The Imbalanced Target Classification Method Based on Mixed Learning of Virtual and Real Data

We proposes a category imbalance classification model based on mixed feature enhancement between virtual and real domains to address the class imbalance problem in maritime target classification applications. In practical maritime target classification tasks, the imbalanced class distribution in real-world data poses challenges such as low accuracy and poor robustness in model training. To overcome this challenge, we combines balanced data generated from virtual simulation environments with imbalanced real-world data, employing a mixed learning approach to improve the model’s generalization ability and robustness. However, the domain shift problem between virtual and real data limits the transferability of the model. To address this challenge, domain adaptation methods are used to adjust the feature distribution discrepancy between the two domains, and a feature augmentation strategy is introduced at the hidden layer level to enhance the representation of tail classes (minority classes). Through experiments on multiple datasets combining balanced and imbalanced domains, the proposed model effectively alleviates the issues caused by class imbalance and improves model generalization performance in mixed data environments.