the proposed algorithm.This experiment uses metrics such as RMSE and PSNR to examine the performance of the proposed algorithms in different datasets,and compares it with benchmark algorithms toevaluate its performance.