Development of a Hairstyle Conversion System Based on Mask R-CNN

Abstract

Interest in hairstyling, which is a means of expressing oneself, has increased, as has the number of people who are attempting to change their hairstyles. A considerable amount of time is required for women to change their hair back from a style that does not suit them, or for women to regrow their long hair after changing their hair to a short hairstyle that they do not like. In this paper, we propose a model combining Mask R-CNN and a generative adversarial network as a method of overlaying a new hairstyle on one's face. Through Mask R-CNN, hairstyles and faces are more accurately separated, and new hairstyles and faces are synthesized naturally through the use of a generative adversarial network. Training was performed over a dataset that we constructed, following which the hairstyle conversion results were extracted. Thus, it is possible to determine in advance whether the hairstyle matches the face and image combined with the desired hairstyle. Experiments and evaluations using multiple metrics demonstrated that the proposed method exhibits superiority, with high-quality results, compared to other hairstyle synthesis models.