Image Inpainting Based on Generative Adversarial Networks

Abstract

Image inpainting has a good application value in image editing, however, traditional image inpainting techniques cannot complete semantic repair in the case of insufficient sample resources. Deep learning neural network have powerful learning capabilities and can extract high-level semantic features. These features can be used to semantically fill missing regions. Ideal image restoration needs to maintain structural consistency and texture clarity. In this paper, using the GAN network structure, we propose a inpainting method which constrains the repair process using the neighborhood loss function and gradient loss. The experimental results show that the repair results can maintain the global consistency of the structure and the clarity of the local texture. It shows that adding gradient loss constraint can further improve the quality of repair.