Generation of realistic images from human drawn sketches using deep learning

Abstract

Processing sketches to produce realistic images is an intriguing idea in the world of emerging Artificial Intelligence. We present a Generative Adversarial Network (GAN) based methodology that creates satisfactory images for the most prevalent categories in our approach. The proposed approach is applicable not just to people, but also to animals, objects and foods. The system takes a sketch and analyzes it using a powerful neural engine to produce new photographs that resemble realistic images. We also used a data augmentation method to dramatically increase the variety of data available for training models. The proposed model has achieved approximately 96.36% accuracy over generating sketch to realistic images of people and 40.63% accuracy for objects and animals. Moreover, about 76.63% accuracy on generating sketches from strokes on an average from people class.