Hybrid Convolution-Transformer for Lightweight Single Image Super-Resolution

The rapid development of deep learning has driven the breakthrough in performance of single image super-resolution (SISR). However, many existing works deepen the network to pursue performance improvement without considering the issue that models with large parameters are not conducive to current production and deployment. Meanwhile, Transformer, relying on its ability to model long-term dependencies, has entered the field of SISR, but the large memory consumption and inference time cannot avoid the above-mentioned problem. In this paper, we propose a Hybrid Convolution-Transformer (HCFormer) for lightweight single image super-resolution. HCFormer effectively combines convolution and Transformer, and its core modules are the super-resolution feature extraction module (SRFEM) and the long-term dependency feature representation module (LD-FRM), respectively composed of a series of light-weight and efficient convolution blocks (LECB) and light-weight and efficient Transformer blocks (LETB). LECB excavates the potential super-resolution features in the input image through multi-scale residual convolutional operations, while LETB performs long-term dependency feature representation on the excavated features through a streamlined and improved Transformer. Extensive experimental results on five benchmark datasets, compared with the state-of-the-art light-weight SISR methods, demonstrate the effectiveness and competitiveness of our proposed method.