Image-to-Image Translation with Conditional Adversarial Networks

Several syntheses of photographs based on label maps, restoration of objects using edge maps, and image colorizing, and many others traditionally requires designing individual/unique loss function for each task. These tasks have one thing in common: they can be treated as a single Image to the Image translation problem. Convolution neural networks (CNNs) have become popular for a wide variety of image-related problems. The Image-to-Image translation problem can be solved with one general approach, which is using the conditional adversarial network. We have implemented conditional adversarial networks. The specialty of this network is they acquire insight on a loss function to train this mapping in addition to learning the mapping to output image from input image.