Research On Image Inpainting Based On Generative Adversarial Networks

In recent years,the research and application of deep learning in the field of computer vision are becoming more and more extensive.As one of the research topics of computer vision,image inpainting has been applied to many fields such as satellite image processing,military public security,video and multimedia system image processing.Its purpose is to inpaint the damaged image according to certain rules,so as to obtain an image similar to the original image.At present,researchers have proposed many image inpainting methods,which can inpaint damaged images to a certain extent,but there are also problems such as blurred image boundaries,unclear image textures,and poor visual effects after inpainting,and satisfactory inpainting effects cannot be obtained.Therefore,this article has done the following research on the above problems.1)This thesis proposes a generative adversarial image inpainting model that combines edge detection and self-attention mechanisms.The model is composed of two parts: an edge complement network and a texture inpainting network.Before the edge completion and texture inpainting of the image,the first step is to generate the edge map of the real image and the damaged image.The current edge detection technology still has the problem of insufficient detail edge information.Based on this,this thesis proposes an improved method based on HED(Holistically-Nested Edge Detection,HED).The improved edge detection method consists of three parts: convolutional network,side output network and fusion layer.First,the edge features of the image are extracted through the convolutional network.Secondly,the feature output of different convolutional layers in the convolutional network is cross-layered fusion,so as to maximize the preservation of the edge detail information of the image,and then output the edge features of each stage through the side output network.Finally,the edge map of the image is obtained by fusing the side edge features.Through the experimental comparison of the methods built in this thesis on the BSDS dataset,the results show that the accuracy of the improved HED edge detection model in this thesis has been improved.See that the image edge detection method proposed in this article can extract edge images with more details.The experimental results prove that the improved edge detection method in this thesis performs better than the existing edge detection methods.2)After the edge detection of the image is completed,the edge completion and texture inpainting of the image are performed.First train an edge completion network to learn the edge information of the image,predict the missing edges,and get the edge completion image.Secondly,input the edge-completed image and the damaged image to the texture inpainting network.When inpainting the missing texture,a self-attention mechanism is added to the network to capture the global information of the image,thereby inpainting the missing information and generating the detailed texture of the image.Through comparative experiments on the models proposed in this thesis on the Celeb A and Places2 datasets,the experimental data shows that: The model proposed in this thesis is significantly higher than the similarity index SSIM value of the context attention model and the edge connection model.The results show that compared with the existing image inpainting model,the model built in this thesis has a significant improvement in performance,and the generated image obtained is closer to the orginal image.