Logo Generation with Generative Adversarial Networks Conditioned on Sentiment Terms

Abstract

Logos are very important to represent the impression of companies and brands. Since many logos are created by designers, there is a problem of high cost both in terms of money and time. Since sentiment terms are important as keywords when designing logos, we propose a logo generation system using sentiment terms. Originally, we should create a dataset of logo images and sentiment terms directly, but it is difficult to collect these logo images and the corresponding sentiment terms. Therefore, we employ the domain adaptation method to the dataset of furniture images and sentiment terms to create dataset of the logos and sentiment terms. Finally, we can generate logos considering 17 sentiment terms using Generative Adversarial Networks. We evaluated the proposed system using MS-SSIM score and confirmed that it can generate logos with diversity. Moreover, we performed subjective experiments to confirm that it can generate logos reflecting sentiment terms.