Multi-granularity Transformer for Image Super-Resolution

Recently, transformers have made great success in computer vision. Thus far, most of those works focus on high-level tasks, e.g., image classification and object detection, and fewer attempts were made to solve low-level problems. In this work, we tackle image super-resolution. Specifically, transformer architectures with multi-granularity transformer groups are explored for complementary information interaction, to improve the accuracy of super-resolution. We exploit three transformer patterns, i.e., the window transformers, dilated transformers and global transformers. We further investigate the combination of them and propose a Multi-granularity Transformer (MugFormer). Specifically, the window transformer layer is aggregated with other transformer layers to compose three transformer groups, namely, Local Transformer Group, Dilated Transformer Group and Global Transformer Group, which efficiently aggregate both local and global information for accurate reconstruction. Extensive experiments on five benchmark datasets demonstrate that our MugFormer performs favorably against state-of-the-art methods in terms of both quantitative and qualitative results.