Active shape model based on sparse representation

Yanqing Guo,Ran He,Wei-Shi Zheng,Xiangwei Kong

DOI: https://doi.org/10.1007/978-3-642-35136-5_12

2012-01-01

Abstract:Active shape model (ASM), as a method for extracting and representing object shapes, has received considerable attention in recent years. In ASM, a shape is represented statistically by a set of well-defined landmark points and its variations are modeled by the principal component analysis (PCA). However, we find that both PCA and Procrustes analysis are sensitive to noise, and there is a linear relationship between alignment error and magnitude of noise, which leads parameter estimation to be ill-posed. In this paper, we present a sparse ASM based on l1-minimization for shape alignment, which can automatically select an effective group of principal components to represent a given shape. A noisy item is introduced to both shape parameter and pose parameter (scale, translation, and rotation), and the parameter estimation is solved by the l1-minimization framework. The estimation of these two kinds of parameters is independent and robust to local noise. Experiments on face dataset validate robustness and effectiveness of the proposed technique.