Optimized highway deep learning network for fast single image super-resolution reconstruction
- PDF / 2,185,858 Bytes
- 10 Pages / 595.276 x 790.866 pts Page_size
- 58 Downloads / 222 Views
SPECIAL ISSUE PAPER
Optimized highway deep learning network for fast single image super‑resolution reconstruction Viet Khanh Ha1 · Jinchang Ren1,2 · Xinying Xu2 · Wenzhi Liao1 · Sophia Zhao1 · Jie Ren3 · Gaowei Yan2 Received: 1 November 2019 / Accepted: 4 April 2020 © Springer-Verlag GmbH Germany, part of Springer Nature 2020
Abstract With the success of the deep residual network for image recognition tasks, the residual connection or skip connection has been widely used in deep learning models for various vision tasks, including single image super-resolution (SISR). Most existing SISR approaches pay particular attention to residual learning, while few studies investigate highway connection for SISR. Although skip connection can help to alleviate the vanishing gradient problem and enable fast training of the deep network, it still provides the coarse level of approximation in both forward and backward propagation paths and thus challenging to recover high-frequency details. To address this issue, we propose a novel model for SISR by using highway connection (HNSR), which composes of a nonlinear gating mechanism to further regulate the information. By using the global residual learning and replacing all local residual learning with designed gate unit in highway connection, HNSR has the capability of efficiently learning different hierarchical features and recovering much more details in image reconstruction. The experimental results have validated that HNSR can provide not only improved quality but also less prone to a few common problems during training. Besides, the more robust and efficient model is suitable for implementation in real-time and mobile systems. Keywords Single image super-resolution · Highway connection · Residual learning · Gating mechanism
1 Introduction Single image super-resolution (SISR) aims to reconstruct a high-resolution (HR) image from its corresponding lowresolution (LR) version. It has attracted increasing interest from both academic and industrial communities for its broad applications in computer vision, face recognition in security * Jinchang Ren [email protected] Xinying Xu [email protected] Jie Ren [email protected] Gaowei Yan [email protected] 1
Department of Electronic and Electrical Engineering, University of Strathclyde, Glasgow G1 1XW, UK
2
College of Electrical and Power Engineering, Taiyuan University of Technology, Taiyuan 030600, China
3
College of Electronics and Information, Xi’an Polytechnic University, Xi’an 710048, China
and surveillance video, medical imaging, and object detection. The image reconstruction approach becomes challenging when the captured images are affected by several factors such as bandwidth, noise, light conditions, and other artifacts. Super-resolution image reconstruction approach is an ill-posed problem since there exist multiple solutions HR for any given LR image. Super-resolution methods can be divided into three main categories, i.e., interpolation-based, reconstruction-based, and learning-based methods. Amo
Data Loading...
