Multi-scale patch-GAN with edge detection for image inpainting

Image inpainting with large missing blocks is quite challenging to obtain visual consistency and realistic effect. In this paper, the multi-scale patch generative adversarial networks with edge detection image inpainting (MPGE) was proposed. Firstly, an edge detector was introduced into the generator of multi-scale generative adversarial networks (GAN) to guide the inpainting of the edge contour in the image inpainting, which improved the inpainting effect of image posture and expression. Secondly, we designed a patch-GAN as the local discriminant to capture high frequency, and a function L2-loss was utilized to keep the high resolution and style of the original image. Thirdly, a multi-head attention mechanism was introduced into the generator and local discriminator to build a multilevel and multi-dimensional dependent network model for image subspaces, which improved the global consistency of the inpainted image. Finally, by finding the minimum data set with similar network expression ability, we quickly obtained the optimal value of multi-head. Thereby, a lot of training time was saved. The experiments conducted on Celeba dataset proved that our proposed algorithm quantitatively and qualitatively outperformed the baselines.