Anime Scene Generator from Real-world Scenario using Generative Adversarial Networks

Abstract

This thesis presents a unique approach for image cartoonization and style transferring: translating an image or video in real life into an aesthetic, anime-like frame. By paying exceptional attention to the animation painting conduct, we propose to separately distinguish three feature maps from pictures: the surface description that contains smooth color shading characteristic of animation pictures, the construction depiction that emulates flattened global content and clear boundaries in a typical anime frame, and the texture representation that reflects high-frequency surface, forms, and details in animation pictures. All the extracted information will be fed into the Generator with the help of a VGG based discriminator to learn how to cartoonize a real-world photo. The learning objectives of our technique are independently based on each extracted feature map, making our model controllable and adjustable. Our solution takes unpaired photos and cartoon/anime images for training which can be fine-tuned for different problems and art styles. It is also incredibly lightweight so as to provide quick and easy inference. Experimental results show that our method can generate high-quality cartoon images from real-world photos and outperforms many existing methods.