Published in


Creating ‘Yuru-Chara’ and a New Class Conditional GAN Approach For Small Datasets

Yuru-chara is the Japanese term for the many “mascot characters” created by local governments and companies to promote activities, regional revitalization, products, etc. This is a long-established tradition — even the Tokyo Metropolitan Police have a yuru-chara in the cute and mouselike Pipo-kun.

The economic impact of yuru-chara is significant, and each year countless yuru-charas from across Japan are ranked by popularity in a gala “Yuru-Chara Grand Prix.“

Creating original yuru-chara however can be time and money consuming, with high designer fees, etc. Moreover, the success of a new yuru-chara cannot be guaranteed.

Machine learning researchers have already analyzed the relationship between yuru-chara character appearance and ranking in the Grand Prix, revealing various popularity, design and colour correlations. Now, researchers have proposed a new and inexpensive method for automatically generating such characters.

The paper YuruGAN: Yuru-Chara Mascot Generator Using Generative Adversarial Networks With Clustering Small Dataset from researchers at Tokyo University of Agriculture and Technology (TUAT) introduces a class conditional GAN approach based on clustering and data augmentation as a stable and efficient method for applying class conditional GANs to small and unlabelled datasets.

Previous research on GANs used class conditions in a dataset for training. This was done to stabilize learning and improve the quality of the generated images. But such class conditional GANs rely on large datasets, and struggle when the original data is insufficient and when a clear class (in this case, a yuru-chara image) is not provided.

The proposed class conditional GAN is a learning stabilization method for improving the quality of generated images. When the category information obtained through various clustering methods is provided to the image generation model together with a yuru-chara image, high-quality and diverse images can be stably generated without mode collapse.

Clustering using X-means
Details of the dataset obtained by clustering

The researchers applied the sum of X mean K-means ++ to the yuru-chara image dataset and created a dataset with classes for comparison. Clustering was performed based on ResNet and edge extraction and features extracted by ResNet from yuru-chara images to create a dataset with classes. Assigning these datasets to a class conditional GAN model enabled the generation of yuru-chara images. When ResNet was used for clustering, it generated high-quality yuru-chara images with characteristics close to the yuru-chara image dataset.

The results illustrate that clustering of datasets based on features obtained by ResNet may be effective for improving the quality of images generated by GANs when the dataset is small. In the future, the researchers say they plan to investigate whether clustering by ResNet is also effective for other small datasets.

The paper YuruGAN: Yuru-Chara Mascot Generator Using Generative Adversarial Networks With Clustering Small Dataset is on arXiv.

Author: Xuehan Wang | Editor: Michael Sarazen

Thinking of contributing to Synced Review? Synced’s new column Share My Research welcomes scholars to share their own research breakthroughs with global AI enthusiasts.

We know you don’t want to miss any story. Subscribe to our popular Synced Global AI Weekly to get weekly AI updates.

Need a comprehensive review of the past, present and future of modern AI research development? Trends of AI Technology Development Report is out!

2018 Fortune Global 500 Public Company AI Adaptivity Report is out!
Purchase a Kindle-formatted report on Amazon.
Apply for Insight Partner Program to get a complimentary full PDF report.



Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store