RSRGAN: computationally efficient real-world single image super-resolution using generative adversarial network
- PDF / 3,461,249 Bytes
- 18 Pages / 595.276 x 790.866 pts Page_size
- 89 Downloads / 215 Views
ORIGINAL PAPER
RSRGAN: computationally efficient real-world single image super-resolution using generative adversarial network Vishal Chudasama1 · Kishor Upla1 Received: 24 March 2020 / Revised: 19 July 2020 / Accepted: 29 September 2020 © Springer-Verlag GmbH Germany, part of Springer Nature 2020
Abstract Recently, convolutional neural network has been employed to obtain better performance in single image super-resolution task. Most of these models are trained and evaluated on synthetic datasets in which low-resolution images are synthesized with known bicubic degradation and hence they perform poorly on real-world images. However, by stacking more convolution layers, the super-resolution (SR) performance can be improved. But, such idea increases the number of training parameters and it offers a heavy computational burden on resources which makes them unsuitable for real-world applications. To solve this problem, we propose a computationally efficient real-world image SR network referred as RSRN. The RSRN model is optimized using pixel-wise L 1 loss function which produces overly-smooth blurry images. Hence, to recover the perceptual quality of SR image, a real-world image SR using generative adversarial network called RSRGAN is proposed. Generative adversarial network has an ability to generate perceptual plausible solutions. Several experiments have been conducted to validate the effectiveness of the proposed RSRGAN model, and it shows that the proposed RSRGAN generates SR samples with more high-frequency details and better perception quality than that of recently proposed SRGAN and SRFeatIF models, while it sets comparable performance with the ESRGAN model with significant less number of training parameters. Keywords Real-world image super-resolution · Generative adversarial network · Perceptual index · Learned perceptual image patch similarity
1 Introduction Single image super-resolution (SISR) refers to reconstruct the high-resolution (HR) image from its corresponding lowresolution (LR) counterpart. It can be applied in many fields such as medical imaging, surveillance and satellite imaging. The real-world LR image can be viewed as, ILR = (IHR ∗ wb ) ↓r +η,
(1)
where, ∗ and ↓r denote the convolution function and downscaling operation with scale factor r , respectively. Here, wb represents the unknown blur kernel and η represents the noise. In Eq. (1), one can notice that there are three different low-level vision tasks named as image deblurring, image denoising and image super-resolution take place for SISR in
B 1
Kishor Upla [email protected] Electronics Engineering Department, Sardar Vallabhbhai National Institute of Technology (SVNIT), Surat, India
real-world scenarios. Hence, it is a very complicated task to reconstruct HR image from real-world LR image. Recently, convolution neural networks (CNNs) obtain improvements in SISR by designing new CNN architectures and loss functions [12,13,17,27–29,33,45,46]. These methods are trained and evaluated on the simulated datasets in which the downsampled
Data Loading...
