Stacked Multi-branch Networks for Single Image Super-Resolution

With the recent superior progress in CNN-based single image super-resolution (SISR), the effectiveness of hierarchical features for SISR has been demonstrated. However, few works concentrate on extracting hierarchical features with multi-scale cases. Furthermore, different practical application scenarios have different requirements on the model size. The realization of resizable model design is still far from ideal. To address these issues, we propose Stacked Multi-branch Network (SMBN) for high-quality SISR. Our proposed multi-branch blocks incorporate features in different scales. We designed two models with different parameter levels based on the proposed method. Both quantitative and qualitative comparisons demonstrated the superior performance and efficiency of our proposed SMBN.