Generative image inpainting via edge structure and color aware fusion

Highlights：

• A novel edge structure and color aware fusion label is introduced into the two-stage generative adversarial network to guide image inpainting more intelligently.

• Improved joint loss functions are introduced to train the multi-stage model more effectively.

• A pixel progressive degradation operator is designed to increase the consistency of the boundary and eliminate artifacts of the final image.

• Experiments on multiple publicly available datasets demonstrate that our method can obtain competitive inpainting effects.

摘要

•A novel edge structure and color aware fusion label is introduced into the two-stage generative adversarial network to guide image inpainting more intelligently.•Improved joint loss functions are introduced to train the multi-stage model more effectively.•A pixel progressive degradation operator is designed to increase the consistency of the boundary and eliminate artifacts of the final image.•Experiments on multiple publicly available datasets demonstrate that our method can obtain competitive inpainting effects.

论文关键词：Deep learning,Image inpainting,Generative adversarial network,Content aware fill,Multi-map fusion

论文评审过程：Received 15 January 2020, Revised 10 April 2020, Accepted 21 June 2020, Available online 27 June 2020, Version of Record 30 June 2020.