Residual shuffle attention network for image super-resolution

The image super-resolution reconstruction methods based on deep learning achieve satisfactory visual quality; however, the majority are difficult to be directly deployed to mobile or embedded devices due to the model complexity. This paper introduces a lightweight residual shuffle attention network for image super-resolution task. Among them, a residual shuffle attention block (RSAB) that fully integrates the information distillation mechanism is designed to extract deep features, which consists of multiple enhanced residual blocks (MERB) and shuffle attention. The MERB is capable of boosting the feature representation, and the shuffle attention can capture critical information extracted by grouping features. Furthermore, the RSAB utilizes multiple skip connection to build the module structure. Extensive experimental results have demonstrated that the network model proposed in this paper outperforms state-of-the-art methods on several benchmarks with acceptable complexity.