Contrastive Learning of Categories Features Representations for the Recognition of Underwater Target with Shot-few Annotations Samples

Abstract Underwater acoustic target recognition based on deep learning is complex owing to the shot-few annotation samples, moreover, the imbalance between classes of annotation samples. In this study, a contrast-based semi-supervised network (CBSN) is proposed for accurately recognizing the underwater targets at near-sea areas where a limited number of annotations samples with imbalance are obtained. This algorithm first designs an unsupervised contrast learning network, which uses the features distinction between positive and negative unannotated samples to contrast learning. The significant recognition features of the unannotated samples are extracted adaptively. Secondly, the feature extraction network is chosen in unsupervised networks representing high-dimensional annotations samples' features. Thus, the practical features for recognition in the annotated samples are obtained. Thirdly, contrast learning loss for underwater acoustic signals is designed to improve recognition accuracy. Extensive experiments on multiple supervised networks and semi-supervised networks demonstrate the superiority of the proposed framework showing that it can extract helpful recognition features for shot-few samples using the mutual guidance between annotated and unannotated samples. Further validation of the feature extraction performance shows that the network has excellent recognition of unbalanced shot-few samples.