Research On Image Inpainting Methods Using Generative Adversarial Network

Image inpainting has been widely used in the fields of relic restoration,photo restoration and watermark removal.With the development of deep learning in machine vision,image inpainting algorithms based on Generative Adversarial Network(GAN)have achieved better results,but there are still many shortcomings.In this thesis,two improved image inpainting algorithms are proposed by using GAN.The main research contents are as follows:(1)An inpainting method based on multi-scale residual gated convolution and GAN is presented.The generator adopts the multi-scale residual gated convolution module combined with attention mechanism as the down sampling module to enlarge the sensory field of the model and improve the overall consistency of the inpainting image.Scale attention block is used in each layer of encoder,which can capture available features of different scales of input images.Aj oint loss function that combines pixel loss,adversarial loss,style loss,content loss and TV loss is used to restrain the model and further enhance the subjective visual effect and authenticity of the result.(2)An improved pluralistic inpainting method based on Vector Quantional Variational AutoEncoder and GAN is proposed.The vector quantization module is used to separate the discrete structural features.The global structure guidance module is embedded in the structure generation network to ensure the structural consistency of the result and guide the autoregressive network to generate discrete structural features,thus achieve the target of pluralistic inpainting.The discrete structural features guide the texture generative adversarial network to complete the filling of the texture features of the image.In addition,a detail enhancement module is embedded in the texture generative adversarial network to improve the overall quality of the inpainting images.(3)The image inpainting method designed in this thesis is applied to realize the restoration of antique scripts,and a system is built to provide services.Firstly,using various font datasets to complete the restoration,and the results are better.Then,based on the restoration process,antique scripts system is build,which has the advantages of well visualization effect and simple operation.In this thesis,the proposed methods is experimented on CelebA-HQ and Paris Street View datasets,subjective and objective results show that the inpainting methods can generate more realistic and sensible images.