Fast image super-resolution with the simplified residual network
- PDF / 2,859,831 Bytes
- 13 Pages / 439.37 x 666.142 pts Page_size
- 91 Downloads / 259 Views
Fast image super-resolution with the simplified residual network Chunmeng Wang 1 & Lingqiang Ran 2 & Chen He 3 Received: 7 June 2020 / Revised: 11 September 2020 / Accepted: 17 September 2020 # Springer Science+Business Media, LLC, part of Springer Nature 2020
Abstract
Recently, the image super-resolution (SR) methods based on residual learning have obtained remarkable quality performance. However, the current residual-learning methods have low computational performance and slow convergence rate. In this paper, we propose a high-efficiency two-level residual network to make the network learn more useful high-frequency information. Only 5 convolution layers in the LR space are used in our residual network, and no parameters are introduced in the other layers. Compared with the long training time up to several hours or days of previous deep residual networks, our simplified network can make the training time reduce to half an hour. Besides, our simplified network achieves satisfactory quality performance. The evaluation on the public datasets shows that our method can process SR of ultra-high definition (UHD) videos in real-time (more than 24 frames per second) on a generic graphical processing unit (GPU). Keywords Super resolution . Convolutional neural networks . Simplified residual network
1 Introduction Single image super-resolution (SISR) is an image restoration problem that reconstructs one high-quality high-resolution (HR) image from one single low-resolution (LR) image. SISR is a highly ill-posed problem because there is an infinite number of solutions from an LR image to an HR image. Recently, deep learning methods provide the promise to recover the lost details in the LR image by using the information of the training data. * Chunmeng Wang [email protected]
1
School of Computer Engineering, Jinling Institute of Technology, Nanjing, Jiangsu 211169, China
2
School of Computer Science and Technology, Shandong University of Finance and Economics, 250014 Jinan, Shandong, China
3
Media and Communication College, Weifang University, Weifang, Shandong 261061, China
Multimedia Tools and Applications
Recently, a variety of SR methods based on convolutional neural networks (CNN) have been proposed since SRCNN [3] introduced deep learning to solve the SISR problem. Among them, deep residual-learning methods [2, 8, 10-12, 18, 27] have provided outstanding reconstruction quality. Instead of directly learning features from the LR image to the HR image, they learned the residual between the HR image and the corresponding LR image. These deep residual-learning methods achieved significantly better quality performance than the standard CNN based method. However, the large number of parameters of these methods made the training time up to many hours or several days and the processing speed far from real-time for UHD videos on one generic GPU. These methods had large number of parameters and low efficiency for the two following reasons: Firstly, many residual-learning methods [8, 18] transformed the LR image to HR s
Data Loading...
