An Improved Active Shape Model for Face Alignment

W Wang,SG Shan,W Gao,B Cao,BC Yin

DOI: https://doi.org/10.1109/icmi.2002.1167050

2002-01-01

Abstract:In this paper, we present several improvements on conventional active shape models (ASM) for face alignment. Despite the accuracy and robustness of ASMs in image alignment, its performance depends heavily on the initial parameters of the shape model, as well as the local texture model for each landmark and the corresponding local matching strategy. In this work, to improve ASMs for face alignment, several measures are taken. First, salient facial features, such as the eyes and the mouth, are localized based on a face detector. These salient features are then utilized to initialize the shape model and provide region constraints on the subsequent iterative shape searching. Secondly, we exploit edge information to construct better local texture models for landmarks on the face contour. The edge intensity at the contour landmark is used as a self-adaptive weight when calculating the Mahalanobis distance between the candidate and reference profile. Thirdly, to avoid unreasonable shift from pre-localized salient features, landmarks around the salient features are adjusted before applying global subspace constraints. Experiments on a database containing 300 labeled face images show that the proposed method performs significantly better than traditional ASMs.