Feature extraction and target classification of side-scan sonar images

Side-scan sonar technology has been used over the last three decades for underwater surveying and imaging. Application areas of side-scan sonar include archaeology, security and defence, seabed classification, and environmental surveying. In recent years the use of autonomous underwater systems has allowed for automatic collection of data. Along with automatic collection of data comes the need to automatically detect what information is important. Automatic target recognition can allow for efficient task planning and autonomous system deployment for security and defence applications. Support Vector Machines (SVMs) are proven general purpose methods for pattern classification. They provide maximum margin classification that does not over fit to training data. It is generally accepted that the choice of kernel function allows for domain specific information to be leveraged in the classification system. In this paper it is shown that for target classification in side-scan sonar, extra feature extraction and data engineering can result in better classification performance compared to parameter optimization alone.