Adaptive order-statistic filters for sea mine classification

Abstract

This paper presents a novel formulation of an adaptive order- statistic filter, and describes the performance enhancements it provides to an automatic sea mine classification system. Non-linear filters based on order statistics (median, 'largest-of,' etc.) have been shown to be effective in suppressing noise with long, heavy-tailed density functions (e.g., Laplacian), and they have also been successfully used to suppress 'salt-and-pepper' noise in image processing, as well as transients and Raleigh-distributed speckle noise in ultrasound imaging. Such 'order-statistic' filters can be adaptively generalized and optimized, for a given data set, by finding the weights that, operating on ordered data samples, minimize filter output power while preserving signals that are constant within the filter window. Morphological filters can also be optimized in this manner, since they have been shown to consist of combinations of order-statistic filters. A new adaptive order-statistic filter formulation, enabling the preservation of signals that are not constant within the filter window, has been developed and its efficacy demonstrated with side-scan sonar imagery data. Using these filters as a non-linear 'corrector' of the outputs of the linear clutter-filtering stage of a sea mine classification system, reduced the number of false alarms by an order of magnitude.