Semantic image inpainting with boundary equilibrium GAN

Abstract

Recently, due to the vigorous development of deep learning, many methods in the field of image inpainting have been proposed which are different from the traditional image inpainting methods. This paper uses the high-quality image generation technology of BEGAN to complete the image inpainting task. Firstly, the image generation model is obtained by pretraining the generator and discriminator of BEGAN. Then this paper redesigns the loss function and finds the generated image suitable for the image inpainting task via gradient descent algorithm. By using the information contained in the undamaged part of the original image to be repaired, the BEGAN model can generate an image that is closest to the original image. Finally, the generated image is used to fill the lost area of the original image to be repaired, and the image inpainting task is completed. This paper confirms the validity of the method through the experiments on the CelebA and LFW datasets.