Numerical experiments of training Cascade detectors with synthetic sonar imagery


  1. Fawcett, J.A.
Corporate Authors
Defence Research and Development Canada, Atlantic Research Centre, Halifax NS (CAN)
Most automated classification and some detection schemes for sidescan or synthetic aperture sonar images are trained methods. The training procedure relies upon the availability of sonar images with a sufficient number of images containing the possible targets of interest. In this report we consider the use of artificially-generated target images to train Haar- and Local Binary Pattern (LBP)-cascade detectors. The advantage of this technique is that one can generate thousands of target images covering a variety of geometries and possible target types. In this report, it is shown that in a straightforward fashion, one can train the detectors with a resulting performance comparable to a method trained with real sonar imagery. Various permutations of the training and testing data, in terms of the target and clutter images, from two different sea trials are considered illustrating the gains and loss of performance from having some knowledge from a new environment. Finally, we examine the possibility of using the synthetic targets injected into real imagery to predict automated target recognition performance for a sonar in a given environment.

Il y a un résumé en français ici.

automatic target recognition;sonar;sidescan
Report Number
DRDC-RDDC-2015-R132 — Scientific Report
Date of publication
01 Jul 2015
Number of Pages
Electronic Document(PDF)

Permanent link

Document 1 of 1

Date modified: