Multi-Shape - Hierarchical Active Shape Models

Active Shape Models (ASMs) have become one of the most widespread segmentation paradigms since their inception in the early nineties. However, their capability to capture and model the shapes’ variability is highly conditioned by the training set used. Trying to overcome this limitation, this paper presents a new hierarchical formulation of classical ASMs. Using the wavelet transform, a new complete tree wavelet packet is used to decompose the shape into small pieces of information, which are easier to model even with a small number of training shapes. Unlike previous hierarchical approaches, this new decomposition scheme and the matrix notation introduced allow the new hierarchical segmentation algorithm to deal with complex multi-shape structures in both, 2D and 3D spaces, maintaining the versatility of classical ASMs. The advantages of the new segmentation algorithm in terms of both accuracy and robustness with the number of training shapes have been successfully tested with two completely different databases containing multi-shape structures.