Facial Feature Localization Using Robust Active Shape Model and POEM Descriptors

Abstract

Active Shape Model (ASM) has been shown to be a powerful tool for image interpretation, especially in feature points localization for face image. The original ASM model parameter estimation is based on the assumption that the profiles follow a Gaussian distribution. Its performance is always vulnerable to distortion due to pose, illumination and expression variations. In this paper, the improvement of ASM model concerns the following two aspects. Firstly, an adaptive parameter estimation method are proposed by defining a rotation factor. Secondly, local appearances of landmarks are originally represented by Patterns of Oriented Edge Magnitudes (POEM) descriptors, which can provide more robust and accurate guidance for searching than intensity profiles. The simulations are carried out using the IMM dataset, which contains 240 face images. Experimental results show that the proposed method significantly outperforms the original ASM and ASM plus LBP method under exterior variations.