Sketch-to-Image Translation Using GANs

Abstract

Any image-to-image translation (I2I) has as its primary purpose the transfer of images from one domain to another while maintaining their content representations. Image synthesis, segmentation, style transfer, restoration, and posture estimation are just a few of the many computer vision and image processing difficulties that it can solve. I2I has gotten a lot of attention and made a lot of progress in recent years.In this Project, we look into Sketch to Real image translation utilizing Contextual GANs, Conditional GANs, and we propose Ensemble Learning on these two GANs, which enhanced the results.Normal GANs are unable to provide standard output when the input sketch is not clearly drawn or does not include a lot of useful information.We solve this challenge by employing an unique joint image completion method, in which the sketch serves as the picture context for completing or creating the output image.And by using Conditional GAN too which learns the mapping between input and output image and also learns the loss function to train this mapping. Because both GANs have advantages over each other, we used ensemble learning techniques to combine the best features of both GANs to make the network more powerful for translation.After completion of 150 epochs we have achieved the inception score of 8.3 using combination of Conditional GAN and Contextual GAN.