Scale-wise Upsampling for Efficient Single Image Super-Resolution

Xiaole Zhao, Xinkun Wu, Xiangsong Jia, Xun Gong, Xiaobo Zhang, Tianrui Li, Xun Xu

Abstract: Multiscale information is a commonly-used prior for various computer vision tasks due to the universal self-similarity of images at different scales. In this work, we present a simple yet effective upsampling strategy for the task of efficient single image super-resolution (SISR), which is increasingly attracting attention in the field of consumer electronics. It introduces two parallel auxiliary branches in the upsampling, one of which infuses static downscale prior and the other one incorporates dynamic upscale information related to the target scale. The proposed approach is designated as Scale-wise Upsampling Module (SUM) and favors realizing the interaction and synergism of multiscale priors and enlarge the effective receptive field of the model. In the stage of feature learning, simple gating mechanism is used to enhance the nonlinearity of the model, and a learnable feature integration is conducted to further explore the representational capacity of the model. Our SUM can be used as a plug-and-play module in other SISR models, and achieve consistent performance gains with few extra overheads. Extensive experiments have shown that although our model is built upon naive CNNs, it can achieve superior SR results comparable or superior to Transformer-based models with higher efficiency. Experiments on diverse degradations also show that leveraging multi-scale priors with our SUM enhances model performance and generalization capability, enabling our method with only 1.06M parameters to outperform RDN (22.3M). The code for the proposed model is available at https://github.com/Nerti98/SwiseNet/tree/main.

External IDs:doi:10.1109/tce.2025.3646229