Perceptual improvements for Super-Resolution of Satellite Imagery

Super-resolution of satellite imagery poses unique challenges. We propose a hybrid method comprising two existing deep network super-resolution approaches, namely a feedforward network called DeepSUM, and ESRGAN, a GAN-based approach, to super-resolve multiple low-resolution images by a factor of four to obtain a single high-resolution image. We also introduce a novel loss function, called variation loss, to better define edges and textures to create a sharper, perceptually better output. Using our hybrid, we inherit some of the advantages of both deep learning approaches, resulting in super-resolved images that better show boundaries, textures, and details.