A deep learning network based end-to-end image composition

作者:

Highlights:

• We propose an image retrieval method that is based on the attention mechanism and focuses on the background of the target image, directly contributing to obtaining semantically similar images from the material images.

• We propose a depth estimation-based method that can extract and optimize the COI of the material image.

• We propose the double-sieving procedure to locate the composition area.

• We propose an end-to-end image composition model that is free from interacting with the user and prove its efficiency via experiments.

摘要

•We propose an image retrieval method that is based on the attention mechanism and focuses on the background of the target image, directly contributing to obtaining semantically similar images from the material images.•We propose a depth estimation-based method that can extract and optimize the COI of the material image.•We propose the double-sieving procedure to locate the composition area.•We propose an end-to-end image composition model that is free from interacting with the user and prove its efficiency via experiments.

论文关键词:Image composition,End-to-end,Background retrieval,Instance optimization,Double-sieving region location

论文评审过程:Received 12 September 2020, Revised 1 July 2021, Accepted 31 October 2021, Available online 17 November 2021, Version of Record 25 November 2021.

论文官网地址:https://doi.org/10.1016/j.image.2021.116570