LEARNING TO RECOGNIZE VISUAL CONCEPTS: Development and Implementation of a Method for Texture Concept Acquisition Through Inductive Learning

LEARNING TO RECOONIZE VISUAL CONCEPTS: DEVELOPMENf AND IMPLEMENTATION OF A METIIOD FOR TEXTURE CONCEPTS ACQUISITION TIIROUGH INDUCTIVE LEARNING JERZY WOJCIECH BALA, Ph.D. George Mason University, May 1993 Dissertation Director: Dr. Ryszard S. Michalski The goal of this research is to explore the application of symbolic learning methods to problems of computer vision. The research presented in this thesis has been concerned primarily with the development of methods for inductive learning of texture descriptions. Texture description learning is done in the following phases: (i) data pre-processing and attribute extraction, (ii) acquisition of texture concept descriptions, (iii) optimization of acquired descriptions, and (iv) recognition of unknown texture samples. The methodology adapted to the acquisition and recognition of complex vision data is based on an extension ofAQ [Michalski, 1986], a learning from-examples algorithm. This approach for inductive learning of texture descriptions was originally proposed by Michalski [1973] and was initially applied using ILL lAC III image recognition computer facilities. This research presents a novel extension to the initial approach, which is called Multilevel Logical Templates. The novelty lies in multilevel symbolic image transformations, new advanced concept description optimization methods for noise-tolerant learning, and a multistrategy approach to learning from vision data. An important contribution of the research is the experimental demonstration that symbolic inductive learning methods can be successfully applied to the domain of continuous attributes of low level vision in which non symbolic methods have been traditionally employed.