Research And Application On Particular Scene Generation Based On Generative Adversarial Network

In recent years,the generative adversarial network has become one of the most popular research directions in the field of deep learning,especially in the field of image generation.Particular scene generation is a kind of image generation task.Unlike the most generative algorithms only generate the same or similar images as the original data set.Particular specific scene generation is a more challenging task.Because the particular scene generation task requires to reconstruct an input scene,it is necessary to integrate the input scene into a new scene and generate a variety of reasonable scenes that containing the object and maintaining its attribute characteristics,which requires complex methods to solve.In this paper,a multi-domain particular scene generation algorithm for semantic control based on generative adversarial network is studied,the main works of this paper are as follows:Firstly,this paper reviews different GAN models and its applications in the field of computer vision.From the theoretical view,GAN's principles,advantages and disadvantages are analyzed,and comparative experiments are carried out on different models.The core ideas,method features and usage scenarios of each model are analyzed.The evolution of the GAN model and introduce the application of GAN in computer vision are concluded.Experiments show that different GAN methods have various improvements on the quality of the generation,but there have other shortcomings.Secondly,to solve the model collapse of GAN and the generated images lack of diversity.This paper proposes a high-quality scene generation algorithm named SPSceneGAN based on spectral regularization.By applying the spectral normalization to generator and discriminator of SPSceneGAN,the algorithm is more stable,the quality of generated scene images has improved greatly,and the generated images are rich in diversity.SPSceneGAN is based on condition GAN,and the particular target scene is used as a constraint to guide the scene generation.Besides,we combine cross entropy loss,L1 Loss and spectral regularization losses to generatehigh quality scene images.SPSceneGAN not only realizes the generation of specific scenes,but also solves the common problems of GAN,such as model collapse,training difficulty and lack of diversity.Experiments show that this method generates higher quality than other similar models.Finally,we propose a multi-domain particular scene generation algorithm named MPSceneGAN based on semantic control.The algorithm realizes the generation of multiple specific scenes in a model through the control of semantic tags.In order to achieve multi-domain scene generation,the semantic label describing the scene is encoded in one-hot form when training the model,and then a dual condition generative adversarial network MPSceneGAN is designed to input the semantic label and the corresponding target scene separately to the generator and the discriminator.The generator of MPSceneGAN extracts image features and domain label information at the same time,and learns the mapping relationship between them.Experiments show that our model can generate corresponding reasonable scene images according to semantic label accurately.In summary,based on the GAN network,we propose two algorithms to realize high-quality specific scene generation and semantic control for multi-domain scene generation.The problem of multi-domain scene generation is solved by introducing semantic labels,which further expands the application scope of generating images and can solve a kind of scene generation problem.Experiments show that the algorithm has great advantages in generating high quality images and domain accuracy rate compared with existing methods.