Active deep belief networks for ship recognition based on BvSB

During remote image classification, accurately classifying ships with insufficient label data is a well-known challenge. In this paper, we propose an intelligent, semi-supervised learning algorithm called active deep network based on BvSB (BvSB-ADN). BvSB-ADN is initially constructed based on the structure of restricted Boltzmann machines (RBM), then active learning is used to identify samples which can be labeled as training data. In the sample-identification phase, the best versus second-best (BvSB) rule is applied to determine the most useful samples; the labeled samples as-selected and all unlabeled samples are then combined to train the BvSB-ADN architecture. The BvSB and classifier are based on the same architecture, which makes selecting the most important samples relatively very simple. We applied BvSB-ADN to a ship classification task to verify its effectiveness and feasibility, and found that it outperforms other classification methods. BvSB-ADN also showed impressive performance on the MNIST dataset.