A Hierarchical Net-Structure Learning System for Pattern Description

Abstract

This thesis discusses a computer program that recognizes and describes two-dimensional patterns and the subpatterns composing those patterns, outputting names, locations and sizes of both patterns and subpatterns. The program also recognizes all patterns in a scene consisting of several patterns . Patterns are stored in a hierarchical, net-structure permanent memory, which is completely learned as a result of simple feedback from a trainer. Weighted links between memory nodes represent subpattern/pattern relationships. The memory is homogeneous, for subpatterns are represented in terms of primitive features in the same manner that patterns are represented in terms of subpatterns. A short term memory is used to store instances of permanent memory information during recognition. Pattern recognition is accomplished with a serial heuristic-search algorithm, unusual for a pattern recognition program, which attempts to search memory and compute input properties efficiently. Without special processing, the program can be asked to look for all occurrences af a specified pattern in an input scene.