A NEURAL NETWORK APPROACH FOR UNSUPERVISED IMAGE CLASSIFICATION

THIS PHASE I RESEARCH PROJECT INVESTIGATES THE PROBLEM OF UNSUPERVISED IMAGE CLASSIFICATION USING NEURAL NETWORK TECHNOLOGY. THE OBJECTIVES INCLUDE THE DEVELOPMENT OF FEATURE EXTRACTION TECHNIQUES, NEURAL NETWORK LEARNING/CLUSTERING PROCEDURE, AND PROOF-OF-CONCEPT SIMULATION. THE IMAGE FEATURES ARE FIRST EXTRACTED TO BECOME THE INPUT PATTERNS TO THE NEURAL NETWORK. THE LEARNING PROCEDURE IS ITERATIVE AND BASED ON THE BACK PROPAGATION PARADIGM. THE CLUSTERING PROCESS IS CARRIED OUT BY MERGING REGIONS BASED ON A MINIMUM ERROR CRITERION. THE CLASSIFIER AND THE LEARNING ALGORITHMS DEVELOPED IN THISPHASE I RESEARCH WILL BE USED FOR MANY IMAGE ANALYSIS TASKS ESPECIALLY IN IMAGE CLUSTERING AND SEGMENTATION. THE PHASE II EFFORTS WILL BE MORE COMPREHENSIVE AND THOROUGH INCLUDINGTHE DEVELOPMENT OF ALGORITHMS TO UTILIZE MULTISPECTRAL AND SPATIAL IMAGE CONTENTS, THE DESIGN OF A GENERIC NEURAL NETWORK CLASSIFIER, AND THE CONSTRUCTION OF A COMPLETE IMAGEANALYZER.