Detail-preserving image super-resolution via recursively dilated residual network

摘要

Convolutional neural network (CNN) methods have been successfully applied in single image super-resolution (SR). However, existing very deep CNN based SR methods face with the challenge of memory footprint and computational complexity for real-world applications. Besides, many previous methods lack flexible ability to emphasize local spatial informative areas, which is limited to recover the high-frequency detail of LR input. In this paper, to address these problems, we implement a spatial modulated residual unit (SMRU) upon the dilated residual unit and propose a recursively dilated residual network (RDRN) to reconstruct high-resolution (HR) images from low-resolution (LR) observations. The proposed RDRN can effectively exploit the contextual information over larger regions and pay attention to the high-frequency parts for image detail recovery. Furthermore, such spatial modulation mechanism (SPM) in SMRU can incorporate well with existing SR models for better reconstruction performance. Extensive evaluations on public benchmark datasets demonstrate that our proposed method achieves superior performance in terms of quantitative and qualitative assessments. (C) 2019 Elsevier B.V. All rights reserved.