The goal of image super-resolution is to reconstruct a high-resolution version of a low-resolution image. At present, image super-resolution plays an important role in many fields such as security monitoring, medical images, and remote sensing images, and has become a research hotspot. However, general super-resolution networks based on deep learning are limited by high computational complexity or low precision. For the problem of generic super-resolution networks, in this thesis proposes a fully convolutional autoencoder combining residual learning and an improved U-Net network to achieve end-to-end generation of high-resolution images. It achieves high PSNR and SSIM results on DIV2K and Set14 test sets.