Abstract
Currently, deep-learning-based methods have been the most popular super-resolution techniques owing to the improvement of super-resolution performance. However, they are still lack perceptual fine details and thus result in unsatisfying visual quality. This article proposes a novel method for high-quality perceptual super-resolution imaging, named SRLRGAN-SN. It aims to recovery visually plausible images with perceptual texture details by using the least squares relativistic generative adversarial network (GAN). The method applies the spectral normalization on the network with the target of enhancing the performance of GAN for super-resolution task. The least squares relativistic discriminator is designed to drive reconstruction images approximating high-quality perceptual manifold. Besides, a novel perceptual loss assembly is proposed to preserve structural texture details as much as possible. Results of experiment show that our method can not only recovery more visually realistic details, but also outperforms other popular methods regarding to quantitative metrics and perceptual evaluations.
Highlights
In order to ensure that the resolved image is similar to the real sample in pix-wise content, this study explores the metric perceptual index (PI) couple with peak signal-to-noise ratio (PSNR) to measure the reconstructed performance
A novel method named SRLRGAN-SN for super-resolution imaging is proposed combined with spectral normalization and least squares relativistic discriminator
The spectral normalization is applied on the network to enhance the performance of generative adversarial network (GAN) for super-resolution task
Summary
As a typical ill-posed issue during restoration [1], [2], super-resolution for imaging reconstruction aims to improve. CNN-based methods always try to minimize the pixel loss to achieve high quantitative value in terms of PSNR and SSIM [21] They are still lack fine details and result in unsatisfying perceptual quality. Odena et al [31] prove that the well-conditioned generator can help to enhance GAN’s performance In view of this conclusion, Zhang et al [32] propose SAGAN by employing spectral normalization in both the generator and the discriminator network for stabilizing training, which owns low computational cost of optimization, and increases the stability of training process. Compared with SRGAN, the spectral normalization is applied on both the generative and discriminative network with the target of enhancing the performance of GAN for super-resolution task.
Sign-in/Register to view highlights & summary Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
