Cartoon Face to Human Face Translation using Contour Loss based CycleGAN

Cartoon to Human Translation transforms a 2D vector cartoon face to a Real Human Face. The mapping is based on semantic similarity of both the input domains. This is an image$\rightarrow$mage translation problem that finds its applications in the entertainment and animation industry. Cartoon movies evolved from 2D animations in 1930 and became more lifelike with timeline. In image synthesis, audio, and other sorts of data, Generative Adversarial Networks have demonstrated promising outcomes. They also produce excellent results when translating images to images. In this research, a CycleGAN based methodology for generating target Human Faces from source Cartoon Faces is proposed, preserving the facial characteristics i.e. face shape, eyebrow alignment and hair style. In order to improve the mapping we have used contour loss along with cycle consistency loss in our model and patch discriminator is used with L2 norm.