Translator Disclaimer
21 September 2004 Robust score-based feature vectors for algorithm fusion
Author Affiliations +
A new technique for constructing score-based feature vectors for Algorithm fusion is presented. Algorithm fusion is the combining (fusing) of the outputs of multiple detection and classification (D/C) algorithms to generate a final target/nontarget decision. A 2-class D/C problem is considered: Target and Nontarget. It is assumed that multiple D/C algorithms are used to process that same raw sensor data. It is further assumed that each algorithm produces a nonnegative score for each detected object that is a measure of the degree of "target-likeness." (A score of zero is assigned to any D/C algorithm that did not detect a particular object that had been detected by another algorithm.) Several new score-based feature vectors are constructed using only the scores of the individual D/C algorithms. The feature vectors can be used as input to any feature-based classifier; for this paper, the 2-class linear classifier based on maximizing the Fisher ratio criterion has proven very effective. The different score-based feature vectors have different dimensionality. In light of Bellman's curse of dimensionality, this permits one to select the feature vector whose size is most compatible with the size of the training data set. Consequently, robust performance can be achieved.
© (2004) COPYRIGHT Society of Photo-Optical Instrumentation Engineers (SPIE). Downloading of the abstract is permitted for personal use only.
Gerald J. Dobeck "Robust score-based feature vectors for algorithm fusion", Proc. SPIE 5415, Detection and Remediation Technologies for Mines and Minelike Targets IX, (21 September 2004);


Back to Top