Research On Image Inpainting Method Based On Generative Adversarial Network

Since the generative adversarial network was proposed,it has been widely loved by researchers in the field of computer vision because of its powerful self-generating ability.At present,generative adversarial networks are gradually applied to various fields,such as cultural relic protection,old film restoration,medical imaging,aerospace,public security criminal investigation.However,when the dataset is small or the image structure is complex,the boundary artifacts and image authenticity of the restored results still need to be further solved.Based on the above problems,this paper improves generative adversarial network from two perspectives of discriminator and generator.The face dataset with a small amount of data is repaired,and the inpainting results of different inpainting methods and the influence of each module on the inpainting effect are analyzed.The main work is as follows:1.Based on discriminator,an improved multi-scale discriminator model is proposed.The generator uses the structure of U-Net,and introduces dliated convolution to get the feature information of larger scale field for stitching,and transmits it to the same size position of decoder.A multi-scale discriminator is added to the discriminator to obtain more real image features from a higher resolution point of view.The original discriminator discriminates generated images with the same size as the real image to ensure the context consistency of the results and enhance the generation ability of the model.The combined loss of counter loss,MSE loss and TV loss is used for training,and each loss is composed of multi-scale loss and global loss.The experiments prove the superiority of the network by comparing the inpainting results of different methods,and analyze the influence of multi-scale model and dilated convolution on the network.The results show that the proposed model can shorten the training period of the model to obtain more realistic generation results.2.Based on generator,a dense convolution generative adversarial network image inpainting method with SENet is proposed.Firstly,using the idea of generative adversarial network,the generator uses dense convolution blocks to capture the semantic information of the missing part of the image for reuse.Secondly,the transition layer between dense convolution blocks is cancelled,and the SE module of SENet attention mechanism is introduced to obtain the importance degree of features and enhance the guidance ability of feature information.Thirdly,skip connection is used between encoder and decoder to reduce information loss caused by downsampling.Finally,the stability of network is enhanced by introducing counter loss,MSE loss and TV loss.The discriminator uses DCGAN network.The experiments prove the superiority of the network by comparing the inpainting results of different methods,and analyzes the influence of dense convolution block and SE module on the network.The results show that the proposed algorithm has good performance in image semantics,peak signal-to-noise ratio and structure similarity index.