Multi-scale gated network for efficient image super-resolution

Xuan Miao, Shijie Li, Zheng Li, Wenzheng Xu, Ning Yang

Abstract: Remarkable progress has been made in the field of single-image super-resolution (SISR), with convolutional neural network being widely adopted to achieve state-of-the-art performance. Recently, researchers have been increasingly interested in exploring the application of Transformer in SISR. However, the high computational cost of Transformer poses a challenge to its deployment on mobile devices. To address this issue, we propose a novel lightweight multi-scale gated network (MSGN) by exploring the variant of the Transformer which is built upon its general structure. MSGN utilizes efficient multi-scale gated block (EMGB) as the token mixer for the Transformer. Specifically, EMGB uses multi-scale filtering block and gating mechanism to extract and augment various features at multiple granularities. In addition, the simplified channel attention is used to extract channel global information. Furthermore, an enhanced multi-layer perceptron is employed instead of the MLP layer in Transformer to further improve the performance of the network. Our extensive experimental results demonstrate that MSGN achieves the best performance among the state-of-the-art efficient image SR models while utilizing the least number of parameters and FLOPs.