Target Classification in SAS Imagery using Orthogonal Basis Selection

This work proposes an approach that finds efficient representations for training and classification of different mine like objects (MLOs) in underwater imagery, e.g. side scan sonar and synthetic aperture sonar (SAS). The focus is on the design and selection of a compact, optimal and a non linear snbspace, a dictionary, based on the gradient and curvature models in 2D images. Here, the traditional sparse approximation formulation is decoupled and modified by an additional discriminating objective function and a corresponding selection strategy is proposed. During training, using a set of labelled sonar images, a single optimised discriminatory dictionary is learnt which can then be used to represent MLOs. During classification, this dictionary together with optimised coefficient vectors is used to label scene entities. Evaluation of our approach has resulted in classification accuracies of 95% and 94% on realistic synthetic side-scan images and real CMRE SAS imagery, respectively.