Human-in-the-loop for autonomous underwater threat recognition

In this paper, human expert operators and automated classification algorithms are charged with the task of analyzing sonar images collected during real mine countermeasures exercises in order to detect and classify targets. Images are collected using synthetic aperture sonar (SAS) and side scan sonar (SSS), covering a test area on the Belgian Continental Shelf, between the Thorton bank and the Goote Bank. A seafloor segmentation map of this area, calculated using lacunarity and representing how difficult or how benign the seafloor is for object-recognition, is used as a new strategy in order to divide the database between operator and computer. Results demonstrate the utility of considering the human operator as an integral part of the automatic underwater object recognition process, and demonstrate how automated algorithms can extend and complement human performances.