A model-based multi-view image registration method for SAS images

Abstract: Autonomous underwater vehicles equipped with high-resolution synthetic aperture sonar (SAS), and automatic target recognition (ATR) algorithms show great potential for the task of search, classify and map. The level of detail recorded by the sonar is typically on the order of hundreds of pixels on an underwater object, valuable for improving classification performance. However, a weakness that is still not under control is the object aspect angle, which when unfavourable causes misclassification leading to significantly increased false alarm rates. The natural solution to this problem is to view the object from more than one direction. For this approach, referred to as multi-view ATR, different sonar views are collected, associated, registered and combined. The registration is not trivial, because sonar resolution surpasses object localisation accuracy. The registration approach proposed here is model based. Image templates from a database of objects are matched to the single-view images, which generates both a matching score and optimal registration shift for each object. Model-based registration is shown to be a robust technique, enabling automatic multi-view SAS image fusion