Single Image Super-Resolution via Laplacian Information Distillation Network

Abstract

Recently, deep convolutional neural networks (CNNS) have been revealed significant progress on single image super-resolution (SISR). Nevertheless, as the depth and width of the networks increase, CNN-based super-resolution (SR) methods have been confronted with the challenges of computational complexity and memory consumption in practice. In order to solve the above issues, we combine the Laplacian Pyramid with the previous methods to propose a convolutional neural network, which is able to reconstruct the HR image from low resolution image step by step. Our Laplacian-Pyramid structure allows each layer to share common parameters with other layers as well as its inner structure; this kind of characteristic reduces the number of parameters dramatically while still extracts sufficient features at the same time. In experiment part, we compare our method with the state-of-art methods. The results demonstrate that the proposed method is superior to the previous methods, furthermore our x2 model also gains an ideal effect.