GANToon - Generative Adversarial Network (GAN) With Dual Discriminators for Cartoonization

Abstract

Cartoons quickly grab one's attention and are widely present in our daily lives. Creating these cartoons manually requires high artistic skills and is a time-consuming activity. In recent years, many techniques have been identified to transform real photos into cartoon images but they have a few drawbacks like the generated images being too dark, the colors not matching, and the textures having sharp edges. This research work focuses on overcoming these shortcomings by proposing the following: (1) a novel lightweight Generative Adversarial Network (GAN) based architecture, named GANToon, with one generator and two discriminators (2) five different loss functions to train the generator, namely, the adversarial loss, the content loss, the texture loss, the color loss, and the total variation regularizer loss. The goal is to generate images with cartoon-like textures that follow the color pattern of the input images. Ablation experiments are conducted to show the effectiveness of each of the loss functions. The results are quantitatively and qualitatively compared with other stateof-the-art models. Through extensive experiments, it is demonstrated that GANToon is lightweight and generates high-quality images outperforming the other models.