Cross-sensor Maritime Target Recognition Algorithm Based on Heterogeneous Sensor Shared Representation Learning
Yihai Liu · 2023
We investigate a cross-sensor recognition algorithm for heterogeneous sensors based on shared representation learning to address the difficulty of non-cooperative object recognition in C4ISR systems due to modal barriers, imbalanced data distribution, and few-shot labeled samples. First, deep multiset canonical correlation analysis is used to learn the shared space among heterogeneous sensors by only unlabeled spatiotemporal correlation data from sensors. Second, scattered labeled data from multiple sensors are concentrated in shared space, which removes the modality barrier of each sensor and produces big data samples for training the classification network. Finally, the classification model is trained and directly used to classify the new observations for each sensor in the shared space, which constitutes a general-purpose cross-sensor target recognition. Experiments based on target recording data show that the proposed algorithm can achieve 60-78% accuracy in cross-sensor recognition for radar, ESM, optical, and sonar sensors, indicating a significant engineering value.