Automatic target classification using multiple sidescan sonar images of different orientations

Abstract

In this report, the target classification performance of a multiple view sidescan sonar is investigated. The classification statistics are estimated using model based automatic classifiers. The guidelines to the design of efficient classification algorithms are defined. The shadow is retained as the basic information for target classification. The input feature vector of the automatic classifier is the cross-section (or height profile) of the target estimated from its shadow. The concept of multiple view sidescan sonar is presented and compared to other techniques for recording multiple aspects of a target. Several ways to

modify a single view based classifier to process multiple aspects are identified and implemented. The task of the classifier is to recognize 10 target shapes, corresponding to proud mines and sinkers of moored mines. The classification results, expressed by ROC curves and confusion matrixes, are computed on a larger number of natural and manufactured object images, generated by modelling software. The classification and identification of targets are closely related to the capacity

of the feature vectors to discriminate between a given target, viewed with a given orientation, from all the other targets, irrespective of orientation. Using height profiles as feature vectors, the capacity to discriminate is established for a single view on target. Different configurations using up to three views with different angular intervals are subsequently compared