CASR: a context-aware residual network for single-image super-resolution
- PDF / 2,550,009 Bytes
- 16 Pages / 595.276 x 790.866 pts Page_size
- 15 Downloads / 241 Views
(0123456789().,-volV)(0123456789(). ,- volV)
DEEP LEARNING APPROACHES FOR REALTIME IMAGE SUPER RESOLUTION (DLRSR)
CASR: a context-aware residual network for single-image super-resolution Yirui Wu1,2 • Xiaozhong Ji2 • Wanting Ji3
•
Yan Tian4 • Helen Zhou5
Received: 2 April 2019 / Accepted: 8 November 2019 Springer-Verlag London Ltd., part of Springer Nature 2019
Abstract With the significant power of deep learning architectures, researchers have made much progress on super-resolution in the past few years. However, due to low representational ability of feature maps extracted from nature scene images, directly applying deep learning architectures for super-resolution could result in poor visual effects. Essentially, unique characteristics like low-frequency information should be emphasized for better shape reconstruction, other than treated equally across different patches and channels. To ease this problem, we propose a lightweight context-aware deep residual network named as CASR network, which appropriately encodes channel and spatial attention information to construct contextaware feature map for single-image super-resolution. We firstly design a task-specified inception block with a novel structure of astrous filters and specially chosen kernel size to extract multi-level information from low-resolution images. Then, a Dual-Attention ResNet module is applied to capture context information by dually connecting spatial and channel attention schemes. With high representational ability of context-aware feature map, CASR can accurately and efficiently generate high-resolution images. Experiments on several popular datasets show the proposed method has achieved better visual improvements and superior efficiencies than most of the existing studies. Keywords Context-aware residual network Channel and spatial attention scheme Inception block Single-image superresolution
1 Introduction Super-resolution (SR) is generally defined as a process to obtain high-resolution (HR) images form inputs of lowresolution (LR) observations. There exists a rough but classical taxonomy way to category SR methods based on
& Wanting Ji [email protected]
number of input LR images: single-image super-resolution (SISR) and multiple images super-resolution (MISR). Being a highly ill-posed problem, SISR is more challenging than MISR, since it requires to hallucinate missing image details by learning the relationship between LR and HR from a training dataset.
1
College of Computer and Information, Hohai University, Nanjing, China
Yirui Wu [email protected]
2
National Key Lab for Novel Software Technology, Nanjing University, Nanjing, China
Xiaozhong Ji [email protected]
3
School of Natural and Computational Sciences, Massey University, Auckland, New Zealand
Yan Tian [email protected]
4
Zhejiang Gongshang University, Hangzhou, China
5
School of Engineering, Manukau Institute of Technology, Auckland, New Zealand
Helen Zhou [email protected]
123
Neural Computing and Applications
Fig. 1 Comparisons o
Data Loading...
