Proceedings Article | 22 July 1997
KEYWORDS: Data fusion, Image fusion, Classification systems, Magnetism, Mining, Neural networks
Research in FY-95 first addressed the problem of combining high-frequency (HF) side-scan sonar imagery, low-frequency (LF) side-scan Sonar imagery, and magnetic gradiometer data in order to detect/classify undersea mines. The first approach developed, termed the "Blob-Pair" based acoustic/magnetic (AM) Fusion system architecture, implicitly assumed that a target manifests itself in both HF,LF imagery, and was based on the fusion of single-sensor derived neural network classifier discriminants at a collection of three "decision" nodes, identified with magnetic (M), HF/LF, and HF/LF/M - data fusion cases, respectively. In order, to remove the restrictive assumption of a target manifesting in both }IF,LF data, the "Generalized" AM Fusion Architecture was developed, with a total of 7 "decision" nodes, identified with M, HF, HF/LF, LF, HF/M, HF/LFIM, and LF/M data fusion cases, respectively. However, the "Generalized" AM-Fusion architecture was found empirically to have significantly increased number of false alarms, relative to the "Blob-Pair" based system. Hence, through two-additional AM-Fusion architecture varaints, involving first the use of Classification Token "Post-Processing", and then both Token "Post-Processing" and decision node statistic modification, the performance "gap" between "Blob-Pair" and "Generalized" AM-Fusion Architecture performance was closed.
