Generative Image Inpainting with Submanifold Alignment

  • 2019-08-01 05:01:17
  • Ang Li, Jianzhong Qi, Rui Zhang, Xingjun Ma, Kotagiri Ramamohanarao
  • 15


Image inpainting aims at restoring missing regions of corrupted images, whichhas many applications such as image restoration and object removal. However,current GAN-based generative inpainting models do not explicitly exploit thestructural or textural consistency between restored contents and theirsurrounding contexts.To address this limitation, we propose to enforce thealignment (or closeness) between the local data submanifolds (or subspaces)around restored images and those around the original (uncorrupted) imagesduring the learning process of GAN-based inpainting models. We exploit LocalIntrinsic Dimensionality (LID) to measure, in deep feature space, the alignmentbetween data submanifolds learned by a GAN model and those of the originaldata, from a perspective of both images (denoted as iLID) and local patches(denoted as pLID) of images. We then apply iLID and pLID as regularizations forGAN-based inpainting models to encourage two levels of submanifold alignment:1) an image-level alignment for improving structural consistency, and 2) apatch-level alignment for improving textural details. Experimental results onfour benchmark datasets show that our proposed model can generate more accurateresults than state-of-the-art models.


Quick Read (beta)

This feature is not avaialbe for this paper.