A Methodology for Multi-view Side-scan Sonar Target Recognition

Abstract

Side-scan sonar is widely used for seabed imaging, but automated target recognition in such data remains challenging because of acoustic noise, stripe artifacts, view-dependent target appearance, navigation errors, and repeated observations of the same physical object during different passes or missions. The aim of this paper is to develop a methodology for multi-view target recognition in side-scan sonar data that enables a transition from individual sonar images and detections to the level of physical targets on the seafloor. The proposed methodology integrates mission data preparation, inter-mission change detection, estimation of the number of physical targets, detection clustering, localization, visual similarity modeling, deep model training, and multi-view classification. The input data include side-scan sonar missions, trajectory data, sonar images, metadata, and object detections, while the output consists of information about inter-mission changes, hypotheses about physical targets, their coordinates, and class labels. Individual components of the methodology were evaluated on synthetic and real datasets using task-specific quality metrics. The results showed improved robustness of change detection on noisy sonar images, estimation of the number of physical targets with an error not exceeding one object, improved clustering and localization due to the integration of spatial, topological, and visual information, and higher quality of object-level multi-view classification. The methodology forms a unified modular workflow for side-scan sonar data analysis, in which the output of one stage can be used as input or as a constraint for subsequent stages. It can serve as a basis for automated sonar data analysis systems for underwater monitoring, object search, humanitarian demining, infrastructure inspection, and operator decision support.