An experiment of surface recognition by neural trees

The recognition process of objects represented by range data is based on the description of their surfaces. The most popular method for doing that consists in decomposing the surface into regions holding the same differential properties. After successfully performing that task, a high level vision procedure for relating the various morphological segments has to be constructed. The decomposition of the surface is generally performed by calculating the functions K and H in any point and labeling the surface pixels according to the values of those functions. This paper describes the main lines of a surface recognition system based on computing structures called neural trees. Encodings of local samples of surfaces are used as input to a neural tree generator which is subsequently used to forecast global contours from local samples. Various noise levels are used in the training exercise. Experiments in varying the training order, the tree structure and the surface sampling method are performed in order to determine the resilience of such structures as global recognizers. Tree fan-out is studied in some detail. Binary and multi-class tree organizations are studied as well as a hybrid tree structure which combines sub-nets which perform n-way classification followed by binary sub-nets which deal with classified and misclassified patterns.