Enhancing Generalization of Active Sonar Classification Using Semisupervised Anomaly Detection With Multisphere for Normal Data

Anomaly detection is suitable for active sonar classification due to its ability to handle the challenges posed by small imbalanced data sets. Recently, a modified anomaly detection approach called bisphere anomaly detection (BiSAD) has been developed for active sonar classification and has demonstrated improved generalization performance over conventional deep-learning-based methods. However, BiSAD has some limitations: multimodalities of clutter distribution induce unnecessary redundancy in the clutter manifold, and the inconsistency of two encoder outputs causes instability during learning. We propose a modified version of BiSAD called multisphere anomaly detection (MulSAD), which incorporates clustering and regularization. Clustering is used to model the multimodal distribution of the clutter samples, whereas regularization ensures consistency in the manifold learning of the two encoders. Two active sonar data sets generated in two different ocean experiments with different environments are used alternatively as the training/validation and test data sets. The efficacy of the modifications is confirmed by analyzing the classification performance according to hyperparameters. In the generalization test, MulSAD outperforms both the supervised-learning-based deep learning methods and BiSAD. Furthermore, MulSAD is more robust to mislabeled data samples in the training data sets.