{Tammy Lee Hanul Shin Youngchan Song Soo Min Kang}

Abstract

The goal of image restoration is to recover a high-quality image from its degraded input. While impressive results on various image restoration tasks have been achieved using CNNs, the convolution operation has limited its ability to utilize information outside of its receptive field. Transformers, which use the self-attention mechanism to model long-range dependencies of its input, have demonstrated promising results in various high-level vision tasks. In this paper, we propose intra-inter Transformer (iiTransformer) by explicitly modelling long-range dependencies at the pixel- and patch-levels since there are benefits to considering both local and non-local feature correlations. In addition, we provide a boundary artifact-free solution to support images with arbitrary sizes. We demonstrate the potential of iiTransformer as a general purpose backbone architecture through extensive experiments on various image restoration tasks.