Lightweight Image Super-Resolution with Pyramid Clustering Transformer

Recently, Transformer-based methods have demonstrated satisfactory results on lightweight Image Super-Resolution. However, most of them limit the computational range of Transformer within a local neighbourhood, thus missing much global information. In addition, exploring Transformer on only one scale seems less powerful. To address these problems, we propose a concise and powerful Pyramid Clustering Transformer Network (PCTN) for lightweight image super-resolution. PCTN is constructed by multiple stacked Pyramid Clustering Transformer Blocks (PCTBs). Each PCTB is composed of two parts: Information Recurrent Distillation Block (IRDB) and Pyramid Clustering Transformer Attention (PCTA). Specifically, we first employ an IRDB to extract local structural information effectively, which can generate a larger receptive field without introducing additional learnable parameters. On the heels of that, we design a PCTA covering the most informative and relevant locations globally at different scales with less GPU memory and computational cost. Extensive experiments show that the proposed PCTN outperforms state-of-the-art lightweight SR algorithms in terms of visual quality and computational complexity.