Abstract:
In the realm of remote sensing (RS), the demand for improved spatial resolution has per sisted across multiple domains despite the exponential rise in satellite images. Single Im age Super-Resolution (SISR) aims to bridge this gap, and many state-of-the-art (SOTA)
deep learning models have been utilized for this purpose. However, the generation of
Low Resolution (LR) image pairs from High Resolution (HR) images is questionable, as
most datasets are synthetically generated either by bicubic down-sampling or by mod eling degradation with blur kernels and imaging noises. Therefore, SISR models trained
on such datasets cannot perfectly model real-world scenarios. This study proposes a
novel Real World Super-Resolution GAN (RWSRGAN) which is powered by Attention
Enhanced Residual in Residual Network (ARRDNet) along with a weighted combi nation loss function. It also presents a comparative analysis of different CNN-based,
GAN-based, Transformer-based SR models, and Image-to-Image Translation models for
the SISR task. These methods are trained on the WorldStrat dataset, a real-world
RS dataset comprising LR and HR pairs from two different satellites, Sentinel-2 and
SPOT 6/7, respectively, with a resolution ratio of approximately 6. Our experiments
show that the proposed RWSRGAN yield better qualitative and quantitative results as
compared to the existing SOTA models, and outperform most of them on a variety of
distortion-based and perception-based metrics.
