Deep convolutional neural networks for semi-supervised learning from synthetic aperture sonar (SAS) images

Abstract

Advancements in deep neural networks for computer-vision tasks have the potential to improve automatic target recognition (ATR) in synthetic aperture sonar (SAS) imagery. Many of the recent improvements in computer vision have been made possible by densely labeled datasets such as ImageNet. In contrast, SAS datasets typically contain far fewer labeled samples than unlabeled samples—often by several orders of magnitude. Yet unlabeled SAS data contain information useful for both generative and discriminative tasks. Here results are shown from semi-supervised ladder networks for learning to classify and localize in SAS images from very few labels. We perform end-to-end training concurrently with unlabeled and labeled samples and find that the unsupervised-learning task improves classification accuracy. Ladder networks are employed to adapt fully convolutional networks used for pixelwise prediction based on supervised training to semi-supervised semantic segmentation and target localization by pixel-level classification of whole SAS images. Using this approach, we find improved segmentation and better generalization in new SAS environments compared to purely supervised learning. We hypothesize that utilizing large unsupervised data in conjunction with the supervised classification task helps the network generalize by learning more invariant hierarchical features. [Work supported by the Office of Naval Research.]