A general model compression method for image restoration network

作者：

Highlights：

• We propose the first general model compressed method for low-level vision tasks.

• Side window kernel and convolution factorization aim at compress networks.

• Symmetric dilated convolutions and attention block aim at reduce performance loss.

• Experiments have shown the proposed method outperforms other compact modules.

摘要

•We propose the first general model compressed method for low-level vision tasks.•Side window kernel and convolution factorization aim at compress networks.•Symmetric dilated convolutions and attention block aim at reduce performance loss.•Experiments have shown the proposed method outperforms other compact modules.

论文关键词：Model compression,Image restoration,Deformable convolution kernel,Symmetric dilated convolutions,Attention mechanism

论文评审过程：Received 21 July 2020, Revised 23 November 2020, Accepted 31 December 2020, Available online 9 January 2021, Version of Record 11 January 2021.

论文官网地址：https://doi.org/10.1016/j.image.2021.116134