Performance Evaluation of T-Shirt Image Generation System Considering User’s Preference

Abstract

In this study, we propose a system to generate the T-shirt design images that match a user’s preference based on generative adversarial networks (GAN) and a Kansei agent. The generator of GAN can be used to generate lifelike images similar to the learned data. The Kansei agent is a model that learns the user’s Kansei evaluation and retrieves images that the user likes. The Kansei agent is constructed using two neural networks. The first network extracts features from given T-shirt images. The second one estimates the user’s evaluation of T-shirt design using the extracted features. We examined the performance of the proposed system according to the results of the evaluation experiment. The results of the experiment showed that the proposed system finally presented most users T-shirt design images that were highly evaluated.