PHOTO CARTOONIZATION USING GAN

Abstract

In this report we are proposing a model capable of converting an image to cartoonized version. This functionality has many applications across anime, cartoon industry. The proposed model is one of the recent popular learning-based methods for stylizing images in artistic forms like paintings. However, in existing methods, (1) the style of comics has a unique feature of strong simplification and abstraction, which tends to have relatively simple textures and therefore does not give satisfactory results, poses a significant challenge to the popular loss functions for generator defined in recent developed methods. In this project, a Generative Adversarial Network (GAN) model is proposed for styling images very similar to cartoons. The model takes edge-smoothed cartoon images and for training. Other than that, to measure the model two new loss functions are also proposed. (1) loss of semantic content, its value indicates if converted image looks like cartoon images or not (i.e., clear and visible edges in the image), and (2) edge-promoting clarity loss to maintain a good edge. Also, the proposed model is capable of being trained much efficiently than existing methods proposed so far. With the experimental results we can tell that model performs good and is able to develop cartoon images from normal input image with high-quality.