False Alarm Reduction by LS-SVM for Manmade Object Detection from Sidescan Sonar Images

Automatic manmade object detection from sidescan sonar image is a key step for automatic mine countermeasure. A high detection rate and low false alarm rate are required if the mine countermeasure system can find the target and destroy it without operator interfering. In our previous work, an algorithm for manmade object detection has been developed for after-mission processing. In order to maintain a high detection rate for operator making further decision, the false alarm rate is more than 50%, which is too high for automatic mine countermeasure. In this work, we propose to use the pattern recognition technique of support vector machines (SVM) to reduce false detections in our previous work. The SVM is selected because of its open structure. It is thus possible to adjust the parameters to achieve the requirement of maintaining high detection rate. We have studied the features of the correct detections and the false detections and proposed to use 2 features to classify the false detection and correct detection. After applying the SVM, more than 50% of false detections have been reduced and above 80% correct detections are maintained.