Genetic lips extraction method with flexible search domain control

Abstract

In this paper, a lips extraction method by genetic algorithms (GAs), which has simple, high speed and high accuracy properties is described. This method is a visual front end of audio-visual speech recognition systems for mobile devices. Our method is based on template matching with GA. GA can use a significant amount of computation time, because the calculation and test of fitness values for generation × population is necessary. Furthermore GA is basically a global optimization algorithm. So, premature convergence in GA often causes it to fail to reach the global optimum. To extract lips region at high speed with high accuracy, the population must be small and local search must be more efficient. In order to overcome these problems, we propose the method to control search domain flexibly. We compare this method with the previous method by small population. The experimental results show the proposed method is effective. © Springer-Verlag 2004.