Research of Facial Features Extraction Based on Improved ASM
Traditional ASM (Active Shape Model) methods usually depend heavily on the initial facial landmark positions located by the face detector. During the fitting process the ASM model may be trapped into local optima instead of the global optima due to the inaccurate initial position selection. In this paper we present that this problem can be solved by incorporating pupil and chin localizations in adjusting the initial landmark positions, together with adaptive step length selection based on two-dimensional gradient characteristic in a multi-scale ASM model. Experiments on a face database consisting of 1000 individuals show that this method is practically effective.
G. F. Zhai et al., "Research of Facial Features Extraction Based on Improved ASM", Advanced Materials Research, Vols. 271-273, pp. 131-136, 2011