A critical part of automatic classification algorithms is the extraction of features which distinguish targets from background noise and clutter. The focus of this paper is the use of variational methods for improving the classification of sea mines from both side-scan sonar and laser line-scan images. These methods are based on minimizing a functional of the image intensity. Examples include Total Variation Minimization (TVM) which is very effective for reducing the noise of an image without compromising its edge features, and Mumford-Shah segmentation, which in its simplest form, provides an optimal piecewise constant partition of the image. For the sonar side-scan images it is shown that a combination of these two variational methods, (first reducing the noise using TVM, then using segmentation) outperforms the use of either one individually for the extraction of minelike features. Multichannel segmentation based on a wavelet decomposition is also effectively used to declutter a sonar image. Finally, feature extraction and classification using segmentation is demonstrated on laser line-scan images of mines in a cluttered sea floor.

Date Published: 4 September 1998
PDF: 11 pages
Proc. SPIE 3392, Detection and Remediation Technologies for Mines and Minelike Targets III, (4 September 1998); doi: 10.1117/12.324200

Published in SPIE Proceedings Vol. 3392:
Detection and Remediation Technologies for Mines and Minelike Targets III
Abinash C. Dubey; James F. Harvey; J. Thomas Broach, Editor(s)