An image super-resolution deep learning network based on multi-level feature extraction module
- PDF / 638,535 Bytes
- 13 Pages / 439.37 x 666.142 pts Page_size
- 95 Downloads / 189 Views
An image super-resolution deep learning network based on multi-level feature extraction module Xin Yang 1
1
1
& Yifan Zhang & Yingqing Guo & Dake Zhou
1
Received: 18 January 2020 / Revised: 28 August 2020 / Accepted: 22 September 2020 # Springer Science+Business Media, LLC, part of Springer Nature 2020
Abstract
Due to the lack of depth of the super-resolution (SR) method based on shallow networks, the feature maps of different convolutional layers have similar receptive fields, so that the performance improvement is not obvious. To solve this problem effectively, we propose an image SR reconstruction deep model based on a new multi-level feature extraction module in this paper. The method constructs an improved multi-level feature extraction module using the dense connection to obtain a deeper network and richer hierarchical feature maps for the SR task. In addition, we apply the loss function combined with the perceptual characteristics to improve the visual effect of the reconstructed image. Experiments show that the proposed method works well at reconstructed images with different magnification. Keywords Super-resolution . Deep learning . Convolutional neural network . Loss function
1 Introduction Image super-resolution (SR) technology has attracted considerable attention from both academia and industry in recent years. SR aims to reconstruct a corresponding high-resolution (HR) image with excellent quantitative measurement performance and visual effects from one or
* Xin Yang [email protected] Yifan Zhang [email protected] Yingqing Guo [email protected] Dake Zhou [email protected]
1
College of Automation Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing 210016, China
Multimedia Tools and Applications
multiple low-resolution (LR) images. SR has been widely used in many fields, such as target detection [2], image fusion [29], face recognition [9, 26] medical imaging [10], remote sensing image interpretation [24], PMMW image processing [44], forensic medicine [30], etc. In the past few decades, many traditional machine learning algorithms have been developed for SR, such as the SR method based on sparse representation [9], dual geometry neighborhood embedding (DGNE) [38], self-example learning [4] and so on. Its basic principle is to obtain the mathematical mapping relationship between the corresponding HR images and LR images through a large number of data training. Then SR can be achieved by the mapping. With the rise of artificial intelligence technology, the latest SR algorithm obtains the fine details required in the SR reconstruction process by constructing a data-driven deep learning model. Compared with the traditional algorithms, it has better objective criterion and visual effect. As a branch of machine learning, deep learning uses a multi-level structure to process information in a nonlinear manner. It can automatically learn the relationship between input data and output data. In addition to SR, deep learning has also obtained significant achievements in other areas
Data Loading...
