Multistage attention network for image inpainting

作者：

Highlights：

• The multistage attention module is introduced to ensure better results in multi-scale.

• Partial convolution is adopted to avoid the interference of mask areas on generated results.

• We use the joint loss to provide better details and consistent style for inpainting results.

• Experimental results demonstrate the superiority of our proposed approach.

摘要

•The multistage attention module is introduced to ensure better results in multi-scale.•Partial convolution is adopted to avoid the interference of mask areas on generated results.•We use the joint loss to provide better details and consistent style for inpainting results.•Experimental results demonstrate the superiority of our proposed approach.

论文关键词：Image inpainting,Irregular mask,Deep learning,Attention mechanism,Unet-like network

论文评审过程：Received 22 January 2020, Revised 31 March 2020, Accepted 13 May 2020, Available online 24 May 2020, Version of Record 30 May 2020.

论文官网地址：https://doi.org/10.1016/j.patcog.2020.107448