Attention augmented multi-scale network for single image super-resolution
- PDF / 6,658,958 Bytes
- 17 Pages / 595.224 x 790.955 pts Page_size
- 2 Downloads / 263 Views
Attention augmented multi-scale network for single image super-resolution Chengyi Xiong1,2 · Xiaodi Shi1 · Zhirong Gao3 · Ge Wang4
© Springer Science+Business Media, LLC, part of Springer Nature 2020
Abstract Multi-scale convolution can be used in a deep neural network (DNN) to obtain a set of features in parallel with different perceptive fields, which is beneficial to reduce network depth and lower training difficulty. Also, the attention mechanism has great advantages to strengthen representation power of a DNN. In this paper, we propose an attention augmented multi-scale network (AAMN) for single image super-resolution (SISR), in which deep features from different scales are discriminatively aggregated to improve performance. Specifically, the statistics of features at different scales are first computed by global average pooling operation, and then used as a guidance to learn the optimal weight allocation for the subsequent feature recalibration and aggregation. Meanwhile, we adopt feature fusion at two levels to further boost reconstruction power, one of which is intra-group local hierarchical feature fusion (LHFF), and the other is inter-group global hierarchical feature fusion (GHFF). Extensive experiments on public standard datasets indicate the superiority of our AAMN over the state-of-the-art models, in terms of not only quantitative and qualitative evaluation but also model complexity and efficiency. Keywords Single image super-resolution · Attention mechanism · Multi-scale convolution · Feature recalibration and aggregation · Local hierarchical feature fusion · Global hierarchical feature fusion.
1 Introduction Single Image Super-Resolution (SISR), which works to construct a high-quality high-resolution (HR) image based Chengyi Xiong
[email protected] Xiaodi Shi [email protected] Zhirong Gao [email protected] Ge Wang [email protected] 1
School of Electronic and Information Engineering, SouthCentral University for Nationalities, Wuhan, 430074 China
2
Hubei Key Laboratory of Intelligent Wireless Communication, South-Central University for Nationalities, Wuhan, 430074 China
3
School of Computer Science, South-Central University for Nationalities, Wuhan, 430074 China
4
Department of Biomedical Engineering, Rensselaer Polytechnic Institute, Troy, NY 12180 USA
on a corresponding low-resolution (LR) one, has attracted increasing attention in both academia and industry. With an efficient and cost-saving SISR technology, the obtained HR image is expected to equip itself with pleasing visual comfort and detailed information, which is needed in many fields, such as security and surveillance imaging [56], medical or satellite imaging [27, 44] etc. There could be many super-resolution results for a given LR one due to the lack of information, thus SISR is a challenging and ill-posed problem. So far, a large number of SISR schemes were proposed, which are generally classified as interpolationbased [3, 6, 29, 34], reconstruction-based [7, 53] and learning-based methods [2, 9–11, 20, 23,
Data Loading...
