Fast and Accurate Image Super-Resolution with Deep Laplacian Pyramid Networks

摘要

Convolutional neural networks have recently demonstrated high-qualityreconstruction for single image super-resolution. However, existing methodsoften require a large number of network parameters and entail heavycomputational loads at runtime for generating high-accuracy super-resolutionresults. In this paper, we propose the deep Laplacian Pyramid Super-ResolutionNetwork for fast and accurate image super-resolution. The proposed networkprogressively reconstructs the sub-band residuals of high-resolution images atmultiple pyramid levels. In contrast to existing methods that involve thebicubic interpolation for pre-processing (which results in large feature maps),the proposed method directly extracts features from the low-resolution inputspace and thereby entails low computational loads. We train the proposednetwork with deep supervision using the robust Charbonnier loss functions andachieve high-quality image reconstruction. Furthermore, we utilize therecursive layers to share parameters across as well as within pyramid levels,and thus drastically reduce the number of parameters. Extensive quantitativeand qualitative evaluations on benchmark datasets show that the proposedalgorithm performs favorably against the state-of-the-art methods in terms ofrun-time and image quality.