Recently, generative adversarial network-based image super resolution has been investigated, and it has been shown to lead to overwhelming improvements in subjective quality. However, it also leads to checkerboard artifacts and the unpleasing high-frequency (HF) components. In this paper, we propose a multi-discriminators-based image super resolution method that distinguishes those artifacts from various perspectives. First, the DCT perspective discriminator is proposed because the checkerboard artifacts are easily separated on the frequency domain. Second, the gradient perspective discriminator is proposed, because the unpleasing HF components can be discriminated on the gradient magnitude distribution. These proposed multi-perspective discriminators can easily identify artifacts, and they can help the generator reproduce artifact-less SR images. The experimental results show that the proposed SR-GAN with multiperspective discriminators achieves objective and subjective quality improvements in terms of PSNR, SSIM, PI and MOS, as compared to the conventional SR-GAN by reducing the aforementioned artifacts. INDEX TERMS Image super-resolution, deep learning for super resolution, SR GAN, multi-discriminators.

No abstract

Self-similarity has been popularly exploited for image super resolution in recent years. Image is decomposed into LF (low frequency) and HF (high frequency) components, and similar patches are searched in the LF domain across the pyramid scales of the original image. Once a similar LF patch is found, the LF is combined with the corresponding HR patch, and we reconstruct the HR (high resolution) version. In this paper, we separately search similar LR and HR patches in the LF and HF domains, respectively. In addition, self-similarity based SR is applied to the new structure-texture domain instead of the existing LF and HF. Experimental results show that the proposed method outperforms several conventional SR algorithms based on self-similarity.

No abstract

Bayesian based Multi-Frame Super-Resolution (MF-SR) has been used as a popular and effective SR model. On the other hand, the texture region is not reconstructed sufficiently because it works on the spatial domain. In this study, the MF-SR method was extended to operate on the frequency domain to improve HF information as much as possible. For this, a spatially weighted bilateral total variation model was proposed as a regularization term for a Bayesian estimation. The experimental results showed that the proposed method can recover the texture region more realistically with reduced noise, compared to conventional methods