NASA spacecraft and scientific instruments have collected a large volume of image data which are too complex to be manually analyzed. Therefore, NASA is expending a considerable amount of research effort in applying machine learning and pattern recognition techniques to the task of digital image analysis. This proposal presents a novel learning approach for recognizing relevant geological entities in raw pixel images. NASA's current existing tool will be used to extract a wide variety of possible features from the images. The newly proposed learning approach will then use these features to create an effective set of separately trained neural networks. Combining the output of multiple trained neural networks is an extremely powerful classification technique as long as there is some disagreement among the network's predictions. Training on different subsets of features should promote an appropriate level of disagreement.The proposed ensemble technique naturally applies to the problem of feature selection - the central problem faced by digital image analysis. This algorithm builds on two well established and thoroughly evaluated research fields: (i) backpropagation training of neural networks, and (ii) optimization in genetic algorithms. Since this proposed approach has proven to be successful and is supported by a solid foundation of thoroughly tested research, it promises to be successful at (i) improving the ability of finding a desired geological entity in a digital image, and (ii) advancing the state-of-the-art in machine learning.
Mail: | Dr. David W. Opitz |
Computer Science Department | |
University of Montana | |
Missoula, MT 59812 |
E-mail: | opitz@cs.umt.edu |
Phone: | (406)-243-2831 |
FAX: | (406)-243-4076 |