Implementation of Cycle Consistent GANs for Map to Image Translation using Deep Learning Techniques

With the rapid urbanization and population growth, locating suitable geographical areas to accommodate these drastic changes is paramount. This study explores the use of Generative Adversarial Networks (GANs), specifically Cycle- Consistent Generative Adversarial Networks (Cycle GAN) and Conditional GAN(C-GAN), to generate a map image from a satellite image, that is easy to read and understand. By conditioning the generator and discriminator networks on the input satellite images and target map domain, the model picks up on certain distinguishable features from the satellite image and the map image to generate a map image. A Cycle-GAN doesn’t require the use of paired datasets, unlike C-GAN. This research study also briefly discusses about the various features that the Cycle-GAN applies to generate an accurate image. Various loss functions such as GAN loss, identity loss, and Cycle consistency loss that the GAN employs to improve the resolution of the image are also discussed in the study.