Group Members – Casey, Eric
Background
Generative Adversarial Network(GAN) is a mature and widely applied machine learning technique, in topics including but not limited to Image generation/translation/style transfer, constantly proving GAN’s power in generative work. Research has been done on realizing style transfer from real image to cartoon. CartoonGAN[1] is a powerful network that performances such style transfer on high resolution images, powered by TensorFlow 2.0 Alpha.
Motivation
Albeit its promising outcome, the current CartoonGAN can only be performed locally, we hope to bring this project a few steps further. With TensorFlow.js making performing deep learning projects within the browser possible, our goal is to wrap CartoonGAN with ml5.js, so that it can be utilized on the web.
Following the goal for the first stage, we also want to develop a web application implementing CartoonGAN, allowing users to perform style transfer on their custom image input.
Reference
[1] CartoonGAN: Generative Adversarial Networks for Photo Cartoonization