Multiframe Disentangling and Unifying Graph Neural Network for Maritime Radar Target Detection

Detecting maritime targets with radar in complicated sea conditions is challenging due to the strong sea clutter. To address this problem, we propose a multiframe disentangling and unifying graph neural network. In this method, multiple radar frames are first constructed as graph data. To capture multilevel relationships between frames, we decompose the multiframe radar graph into intraframe and interframe subgraphs, representing spatial and temporal dependencies, respectively. Through feature correlation coefficient calculation and a weighted message-passing mechanism, the intraframe and interframe subgraphs are updated separately. Subsequently, these subgraphs are unified to extract feature differences between targets and sea clutter from a global perspective. Finally, a segmentation head maps the features to segmentation masks, and the detection results are obtained by determining whether the mask values exceed a predefined threshold. Experiments on simulated and measured data demonstrate that the proposed method substantially improves target detectability while suppressing sea clutter, thereby improving detection performance.