Using a hierarchical approach to avoid over-fitting in early vision

Abstract

The ALISA system is an adaptive learning image analysis system whose hierarchical design allows learning at two levels: texture and geometry. Earlier experiments using only the texture level were repeated using the combination of the texture and geometry modules to demonstrate the advantages of learning without resorting to inventing application-specific features which over-fit the domain. The two-level approach achieves quantitative results comparable with the single-level approach, but requires far fewer training examples and uses simple general-purpose features. The hierarchical approach also generates output class maps that are isomorphic with the original image and preserve important structures, and which therefore may be used for further processing.