A Lightweight Multi-level feature Mixing network For Image Super-Resolution

With the development of digital image technology, single image super-resolution (SISR) has made unprecedented progress. However, the existing SISR technology based on convolutional neural network (CNN) still faces great challenges in the application of resource-constrained mobile devices. For lightweight SISR, we propose a multi-level feature mixing network (MFMN), which achieves high-quality super-resolution image reconstruction at low computational cost. Based on re-parameterization, a multi-level feature generation module (MFGM) is designed to extract feature mappings from different perspectives. Then, to enhance feature details, a feature refinement module (FRM) is constructed to aggregate high-frequency information, which can assign weight values to pixels in feature maps according to the importance of different hierarchical features. Through many experiments, our proposed method shows excellent image reconstruction results in a lightweight way on multiple benchmark datasets and exceeds the current advanced methods.