Image to Image translation for Hair-style manipulation

Abstract

Editing hairstyles in images has several practical applications. One of such applications, which is important in the beauty industry, is virtual hairstyle try-on products. This project explores hairstyle editing using image-to-image translation. In this work, consideration is given to the simple case of removing (or adding) hair from (or to) the image of a person’s head. This problem is approached using a labeled dataset that is generated using a Generative Adversarial Network. This enables the trained model to achieve state-of-the-art qualitative results, as demonstrated through extensive evaluation. This report also demonstrates that Generative Adversarial Networks can generate high fidelity paired datasets that can be used as training data.