Robust Object Classification in Underwater Sidescan Sonar Images by Using Reliability-Aware Fusion of Shadow Features

Detecting and classifying objects in sidescan sonar images is an important underwater application with relevance to naval transportation and defense. Properties of the imaging modality, in this case, often introduce large intraclass variabilities reducing the discriminative power of any classification algorithm and limiting the possibilities of improving classification accuracy by advances in pattern recognition only. In this work, we investigate the role of an ancillary feature set computed on object shadows and propose a scheme for exploiting this useful, but variedly reliable information for object classification. A mean-shift-clustering-based segmentation technique is used for isolating highlight and shadow segments from the images. We show the results of reliability-aware fusion of features computed on highlight and shadows on three different data sets of sidescan sonar images, to illustrate under what conditions such information might be useful.