SatGAN: Satellite Image Generation using Conditional Adversarial Networks

Mapping accurate satellite images from street view is a challenging image-to-image translation task. Recent advances in Generative Adversarial Networks (GANs) have shown promising results in image-to-image translation. Pix2Pix is a generalized framework for image-to-image translation which uses Conditional Adversarial Networks. It has performed well on a large number of datasets. However, it uses a simple pixel-level reconstruction loss due to which the output suffers distortion in some cases. In this paper, we propose SatGAN, which is based on Pix2Pix but adds the perceptual reconstruction loss with the pixel-level reconstruction loss to produce colourful and blur free images. The perceptual reconstruction loss forces the Generator to generate samples having a similar feature representation with that of the ground truth. We experimentally prove that SatGAN produces both qualitatively and quantitatively better results than Pix2Pix for street-view to satellite dataset.