Unsupervised Domain Adaption on Category Level by Pseudo-Label for Radar Target Detection

Due to changes in sea conditions and radar parameters, the statistical parameters of sea clutter can obviously change. The neural network (NN) trained on the specific statistical parameter dataset (source domain) will face significant performance degradation in other statistical parameter datasets (target domain). Unsupervised domain adaptation (UDA) can effectively mitigate this problem by making the two domain distributions consistent in the feature space, without supervised information from the target domain. However, the similarity between clutter (e.g., sea spike) and the target leads to category confusion when aligning the two domains globally. Therefore, we propose a category-level joint loss (CJL) consisting of cross-entropy, local maximum mean discrepancy (LMMD), and git loss in our UDA radar target detection framework. Among them, based on the pseudo-labels provided by the NN during the training process, the LMMD is used to reduce the feature distance of the same category between two domains. Furthermore, we adopt git loss to increase the depth feature distance of interclass samples, while encouraging the aggregation of intraclass features. Finally, we analyze the performance loss of NN detectors and demonstrate that the proposed method has the best alignment and classification performance compared to existing methods on the constructed shore-based radar datasets.