Research On Image Inpainting Technology Based On Generative Adversarial Networks

Image inpainting technology plays an increasingly important role in the field of modern image processing.The traditional image inpainting algorithms can repair the missing background and simple texture of the image well,but it is not suitable for the restoration of complex natural images such as face and scenery.Deep learning-based image inpainting algorithms can make up for the shortcomings of traditional image inpainting algorithms,but most of the existing deep learning-based image inpainting algorithms are still difficult to repair the large range of image missing and complex image structure information.In view of the shortcomings of existing methods,the research contents and main work of this paper are as follows(1)An image inpainting model based on Generative Adversarial Networks is studied Aiming at the problem that the image is difficult to repair in a large range,we propose an image inpainting model based on Generative Adversarial Networks.The model includes a generative model and a discriminative model.The architecture of the generative model is an auto-encoder structure with skip-connection.The main task is to encode the image to be repaired and then decode and generate the restored image.The discriminative model is composed of a local discriminator and a global discriminator,in which the local discriminator is responsible for determining whether the semantic correctness of the restoration results is correct or not,while the global discriminator is concerned about whether the whole image after the restoration has overall coherence.In this study,a combined loss training model for adversarial loss and reconstruction loss was used.Experiments prove that the model proposed in this paper is better than some existing models in the visual and performance indicators of the repair results.(2)An edge-guided image inpainting model based on Generative Adversarial Networks is studied.This study aims to make up for the difficulty in repairing the complex structure of the model in the first study.This model consists of an edge-inpainting model and a content-filling model.The edge-inpainting model is based on Generative Adversarial Networks.The generator of the edge-inpainting model is an auto-encoder with a refine network,and the discriminator of the edge-inpainting model is a simple convolutional neural network.The architecture of the content-filling model is similar to the edge-inpainting model but without the refine network.In this study,we train the edge-inpainting model with adversarial loss and feature matching loss and train the content-filling model with reconstruction loss,adversarial loss and style loss.Experiments show that the proposed model is superior to most of the existing image inpainting models(3)An online image processing platform is designed and implemented.Users enter the system through registration and login,and use the image inpainting and image generation services provided by the online image processing platform.When the administrator logs into the system,he can use the model to call the monitoring service more often.To sum up,this paper proposes two image inpainting models based on Generative Adversarial Networks,and experiments show that the proposed model is better than some existing models.This article also deployed the model to the server and developed an online image processing platform invocation model for use by users.