A lightweight multi-scale residual network for single image super-resolution

Single image super-resolution (SISR) technology based on deep learning provides an effective solution for improving image resolution. However, the computational complexity brought by the deep models also makes SISR technology face challenges in practical applications. In order to overcome the limitations of hardware devices and alleviate the memory and computational overhead brought by the deep models, we propose a lightweight multi-scale residual network for SISR. In detail, we designed a dilated residual block with new channel attention module to explore multi-scale feature information with less parameter cost and aggregate them in a weighted way to enhance the discriminative ability of the network for different scales. Meanwhile, local dense cascade is utilized to make better use of hierarchical features, which further takes advantage of the multi-scale representation. Both qualitative and quantitative experiments demonstrate that the proposed model achieves superior performance with lower model complexity.