Image Inpainting for Irregular Holes Using Partial Convolutions

• DOI: 10.48550/arxiv.1804.07723
• Publication Date: 2018-01-01
• Authors: Guilin Liu; Fitsum A. Reda; Kevin J. Shih; Ting-Chun Wang; Andrew Tao; Bryan Catanzaro
• Keywords: Inpainting; Convolution (computer science); Value (mathematics)

Existing deep learning based image inpainting methods use a standard convolutional network over the corrupted image, using filter responses conditioned on both valid pixels as well substitute values in masked holes (typically mean value). This often leads to artifacts such color discrepancy and blurriness. Post-processing is usually used reduce artifacts, but are expensive may fail. We propose of partial convolutions, where convolution renormalized be only pixels. further include mechanism automatically generate an updated mask for next layer part forward pass. Our model outperforms other irregular masks. show qualitative quantitative comparisons with validate our approach.