Learning Deformable and Attentive Network for image restoration
作者：
Highlights：
• We propose a Deformable and Attentive Network (DANet), which consists of two novel blocks (ADEB and AROB) based on DeConv to learn contextual features with a more adaptive receptive field and preserve fine spatial details in the restored images.
• We propose a knowledge distillation scheme to train a light-weighted version of DANet, named DANet-S, which preserves comparable performance with fewer parameters.
• We apply DANet and DANet-S for IR tasks such as synthetic and realistic noise removal, JPEG artifacts removal, and real image super resolution, which achieve SOTA performance.
摘要
•We propose a Deformable and Attentive Network (DANet), which consists of two novel blocks (ADEB and AROB) based on DeConv to learn contextual features with a more adaptive receptive field and preserve fine spatial details in the restored images.•We propose a knowledge distillation scheme to train a light-weighted version of DANet, named DANet-S, which preserves comparable performance with fewer parameters.•We apply DANet and DANet-S for IR tasks such as synthetic and realistic noise removal, JPEG artifacts removal, and real image super resolution, which achieve SOTA performance.
论文关键词：Image restoration,Convolution neural network,Deformable convolution,Attention mechanism,Knowledge distillation,Image denoising,JPEG artifacts removal,Real-world super resolution
论文评审过程：Received 1 March 2021, Revised 30 July 2021, Accepted 10 August 2021, Available online 19 August 2021, Version of Record 31 August 2021.
论文官网地址：https://doi.org/10.1016/j.knosys.2021.107384
