Adaptation of ASM to Lips Edge Detection

Seeing the talker’s lips in addition to audition can improve speech understanding which is rather based on lips shape temporal evolution than on absolute mouth shape. In this article, we propose an adaptation of Active Shape Model (ASM) to the extraction of lips shape over an image sequence. The algorithm does not require any make-up or markers and works under natural lighting conditions. After the definition of a training base, initial mouth model is iteratively deformed under constraints according to spatiotemporal energies depending either on luminance or hue. A robust prior detection of four points of the model is proposed in order to automatically and accurately initialize the egde detection. The success of our approach is tested on many image sequences of multi-speakers with multi-speaking.