IEEE Conference on Applied Imagery Pattern Recognition (AIPR), Washington, DC (October, 2011).
A major challenge for fielding an automated target recognition (ATR) system is training the algorithm to perform well under realistic conditions. An ATR trained a priori on a general set of target imagery often lacks the specificity required for high confidence strike. A promising alternative would be rapid training of the ATR on mission-specific imagery immediately prior to its operational use. This rapid training hinges on the existence of a small, representative training set. To solve the problem of selecting such training sets, and thus to make rapid ATR training a reality, we have developed a novel data mining method. Our technique examines large databases and extracts sparse, representative subsets. For data mining, we use a variant of decision trees, called random decision trees (RDTs). We augment these trees with a distribution modeling component that eliminates redundant information, ignores misrepresentative class distributions in the database, and stops training when decision boundaries are adequately sampled. These augmented random decision trees enable fast construction of reliable, mission-specific training data. Earlier work by this team coupled the training and classifier by using a collection of RDTs, called a random decision forest, to perform the classification. More recent research has shown that this close coupling is not necessary. The training set developed by our technique can be paired with any classifier to realize a full system. We discuss the development of our approach and its application to ATR problems using a range of feature sets and classifier types. The performance results demonstrate the efficiency that is possible when our technique is used to choose the training sample.
1 Charles River Analytics
2 Charles Stark Draper Laboratory
3 US Army AMRDEC
For More Information
(Please include your name, address, organization, and the paper reference. Requests without this information will not be honored.)