Underwater source ranging by Siamese network aided semi-supervised learning

Underwater source ranging based on Deep Learning methods demands a considerable amount of labeled data, which is costly to collect. To alleviate this challenge, semi-supervised learning of the wrapper paradigm is introduced into this task. First, the Siamese networkwhich is costly to collect. To alleviate this challenge, semi-supervised learning of the wrapper paradigm is introduced into this task. First, the Siamese network is used to generate pseudo labels for unlabeled data to expand the labeled dataset. A new effective confidence criterion based on similarity score and similar sample distribution is proposed to evaluate the reliability of pseudo labels. Then the model can be trained more fully with an expanded dataset. Experiments on the SwellEx-96 dataset validate that this method can effectively improve prediction accuracy.