Image Synthesis of Warli Tribal Stick figures using Generative Adversarial Networks

With the advancement in deep learning methods, Generative Adversarial Neural Networks(GANs) have been successful in the application of Computer Vision, Natural Language Processing, and medical. It is observed to have great research potential of using GAN for Image Synthesis. GANs can generate meaningful samples from complex patters or distributions. We present a method to generate Asian Tribal human synthetic stick figures using GAN. With variety of applications the output are limited to low resolution and are unrealistic. The work includes image to image translation where the inputs are low resolution images and as the training progresses, addition of new layers lead to better or high resolution output. For this purpose, it was mandatory to train the generator and discriminator following the principles governing the concept of transfer learning and to fetch evidence of how the generated output resembles the original data set. We also extend additional visualizations of our approach and support our GAN model with images of our generated output. This piece of work includes customized data set as input, where the input images are constructed based on laws of coordinate geometry and spherical polar coordinates and final generated images to illustrate its resemblance with the actual input images. The generated images were quite efficient in faking the discriminator, contributing towards robustness, precision and viability of the stated architecture.