Bayesian Image Matting Using Infrared and Color Cues

Extracting foreground objects from an arbitrary natural image or video sequences has become a classical problem in many photo-editing and video compositing applications. This topic has been deeply studied because it's widely used in filming and image rendering industry. With the recent developments in imaging sensors and computing technologies, image matting has become more attractive not only for professionals but also for general consumers as well. Consequently, various matting techniques and systems have been proposed to extract high quality mattes from both natural images and video. Most of these techniques require however intensive interaction with the user who provides a trimap to guide the matting. In this thesis, we provide a comprehensive review of existing image matting algorithms and systems, with an emphasis on our infrared-color matting system. We examine matting approaches that employ a single sensor such as the closed form solution [1] [2] and the Bayesian matting [3]. The performances of these techniques will be evaluated through some tests on color images. In order to reduce the amount of interaction with the user, we propose a new matting solution that combines the use of a color sensor with an infrared sensor for matting applications involving human actors. The latter facilitates the extraction of the initial trimap and provides additional information for the matte estimation. Our approach will be discussed in details and tested on several examples. The approach proposed in this thesis differs from the techniques proposed in the literature in many aspects. Indeed, it employs thermal information for human actors, which proves to be useful and effective for matting when combined with color information. It also introduces a new technique for automatic trimap construction that is based on the temperature's difference between the foreground actor and the background iii objects. Finally, the matting step is carried out using a Bayesian approach which combines the color and the infrared inputs into a single criterion. The matting results accuracy and speed show that our approach is capable of tackling digital image and video matting problems.