Single image super-resolution using fast sensing block

Abstract

Single image super-resolution (SISR) is a classical task in computer vision. In recent years, convolutional neural network (CNN) has been widely used to solve this problem. CNN-based methods directly learn an end to end mapping between low-resolution (LR) and high-resolution (HR) images to achieve state-of-the-art performance. Recent studies show that larger receptive field in CNN is more beneficial for SISR. However, most CNN-based methods have to pass through a mass of serial convolutional layers to get a large size of receptive field. Consequently, computational efficiency is low. Moreover, it is difficult to fully use multi-scale information. In this paper, a fast sensing super-resolution network (FSSRN) built with parallel Fast Sensing Blocks (FSB) is proposed to extract multi-scale features from LR image more efficiently. Experimental results show that FSSRN achieves significant improvement of efficiency while achieves state-of-the-art performance.