A large number of underwater missions, such as obstacle avoidance, surveying, object recovery, and detection, classification, and recognition of hazards, are simply too dangerous or costly for manned vehicles. Remotely operated vehicles are subject to different limitations, such as communication bandwidth, operator fatigue, and a restricted radius of operation. These considerations make autonomous underwater vehicles (AUV) an increasingly attractive alternative. To be truly autonomous, an underwater vehicle requires scene recognition capabilities. Advances in pattern recognition and the use of increasingly high-resolution underwater sensors hold the promise that such capabilities will be developed in the near future. This paper reports the training and testing of multilayer perceptrons designed to classify specific manmade underwater objects under various environmental conditions, from arbitrary viewing aspects, and in highly cluttered environments. The trained classifiers have been tested against difficult side-scan sonar imagery and appear to work as well as a trained human analyst. Feature sets that account for the sensor response to range and that adapt to environmental variations improve performance and make the design robust. Receiver Operating Curves (ROC) show up to a 96% detection rate for a 2% false alarm rate. The set of multilayer perceptron networks have been demonstrated on special-purpose hardware and run in real time.
展开▼
