Objective: To develop a system of searching image databases for patterns of interest to the scientist using a neural network pattern recognition algorithm. The patterns can be spatial, spectral, or temporal in nature and in any kind of data. Images or parts of images with likely matches would be returned to the user for further examination. The idea is to run this procedure remotely and download only those images that have a high probability of containing the pattern of interest.
Approach: A neural network pattern recognition system was implemented in the C programming language on a serial workstation. A PVM (Parallel Virtual Machine) version was developed later to run the algorithm on a cluster of workstations. The software takes sample patterns provided by the user, trains a neural network until some minimal error is achieved, and then uses the trained network to search a database of the same type of data.
Accomplishments: The neural network algorithm has been implemented both on a single workstation and in the PVM environment on multiple workstations with near optimal speed-up. The algorithm has been shown to work well for some sample spatial and spectral patterns in SAR (synthetic aperture radar) and Landsat multispectral data.
Significance: The neural network method is easily adaptable to any type of data. The same algorithm (and thus the same code) can be used for vastly difference search criteria. Also, once the network is trained and being used for searching, it is compact and fast. Thus, it has the potential for being a flexible search engine for Earth science data sets.
Status/Plans: The search algorithm has been implemented in PVM. The next step is to try to increase the performance of the training algorithm by implementing it in PVM as well. Also, the searching will be attempted over an ATM network to examine the effect of network speed on overall search performance.
Points of Contact:
Dr. Robert A. Schowengerdt
Justin D. Paola
University of Arizona
schowengerdt@ece.arizona.edu, 520-621-2706
paola@ece.arizona.edu, 520-621-4554
URL: http://dialsparc10.ece.arizona.edu/dial-home.html