Cross Transformer Network for Scale-Arbitrary Image Super-Resolution

Abstract

Since implicit neural representation methods can be utilized for continuous image representation learning, pixel values can be successfully inferred from a neural network model over a continuous spatial domain. The recent approaches focus on performing super-resolution tasks at arbitrary scales. However, their magnified images are often distorted and their results are inferior compared to single-scale super-resolution methods. This work proposes a novel CrossSR consisting of a base Cross Transformer structure. Benefiting from the global interactions between contexts through a self-attention mechanism of the Cross Transformer, the CrossSR could efficiently exploit cross-scale features. A dynamic position-coding module and a dense MLP operation are employed for continuous image representation to further improve the results. Extensive experimental and ablation studies show that our CrossSR obtained competitive performance compared to state-of-the-art methods, both for lightweight and classical image super-resolution.