Generative Adversarial Network (GANs) for Image Translation

Image to Image translation deals with converting images from one form of representation to another automatically, while keeping the original structure or semantics. For example, the conversion of images represented by semantic label maps to color images, or the conversion of edge maps to full images. Most image transformation tasks require significant human involvement and artistic skill. Colorization of black and white images, is currently done by hand, a process that is both time consuming and expensive. However, due to the multitudes of coloring schema that exist for a single black and white image, the automatic colorization of images is not a trivial problem, as the mapping is notone-to-one.This paper deals with image to image translation using Generative Adversarial Networks (GANs), a technique for image synthesis and prediction which is time and cost effective. This work proposes a comparison between the proposed model (GAN) and a CNN, both were made to run on one image translation problem. The chosen problem was grayscale to color translation and the dataset was a subset of the Imagenet dataset. It was observed that the GAN took around eight hours lesser than the CNN model and the quality of images generated by the GAN were much superior than the one generated by the CNN.