Multi-scale spatial adaptive kernel selection and feature modulation network for lightweight single image super-resolution

Although, numerous methods have proposed to improve the quality and running-time of single image super-resolution (SR) reconstruction tasks, most of them cannot be deployed on edge devices because of the huge computation consumption and long running-time. In this paper, a lightweight SR method named multi-scale spatial adaptive kernel selection and feature modulation network (MKSFMN) is proposed to address such problems, efficiently. In detail, a global-to-local feature extraction module (AFHEM) like transformer is proposed, which can realize long range dependence and capture high-frequency information. Within it, the spatially-adaptive kernel selection and feature modulation module (SKFM) is introduced to realize channel mixer, dynamic spatial kernel selection and feature modulation in adaptation. What's more, by applying pixel attention, an enhanced convolutional channel mixer (ECCM) and a multi-scale progressive feature extraction module (MPEFM). Extensive experimental results show that the proposed method is outperforms in PSNR, SSIM, parameters and running-time, which is suitable for deploying on edge devices.