Image super-resolution using supervised multi-scale feature extraction network
- PDF / 4,528,043 Bytes
- 14 Pages / 439.642 x 666.49 pts Page_size
- 48 Downloads / 211 Views
Image super-resolution using supervised multi-scale feature extraction network Yemei Sun1 · Yan Zhang1 · Shudong Liu1 · Weijia Lu1 · Xianguo Li2 Received: 19 February 2020 / Revised: 13 July 2020 / Accepted: 28 July 2020 / © Springer Science+Business Media, LLC, part of Springer Nature 2020
Abstract Image super-resolution using deep convolutional networks have recently achieved great successes. However, previous studies have failed to consider the spatial information by simply using a single-size filter, and they do not take full advantage of hierarchical features from low-resolution images, thereby these results are unsatisfactory. In this paper, the supervised convolutional network with multi-scale feature extraction is presented to further improve accuracy. First, the spatial information of the image can be better utilized by different filter sizes. This enhances the adaptability of the network. Second, dense connections are introduced to alleviate the vanishing-gradient problem and accelerate the convergence speed. Third, by adding auxiliary supervised connections to these intermediate layers, they provide additional regularization and increase the backpropagation gradient signal. Extensive experiments on the open challenge datasets confirm the effectiveness of proposed network. Our algorithm can restore high-quality high-resolution images quickly and outperform other methods by a large margin. Keywords Super-resolution · Multi-scale feature extraction · Supervised · Vanishing gradients
1 Introduction Image super-resolution (SR) is an active topic in computer vision. It aims to restore a visually pleasing high-resolution (HR) image from its degraded low-resolution (LR) image. It has been widely used in a variety of computer vision applications, such as medical imaging and surveillance imaging [2]. While image SR is an ill-posed inverse procedure, since there exists a multitude of solutions for any LR input. To address this problem, many SR learning algorithms have been studied, including interpolation-based [41], reconstruction-based [38], and learning-based methods [7, 15, 20, 26, 31–33, 42]. Yan Zhang
[email protected] 1
School of Computer and Information Engineering, Tianjin Chengjian University, Tianjin, China
2
Tianjin Key Laboratory of Optoelectronic Detection Technology and System, Tianjin, China
Multimedia Tools and Applications
Recently, deep convolutional neural network (CNN) with strong learning ability has achieved great success. This makes CNN the dominant learning method in computer vision. Since Dong et al. [8] firstly introduced a shallow three-layer convolutional network (SRCNN) to implement image SR, deep learning based methods have attracted extensive attention. The benefits of these methods mainly come from three key factors: increasing depth [20], using skip connections (residual) [12, 18, 27, 39] and using effective building modules [14, 23, 30, 40]. Although these methods are flexible to produce relative highquality HR images, they still suffer from some drawbacks: (1) the c
Data Loading...
