Multi-view target classification in synthetic aperture sonar imagery

This work proposes an elegantly simple solution to the general task of classifying the shape of an object that has been viewed multiple times. Specifically, this problem is addressed in the context of underwater mine classification where the objective is to discriminate targets (i.e., mines) from benign clutter (e.g., rocks) when each object is observed in an arbitrary number of synthetic aperture sonar (SAS) images. The proposed multi-view classification algorithm is based on finding the single highest maximum correlation between (i) a set of views of a training shape of interest and (ii) a set of views of a given testing object. Classification is performed by using this measure of similarity, which we term the affinity, directly. This approach obviates the need for explicit feature extraction and classifier construction. Moreover, the framework induces no constraints on the number of views that each object can possess. Promising experimental results using real SAS imagery demonstrate the feasibility of the proposed approach for multi-view classification of underwater mines. In particular, it is shown that classification performance improves dramatically as the number of views of the objects increases.