Research On Image Inpainting Algorithm Based On Generative Adversarial Nets

Digital image inpainting technology refers to inpainting images via neighborhood information of the missing image,according to certain rules to inpaint image,so that the observer cannot visually noticed that the image has been damaged or has been repaired.The technology is widely used in many fields such as restoration of ancient artfacts,video and image restoration,removal of obstacles,film,television special effects production and so on.In the image inpainting,when the image semantics are missing,the traditional image inpainting algorithm based on structural and texture cannot achieve satisfactory results in a single-source image using neighborhood information of defect area.In this paper,We proposed an image inpainting algorithm based on wasserstein generative adversarial networks.The innovation of the algorithm is that it can obtain advanced context information from multi-source correlated images,which not only improves the traditional image inpainting algorithm,but also improves the effect of image inpainting.In this work,firstly,we train WGAN on the Celeb A dataset,then the image to be restored is input into WGAN to generate a series of similar fake images.Secondly,in order to obtain the best fake picture from the fake picture set,we propose a loss function consisting of contextual loss and perceptual loss.The contextual loss ensures the content similarity between the image to be repaired and the restored image,and the perceptual loss guarantees a complete and realistic image.Finally,the back propagation algorithm of the loss function is used to map to a smaller potential space for the image to be repaired,and the mapped vector is input into WGAN to generate the best fake image for the image to be repaired,so that the image inpainting can be realized.In this paper,A missing area of facial image up to 80%on LFW face dataset is repaired successfully.By the subjective and objective evaluation,it is proved that the method can not only successfully predict the semantic absence of image,but also can produce complete and realistic image.