Temporal-SNE-based Underwater Target Recognition Approach for Class-imbalanced Datasets

Aiming at the problem of low target recognition accuracy due to the small number of underwater target samples and the imbalance of dataset classes, a feature selection method with the name of t-distributed stochastic neighbor embedding based on temporal information (Temporal-SNE) is proposed, which describes the potential relationship between features and classes to enhance the essential representation of underwater target features. The method uses underwater target radiation noise to construct the feature space, and fuses the time-series information into the feature space in order to optimize the features with strong consistency of similarity and high differentiation of dissimilarity. The experiments use class-imbalanced underwater target datasets to compare the feature discrimination and recognition accuracy between Temporal-SNE method and the traditional methods. The results show that the introduction of temporal sequence improves the distinguishability of underwater target features and significantly increases the classification accuracy compared with the traditional methods.