Conditional Inpainting Generative Flow.

Buliao Huang,Yunhui Zhu,Xiren Zhou,Huanhuan Chen

DOI: https://doi.org/10.1109/ssci51031.2022.10022305

2022-01-01

Abstract:Image inpainting tasks have attracted much attention in computer vision. Many existing methods propose to generate the corrupted images by employing generative adversarial networks (GAN) to sample from the data distribution conditioned on the observed regions. However, due to the slowly growing receptive field of the convolutional layers in GAN, it is difficult for these GAN-based methods to effectively model long-range interactions, which is critical for inpainting. Besides, existing approaches generally introduce joint losses to render visually realistic results, while the weights for different losses are fixed during training and may not adapt to the dynamical learning process. To address the above issues, this paper proposes a novel conditional inpainting generative flow (CIFlow). CIFlow is a flow-based conditional generative method with pixel-wise permutations and channel-wise permutations performed before traditional convolutional layers to effectively model long-range interactions. Moreover, the weights for the joint losses in CIFlow are dynamically adjusted based on the current learning state to provide gradients of the same scale for different losses. Experiments are conducted on real-world datasets and the results demonstrate the robustness and effectiveness of the proposed algorithm.