Semantic Image Inpainting with Deep Generative Models

Raymond A. Yeh, Chen Chen, Teck Yian Lim, Alexander G. Schwing, Mark Hasegawa‐Johnson, N. Minh

arXiv (Cornell University) · 2016 · 69 citations · 0 references

DOIFull text

Open access

Concepts

TL;DR

Semantic image inpainting requires filling large missing regions using available visual data, yet single‑image methods often fail due to insufficient high‑level context. The study proposes a novel method that generates missing content by conditioning on the available data. Using a trained generative model, the method locates the nearest latent encoding of the corrupted image via context and prior losses, then decodes it to infer missing content, enabling inference regardless of hole structure without requiring hole‑specific training. Experiments on three datasets demonstrate that the method accurately predicts large missing regions, achieves pixel‑level photorealism, and significantly outperforms state‑of‑the‑art approaches.

Abstract

Semantic image inpainting is a challenging task where large missing regions have to be filled based on the available visual data. Existing methods which extract information from only a single image generally produce unsatisfactory results due to the lack of high level context. In this paper, we propose a novel method for semantic image inpainting, which generates the missing content by conditioning on the available data. Given a trained generative model, we search for the closest encoding of the corrupted image in the latent image manifold using our context and prior losses. This encoding is then passed through the generative model to infer the missing content. In our method, inference is possible irrespective of how the missing content is structured, while the state-of-the-art learning based method requires specific information about the holes in the training phase. Experiments on three datasets show that our method successfully predicts information in large missing regions and achieves pixel-level photorealism, significantly outperforming the state-of-the-art methods.