Information Theoretic Investigation of Automatic Object Recognition and ImageFusion

摘要

This work involves development of an information theoretic viewpoint for themodel-based ATR problem, bringing the problems of ATR and pose estimation (PE) into the framework of signal detection/demodulation theory. Specific goals are to predict maximum achievable ATR/PE performance for a given object/scene scenario, and to quantify the tradeoff between performance and system complexity. Progress has been made toward investigating the feasibility of developing useful system performance bounds. Extensions of classical results by Blabut and Fano have shown that tight upper performance bounds can be developed; however, the bounds thus far produced require more information and computation than is practical. Progress has been made toward the long-term goal of designing ATR strategies based directly on information theoretic considerations. A Hilbert Space Decision Tree approach has been developed and tested. This ATR algorithm consists of a series of binary decisions, each of which occurs in the native object image space and maximizes the amount of object discrimination information gained by the decision. In this way, heuristic assumptions regarding the importance of model and image information are avoided. This technique shows great promise for guiding the design of ATR systems.