Long-Tailed Distributed Radar Emitter Signal Automatic Modulation Recognition Based on Decoupled Training

The existing radar emitter modulation recognition methods typically assume that the data distribution across different types is balanced. But in reality, the number of signals of various kinds often follows a long-tail distribution, leading to model overfitting for the head classes and underfitting for the tail classes. As a result, the overall recognition performance of models under such data imbalances is suboptimal. A long-tail distribution automatic modulation recognition method based on decoupled training is proposed to address this issue. Based on the ResNeXt network, the proposed method decouples the model training process into two stages: a feature extraction phase under the imbalanced dataset and the classifier learning stage under a balanced dataset. The classifier boundary is fine-tuned by $\tau$-normalization method. Compared to existing radar emitter modulation recognition frameworks, the proposed method achieves an overall recognition accuracy of 86.8% when the data imbalance factor is 0.01, surpassing the baseline model by 5%, and improves the performance of radar emitters modulation recognition in the real environment.