Image Inpainting with Learnable Feature Imputation

A regular convolution layer applying a filter in the same way over known and unknown areas causes visual artifacts inpainted image. Several studies address this issue with feature re-normalization on output of convolution. However, these models use significant amount learnable parameters for [41, 48], or assume binary representation certainty an [11, 26]. We propose (layer-wise) imputation missing input values to In contrast learned our method is efficient introduces minimal number parameters. Furthermore, we revised gradient penalty image inpainting, novel GAN architecture trained exclusively adversarial loss. Our quantitative evaluation FDF dataset reflects that alternative improves generated quality significantly. present comparisons CelebA-HQ Places2 current state-of-the-art validate model. (Code available at: github.com/hukkelas/DeepPrivacy . Supplementary material can be downloaded from: folk.ntnu.no/haakohu/GCPR_supplementary.pdf )