Research On Image Generation Based On Generative Adversarial Networks

Image generation is an important research direction in the field of image processing.Image generation technology has a wide range of applications and complex algorithm design.The traditional image processing methods can not meet the requirements of practical application.At present,a large number of researchers have carried out research on image generation algorithm based on deep learning for different scenes or special requirements.The existing image generation methods have certain universality.These general methods can deal with multi category image processing problems,such as generating a specified image from random noise or image translation.However,these general methods are difficult to achieve satisfactory results in specific scenes,such as consistent image generation,sequence image generation and so on.The existing general image generation technology still has some problems,such as insufficient details of the generated image.Therefore,it is of practical significance to carry out the research on image generation algorithm for different types of application scenarios,which can improve the image generation quality and efficiency in specific scenes.Based on the generative adversarial network,this paper studies three image generation problems,proposes three efficient image generation algorithms,and applies them to three different scenes.· First,we explore coherent image generation.Aiming at the problem of color consistent coloring in image generation,a method based on generative adver-sarial network is proposed to realize the high-quality coloring of sketch image.We construct a loss function so that a certain kind of sketch image can be painted with consistent color,and then generate coherent image.We use the online sketch of comics,games and other fields to verify the effectiveness of our method,and summarize the advantages and disadvantages of our method through the analysis of various angles.· Second,we study sequence image generation,explore the role of generation against latent codes in network,and apply the exploration results to the codes of spatio-temporal sequence images.We find that latent codes can realize the fast coding of sequence images,improve the efficiency of network transmission of sequence images,and then improve the experience of users when analyz-ing spatio-temporal image data.We also discuss the use of sequence image generation technology to analyze the changes of sequence images,and try to complete the missing data.· Finally,we explore the impact of detail perception on the quality of image generation.Taking Chinese painting image generation as an example,we introduce the attention mechanism,that is,transformer structure,into the generative network to generate high-quality Chinese painting images.By ad-justing the parameters and analyzing the results,we summarize the influence of attention mechanism on image generation.