Single-image super-resolution with multilevel residual attention network
- PDF / 2,380,823 Bytes
- 14 Pages / 595.276 x 790.866 pts Page_size
- 35 Downloads / 239 Views
(0123456789().,-volV)(0123456789(). ,- volV)
ORIGINAL ARTICLE
Single-image super-resolution with multilevel residual attention network Ding Qin1 • Xiaodong Gu1 Received: 14 September 2019 / Accepted: 25 March 2020 Ó Springer-Verlag London Ltd., part of Springer Nature 2020
Abstract Recently, a great variety of image super-resolution (SR) algorithms based on convolutional neural network (CNN) have been proposed and achieved significant improvement. But how to restore more high-frequency details such as edges and textures is still an unsolved issue. The low-frequency information is similar in a pair of low-resolution and high-resolution images. So the SR model is supposed to pay more attention to the high-frequency features to restore more realistic images. But most CNN-based methods don’t consider the different types of features and think the features in different channels and regions contribute equally to the reconstruction performance, which limits the representation capacity of the model. In the meantime, most of these deep networks only simply stack blocks like residual block, which only capture the local features. In this paper, we propose a deep multilevel residual attention network (MRAN) for image SR to focus on the highfrequency features and improve the flow of information. Specially, we propose a channel-wise attention module and a spatial attention module to rescale the channel-wise and spatial weights adaptively, which makes our MRAN focus more on the high-frequency information. Meanwhile, to improve the flow of information and ease the training process, the multilevel residual learning is adopted. Extensive experimental results on five benchmark datasets demonstrate that our MRAN is superior to those state-of-the-art methods for both accuracy and visual comparisons. Keywords Image super-resolution Deep convolutional neural network Attention mechanism Residual learning
1 Introduction Single-image super-resolution (SISR) aims to reconstruct a high-resolution image ISR from a low-resolution image ILR. As a low-level computer vision task, image SR has various applications and attracts a lot of researchers’ attention. While image SR is an ill-posed problem, one LR image may correspond to a multitude of SR images. Therefore, plenty of SR algorithms have been proposed including interpolation-based [1], reconstruction-based [2] and learning-based algorithms [3–5]. After Dong et al. [3] first introduced convolutional neural network into image SR to learn the nonlinear mapping from the bicubic-upsampled image to the highresolution image, the performance of CNN-based SR & Xiaodong Gu [email protected] 1
Department of Electronic Engineering, Fudan University, Shanghai 200433, China
algorithms has been improved significantly. From then on, a variety of networks have been proposed to solve this problem. To alleviate gradients vanishing and capture larger receptive fields, Kim et al. proposed VDSR [4] and DRCN [6]. VDSR is a 20-layer network using skip connection and DRCN first introduced recursive neural
Data Loading...
