Multi-scale skip-connection network for image super-resolution
- PDF / 3,262,971 Bytes
- 16 Pages / 595.276 x 790.866 pts Page_size
- 77 Downloads / 215 Views
SPECIAL ISSUE PAPER
Multi‑scale skip‑connection network for image super‑resolution Jing Liu1 · Jianhui Ge1 · Yuxin Xue1 · Wenjuan He1 · Qindong Sun1 · Shancang Li1,2 Received: 1 June 2020 / Accepted: 26 October 2020 © Springer-Verlag GmbH Germany, part of Springer Nature 2020
Abstract A skip-connection learning framework-based convolution neural network (CNN) has recently achieved great success in image super-resolution (SR). However, most CNN models based on the skip-connection learning framework do not fully make use of potential multi-scale features of images. In this paper, we propose a multi-scale skip-connection network (MSN) to improve the visual quality of the image SR. First, convolution kernels with different sizes are exploited to capture the multi-scale features of LR images. All the feature-maps captured by convolution kernels of the same size are direct input into a multi-scale hybrid group (MHG); second, the convolution layers of each MHG are composed of dilated convolutions and standard convolutions. The hybrid convolutions can fully train feature details obtained from preceding and current scale convolution layers; three, the output of each hybrid convolution layer is fed into subsequent hybrid convolution layers by skip-connections, thus producing dense connections; lastly, the meta-upscale module is used as the upscale module, which can magnify the trained feature maps arbitrary scale factors. By being evaluated on a wide variety of images, the proposed MSN network achieves an advantage over the state-of-the-art methods in terms of both numerical results and visual quality. Keywords Super-resolution image · Skip connection · Convolution neural network · Multi-scale feature · Hybrid convolution
1 Introduction The aim of image super-resolution (SR) is to construct a high-resolution (HR) image from a low-resolution (LR) image. It is widely studied problem in the fields such as remote surveillance, medical imaging, military reconnaissance, and video supervision [1–3]. Over the past few decades, lots of image SR methods have been proposed, which are generally divided into the three categories: interpolation methods [4, 5], reconstruction methods [6, 7], and learning methods [8–24, 26–31, 34–37]. Interpolation-based methods estimate new pixels by the known adjacent pixels so that the image can have higher resolution. The limitation of the kind of methods is that they will produce noticeably blurring or artifacts at the edges. Reconstruction-based methods adopt image statistics characters to predict new pixels. The * Jing Liu [email protected] 1
Faculty of Computer Science and Engineering, Xi’an University of Technology, Xi’an, China
Department of Computer Science and Creative Technologies, University of the West of England, Bristol, UK
2
basic idea of learning methods is to learn mapping functions between LR and SR image pairs according to an external training dataset. Those learning strategies are various, ranging from sparse coding [8], support vector regression (SVR) [9], neighbor
Data Loading...
