Conditional generative adversarial network with densely-connected residual learning for single image super-resolution
- PDF / 3,098,469 Bytes
- 15 Pages / 439.642 x 666.49 pts Page_size
- 76 Downloads / 276 Views
Conditional generative adversarial network with densely-connected residual learning for single image super-resolution Jiaojiao Qiao1 · Huihui Song1 · Kaihua Zhang1
· Xiaolu Zhang1
Received: 7 October 2019 / Revised: 20 August 2020 / Accepted: 2 September 2020 / © Springer Science+Business Media, LLC, part of Springer Nature 2020
Abstract Recently, generative adversarial network (GAN) has been widely employed in single image super-resolution (SISR), achieving favorably good perceptual effects. However, the SR outputs generated by GAN still have some fictitious details, which are quite different from the ground-truth images, resulting in a low PSNR value. In this paper, we leverage the groundtruth high-resolution (HR) image as a useful guide to learn an effective conditional GAN (CGAN) for SISR. Among it, we design the generator network via residual learning, which introduces dense connections to the residual blocks to effectively fuse low and high-level features across different layers. Extensive evaluations show that our proposed SR method performs much better than state-of-the-art methods in terms of PSNR, SSIM, and visual perception. Keywords Super-resolution · Conditional generative adversarial network · Residual network · Deep convolutional neural network
1 Introduction Predicting its HR image from a low-resolution (LR) image is known as SISR [4, 18, 19, 21, 22, 30, 35], which has been widely applied in various fields, such as security monitoring and identification, medical digital imaging diagnosis, and satellite remote sensing [37]. In recent years, approaches based on convolutional neural networks (CNNs) have made a significant improvement in the field of SR [1, 4, 12, 18, 19, 21, 22, 27, 35]. Since the introduction of SRCNN by Dong et al. [3], an increasing number of works based on deep CNNs have been proposed. Kim et al.’s VDSR [19] and DRCN [18] demonstrate the importance of network depth for SR. Although SRCNN and VDSR have demonstrated excellent Kaihua Zhang
[email protected] 1
Jiangsu Key Laboratory of Big Data Analysis Technology (B-DAT) and Jiangsu Collaborative Innovation Center of Atmospheric Environment and Equipment Technology (CICAEET), Nanjing University of Information Science & Technology, Nanjing, China
Multimedia Tools and Applications
performance, they all first use interpolation to enlarge the LR input images and then input them into the network for feature learning. This not only increases computational complexity, but also loses some low frequency information [27]. To address these issues, Shi et al. [27] first perform non-linear mappings in low-dimensional space and then design an efficient sub-pixel convolution layer to upsample the feature maps at the end of the SR models. Afterwards, Dong et al. propose FSRCNN [4] which employs a learnable transposed convolution layer for post-upsampling. Lai et al. [20] present Lap-SRN to progressively reconstruct higher-resolution images. Since then, SISR models tend to be more and more deeper and wider. Zhang et al. [39] propose an over
Data Loading...