Learning multiscale image models of 2D object classes

This paper is concerned with learning the canonical gray scale structure of the images of a class of objects. Structure is defined in terms of the geometry and layout of salient image regions that characterize the given views of the objects. The use of such structure based learning of object appearence is motivated by the relative stability of image structure over intensity values. A multiscale segmentation tree description is antomatically extracted for all sample images which are then matched to construct a single canonical representative which serves as the model 0fthe class. Different images are selected as prototypes, and each prototype tree is refined to best match the rest of the class. The model tree for the class is that tree which is best supported over all the initializations with different prototypes. Matching is formulated as a problem of finding the best mapping from regions of example images to those of the model tree, and implemented as a problem in incremental refinement of the model tree using a learning approach. Experiments are reported on a face image database. The results demonstrate that a reasonable model of facial geometry and topology is learnt which includes prominent facial features.