Some advantages and disadvantages of training based object detection/classification in SAS images

In this paper, rule based and training based fusion strategies for ATR will be compared based on selected methods for object detection in Side Scan Sonar Images. Test results were gained using SAS data recorded in the Baltic Sea with the MUSCELS AUV in different adjacent areas. The considered training based fusion methods normally outperform rule based fusion methods, if the training and test data set are from the same area. However, if this is not the case training based fusion may show a poor performance, which might be even worse than the performance of a single used detector. Potential reasons for such a drop in performance will be discussed.