Hierarchical multi-resolution decomposition of statistical shape models

Abstract

Active shape models (ASMs) are some of the most actively used model-based segmentation approaches, whose usefulness has been successfully demonstrated in a wide variety of applications. However, they suffer from two important drawbacks: (1) the large number of training samples required to adequately model the object of interest and (2) their dependence on the initialization. Partial solutions for both of these problems have been proposed but come at the expense of a significant increase in computation time, which is of great importance in current applications. Based on the wavelet transform, the present paper proposes a new full multi-resolution framework that allows not only to hierarchically decompose any type of shape into a set of bands of information, but also to create instances of the shape with different degrees of detail. This is a multi-resolution decomposition of the shape space. Combining this new approach with a multi-resolution Gaussian pyramid of the image, the result is a new full multi-resolution hierarchical ASM that reduces both the size of the training set required and the initialization dependence while improving the classical segmentation algorithms accuracy and computational complexity. The advantages of this new algorithm have been successfully verified over two different databases.