Using classifiers as heuristics to describe local structure in Active Shape Models with small training sets

• We show a method for small data sets in ASMs called Heuristic Local Appearance Model (HLAM).
• We extend the procedure to multi-channel images.
• We compare HLAM to the Mahalanobis appearance model and the edge search method.
• We use therefore a synthetic and a real dataset of images.
• The results show that HLAM outperforms the other two approaches.

Active Shape Models (ASM) are a successful image segmentation technique that is widely used by the image processing community. This technique is very appealing when the results of the segmentation are going to be used to perform some kind of classification, as it provides a mathematical model of the segmented contours. Nevertheless, little attention has been paid to the development of general local appearance models for small image training sets and most researchers have resorted to ad hoc solutions. In this paper we propose a heuristic approach to this problem. A general procedure for the use of heuristics to guide the ASM search algorithm and an implementation using machine learning classifiers is presented. This procedure is also extended to cope with multichannel images. Tests are carried out over small synthetic and real image datasets. The performance of this approach is compared to the most commonly used Mahalanobis appearance model and the simpler edge search strategy. The results show that the heuristic approach performs better than the other two procedures.