Saliency and Anomaly: Transition of Concepts from Natural Images to Side-Scan Sonar Images

An autonomous underwater vehicle (AUV) or a multi-AUV system performing autonomous seafloor exploration missions by a side-scan sonar need to perceive their environment in order to replan the mission if they detect interesting objects in sensor data. Several anomalous/salient object detection methods mostly used for natural images are here applied to sonar images. All methods were firstly benchmarked on a 1500 simulated side-scan sonar images dataset. Precision-recall and processing time analysis was conducted in order to choose the best suited method in such controlled conditions. The performance of the best performing detection method was then validated on a 350 real side-scan sonar images dataset. This method was then implemented and optimized for the computer onboard an AUV. It turned out to be fast enough for online processing of large volumes of sonar data. Copyright (C) 2020 The Authors.