Multi-Template ASM and Its Application in Facial Feature Points Detection

Active shape model is one of the most popular methods for facial image alignment. To improve its accuracy in facial feature points detection in facial images with various expressions and under nonlinear illumination, a multi-template ASM method, which integrates Gabor features and combines local ASM and global ASM, is proposed in this paper. Human faces often have various kinds of expression, such as smiling, surprising, anger, being confused and so on. For eyes, it can be open or closed, while for mouths, it can be smiling, widely open, tightly closed and “O” shape with surprise. Such two different kinds of eye states and four various kinds of mouth states give great nonlinear transformation to their ordinary shapes. Therefore, they can’t be processed simply in a single model. In this paper, two local templates for eyes, four local templates for mouths, and some global templates for the whole face are created. Under the assumption that the locations of two inner corners of eyes and two outer corners of mouths are known, in the new method, the approximate area for the two eyes is found first, and then the state of eyes in this area is determined by using the eyes local templates and the Hausdorff distance. Similarly, the state of the mouth can also be known. Finally the whole face contour is searched by global template corresponding to the estimated eyes and mouth states. The experiment shows that the method can achieve much higher detection rate than the standard ASM method.