Beauty with Variational Methods: An Optic Flow Approach to Hairstyle Simulation

摘要

Although variational models offer many advantages in image analysis, their successful application to real-world problems is documented only for some specific areas such as medical imaging. In this paper we show how well-adapted variational ideas can solve the problem of hairstyle simulation in a fully automatic way: A customer in a hairdresser's shop selects a new hairstyle from a database, and this hairstyle is automatically registered to a digital image of the customer's face. Interestingly already a carefully modified optic flow method of Horn and Schunck turns out to be ideal for this application. These modifications include an extension to colour sequences, an incorporation of warping ideas in order to allow large deformation rates, and the inclusion of shape information that is characteristic for human faces. Employing classical numerical ideas such as finite differences and SOR iterations offers sufficient performance for real-life applications. In a number of experiments we demonstrate that our variational approach is capable of solving the hairstyle simulation problem with high quality in a fully practical setting.